Gst-nvtracker

The Gst-nvtracker plugin allows the DS pipeline to use a low-level tracker library to track the detected objects over time persistently with unique IDs. It supports any low-level library that implements NvDsTracker API, including the reference implementations provided by the NvMultiObjectTracker library: IOU, NvSORT, NvDeepSORT and NvDCF trackers. As part of this API, the plugin queries the low-level library for capabilities and requirements concerning the input format, memory type, and additional feature support. Based on these queries, the plugin then converts the input frame buffers into the format requested by the low-level tracker library. For example, the NvDeepSORT and NvDCF trackers use NV12 or RGBA, while IOU and NvSORT requires no video frame buffers at all.

Based on the queries, the Gst-nvtracker plugin constructs the input data to the low-level tracker library, which consists of video frames and detected objects across multiple streams that are fed to the low-level library in a single data structure (i.e, a frame batch) through a single API call. This API design allows all the input data from multiple streams in the low-level tracker library processed in a batched processing mode (similar to the batch mode in cuFFT™, cuBLAS™, etc.) for potentially more efficient execution especially when accelerated on GPU. There for, it is required that the low-level tracker supports batch processing when using NvMOT_Process API.

The Gst-nvtracker plugin supports retrieval of the user-defined miscellaneous data from the low-level tracker library through NvMOT_RetrieveMiscData API, which includes useful object tracking information other than the default data for the current frame targets; for example, past-frame object data, targets in shadow tracking mode, full trajectory of terminated targets and re-identification features. More details on the types of miscellaneous data and what they means can be found in Miscellaneous Data Output section. The users are allowed to define other types of miscellaneous data in NvMOTTrackerMiscData.

More details on all these tracker APIs is discussed in NvDsTracker API for Low-Level Tracker Library section.

The plugin accepts NV12- or RGBA-formatted frame data from the upstream component and scales (and/or converts) the input buffer to a buffer in the tracker plugin based on the format required by the low-level library, with the frame resolution specified by tracker-width and tracker-height in the configuration file’s [tracker] section. The path to the low-level tracker library is to be specified via ll-lib-file configuration option in the same section. The low-level library to be used may also require its own configuration file, which can be specified via ll-config-file option. If ll-config-file is not specified, the low-level tracker library may proceed with its default parameter values.

The reference low-level tracker implementations provided by the NvMultiObjectTracker library support various types of multi-object tracking algorithms:

• IOU Tracker: The Intersection-Over-Union (IOU) tracker uses the IOU values among the detector’s bounding boxes between the two consecutive frames to perform the association between them or assign a new target ID if no match found. This tracker includes a logic to handle false positives and false negatives from the object detector; however, this can be considered as the bare-minimum object tracker, which may serve as a baseline only.

• NvSORT: The NvSORT tracker is the NVIDIA®-enhanced Simple Online and Realtime Tracking (SORT) algorithm. Instead of a simple bipartite matching algorithm, NvSORT uses a cascaded data association based on bounding box (bbox) proximity for associating bboxes over consecutive frames and applies a Kalman filter to update the target states. It is computationally efficient since it does not involve any pixel data processing.

• NvDeepSORT: The NvDeepSORT tracker is the NVIDIA®-enhanced Online and Realtime Tracking with a Deep Association Metric (DeepSORT) algorithm, which uses the deep cosine metric learning with a Re-ID neural network for data association of multiple objects over frames. This implementation allows users to use any Re-ID network as long as it is supported by NVIDIA’s TensorRT™ framework. NvDeepSORT also uses a cascaded data association instead of a simple bipartite matching. The implementation is also optimized for efficient processing on GPU.

• NvDCF: The NvDCF tracker is an online multi-object tracker that employs a discriminative correlation filter for visual object tracking, which allows independent object tracking even when detection results are not available. It uses the combination of the correlation filter responses and bounding box proximity for data association.

More details on each algorithm and its implementation details can be found in NvMultiObjectTracker : A Reference Low-Level Tracker Library section.

Note

The source code of the Gst-nvtracker plugin is provided as a part of DeepStream SDK package under sources/gst-plugins/gst-nvtracker/ when installed on a system This is to allow users to make direct changes in the plugin whenever needed for their custom applications and also to show the users as to how the low-level libraries are managed and how the metadata is handled in the plugin.

Sub-batching (Alpha)

The Gst-nvtracker plugin works in the batch processing mode by default. In this mode, the input frame batch is passed to and processed by a single instance of low-level tracker library. The advantage of batch processing mode is to allow GPUs to work on bigger amount of data at once, potentially increasing the GPU occupancy during execution and reducing the CUDA kernel launch overhead. Depending on the use cases, however, a potential issue is that there is a possibility that GPU could be idling (also referred to as GPU bubble) in some compute stages in the tracker unless the end-to-end operation within the module is carried out solely on the GPU. This is indeed the case if some of the compute modules in the tracker runs on CPU. If there are other components in the DeepStream pipeline that uses GPU (e.g., GPU-based inference in PGIE and SGIE), such CPU blocks in tracker can be hidden behind them, not affecting the overall throughput of the pipeline.

The newly-introduced Sub-batching feature allows the plugin to split the input frame batch into multiple sub-batches (for example, a four-stream pipeline can use two sub-batches in the tracker plugin, each of which takes care of two streams). Each sub-batch is assigned to a separate instance of low-level tracker library, where the input to the corresponding sub-batch is processed separately. Each instance of low-level tracker libraries runs on a dedicated thread running independently, allowing parallel processing of sub-batches and minimizing the GPU idling due to CPU compute blocks, which eventually results in higher resource utilization.

Because sub-batching assigns separate low-level tracker library instances to different sub-batches, it allows the user to configure each individual sub-batch differently with different low-level tracker library configuration files. This can be utilized in multiple ways like setting varied compute backends across sub-batches, using varied tracking algorithms across sub-batches or modifying any other configuration that is supported in low-level tracker configuration file. More detailed example use-cases are discussed in Setup and Usage of Sub-batching (Alpha) section.

Inputs and Outputs

This section summarizes the inputs, outputs, and communication facilities of the Gst-nvtracker plugin.

• Input

  • Gst Buffer

    • A frame batch from available source streams

  • NvDsBatchMeta

    • Includes the detected object info from primary inference module

    • More details about NvDsBatchMeta can be found in the link.

The color formats supported for the input video frame by the NvTracker plugin are NV12 and RGBA. A separate batch of video frames are created from the input video frames based on the color format and the resolution that is required to the low-level tracker library.

• Output

  • Gst Buffer

    • Same as the input. Unmodified.

  • NvDsBatchMeta

    • Updated with additional data from tracker low-level library

There’s no separate data structure for the output. Instead, the tracker plugin just adds/updates the data in the existing NvDsBatchMeta (and its NvDsObjectMeta) with the output data from the tracker low-level library, including tracked object coordinates, tracker confidence, and object IDs. There are some other miscellaneous data that can be attached as user-meta, which is covered in Miscellaneous Data Output section.

Note

If the tracker algorithm does not generate confidence value, then tracker confidence value will be set to the default value (i.e., 1.0) for tracked objects. For IOU, NvSORT and NvDeepSORT trackers, tracker_confidence is set to 1.0 as these algorithms do not generate confidence values for tracked objects. NvDCF tracker, on the other hand, generates confidence for the tracked objects due to its visual tracking capability, and its value is set in tracker_confidence field in NvDsObjectMeta structure.

Note that there are separate parameters in NvDsObjectMeta for detector’s confidence and tracker’s confidence, which are confidence and tracker_confidence, respectively. More details can be found in New metadata fields

The following table summarizes the features of the plugin.

Gst-nvtracker plugin features

Feature	Description	Release
Configurable tracker width/height	Frames are internally scaled in NvTracker plugin to the specified resolution for tracking and passed to the low-level lib	DS 2.0
Multi-stream CPU/GPU tracker	Supports tracking on batched buffers consisting of frames from multiple sources	DS 2.0
NV12 Input		DS 2.0
RGBA Input		DS 3.0
Configurable GPU device	User can select GPU for internal scaling/color format conversions and tracking	DS 2.0
Dynamic addition/deletion of sources at runtime	Supports tracking on new sources added at runtime and cleanup of resources when sources are removed	DS 3.0
Support for user’s choice of low-level library	Dynamically loads user selected low-level library	DS 4.0
Support for batch processing exclusively	Supports sending frames from multiple input streams to the low-level library	DS 4.0
Multiple buffer formats as input to low-level library	Converts input buffer to formats requested by the low-level library, for up to 4 formats per frame	DS 4.0
Enabling tracking-id display	Supports enabling or disabling display of tracking-id	DS 5.0
Tracking ID reset based on event	Based on the pipeline event (i.e., GST_NVEVENT_STREAM_EOS and GST_NVEVENT_STREAM_RESET), the tracking IDs on a particular stream can be reset to start from 0 or new IDs.	DS 6.0
Miscellaneous data	Supports outputting user defined miscellaneous data (including the past-frame data, a list of terminated tracks, etc. ) if the low-level library supports the capability	DS 6.3
Re-ID tensor output	Supports outputting objects’ Re-ID features (i.e., embeddings) if the low-level library uses a Re-ID model	DS 6.3
Support for NVIDIA’s VPI™ based Crop-scaler and DCF-Tracker algorithms in NvDCF tracker (Alpha feature)	Configuration options provided in NvDCF tracker which allow the user to switch to NVIDIA’s VPI™ implementation of Crop-scaler and DCF-Tracker. The user can also configure the compute backend to be used amongst the backends supported by VPI™	DS 6.4
PVA-backend for NvDCF via VPI™’s unified API (Alpha feature)	Allow PVA-based execution of a significant part of NvDCF on Jetson, resulting in lower GPU utilization	DS 6.4
Sub-batching (Alpha feature)	Supports splitting of a batch of frames in sub-batches which are internally processed in parallel resulting in higher resource utilization. This feature also enables specification of a different config file for each sub-batch.	DS 6.4
Single-View 3D Tracking (Alpha feature)	Allow 3D world coordinate system based object tracking when camera/model info (3x4 projection matrix and 3D human model info) is provided for better handling of partial occlusion	DS 6.4

Gst Properties

The following table describes the Gst properties of the Gst-nvtracker plugin.

Gst-nvtracker plugin Gst Properties

Property	Meaning	Type and Range	Example Notes
tracker-width	Frame width at which the tracker is to operate, in pixels. (To be a multiple of 32 when visualTrackerType: 1 or reidType is non-zero with useVPICropScaler: 0)	Integer, 0 to 4,294,967,295	tracker-width=640
tracker-height	Frame height at which the tracker is to operate, in pixels. (To be a multiple of 32 when visualTrackerType: 1 or reidType is non-zero with useVPICropScaler: 0)	Integer, 0 to 4,294,967,295	tracker-height=384
ll-lib-file	Pathname of the low-level tracker library to be loaded by Gst-nvtracker.	String	ll-lib-file=­/opt/nvidia/­deepstream/­deepstream/­lib/libnvds_nvmultiobjecttracker.so
ll-config-file	Configuration file for the low-level library if needed. (Alpha feature) A list of configuration files can be specified when the property sub-batches is configured.	Path to configuration file OR A list of paths to configuration files delimited by semicolon	ll-config-file=config_tracker_NvDCF_perf.yml ll-config-file=config_tracker_NvDCF_perf1.yml; config_tracker_NvDCF_perf2.yml
gpu-id	ID of the GPU on which device/unified memory is to be allocated, and with which buffer copy/scaling is to be done. (dGPU only.)	Integer, 0 to 4,294,967,295	gpu-id=0
tracking-surface-type	Set surface stream type for tracking. (default value is 0)	Integer, ≥0	tracking-surface-type=0
display-tracking-id	Enables tracking ID display on OSD.	Boolean	display-tracking-id=1
compute-hw	Compute engine to use for scaling. 0 - Default 1 - GPU 2 - VIC (Jetson only)	Integer, 0 to 2	compute-hw=1
tracking-id-reset-mode	Allow force-reset of tracking ID based on pipeline event. Once tracking ID reset is enabled and such event happens, the lower 32-bit of the tracking ID will be reset to 0 0: Not reset tracking ID when stream reset or EOS event happens 1: Terminate all existing trackers and assign new IDs for a stream when the stream reset happens (i.e., GST_NVEVENT_STREAM_RESET) 2: Let tracking ID start from 0 after receiving EOS event (i.e., GST_NVEVENT_STREAM_EOS) (Note: Only the lower 32-bit of tracking ID to start from 0) 3: Enable both option 1 and 2	Integer, 0 to 3	tracking-id-reset-mode=0
input-tensor-meta	Use the tensor-meta from Gst-nvdspreprocess if available for tensor-meta-gie-id	Boolean	input-tensor-meta=1
tensor-meta-gie-id	Tensor Meta GIE ID to be used, property valid only if input-tensor-meta is TRUE	Unsigned Integer, ≥0	tensor-meta-gie-id=5
sub-batches (Alpha feature)	Configures splitting of a batch of frames in sub-batches. There are two ways to configure sub-batches. First option allows static mapping of each source id to individual sub-batch. Second option lets the user configure the sub-batch sizes. Mapping of individual streams to sub-batch happens dynamically at runtime.	Option 1 : Semicolon delimited integer array where each number corresponds to source id. Must include all values from 0 to (batch-size -1) where batch-size is configured in [streammux]. Option 2 : Colon delimited integer array where each number corresponds to size of a sub-batch (i.e. max number of stream a sub-batch can accommodate)	Option 1 : sub-batches=0,1;2,3 In this example, a batch size of 4 is split into two sub-batches where the first sub-batch consists of source ids 0 & 1 and second sub-batch consists of source ids 2 & 3 Option 2 : sub-batches=2:1 The above example indicates that there are two sub-batches, first one can accommodate 2 streams and second one can accommodate 1.
sub-batch-err-recovery-trial-cnt (Alpha feature)	Configure the number of times the plugin can try to recover when the low level tracker in a sub-batch returns with a fatal error. To recover from the error, the plugin reinitializes the low level tracker library.	Integer,≥-1 where, -1 corresponds to infinite trials	sub-batch-err-recovery-trial-cnt=3
user-meta-pool-size	The size of tracker miscellaneous data buffer pool	Unsigned Integer, >0	user-meta-pool-size=32

NvDsTracker API for Low-Level Tracker Library

A low-level tracker library can be implemented using the API defined in sources/includes/nvdstracker.h. Parts of the API refer to sources/includes/nvbufsurface.h. The names of API functions and data structures are prefixed with NvMOT, which stands for NVIDIA Multi-Object Tracker. Below is the general flow of the API from a low-level library’s perspective:

1. The first required function is:

   NvMOTStatus NvMOT_Query (
        uint16_t customConfigFilePathSize,
        char* pCustomConfigFilePath,
        NvMOTQuery *pQuery
   );
   

   The plugin uses this function to query the low-level library’s capabilities and requirements before it starts any processing sessions (i.e., contexts) with the library. Queried properties include the input frame’s color format (e.g., RGBA or NV12) and memory type (e.g., NVIDIA^® CUDA^® device or CPU-mapped NVMM).

   The plugin performs this query once during initialization stage, and its results are applied to all contexts established with the low-level library. If a low-level library configuration file is specified, it is provided in the query for the library to consult. The query reply structure, NvMOTQuery, contains the following fields:

   • NvMOTCompute computeConfig: Report compute targets supported by the library. The plugin currently only echoes the reported value when initiating a context.

   • uint8_t numTransforms: The number of color formats required by the low-level library. The valid range for this field is 0 to NVMOT_MAX_TRANSFORMS. Set this to 0 if the library does not require any visual data.

     Note

     0 does not mean that untransformed data will be passed to the library.

   • NvBufSurfaceColorFormat colorFormats[NVMOT_MAX_TRANSFORMS]: The list of color formats required by the low-level library. Only the first numTransforms entries are valid.

   • NvBufSurfaceMemType memType: Memory type for the transform buffers. The plugin allocates buffers of this type to store color- and scale-converted frames, and the buffers are passed to the low-level library for each frame. The support is currently limited to the following types:

     dGPU:

     NVBUF_MEM_CUDA_PINNED
     NVBUF_MEM_CUDA_UNIFIED
     

     Jetson:

     NVBUF_MEM_SURFACE_ARRAY
     

   • bool supportBatchProcessing: True if the low-level library supports the batch processing across multiple streams; otherwise false.

   • bool supportPastFrame: True if the low-level library supports outputting the past-frame data; otherwise false.

2. After the query, and before any frames arrive, the plugin must initialize a context with the low-level library by calling:

   NvMOTStatus NvMOT_Init (
        NvMOTConfig *pConfigIn,
        NvMOTContextHandle *pContextHandle,
        NvMOTConfigResponse *pConfigResponse
   );
   

The context handle is opaque outside the low-level library. In the batch processing mode, the plugin requests a single context for all input streams. In per-stream processing mode, on the other hand, the plugin makes this call for each input stream so that each stream has its own context. This call includes a configuration request for the context. The low-level library has an opportunity to:

• Review the configuration and create a context only if the request is accepted. If any part of the configuration request is rejected, no context is created, and the return status must be set to NvMOTStatus_Error. The pConfigResponse field can optionally contain status for specific configuration items.

• Pre-allocate resources based on the configuration.

Note

• In the NvMOTMiscConfig structure, the logMsg field is currently unsupported and uninitialized.

• The customConfigFilePath pointer is only valid during the call.

3. Once a context is initialized, the plugin sends frame data along with detected object bounding boxes to the low-level library whenever it receives such data from upstream. It always presents the data as a batch of frames, although the batch can contain only a single frame in per-stream processing contexts. Note that depending on the frame arrival timings to the tracker plugin, the composition of frame batches could either be a full batch (that contains a frame from every stream) or a partial batch (that contains a frame from only a subset of the streams). In either case, each batch is guaranteed to contain at most one frame from each stream.

The function call for this processing is:

NvMOTStatus NvMOT_Process (
     NvMOTContextHandle contextHandle,
     NvMOTProcessParams *pParams,
     NvMOTTrackedObjBatch *pTrackedObjectsBatch
);

, where:

• pParams is a pointer to the input batch of frames to process. The structure contains a list of one or more frames, with at most one frame from each stream. Thus, no two frame entries have the same streamID. Each entry of frame data contains a list of one or more buffers in the color formats required by the low-level library, as well as a list of object attribute data for the frame. Most libraries require at most one-color format.

• pTrackedObjectsBatch is a pointer to the output batch of object attribute data. It is pre-populated with a value for numFilled, which is the same as the number of frames included in the input parameters.

• If a frame has no output object attribute data, it is still counted in numFilled and is represented with an empty list entry (NvMOTTrackedObjList). An empty list entry has the correct streamID set and numFilled set to 0.

  Note

  The output object attribute data NvMOTTrackedObj contains a pointer to the detector object (provied in the input) that is associated with a tracked object, which is stored in associatedObjectIn. You must set this to the associated input object only for the frame where the input object is passed in. For a pipeline with PGIE interval=1, for example:

  • Frame 0: NvMOTObjToTrack X is passed in. The tracker assigns it ID 1, and the output object’s associatedObjectIn points to X.

  • Frame 1: Inference is skipped, so there is no input object from detector to be associated with. The tracker finds Object 1, and the output object’s associatedObjectIn points to NULL.

  • Frame 2: NvMOTObjToTrack Y is passed in. The tracker identifies it as Object 1. The output Object 1 has associatedObjectIn pointing to Y.

4. Depending on the capability of the low-level tracker, there could be some user-defined miscellaneous data to report to tracker plugin. batch_user_meta_list in NvDsBatchMeta as a user-meta:

   NvMOTStatus NvMOT_RetrieveMiscData (
        NvMOTContextHandle contextHandle,
        NvMOTProcessParams *pParams,
        NvMOTTrackerMiscData *pTrackerMiscData
   );
   

where:

• pParams is a pointer to the input batch of frames to process. This structure is needed to check the list of stream ID in the batch.

• pTrackerMiscData is a pointer to the output miscellaneous data for the current batch to be filled by the low-level tracker. The data structure NvMOTTrackerMiscData is defined in nvdstracker.h.

5. In case that a video stream source is removed on the fly, the plugin calls the following function so that the low-level tracker library can remove it as well. Note that this API is optional and valid only when the batch processing mode is enabled, meaning that it will be executed only when the low-level tracker library has an actual implementation for the API. If called, the low-level tracker library can release any per-stream resource that it may be allocated:

   void NvMOT_RemoveStreams (
        NvMOTContextHandle contextHandle,
        NvMOTStreamId streamIdMask
   );
   

6. When all processing is complete, the plugin calls this function to clean up the context and deallocate its resources:

   void NvMOT_DeInit (NvMOTContextHandle contextHandle);
   

NvMultiObjectTracker : A Reference Low-Level Tracker Library

Multi-object tracking (MOT) is a key building block for a large number of intelligent video analytics (IVA) applications that requires spatio-temporal analysis of objects of interest to draw further insights about the objects’ behaviors in long term. Given a set of detected objects from the Primary GIE (PGIE) module on a single or multiple streams and with the APIs defined to work with the tracker plugin, the low-level tracker library is expected to carry out actual multi-object tracking operations to keep persistent IDs to the same objects over time.

DeepStream SDK provides a single reference low-level tracker library, called NvMultiObjectTracker, that implements all four low-level tracking algorithms (i.e., IOU, NvSORT, NvDeepSORT, and NvDCF) in a unified architecture. It supports multi-stream, multi-object tracking in the batch processing mode for efficient processing on CPU and GPU (and PVA for Jetson). The following sections will cover the unified tracker architecture and the details of each reference tracker implementation.

Unified Tracker Architecture for Composable Multi-Object Tracker

In NvMultiObjectTracker low-level tracker library, different types of multi-object trackers share common modules when it comes to basic functionalities (e.g., data association, target management, state estimation, etc.), while differing in other core functionalities (e.g., visual tracking for NvDCF and deep association metric for NvDeepSORT). The NvMultiObjectTracker library employs a unified architecture to allow the composition of a multi-object tracker through configuration by enabling only the modules required for a particular object tracker. The IOU tracker, for example, requires a minimum set of modules that consist of data association and target management modules. On top of that, NvSORT adds a state estimator for more accurate motion estimation & prediction, and NvDeepSORT further introduces a deep Re-ID network to integrate appearance information into data association. Instead of the deep neural network-based Re-ID features in NvDeepSORT, NvDCF employs a Discriminative Correlation Filter (DCF)-based visual tracking module that uses conventional feature descriptors for more efficient tracking. However, NvDCF can still allow the use of Re-ID module for target re-association for longer-term robustness.

The table below summarizes what modules are used to compose each object tracker, showing what modules are shared across different object trackers and how each object tracker differs in module composition:

Tracker Type	State Estimator	Target Management	Visual Tracker	Target Re-Association		Data Association Metric
				Spatio- temporal	Re-ID	Proximity & Size	Visual Similarity	Re-ID
IOU		O				O
NvSORT	O	O				O
NvDeepSORT	O	O				O		O
NvDCF	O	O	O	O	O	O	O

In the following sections, we will first discuss the general work flow of the NvMultiObjectTracker library and its core modules, and then each type of object trackers in more details with explanations on the config params in each module.

Workflow and Core Modules in The NvMultiObjectTracker Library

The input to a low-level tracker library consists of (1) a batch of video frames from a single or multiple streams and (2) a list of detector objects for each video frame. If the detection interval (i.e., interval in Primary GIE section) is set larger than 0, the input data to the low-level tracker would have the detector object data only when the inferencing for object detection is performed for a video frame batch (i.e., the inferenced frame batch). For the frame batches where the inference is skipped (i.e., the uninferenced frame batch), the input data would include only the video frames.

Note

• A detector object refers to an object that is detected by the detector in PGIE module, which is provided to the multi-object tracker module as an input.

• A target refers to an object that is being tracked by the object tracker.

• An inferenced frame is a video frame where an inference is carried out for object detection. Since the inference interval can be configured in setting for PGIE and can be larger than zero, the frameNum of two consecutive inferenced frames may not be contiguous.

For carrying out multi-object tracking operations with the given input data, below are the essential functionalities to be performed. Multithreading is deployed to optimize their performance on CPU.

• Data association between the detector objects from a new video frame and the existing targets for the same video stream

• Target management based on the data association results, including the target state update and the creation and termination of targets

Depending on the tracker types, there could be some addition processing before data association. For example, NvDeepSORT extracts Re-ID features from all the detector objects and computes the similarity, while NvDCF performs the visual tracker based localization so the targets’ predicted locations in a new frame can be used for data association. More details will be covered in each tracker’s section.

Data Association

For data association, various types of similarity metrics are used to calculate the matching score between the detector objects and the existing targets, including:

• Location similarity (i.e., proximity)

• Bounding box size similarity

• Re-ID feature similarity (specific to NvDeepSORT tracker)

• Visual appearance similarity (specific to NvDCF tracker)

For the proximity between detector objects and targets, IOU is a typical metric that is widely used, but it also depends on the size similarity between them. The similarity of the box size between two objects can be used explicitly, which is calculated as the ratio of the size of the smaller box over the larger one.

The total association score for a pair of detector object and target is the weighted sum of all the metrics:

\[totalScore=w_1*IOU+w_2*sizeSimilarity+w_3*reidSimilarity+w_4*visualSimilarity\]

where \(w_i\) is the weight for each metric set in config file. Users can also set a minimum threshold for each similarity and the total score.

During the matching, a detector object is associated with a target that belongs to the same class by default to minimize the false matching. However, this can be disabled by setting checkClassMatch: 0, allowing objects can be associated regardless of their object class IDs. This can be useful when employing a detector like YOLO, which can detect many classes of objects, where there could be false classification on the same object over time.

Regarding the matching algorithm, users can set associationMatcherType as 0 to employ an efficient greedy algorithm for optimal bipartite matching with similarity metrics defined above, or 1 for a newly introduced method named cascaded data association for higher accuracy. The cascaded data association consists of multi-stage matching, assigning different priorities and similarity metrics based on detection and target confidence. Detector objects are split into two sets, confirmed (confidence between [tentativeDetectorConfidence, 1.0]) and tentative (confidence between [minDetectorConfidence, tentativeDetectorConfidence]). Then three stage matching are performed sequentially:

• Confirmed detections and validated (both active and inactive) targets

• Tentative detections and active targets left

• Confirmed detections left and tentative targets

The first stage uses the joint-similarity metrics defined above, while the later two stages only considers the IOU similarity, because proximity can be a more reliable metric than visual similarity or Re-ID when the detection confidence is low due to, say, partial occlusions or noise. Each stage takes different sets of bboxes as candidates and uses the efficient greedy algorithm for matching. The matched pairs are produced from each stage and combined together.

The output of the data association module consists of three sets of objects/targets:

• The unmatched detector objects

• The matched pairs of the detector objects and the existing targets

• The unmatched targets

The unmatched detector objects are among the objects detected by a PGIE detector, yet not associated with any of the existing targets. An unmatched detector object is considered as a newly observed object that needs to be tracked, unless they are determined to be duplicates to any of the existing target. If the maximum IOU score of a new detector object to any of the existing targets is lower than minIouDiff4NewTarget, a new target tracker would be created to track the object since it is not a duplicate to an existing target.

Target Management and Error Handling

Although a new object is detected by the detector (i.e., a detector object), there is a possibility that this may be a false positive. To suppress such noise in detection, the NvMultiObjectTracker tracker library employs a technique called Late Activation, where a newly detected object is examined for a period of time and activated for long-term tracking only if it survives such a period. To be more specific, whenever a new object is detected, a new tracker is created to track the object, but the target is initially put into the Tentative mode, which is a probationary period, whose length is defined by probationAge under TargetManagement section of the config file. During this probationary period, the tracker output will not be reported to the downstream, since the target is not validated yet; however, those unreported tracker output data (i.e., the past-frame data) are stored within the low-level tracker for later report.

The same target may be detected for the next frame; however, there could be false negative by the detector (i.e., missed detection), resulting in a unsuccessful data association to the target. The NvMultiObjectTracker library employs another technique called Shadow Tracking, where a target is still being tracked in the background for a period of time even when the target is not associated with a detector object. Whenever a target is not associated with a detector object for a given time frame, an internal variable of the target called shadowTrackingAge is incremented. Once the target is associated with a detector object, then shadowTrackingAge will be reset to zero.

If the target is in the Tentative mode and the shadowTrackingAge reaches earlyTerminationAge specified in the config file, the target will be terminated prematurely (which is referred to as Early Termination). If the target is not terminated during the Tentative mode and successfully assocciated with a detector object, the target is activated and put into the Active mode, starting to report the tracker outputs to the downstream. If the past-frame data is enabled, the tracked data during the Tentative mode will be reported as well, since they were not reported yet. Once a target is activated (i.e., in Active mode), if the target is not associated for a given time frame (or the tracker confidence gets lower than a threshold), it will be put into the Inactive mode, and its shadowTrackingAge will be incremented, yet still be tracked in the background. However, the target will be terminated if the shadowTrackingAge exceeds maxShadowTrackingAge.

The state transitions of a target tracker are summarized in the following diagram:

The NvMultiObjectTracker library can generate a unique ID to some extent. If enabled by setting useUniqueID: 1, each video stream will be assigned a 32-bit long random number during the initialization stage. All the targets created from the same video stream will have the same upper 32-bit of the uint64_t-type target ID set by the per-stream random number. In the meantime, the lower 32-bit of the target ID starts from 0. The randomly generated upper 32-bit number allows the target IDs from a particular video stream to increment from a random position in the possible ID space. If disabled (i.e., useUniqueID: 0, which is the default value), both the upper and lower 32-bit will start from 0, resulting in the target ID to be incremented from 0 for every run.

Note that the incrementation of the lower 32-bit of the target ID is done across the whole video streams in the same NvMultiObjectTracker library instantiation. Thus, even if the unique ID generation is disabled, the tracker IDs will be unique for the same pipeline run. If the unique ID generation is disabled, and if there are three objects for Stream 1 and two objects for Stream 2, for example, the target IDs will be assigned from 0 to 4 (instead of 0 to 2 for Stream 1 and 0 to 1 for Stream 2) as long as the two streams are being processed by the same library instantiation.

preserveStreamUpdateOrder controls whether to use single or multiple threads to update targets. If it is enabled, new IDs are generated sequentially following input stream ID order in each batch using a single thread, i.e. the objects for Stream 1 and 2 will have IDs from 0 to 2 and 3 to 4 respectively. By default, this option is disabled so target management is done with multi-threads to enable better performance but the ID order is not preserved. If the user needs consistent IDs over multiple runs for the same video source, please set preserveStreamUpdateOrder: 1 and batched-push-timeout=-1 in deepstream-app config.

The NvMultiObjectTracker library pre-allocates all the GPU memories during initialization based on:

• The number of streams to be processed

• The maximum number of objects to be tracked per stream (denoted as maxTargetsPerStream)

Thus, the CPU/GPU memory usage by the NvMultiObjectTracker library is almost linearly proportional to the total number of objects being tracked, which is (number of video streams) × (maxTargetsPerStream), except the scratch memory space used by dependent libraries (such as cuFFT™, TensorRT™, etc.). Thanks to the pre-allocation of all the necessary memory, the NvMultiObjectTracker library is not expected to have memory growth during long-term run even when the number of objects increases over time.

Once the number of objects being tracked reaches the configured maximum value (i.e., maxTargetsPerStream), any new objects will be discarded until some of the existing targets are terminated. Note that the number of objects being tracked includes the targets that are being tracked in the shadow tracking mode. Therefore, NVIDIA recommends that users set maxTargetsPerStream large enough to accommodate the maximum number of objects of interest that may appear in a frame, as well as the objects that may have been tracked from the past frames in the shadow tracking mode.

The minDetectorConfidence property under BaseConfig section in a low-level tracker config file sets the confidence level below which the detector objects are filtered out.

State Estimation

The NvMultiObjectTracker library employs two types of state estimators, both of which are based on Kalman Filter (KF): Simple-bbox KF, Regular-bbox KF, and Simple-location KF. The Simple-bbox KF has 6 states defined, which are {x, y, w, h, dx, dy}, where x and y indicate the coordinates of the top-left corner of a target bbox, while w and h the width and the height of the bbox, respectively. dx and dy denote the velocity of x and y states. The Regular-bbox KF, on the other hand, have 8 states defined, which are {x, y, w, h, dx, dy, dw, dh}, where dw and dh are the velocity of w and h states and the rest is the same as the Simple-bbox KF. The Simple-location KF has 4 states only, which are {x, y, dx, dy}. Unlike the the two KFs that estimate the bbox attributes, note that the Simple-location KF is meant to estimate the object location in either 2D camera image plane or 3D world ground plane.

All the Kalman Filter types employ a constant velocity model for generic use. The measurement vector for the Simple-bbox and Regular-bbox KF is defined as {x, y, w, h}, which represents the bbox attributes, while that for the Simple-location KF is defined as {x, y}. There is an option to use bbox aspect ratio a and its velocity da instead of w and dw when useAspectRatio is enabled, which is specially used by NvDeepSORT. In case the state estimator is used for a generic use case (like in the NvDCF tracker), the process noise variance for {x, y}, {w, h}, and {dx, dy, dw, dh} can be configured by processNoiseVar4Loc, processNoiseVar4Size, and processNoiseVar4Vel, respectively.

When a visual tracker module is enabled (like in the NvDCF tracker), there could be two different measurements from the state estimator’s point of view: (1) the bbox (or location) from the detector at PGIE and (2) the bbox (or location) from the tracker’s localization. This is because the NvDCF tracker module is capable of localizing targets using its own learned filter. The measurement noise variance for these two different types of measurements can be configured by measurementNoiseVar4Detector and measurementNoiseVar4Tracker. These parameters are expected to be tuned or optimized based on the detector’s and the tracker’s characteristics for better measurement fusion.

The usage of the state estimator in the NvDeepSORT tracker slightly differs from that for the aforementioned generic use case in that it is basically a Regular KF, yet with a couple of differences as per the original paper and the implementation (Check the references in NvDeepSORT Tracker section):

• Use of the aspect ratio a and the height h (instead of w and h) to estimate the bbox size

• The process and measurement noises that are proportional to the bounding box height (instead of constant values)

To allow these differences, the state estimator module in the NvMultiObjectTracker library has a set of additional config parameters:

• useAspectRatio to enable the use of a (instead of w)

• noiseWeightVar4Loc and noiseWeightVar4Vel as the proportion coefficients for the measurement and velocity noise, respectively

Note that if these two parameters are set, the fixed process noise and measurement noise parameters for the generic use cases will be ignored.

Object Re-Identification

Re-identification (Re-ID) uses TensorRT™-accelerated deep neural networks to extract unique feature vectors from detected objects that are robust to spatial-temporal variance and occlusion. It has two use-cases in NvMultiObjectTracker: (1) In NvDeepSORT, the Re-ID similarity is used for data association of objects over consecutive frames.; (2) In target re-association (which will be described in more detail in the following section), the Re-ID features of targets are extracted and kept, so that they can be used for re-association with the same target if they are seemingly lost. reidType selects the mode for each aforementioned use-case.

In the Re-ID module, the detector objects are cropped and resized into the configured input size of the Re-ID network. The parameter keepAspc controls whether the object’s aspect ratio is preserved after cropping. Then NVIDIA TensorRT™ creates an engine from the network, which processes the input in batches and outputs a fixed-dimensional vector for each detector object as the Re-ID feature. The cosine similarity function requires each feature’s L2 norm normalized to 1. Check Re-ID Feature Output on how to retrieve these features in the tracker plugin and downstream modules. For each target, a gallery of its Re-ID features in most recent frames are kept internally. The size of the feature gallery can be set by reidHistorySize.

Note

config_tracker_NvDeepSORT.yml and config_tracker_NvDCF_accuracy.yml configs use ReIdentificationNet by default, which is a ResNet-50 Re-ID network in NVIDIA TAO toolkit on NGC. Users need to follow instructions in Setup Sample Re-ID Models to setup, or check Customize Re-ID Model for more information on adding a custom Re-ID model for object tracking with different architectures and datasets.

The Re-ID similarity between a detector object and a target is the cosine similarity between the detector object’s Re-ID feature and its nearest neighbor in the target’s featue gallery, whose value is in range [0.0, 1.0]. Specifically, each Re-ID feature in the target’s gallery takes the dot product with the detector object’s Re-ID feature. The maximum of all the dot products is the similarity score, i.e.

\[score_{ij}=\max_{k}(feature\_det_{i}\cdot feature\_track_{jk})\]

where:

• \(\cdot\) denotes the dot product.

• \(feature\_det_{i}\) denotes the i-th detector object’s feature.

• \(feature\_track_{jk}\) denotes the k-th Re-ID feature in the j-th target’s feature gallery. \(k\) =[1, reidHistorySize].

The Re-ID has a spatial-temporal constraint. If an object moves out of frame or gets occluded beyond maxShadowTrackingAge, it will be assigned a new ID even if it returns into the frame.

The extracted Re-ID features (i.e., embeddings) can be exported to the metadata, which is explained in a separate section in Re-ID Feature Output.

Target Re-Association

The target re-association algorithm enhances the long-term robustness of multi-object tracking by jointly using the Re-ID and spatio-temporal (i.e., motion) features. It addresses one of the major tracking failure cases that occurs in the situation where objects undergo partial- or full-occlusions in a gradual or abrupt manner. During this course of action, the detector at PGIE module may capture only some part of the objects (due to partial visibility), resulting in ill-sized, ill-centered boxes on the target. Later, the target cannot be associated with the object appearing again due to the size and location prediction errors, potentially causing tracking failures and ID switches. Such a re-association problem can typically be handled as a post-processing; however, for real-time analytics applications, this is often expected to be handled seamlessly as a part of the real-time multi-object tracking.

The target re-association takes advantage of the Late Activation and Shadow Tracking in target management module. It tries to associate the newly-appeared targets with previously lost targets based on motion and Re-ID similarity in a seamless, real-time manner by the following steps:

Tracklet Prediction: Whenever an existing target is not associated with a detector object for a prolonged period (same as probationAge), it is considered that the target is lost. While the visual tracker module keeps track of the target in the shadow tracking mode, a length of the predicted tracklet (configured by trajectoryProjectionLength) is generated using some of the recently matched tracklet points (whose length is set by prepLength4TrajectoryProjection) and stored into an internal database until it is matched again with a detector object or re-associated with another target.

Re-ID Feature Extraction: Before a target is lost, the Re-ID network extracts its Re-ID feature with the frame interval of reidExtractionInterval and stores them in the feature gallery. These features will be used to identify target re-appearance in the tracklet matching stage.

Target ID Acquisition: When a new target is instantiated, its validity is examined for a few frames (i.e., probationAge) and a target ID is assigned only if validated (i.e., Late Activation), after which the target state report starts. During the target ID acquisition, the new target is examined if it matches with one of the predicted tracklets from the existing targets in the internal database where the aforementioned predicted tracklets are stored. If matched, it would mean that the new target is actually the re-appearance of a disappeared target in the past. Then, the new target is re-associated with the existing target and its tracklet is fused into that as well. Otherwise, a new target ID is assigned.

Tracklet Matching: During the tracklet matching process in the previous step, the valid candidate tracklets are queried from the database based on the feasible time window configured by maxTrackletMatchingTimeSearchRange. For the new target and each candidate, both the motion and Re-ID similarity are taken into account for tracklet matching. The motion similarity is the average IOU along the tracklet with various criteria including the minimum average IOU score (i.e., minTrackletMatchingScore), maximum angular difference in motion (i.e., maxAngle4TrackletMatching), minimum speed similarity (i.e., minSpeedSimilarity4TrackletMatching), and minimum bbox size similarity (i.e., minBboxSizeSimilarity4TrackletMatching) computed by a Dynamic Time Warping (DTW)-like algorithm. The Re-ID similarity is the cosine distance between the new target’s Re-ID feature and its nearest neighbor in the candidate’s feature gallery. The total similarity score is the weighted sum of both metrics:

\[totalScore=w_1*IOU+w_2*reidSimilarity\]

where \(w_i\) is the weight for each metric set in config file. Users can also set a minimum threshold for each similarity and the total score.

Tracklet Fusion: Once two tracklets are associated, they are fused together to generate one smooth tracklet based on the matching status with detector and the confidence at each point.

config_tracker_NvDCF_accuracy.yml provides an example to enable this feature. Since Re-ID is computationally expensive, users may choose to increase reidExtractionInterval to improve performance or set the parameters like below (i.e., disabling Re-ID feature extraction) to use motion-only target re-association without Re-ID.

TrajectoryManagement:
  useUniqueID: 0      # Use 64-bit long Unique ID when assignining tracker ID. Default is [true]
  enableReAssoc: 1    # Enable Re-Assoc
  minMatchingScore4Overall: 0    # min matching score for overall
  minTrackletMatchingScore: 0.5644    # min tracklet similarity score for re-assoc
  matchingScoreWeight4TrackletSimilarity: 1.0    # weight for tracklet similarity score
  minTrajectoryLength4Projection: 36    # min trajectory length required to make projected trajectory
  prepLength4TrajectoryProjection: 50    # the length of the trajectory during which the state estimator is updated to make projections
  trajectoryProjectionLength: 94    # the length of the projected trajectory
  maxAngle4TrackletMatching: 106    # max angle difference for tracklet matching [degree]
  minSpeedSimilarity4TrackletMatching: 0.0967    # min speed similarity for tracklet matching
  minBboxSizeSimilarity4TrackletMatching: 0.5577    # min bbox size similarity for tracklet matching
  maxTrackletMatchingTimeSearchRange: 20    # the search space in time for max tracklet similarity
  trajectoryProjectionProcessNoiseScale: 0.0100    # trajectory projector's process noise scale w.r.t. state estimator
  trajectoryProjectionMeasurementNoiseScale: 100    # trajectory projector's measurement noise scale w.r.t. state estimator
  trackletSpacialSearchRegionScale: 0.2598    # the search region scale for peer tracklet

ReID:
  reidType: 0    # The type of reid among { DUMMY=0, NvDEEPSORT=1, Reid based reassoc=2, both NvDEEPSORT and reid based reassoc=3}

Note

Target re-association can be effective only when the state estimator is enabled, otherwise the tracklet prediction will not be made properly. The parameters provided above is tuned for PeopleNet v2.6.2, and it may not work as expected for other types of detectors.

Bounding-box Unclipping

Another small experimental feature is the bounding box unclipping. If a target is fully visible within the field-of-view (FOV) of the camera but starts going out of the FOV, the target would be partially visible and the bounding box (i.e., bbox) may capture only a part of the target (i.e., clipped by the FOV) until it fully exits the scene. If it is expected that the size of the bbox doesn’t change much around the border of the video frame, the full bbox can be estimated beyond the FOV limit using the bbox size estimated when the target was fully visible. This feature can be enabled by setting enableBboxUnClipping: 1 under TargetManagement module in the low-level config file.

Single-View 3D Tracking (Alpha)

As mentioned earlier, partial occlusion is one of the most challenging problems that object trackers have to deal with and often lead to tracking failures. If the object detectors capture only the visible part of the object (which is often the case), the partial occlusion would cause the detection bboxes to have abrupt or gradual changes in attributes in terms of bbox location, size, aspect ratio, confidence, and most importantly the visual appearance within the bbox. Considering the object trackers rely on the bbox attributes as spatio-temporal measure and the visual appearance (e.g., ReID embedding) extracted within the bbox as visual similarity measure, such changes in bbox attributes is a major source of tracking failures, resulting in more frequent ID switches.

To tackle these challenging problems, DeepStream SDK introduced a new feature called the Single-View 3D Tracking (SV3DT) that allows the object tracking to be carried out in a 3D world coordinate system (instead of the 2D camera image plane) when (1) a 3x4 projection matrix and (2) a 3D model info are provided for a video stream in a camera info file like below.

# camInfo-01.yml

# The 3x4 camera projection matrix (in row-major):
#    996.229 -202.405  -9.121 -1.185
#    105.309  478.174 890.944 1.743
#    -0.170   -0.859   0.481  -1085.484
projectionMatrix_3x4:
- 996.229
- -202.405
- -9.121
- -1.185
- 105.309
- 478.174
- 890.944
- 1.743
- -0.170
- -0.859
- 0.481
- -1085.484

# The cylindrical human model
modelInfo:
  height: 250.0
  radius: 30.0

There are two options (projectionMatrix_3x4 and projectionMatrix_3x4_w2p) in which users can provide corresponding 3x4 camera projection matrices to support different usecase. Please refer to The 3x4 Camera Projection Matrix section for more details.

Note that there are a few assumptions that this algorithm requires:

• A human is modeled as a cylinder with height and radius in 3D world coordinate system, and the height and radius are provided as 3D model info.

• A 3x4 projection matrix (that transforms a 3D world coordinate point to a 2D camera image coordinate point) is provided for a video stream or a camera.

• Video streams are captured from cameras that are mounted higher than the human height.

The third assumption ensures that when a human is partially-occluded, the head is still visible, allowing us to use the top bbox edge as an anchor, which we will touch upon shortly.

For each detection bounding box for a person, SV3DT algorithm tries to fit the 3D human model to the detection bbox in such a way that the bounding box of the projected 3D human model from the world coordinate system to the camera image plane matches with the detection bbox.

A figure below shows how cylindrical 3D human models can be fitted to the input detection bboxes.

In the cases where a person is partially occluded, the top edge of the detection bbox is used as an anchor to align the bbox of the projected 3D human model. Once aligned, we can recover the full-body bbox using the projected 3D human model, as if the person is not occluded. Therefore, if SV3DT is enabled, the input detection bboxes are always first recovered to the full-body bboxes based on the provided 3D model info especially when the input detection bboxes capture only the visible part of the person due to partial occlusions. This greatly enhances the multi-object tracking accuracy and robustness, since the bbox attributes are not altered during the course of partial occlusions.

An animated image below shows how cylindrical 3D human models can be fitted into the input detection bboxes when the persons are partially occluded. The thin, gray bboxes on the persons indicate the input detection bboxes, which capture only the visible part of the objects. The figure demonstrates that the SV3DT algorithm is still able to estimate the accurate foot location of each person. The person trajectories are drawn based on the estimated foot locations, allowing robust spatio-temporal behavior analytics of persons in the scene despite varying degree of occlusions. Some of the persons in this example are barely seen only on the head and shoulders, but they are being successfully tracked as if not occluded at all.

As a derived metric, the ratio between the bbox for the visible part and the bbox for the projected 3D human model can be considered as an approximated visibility of the object, which could be a useful information.

Users can still get access to the corresponding detection bboxes by checking out detector_bbox_info in NvDsObjectMeta.

To enable SV3DT feature, we introduced a new section in tracker config files, ObjectModelProjection, like below:

ObjectModelProjection:
  cameraModelFilepath:  # In order of the source streams
    - 'camInfo-01.yml'
    - 'camInfo_02.yml'
    ...

Every camera view is different, so the 3x4 projection matrix is supposed to be unique to each camera. Therefore, the camera info file (e.g., camInfo-01.yml) is to be provided for each stream, which include the 3x4 projection matrix and the model info that are shown at the beginning of the section.

Once a 3D human model corresponding to an input detection bbox is estimated and located in the world coordinate system, the foot location (i.e., the center of the base of the cylindrical model) of a person on the world ground plane is what we want to keep estimating because it is a physical state that better follows the motion dynamics modeling than the motion of the object on 2D camera image plane. To perform the state estimation of the foot location of the objects on a 3D world ground plane, users need to set the state estimator type as stateEstimatorType: 3 like below:

StateEstimator:
  stateEstimatorType: 3    # the type of state estimator among { DUMMY_ESTIMATOR=0, SIMPLE_BBOX_KF=1, REGULAR_BBOX_KF=2, SIMPLE_LOCATION_KF=3 }

  # [Dynamics Modeling]
  processNoiseVar4Loc: 6810.866    # Process noise variance for location
  processNoiseVar4Vel: 1348.487    # Process noise variance for velocity
  measurementNoiseVar4Detector: 100.000   # Measurement noise variance for detector's detection
  measurementNoiseVar4Tracker: 293.323    # Measurement noise variance for tracker's localization

An additional miscellaneous data that can be useful for debugging or visualization is the convex hull of the projected 3D human model on 2D camera image plane. By using the convex hull data for each object, users can also create the visualization like the figures above.

The additional miscellaneous data that is generated when SV3DT is enabled include (1) visibility, (2) foot location in both world plane and 2D image, and (3) convex hull (human cylinders projected on 2D image). These data can be saved in text files and/or outputted to object meta for downstream usage. To do that, users would need to set outputVisibility: 1, outputFootLocation: 1, outputConvexHull: 1 in ObjectModelProjection section respectively. The sample use cases includes saving in terminated track dump for low level tracker, attaching in KITTI track dump for deepstream-app, and converting them in schema through Gst-nvmsgconv.

To allow users to easily try out and experience SV3DT, a sample usecase for SV3DT has been hosted on GitHub. So, users can just clone and run it with the sample data provided.

The 3x4 Camera Projection Matrix

The 3x4 Camera Projection Matrix is also called as simply the camera matrix, which is a 3x4 matrix that converts a 3D world point to a 2D point on camera image plane based on a pinhole camera model like shown in the figure below:

Pinhole Camera Model. Source: OpenCV Doc and License

More detailed and general information about the camera matrix can be found in various sources that deal with the computer vision geometries and camera calibration, including OpenCV’s documentation on Camera Calibration.

For projectionMatrix_3x4 in a camera model file (e.g., camInfo-01.yml), the principal point (i.e., (Cx, Cy)) in the camera matrix is assumed to be at (0, 0) as image coordinates. But, the optical center (i.e., (Cx, Cy)) is located at the image center (i.e., (img_width/2, img_height/2)). Thus, to move the origin to the left-top of the camera image (i.e., the pixel coordinates), SV3DT internally adds (img_width/2, img_height/2) after the transformation using the camera matrix provided in projectionMatrix_3x4.

In case that the 3x4 camera projection matrix already accounts for such translation of the principal point, users can provide the camera matrix in projectionMatrix_3x4_w2p instead. This assumes that the 3x4 camera projection matrix transforms a 3D world point directly into an actual pixel point whose origin is at the left-top corner of the image, so it does not require any further translation of the principal point.

Configuration Parameters

The following table summarizes the configuration parameters for the common modules in the NvMultiObjectTracker low-level tracker library.

Configuration properties in Common Modules in NvMultiObjectTracker low-level tracker library

Module	Property	Meaning	Type and Range	Default value
Base Config	minDetectorConfidence	Minimum detector confidence for a valid object	Float, -inf to inf	minDetectorConfidence: 0.0
Target Management	preserveStreamUpdateOrder	Whether to ensure target ID update order the same as input stream ID order	Boolean	preserveStreamUpdateOrder: 0
	maxTargetsPerStream	Max number of targets to track per stream	Integer, 0 to 65535	maxTargetsPerStream: 30
	minIouDiff4NewTarget	Min IOU to existing targets for discarding new target	Float, 0 to 1	minIouDiff4NewTarget: 0.5
	enableBboxUnClipping	Enable bounding-box unclipping	Boolean	enableBboxUnClipping: 0
	probationAge	Length of probationary period in #of frames	Integer, ≥0	probationAge: 5
	maxShadowTrackingAge	Maximum length of shadow tracking	Integer, ≥0	maxShadowTrackingAge: 38
	earlyTerminationAge	Early termination age	Integer, ≥0	earlyTerminationAge: 2
	outputTerminatedTracks	Output total frame history for terminated tracks to the tracker plugin for downstream usage	Boolean	outputTerminatedTracks: 0
	outputShadowTracks	Output shadow track state information to the tracker plugin for downstream usage	Boolean	outputShadowTracks: 0
	terminatedTrackFilename	File name prefix to save terminated tracks	String	terminatedTrackFilename: “”
Trajectory Management	useUniqueID	Enable unique ID generation scheme	Boolean	useUniqueID: 0
	enableReAssoc	Enable motion-based target re-association	Boolean	enableReAssoc: 0
	minMatchingScore4Overall	Min total score for re-association	Float, 0.0 to 1.0	minMatchingScore4Overall: 0.4
	minTrackletMatchingScore	Min tracklet similarity score for matching in terms of average IOU between tracklets	Float, 0.0 to 1.0	minTrackletMatchingScore: 0.4
	minMatchingScore4ReidSimilarity	Min ReID score for re-association	Float, 0.0 to 1.0	minMatchingScore4ReidSimilarity: 0.8
	matchingScoreWeight4TrackletSimilarity	Weight for tracklet similarity term in re-assoc cost function	Float, 0.0 to 1.0	matchingScoreWeight4TrackletSimilarity: 1.0
	matchingScoreWeight4ReidSimilarity	Weight for ReID similarity term in re-assoc cost function	Float, 0.0 to 1.0	matchingScoreWeight4ReidSimilarity: 0.0
	minTrajectoryLength4Projection	Min tracklet length of a target (i.e., age) to perform trajectory projection [frames]	Integer, >=0	minTrajectoryLength4Projection: 20
	prepLength4TrajectoryProjection	Length of the trajectory during which the state estimator is updated to make projections [frames]	Integer, >=0	prepLength4TrajectoryProjection: 10
	trajectoryProjectionLength	Length of the projected trajectory [frames]	Integer, >=0	trajectoryProjectionLength: 90
	maxAngle4TrackletMatching	Max angle difference for tracklet matching [degree]	Integer, [0, 180]	maxAngle4TrackletMatching: 40
	minSpeedSimilarity4TrackletMatching	Min speed similarity for tracklet matching	Float, 0.0 to 1.0	minSpeedSimilarity4TrackletMatching: 0.3
	minBboxSizeSimilarity4TrackletMatching	Min bbox size similarity for tracklet matching	Float, 0.0 to 1.0	minBboxSizeSimilarity4TrackletMatching: 0.6
	maxTrackletMatchingTimeSearchRange	Search space in time for max tracklet similarity	Integer, >=0	maxTrackletMatchingTimeSearchRange: 20
	trajectoryProjectionProcessNoiseScale	Trajectory state estimator’s process noise scale	Float, 0.0 to inf	trajectoryProjectionProcessNoiseScale: 1.0
	trajectoryProjectionMeasurement NoiseScale	Trajectory state estimator’s measurement noise scale	Float, 0.0 to inf	trajectoryProjectionMeasurement NoiseScale: 1.0
	trackletSpacialSearchRegionScale	Re-association peer tracklet search region scale	Float, 0.0 to inf	trackletSpacialSearchRegionScale: 0.0
	reidExtractionInterval	Frame interval to extract ReID features per target for re-association; -1 means only extracting the beginning frame per target	Integer, ≥-1	reidExtractionInterval: 0
Data Associator	associationMatcherType	Type of matching algorithm { GREEDY=0, CASCADED=1 }	Integer, [0, 1]	associationMatcherType: 0
	checkClassMatch	Enable associating only the same-class objects	Boolean
	minMatchingScore4Overall	Min total score for valid matching	Float, 0.0 to 1.0	minMatchingScore4Overall: 0.0
	minMatchingScore4SizeSimilarity	Min bbox size similarity score for valid matching	Float, 0.0 to 1.0	minMatchingScore4SizeSimilarity: 0.0
	minMatchingScore4Iou	Min IOU score for valid matching	Float, 0.0 to 1.0	minMatchingScore4Iou: 0.0
	matchingScoreWeight4SizeSimilarity	Weight for size similarity term in matching cost function	Float, 0.0 to 1.0	matchingScoreWeight4SizeSimilarity: 0.0
	matchingScoreWeight4Iou	Weight for IOU term in matching cost function	Float, 0.0 to 1.0	matchingScoreWeight4Iou: 1.0
	tentativeDetectorConfidence	If a detection’s confidence is lower than this but higher than minDetectorConfidence, then it’s considered as a tentative detection	Float, 0.0 to 1.0	tentativeDetectorConfidence: 0.5
	minMatchingScore4TentativeIou	Min iou threshold to match targets and tentative detection	Float, 0.0 to 1.0	minMatchingScore4TentativeIou: 0.0
State Estimator	stateEstimatorType	Type of state estimator among { DUMMY=0, SIMPLE=1, REGULAR=2, SIMPLE_LOC=3 }	Integer, [0,3]	stateEstimatorType: 0
	processNoiseVar4Loc	Process noise variance for bbox center	Float, 0.0 to inf	processNoiseVar4Loc: 2.0
	processNoiseVar4Size	Process noise variance for bbox size	Float, 0.0 to inf	processNoiseVar4Size: 1.0
	processNoiseVar4Vel	Process noise variance for velocity	Float, 0.0 to inf	processNoiseVar4Vel: 0.1
	measurementNoiseVar4Detector	Measurement noise variance for detector’s detection	Float, 0.0 to inf	measurementNoiseVar4Detector: 4.0
	measurementNoiseVar4Tracker	Measurement noise variance for tracker’s localization	Float, 0.0 to inf	measurementNoiseVar4Tracker: 16.0
	noiseWeightVar4Loc	Noise covariance weight for bbox location; if set, location noise will be proportional to box height	Float, >0.0 considered as set	noiseWeightVar4Loc: -0.1
	noiseWeightVar4Vel	Noise covariance weight for bbox velocity; if set, location noise will be proportional to box height	Float, >0.0 considered as set	noiseWeightVar4Vel: -0.1
	useAspectRatio	Use aspect ratio in Kalman Filter’s states	Boolean	useAspectRatio: 0
Object Re-ID	reidType	The type of Re-ID network among { DUMMY=0, NvDEEPSORT=1, Reid based reassoc=2, both NvDEEPSORT and reid based reassoc=3 }	Integer, [0, 3]	reidType: 0
	batchSize	Batch size of Re-ID network	Integer, >0	batchSize: 1
	workspaceSize	Workspace size to be used by Re-ID TensorRT engine, in MB	Integer, >0	workspaceSize: 20
	reidFeatureSize	Size of Re-ID feature	Integer, >0	reidFeatureSize: 128
	reidHistorySize	Size of feature gallery, i.e. max number of Re-ID features kept for one tracker	Integer, >0	reidHistorySize: 100
	inferDims	Re-ID network input dimension CHW or HWC based on inputOrder	Integer, >0	inferDims: [128, 64, 3]
	inputOrder	Re-ID network input order {NCHW=0, NHWC=1}	Integer, [0, 1]	inputOrder: 1
	colorFormat	Re-ID network input color format among {RGB=0, BGR=1 }	Integer, [0, 1]	colorFormat: 0
	networkMode	Re-ID network inference precision mode among {FP32=0, FP16=1, INT8=2 }	Integer, [0, 1, 2]	networkMode: 0
	offsets	Array of values to be subtracted from each input channel, with length equal to number of channels	Comma delimited float array	offsets: [0.0, 0.0, 0.0]
	netScaleFactor	Scaling factor for Re-ID network input after substracting offsets	Float, >0	netScaleFactor: 1.0
	addFeatureNormalization	If Re-ID network’s output Re-ID feature vector is not l2 normalized, explicitly performs l2 normalization	Boolean	addFeatureNormalization: 0
	tltEncodedModel	Pathname of the TAO toolkit encoded model	String	tltEncodedModel: “”
	tltModelKey	Key for the TAO toolkit encoded model	String	tltModelKey: “”
	onnxFile	Pathname of the ONNX model file	String	onnxFile: “”
	modelEngineFile	Absolute path to Re-ID engine file	String	modelEngineFile:””
	calibrationTableFile	Absolute path to calibration table, required by INT8 only	String	calibrationTableFile:””
	keepAspc	Whether to keep aspcect ratio when resizing input objects to Re-ID network	Boolean	keepAspc: 1
	outputReidTensor	Output Re-ID features to user meta for downstream usage	Boolean	outputReidTensor: 0
	useVPICropScaler (Alpha feature)	Use NVIDIA’s VPI™ Crop Scaler algorithm instead of built in implementation	Boolean	useVPICropScaler: 0
Object Model Projection	cameraModelFilepath	A list of file paths to camera info files. A valid camera info file should be provided to each video stream	String	cameraModelFilepath: “”
	outputVisibility	Output object visibility to object meta and file dump	Boolean	outputVisibility: 0
	outputFootLocation	Output object (especially for human) foot location to object meta and file dump	Boolean	outputFootLocation: 0
	outputConvexHull	Output projected object convex hull (especially cylinder for human) to object meta and file dump	Boolean	outputConvexHull: 1
	maxConvexHullSize	Maximum number of points to consist an object convex hull	Integer, >0	maxConvexHullSize: 15

Low-Level Tracker Comparisons and Tradeoffs

DeepStream SDK provides four reference low-level tracker libraries which have different resource requirements and performance characteristics, in terms of accuracy, robustness, and efficiency, allowing the users to choose the best tracker based on their use cases and requirements. See the following table for comparison.

Gst-nvtracker Tracker library comparison

Tracker Type	GPU Compute	CPU Compute	Pros	Cons	Best Use Cases
IOU	No	Very Low	Light weight	No visual features for matching, so prone to frequent tracker ID switches and failures. Not suitable for fast moving scene.	Objects are sparsely located, with distinct sizes Detector is expected to run every frame or very frequently (ex. every alternate frame)
NvSORT	No	Very Low	Light weight Motion estimation and prection with Kalman filter based state estimator Enhanced data association with reasonable tracking accuracy with medium or high accuracy detectors	No visual features for matching, so prone to frequent tracker ID switches and failures.	Medium or high accuracy detectors such as PeopleNet and YOLO Detector is expected to run every frame or very frequently (ex. every alternate frame)
NvDeepSORT	High	Low	Allow custom Re-ID model for appearance matching Less frequent ID switches Highly discriminative on targets with similar appearance	Higher compute cost due to inferencing required for each object Cannot easily adjust accuracy/perf level unless switching Re-ID model	Multi-object, complex scenes even with partial occlusion Objects with similar appearance
NvDCF	Medium	Low	Highly robust against partial occlusions, shadow, and other transient visual changes Less frequent ID switches Can be used with PGIE interval > 0 without significant accuracy degradation Output tracking confidence based on visual similarity Easily adjust params for accuracy-perf tradeoff depending on application requirements	Slower than IOU, NvSORT due to increased computational complexity for visual feature extraction	Multi-object, complex scenes even with partial occlusion PGIE interval > 0

IOU Tracker

The NvMultiObjectTracker library provides an object tracker that has only the essential and minimum set of functionalities for multi-object tracking, which is called the IOU tracker. IOU tracker performs only the following functionalities:

• Greedy data association between the detector objects from a new video frame and the existing targets in the previous video frame

• Target management based on the data association results including the target state update and the creation and termination of targets

The error handling mechanisms like Late Activation and Shadow Tracking are integral part of the target management module of the NvMultiObjectTracker library; thus, such features are inherently enabled in the IOU tracker.

IOU tracker can be used as a performance baseline as it consumes the minimum amount of computational resources. A sample configuration file config_tracker_IOU.yml is provided in DeepStream SDK package.

NvSORT Tracker

NvSORT tracker increases the tracking accuracy while maintaining the high performance on top of IOU tracker with the following improvements:

• State estimation with Kalman filter to better estimate and predict the states of the targets in the current frame.

• Cascaded data association to associate targets and detector objects in multiple stages based on their proximity and confidence, which is more accurate than the simple matching in original SORT tracker.

As it fully relies on the bbox attributes for data association, the NvSORT’s tracking accuracy is solely attributed to the detection accuracy. With a medium or high accuracy detector, NvSORT produces high quality tracking results with minimal computational resources. A sample configuration file config_tracker_NvSORT.yml is provided in DeepStream SDK package.

NvDeepSORT Tracker

NvDeepSORT tracker utilizes deep learning based object appearance information for accurate object matching in different frames and locations, resulting in enhanced robustness over occlusions and reduced ID switches. It applies a pre-trained re-identification (Re-ID) neural network to extract a feature vector for each object, compares the similarity between different objects using the extracted feature vector with a cosine distance metric, and combines it with a state estimator to perform the data association over frames. Before running NvDeepSORT, Re-ID model needs to be set up following Setup Sample Re-ID Models and Customize Re-ID Model.

Data Association

For the data association in the NvDeepSORT tracker, there are two metrics are used:

• Proximity

• Re-ID based similarity

For the proximity score, the Mahalanobis distance between the i-th detector object and the j-th target is calculated using the target’s predicted location and its associated uncertainty:

\[dist_{ij}=(D_i-Y_j)^TS_j^{-1}(D_i-Y_j)\]

where:

• \(D_i\) denotes the i-th detector object in {x, y, a, h} format.

• \(Y_j\) denotes the predicted states {x', y', a', h'} from state estimator for the j-th tracker.

• \(S_j\) denotes the predicted covariance from state estimator for the j-th tracker.

In the original DeepSORT implementation, the maximum threshold of Mahalanobis distance for a valid pair of detector object and target is set 9.4877, representing 95% confidence computed from the inverse Chi-square distribution. Note in NvDeepSORT, the value is configured by thresholdMahalanobis in tracker config to achieve higher accuracy for a particular detector model, such as the PeopleNet v2.6.2, so it may be different from the value in the original implementation.

After filtering out invalid pairs, the Re-ID similarity score is computed as the maximum cosine similarity between a detector object and a target. Then the cascaded data association algorithm is used for high accuracy multi-stage matching.

Configuration Parameters

A sample config file config_tracker_NvDeepSORT.yml is provided in DeepStream SDK package. The following table summarizes the configuration parameters for NvDeepSORT.

Gst-nvtracker NvDeepSORT low-level tracker configuration properties

Module	Property	Meaning	Type and Range	Default value
Data Associator	thresholdMahalanobis	Max Mahalanobis distance based on Chi-square probabilities	Float, >0 considered as set	thresholdMahalanobis: -1.0
	minMatchingScore4ReidSimilarity	Min Re-ID threshold to match targets and tentative detection	Float, 0.0 to 1.0	minMatchingScore4ReidSimilarity: 0.0
	matchingScoreWeight4ReIDSimilarity	Weight for Re-ID similarity term in matching cost function	Float, 0.0 to 1.0	matchingScoreWeight4ReIDSimilarity: 0.0

Implementation Details and Reference

The difference between NvDeepSORT and the original implementation includes:

• For data association, the original implementation sorts the targets in an ascending order based on the tracking age and runs the matching algorithm for each age sequentially, while NvDeepSORT applies the cascaded data association algorithm with higher performance and accuracy.

• NvDeepSORT implementation in the NvMultiObjectTracker library adopts the same target management policy as the NvDCF tracker, which is advanced to the original DeepSORT.

• The cosine distance metric for two features is \(score_{ij}=1-feature\_det_{i}\cdot feature\_track_{jk}\), with smaller values representing more similarity. By contrast, NvDeepSORT directly uses dot product for computational efficiency, so larger values means higher similarity.

Reference: Wojke, Nicolai, Alex Bewley, and Dietrich Paulus. “Simple online and real-time tracking with a deep association metric.” 2017 IEEE international conference on image processing (ICIP). IEEE, 2017. Check Paper and The original implementation on Github.

NvDCF Tracker

NvDCF tracker employs a visual tracker that is based on the discriminative correlation filter (DCF) for learning a target-specific correlation filter and for localizing the same target in the next frames using the learned correlation filter. Such correlation filter learning and localization are usually carried out on per-object basis in a typical MOT implementation, creating a potentially large number of small CUDA kernel launches when processed on GPU. This inherently poses challenges in maximizing GPU utilization, especially when a large number of objects from multiple video streams are expected to be tracked on a single GPU.

To address such performance issues, the GPU-accelerated operations for the NvDCF tracker are designed to be executed in the batch processing mode to maximize the GPU utilization despite the nature of small CUDA kernels in per-object tracking model. The batch processing mode is applied in the entire tracking operations, including the bbox cropping and scaling, visual feature extraction, correlation filter learning, and localization. This can be viewed as a similar model to the batched cuFFT or batched cuBLAS calls, but it differs in that the batched MOT execution model spans many operations in a higher level. The batch processing capability is extended from multi-object batching to the batching of multiple streams for even greater efficiency and scalability.

Thanks to its visual tracking capability, the NvDCF tracker can localize and keep track of the targets even when the detector in PGIE misses them (i.e., false negatives) for potentially an extended period of time caused by partial or full occlusions, resulting in more robust tracking. The enhanced robustness characteristics allow users to use a higher maxShadowTrackingAge value for longer-term object tracking and also allows PGIE’s interval to be higher only at the cost of slight degradation in accuracy.

Unlike NvSORT and NvDeepSORT where the Kalman filter takes the detection bboxes as the only input, the Kalman filter in the NvDCF tracker also takes the localization results from the visual tracking module as an input as well. Once a target is being tracked, the visual tracker keeps trying to localize the same target in the next frames using the learned correlation filter, while there could be matched detector bboxes. The Kalman filter in NvDCF tracker fuses both the DCF-based localization results and the detection bboxes for better target state estimation and prediction.

Visual Tracking

For each tracked target, NvDCF tracker defines a search region around its predicted location in the next frame large enough for the same target to be detected in the search region. The location of a target on a new video frame is predicted by using the state estimator module. The searchRegionPaddingScale property determines the size of the search region as a multiple of the diagonal of the target’s bounding box. The size of the search region would be determined as:

\[ \begin{align}\begin{aligned}SearchRegion_{width}=w+searchRegionPaddingScale*\sqrt{w*h}\\SearchRegion_{height}=h+searchRegionPaddingScale*\sqrt{w*h}\end{aligned}\end{align} \]

, where \(w\) and \(h\) are the width and height of the target’s bounding box, respectively.

Once the search region is defined for each target at its predicted location, the image patches from each of the search regions are cropped and scaled to a predefined feature image size, from which the visual features are extracted. The featureImgSizeLevel property defines the size of the feature image, and its range is from 1 to 5. Each level between 1 and 5 corresponds to 12x12, 18x18, 24x24, 30x30, and 36x36, respectively, for each feature channel. A lower value of featureImgSizeLevel causes NvDCF to use a smaller feature size, increasing GPU performance potentially yet at the cost of accuracy and robustness. Consider the relationship between featureImgSizeLevel and searchRegionPaddingScale when configuring the parameters. If searchRegionPaddingScale is increased while featureImgSizeLevel is fixed, the number of pixels corresponding to the target itself in the feature images will be effectively decreased.

For each cropped image patch, the visual appearance features such as ColorNames and/or Histogram-of-Oriented-Gradient (HOG) are extracted. The type of visual features to be used can be configured by setting useColorNames and/or useHog. The HOG features consist of 18 channels based on the number of bins for different orientations, while The ColorNames features have 10 channels. If both features are used (by setting useColorNames: 1 and useHog: 1), the total number of channels would then be 28. Therefore, if one uses both HOG and ColorNames with featureImgSizeLevel: 5, the dimension of visual features that represents a target would be 28x48x48. The more channels of visual features are used, the higher the accuracy would be, but would increase the computational complexity and reduce the performance. The NvDCF tracker uses NVIDIA’s VPI™ library for extracting those visual features.

The correlation filters are generated with an attention window (using a Hanning window) applied at the center of the target bbox. Users are allowed to move the center of the attention window in the vertical direction. For example, featureFocusOffsetFactor_y: -0.2 would result in the center of the attention window to be at y=-0.2 in the feature map, where the relative range of the height is [-0.5, 0.5]. Consider that typical surveillance or CCTV cameras are mounted at a moderately high position to monitor a wide area of the environment, say, a retail store or a traffic intersection. From those vantage points, more occlusions can occur at the lower part of the body of persons or vehicles by other persons or vehicles. Moving the attention window up a bit may improve the accuracy and robustness for those use cases.

Once a correlation filter is generated for a target, typical DCF-based trackers usually employ an exponential moving average for temporal consistency when the optimal correlation filter is created and updated over consecutive frames. The learning rate for this moving average can be configured by filterLr and filterChannelWeightsLr for the correlation filters and their channel weights, respectively. The standard deviation for Gaussian for the desired response used when creating an optimal DCF filter can also be configured by gaussianSigma.

Compute Backends

Visual tracker module in NvDCF supports multiple compute backends: CUDA/GPU and PVA (Programmable Vision Accelerator). So, the users are allowed to use different compute backends depending on the particular use cases.

PVA is an accelerator in Tegra SOC in NVIDIA’s Jetson devices, which is specialized for image processing and computer vision algorithms with extremely low power consumption. When running DeepStream-based pipeline with tracker on Jetson, it is recommended to use the PVA-backend for DCF operations in NvDCF for better power efficiency. Since the GPU-based processing for DCF operations is offloaded onto PVA, therefore, more GPU resources are made available to the users for any downstream or custom processing that requires GPU-based processing.

To employ PVA-backend, the visual tracker module leverages the APIs provided by NVIDIA’s VPI™, which can be enabled by setting visualTrackerType: 2 (Alpha feature) and vpiBackend4DcfTracker: 2 # {CUDA=1, PVA=2} on Jetson platforms. VPI also has CUDA-backend mode, which can configured by setting vpiBackend4DcfTracker: 1 on any supported HW platforms (Jetson or dGPU platforms).

The PVA-backend implementation of DCF operations by VPI™ currently has the following limitations:

• The maximum number of objects that can be supported by a single instance of tracker library is 512. This translates to the following restrictions in DeepStream configuration:

  • When sub-batching is not enabled, the total number of objects being tracked across all the streams (i.e., the number of streams in a batch * maxTargetsPerStream) should be less than or equal to 512

  • When sub-batching is enabled, the total number of objects being tracked across all the streams in a sub-batch (i.e., the number of streams in a sub-batch * maxTargetsPerStream) should be less than or equal to 512

    • If there are N sub-batches, the total number of objects that can be tracked in the pipeline is effectively N*512.

• Supports only one combination of the these three configurations : useColorNames: 1 and useHog: 1 and featureImgSizeLevel: 3

Another limitation of the DCF APIs provided by VPI™ is that, the max number of streams supported by a single library instance is 33. Hence, to run a DS application with higher batch size (>33) using visualTrackerType: 2, it is recommended to use Sub-batching feature in the tracker plugin such that each sub-batch is of size less than or equal to 33 streams.

Data Association

The association of target IDs across frames for robust tracking typically entails visual appearance-based similarity matching, for which the visual appearance features are extracted at each candidate location. Usually, this is a computationally expensive process and often plays as a performance bottleneck in object tracking. Unlike existing approaches that extract visual features from all the candidate locations and perform feature matching among all the candidate objects, the NvDCF tracker takes advantage of the correlation response (that is already obtained during target localization stage) as the tracking confidence map of each tracker over a search region and simply looks up the confidence values at each candidate location (i.e., the location of each detector object) to get the visual similarity without any explicit computation. By comparing those confidences between trackers, we can identify which tracker has a higher visual similarity to a particular detector object and use it as a part of the matching score for data association. Therefore, the visual similarity matching in the data association process can be carried out very efficiently through a simple look-up table (LUT) operation on existing correlation responses.

In the animated figure below, the left side shows the target within its search region, while the right side shows the correlation response map (where the deep red color indicates higher confidence and deep blue indicates lower confidence). In the confidence map, the yellow cross (i.e., +) around the center indicates the peak location of the correlation response, while the purple x indicate the center of nearby detector bboxes. The correlation response values at those purple x locations indicate the confidence score on how likely the same target exists at that location in terms of the visual similarity.

If there are multiple detector bboxes (i.e., purple x) around the target like the one in the figure below, the data association module will take care of the matching based on the visual similairty score and the configured weight and minimum value, which are matchingScoreWeight4VisualSimilarity and minMatchingScore4VisualSimilarity, respectively.

Configuration Parameters

A few sample configuration files for the NvDCF tracker are provided as a part of DeepStream SDK package, which is named as:

• config_tracker_NvDCF_max_perf.yml

• config_tracker_NvDCF_perf.yml

• config_tracker_NvDCF_accuracy.yml

The first max_perf config file is to configure the NvDCF tracker to consume the least amount of resources, while the second perf config file is for the use case where a decent balance between performance and accuracy is required. The last accuracy config file is to maximize the accuracy and robustness by enabling most of the features to their full capability, especially the target re-association.

The following table summarizes the configuration parameters used in the config files for the NvDCF low-level tracker (except the common modules and parameters already mentioned in an earlier section).

Gst-nvtracker NvDCF low-level tracker configuration properties

Module	Property	Meaning	Type and Range	Default value
Visual Tracker	visualTrackerType	Type of visual tracker among { DUMMY=0, NvDCF=1, NvDCF_VPI=2 } NvDCF_VPI is an Alpha feature.	Int, [0, 1, 2]	visualTrackerType: 0
	useColorNames	Use ColorNames feature	Boolean	useColorNames: 1
	useHog	Use Histogram-of-Oriented-Gradient (HOG) feature	Boolean	useHog: 0
	featureImgSizeLevel	Size of a feature image	Integer, 1 to 5	featureImgSizeLevel: 2
	featureFocusOffsetFactor_y	The offset for the center of hanning window relative to the feature height	Float, -0.5 to 0.5	featureFocusOffsetFactor_y: 0.0
	useHighPrecisionFeature	Whether to use 16 bit high precision feature; otherwise use 8 bit	Boolean	useHighPrecisionFeature: 0
	filterLr	Learning rate for DCF filter in exponential moving average	Float, 0.0 to 1.0	filterLr: 0.075
	filterChannelWeightsLr	Learning rate for weights for different feature channels in DCF	Float, 0.0 to 1.0	filterChannelWeightsLr: 0.1
	gaussianSigma	Standard deviation for Gaussian for desired response	Float, >0.0	gaussianSigma: 0.75
	vpiBackend4DcfTracker	Compute backend among {CUDA=1, PVA=2} Valid when visualTrackerType: 2	Int, [1,2]	vpiBackend4DcfTracker: 1
Target Management	searchRegionPaddingScale	Search region size	Integer, 1 to 3	searchRegionPaddingScale: 1
	minTrackerConfidence	Minimum detector confidence for a valid target	Float, 0.0 to 1.0	minTrackerConfidence: 0.6
Data Assoicator	minMatchingScore4 VisualSimilarity	Min visual similarity score for valid matching	Float, 0.0 to 1.0	minMatchingScore4 VisualSimilarity: 0.0
	matchingScoreWeight4 VisualSimilarity	Weight for visual similarity term in matching cost function	Float, 0.0 to 1.0	matchingScoreWeight4 VisualSimilarity: 0.0

See also the Troubleshooting in Tracker Setup and Parameter Tuning section for solutions to common problems in tracker behavior and tuning.

Miscellaneous Data Output

Miscellaneous data provides a mechanism to return additional data to the user outside of the Gst Buffer or NvDsBatchMeta. The miscellaneous data buffer will only be populated when options are enabled for specific feature. Currently the supported types of miscellaneous data are:

• Terminated Track List

• Shadow Tracking Target Data

• Past-frame Target Data

Each type of data has a unique output variable within the miscellaneous data; however, they do share a common data structure, using the unified NvDsTargetMiscDataBatch data structure. A buffer pool is used for its memory management, whose size can be set with user-meta-pool-size. When the latency for downstream plugins to release the buffers is too long, the buffer pool may be empty so tracker will skip reporting the miscellaneous data for next batch. A warning gstnvtracker: Unable to acquire a user meta buffer will be shown, and users can increase the pool size from default 32 to larger values like 64. Specifics for each type of miscellaneous data is defined below in the following sections.

Terminated Track List

Whenever a target is terminated, the full target trajectory data can be exported to the metadata as part of the miscellaneous data, which is populated in NvDsTargetMiscDataBatch data structure. This data not only informs of the termination event of a target, but also can be useful for a downstream module that performs trajectory-based analysis for each object.

The terminated track list can be saved in a file in either deepstream-app or within the low-level tracker library:

• (Option 1) Saving output in deepstream-app:

  • Add outputTerminatedTracks: 1 in tracker config TargetManagement section

  • Add terminated-track-output-dir=<dir name> in deepstream-app config’s application group properties

  • create the folder <dir name>

  • Run deepstream-app to save the terminated track history into text files in <dir name>

  • Data format is defined below.

frame number

object unique id

class id

blank

blank

bbox left

bbox top

bbox right

bbox bottom

blank

blank

blank

blank

blank

blank

blank

confidence

tracker state

visibility

unsigned int

long unsigned int

unsigned int

int

float

float

float

float

float

float

float

float

float

float

float

float

float

int

float

• A file will be created for each frame in each stream. Sample data is like:

0 7 2 0 0.0 1535.194092 94.266541 1603.132812 301.653625 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.000000 2 1.000000
1 7 2 0 0.0 1535.938232 94.234810 1603.121338 301.769501 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.982758 2 1.000000
...

• (Option 2) Saving output directly from low level tracker:

  • Add outputTerminatedTracks: 1 in tracker config TargetManagement section

  • Add terminatedTrackFilename: <file name prefix> in tracker config TargetManagement section. For example, set terminatedTrackFilename: track_dump and the saved file names will be track_dump_0.txt, track_dump_1.txt, etc.

  • Run deepstream-app to save the terminated track information directly from the low-level tracker library

  • Data format is defined below. The foot location, and convex hull data are only useful if SV3DT output is enabled.

frame number(starting from 1)

object unique id

bbox left

bbox top

bbox right

bbox bottom

confidence

Foot World Position X

Foot World Position Y

blank

class id

tracker state

visibility

Foot Image Position X

Foot Image Position Y

ConvexHull Points (optional)

unsigned int

long unsigned int

int

int

int

int

float

float

float

int

unsigned int

int

float

float

float

int separated by vertical bar

• A file will be created for each stream. Sample data is like:

# SV3DT output disabled
31,48,558,104,31,74,0.949,-1.000,-1.000,-1,0,0.994,-1,-1
32,48,558,104,31,74,0.951,-1.000,-1.000,-1,0,0.995,-1,-1
...

# SV3DT output enabled
31,48,558,104,31,74,0.949,1254.535,2962.867,-1,0,0.994,581,176,-15|-34|-14|-35|-13|-35|-10|-36|-6|-36|-3|-36|0|-36|1|-36|2|-35|16|35|15|35|13|36|9|37|6|37|3|37|0|37|0|36
32,48,558,104,31,74,0.951,1255.602,2968.294,-1,0,0.995,581,176,-14|-34|-14|-35|-13|-35|-10|-35|-6|-36|-3|-36|0|-36|1|-36|2|-35|15|35|15|35|13|36|9|37|6|37|3|37|0|37|0|36
...

---

Shadow Tracking Target Data

As mentioned earlier, even when a target is not being associated with any of the detection bboxes, the target is still being tracked in a Shadow Tracking mode. When tracked in the Shadow Tracking mode, the target data is not being reported to the downstream because the target data may not be reliable.

However, the users are allowed to still report these shadow tracking target data as a part of the miscellaneous data, when outputShadowTracks: 1 is set under TargetManagement section in a tracker config file.

The shadow tracking data can be dumped in to a file when enabled in deepstream-app like below:

• Add shadow-track-output-dir=<dir name> in deepstream-app config’s application group properties

• create the folder <dir name>

• Run deepstream-app to save the Shadow Track history into text files in <dir name>

• Data format is defined below.

frame number

object unique id

class id

blank

blank

bbox left

bbox top

bbox right

bbox bottom

blank

blank

blank

blank

blank

blank

blank

confidence

tracker state

visibility

unsigned int

long unsigned int

unsigned int

int

float

float

float

float

float

float

float

float

float

float

float

float

float

int

float

• A file will be created for each frame in each stream. Sample data is like:

  1 11 2 0 0.0  296.346130 262.343445  333.428864  376.817291 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.407002 2 1.000000
  1 22 2 0 0.0 1663.921875 857.167725 1752.483521 1049.053223 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.912138 2 1.000000
  ...
  

Important

Both the terminated track data and the shadow tracking object data are not retained in the low-level tracker library when subsequent frames arrive. Thus, if users want to use these data, they should retrieve these miscellaneous data using the NvMOT_RetrieveMiscData() API at the end of every frame in the plugin. Otherwise, the data will be discarded within the tracker library.

Past-frame Target Data

Past-frame target data is always reported in miscellaneous data, and is appended to the current frame objects in tracker KITTI dump when enabled.

To enable tracker KITTI dump:

• Add kitti-track-output-dir=<dir name> in deepstream-app config’s application group properties

• create the folder <dir name>

• Run deepstream-app to save tracked objects files in <dir name>.

• Data format is defined below following the KITTI format. The foot location and visibility data will append to the end of each line if SV3DT output is enabled.

object Label

object Unique Id

blank

blank

blank

bbox left

bbox top

bbox right

bbox bottom

blank

blank

blank

blank

blank

blank

blank

confidence

visibility (optional)

Foot Image Position X (optional)

Foot Image Position Y (optional)

string

long unsigned

float

int

float

float

float

float

float

float

float

float

float

float

float

float

float

float

float

float

• A file will be created for each frame in each stream. Sample data is like:

  # SV3DT output disabled
  person 0 0.0 0 0.0 1302.667236 135.852036 1340.975830 241.724579 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.917301
  person 1 0.0 0 0.0 878.249023  195.080475  913.410950 320.695618 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.899029
  ...
  
  # SV3DT output enabled
  person 0 0.0 0 0.0 1302.667236 135.852036 1340.975830 241.724579 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.917301 0.966531 1314.492554 239.495193
  person 1 0.0 0 0.0  878.249023 195.080475  913.410950 320.695618 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.899029 0.930824  899.187500 316.670013
  ...
  

Setup and Usage of Re-ID Model

This section describes how to download and setup Re-ID models in different formats, perform INT8 calibration for performance optimization and output Re-ID features to downstream modules.

Setup Sample Re-ID Models

The supported Re-ID model formats are NVIDIA TAO and ONNX. Multiple ready-to-use sample models are listed below. Scripts and README file for users to setup the model are provided in sources/tracker_ReID.

Note

UFF is no longer supported by TensorRT, so please migrate to TAO or ONNX models.

NVIDIA TAO ReIdentificationNet

NVIDIA pre-trained ReIdentificationNet is a high accuracy ResNet-50 model with feature length 256. It can be downloaded and used directly with command:

mkdir /opt/nvidia/deepstream/deepstream/samples/models/Tracker/
wget 'https://api.ngc.nvidia.com/v2/models/nvidia/tao/reidentificationnet/versions/deployable_v1.0/files/resnet50_market1501.etlt' -P /opt/nvidia/deepstream/deepstream/samples/models/Tracker/

The tracker config file supports this model by default. Note the raw output from this network is not L2 normalized, so addFeatureNormalization: 1 is set to add L2 normalization as a post processing.

ONNX Model

An open-sourced ONNX model sample is Simple Cues Lead to a Strong Multi-Object Tracker , which proposes a new Re-ID model using on-the-fly domain adaptation. The network is based on ResNet-50 with feature length being 512.

mkdir /opt/nvidia/deepstream/deepstream/samples/models/Tracker/
wget 'https://vision.in.tum.de/webshare/u/seidensc/GHOST/ghost_reid.onnx' -P /opt/nvidia/deepstream/deepstream/samples/models/Tracker/

Then update below parameters in tracker config file ReID session (keep reidType unchanged).

ReID:
  batchSize: 100
  workspaceSize: 1000
  reidFeatureSize: 512
  reidHistorySize: 100
  inferDims: [3,384, 128]
  networkMode: 1
  # [Input Preprocessing]
  inputOrder: 0
  colorFormat: 0
  offsets: [109.1250, 102.6000, 91.3500]
  netScaleFactor: 0.01742919
  keepAspc: 1
  # [Paths and Names]
  onnxFile: "/opt/nvidia/deepstream/deepstream/samples/models/Tracker/ghost_reid.onnx"

Customize Re-ID Model

Users can also train a custom Re-ID model in ONNX format, whose output is a single vector for each object. Then the Re-ID similarity score will be computed based on the cosine metric and used to perform the data association in the same way as the official model. The steps are:

• Train a Re-ID network using deep learning frameworks such as PyTorch or TensorFlow.

• Make sure the network layers are supported by TensorRT and convert the model into ONNX. Mixed precision inference is still supported, and a calibration cache is required for INT8 mode.

• Specify the following parameters in tracker config file based on the custom model’s properties. Then run DeepStream SDK with the new Re-ID model.

  • reidFeatureSize

  • reidHistorySize

  • inferDims

  • colorFormat

  • networkMode

  • offsets

  • netScaleFactor

  • addFeatureNormalization

• ONNX model must specify below parameters.

  • onnxFile

Re-ID Feature Output

Objects’ Re-ID features can be accessed in the tracker plugin and downstream modules, which can be used for other tasks such as multi-target multi-camera tracking. Steps to retrieve those features using deepstream-app are:

• Add outputReidTensor: 1 in tracker config ReID section. Add reid-track-output-dir=<dir name> in deepstream-app config’s application group properties and create the folder <dir name>.

• Run deepstream-app to save the Re-ID features in each frame into text files in <dir name>. In each text file, each line’s first integer is object id, and the remaining floats are its feature vector. Users can check write_reid_track_output() in deepstream_app.c to understand how these features are retrieved.

This feature is supported whenever NvDeepSORT or Re-ID based re-association is used. To retrieve Re-ID features for every frame, make sure interval=0 in PGIE config and reidExtractionInterval: 0 if re-association is enabled. Otherwise, the Re-ID features will be extracted at intervals only when PGIE generates bounding boxes and reidExtractionInterval is met.

Setup and Usage of Sub-batching (Alpha)

This section describes how to use the sub-batching feature with multiple low-level tracker config files. Two use cases are explained: the first one with varied tracker algorithms and the second one with varied compute backends.

Here, a deepstream-app pipeline is used as an example.

Use-case 1

This use-case illustrates an application with a batch size of 4 (i.e., 4 streams). The batch is split into 3 sub-batches: the first sub-batch of size 2 and the next two sub-batches of size 1 each. The first sub-batch uses NvDCF tracker, the second sub-batch uses NvSORT tracker, and the third sub-batch uses IOU tracker, respectively. To achieve this, modify the [tracker] section in source4_1080p_dec_infer-resnet_tracker_sgie_tiled_display_int8.txt configuration file that comes with DeepStream package as shown below :

[tracker]
enable=1
tracker-width=960
tracker-height=544
ll-lib-file=/opt/nvidia/deepstream/deepstream/lib/libnvds_nvmultiobjecttracker.so

ll-config-file=config_tracker_NvDCF_accuracy.yml;config_tracker_NvSORT.yml;config_tracker_IOU.yml
sub-batches=0,1;2;3

In this example, sub-batches can also be configured using option 2 i.e. sub-batches=2:1:1. Refer Gst Properties for details.

Use-case 2

This use-case illustrates an application with a batch size of 4, where the batch is split into 2 sub-batches each of size 2: the first sub-batch uses NvDCF tracker with setting visualTrackerType: 1 (i.e., existing DCF module). The second sub-batch uses NvDCF tracker with visualTrackerType: 2 and vpiBackend4DcfTracker: 2 (i.e. NvDCF_VPI tracker with PVA backend). Please note that, since this use-case configures PVA backend, it will run only on Jetson platforms.

Steps are :

• Create a copy of config_tracker_NvDCF_accuracy.yml and name it as config_tracker_NvDCF_accuracy_PVA.yml

• In config_tracker_NvDCF_accuracy_PVA.yml modify the VisualTracker: section as follows :

  VisualTracker:
    visualTrackerType: 2    # the type of visual tracker among { DUMMY=0, NvDCF=1,  NvDCF_VPI=2 }
    vpiBackend4DcfTracker: 2 # the type of compute backend among {CUDA=1, PVA=2}
    ....
  

• Then modify the [tracker] section in source4_1080p_dec_infer-resnet_tracker_sgie_tiled_display_int8.txt as follows :

  [tracker]
  enable=1
  tracker-width=960
  tracker-height=544
  ll-lib-file=/opt/nvidia/deepstream/deepstream/lib/libnvds_nvmultiobjecttracker.so
  
  ll-config-file=config_tracker_NvDCF_accuracy.yml;config_tracker_NvDCF_accuracy_PVA.yml
  sub-batches=0,1;2,3
  

In this example, sub-batches can also be configured using option 2 i.e. sub-batches=2:2. Refer Gst Properties for details.

For further details on implementation of sub-batches feature, please refer to a writeup at the end of nvtracker_proc.cpp in nvtracker plugin’s source code.

The most optimal sub-batches configuration for a pipeline depends on multiple factors like elements in the pipeline, configuration of each element, hardware configuration etc. Increasing the number of sub-batches parallelizes the processing of streams in a batch. But it also adds an overhead. Hence, the number of sub-batches to configure needs to be determined experimentally by comparing GPU/PVA utilization and performance across various sub-batches configurations. A thumb rule is to start with a single batch and keep splitting it into sub-batches until an optimal performance point is reached.

Setup and Visualization of Tracker Sample Pipelines

This section describes how to setup a multi-object tracking pipeline with various NVIDIA® pre-trained detector models and DeepStream multi-object trackers, and provides ready-to-use detector and tracker config files optimized for high accuracy tracking. The optimal tracker configs for People tracking (e.g., config_tracker_NvSORT.yml, config_tracker_NvDeepSORT.yml, config_tracker_NvDCF_accuracy.yml, etc.) are already provided in DeepStream release package, so here we present optimized detector parameters only. Then the visualization of some sample outputs and internal states (such as correlation responses for a few selected targets) are presented to help users to better understand how NvDsTracker works, especially on the visual tracker module. In addition, we present detector config params and tracker config params for vehicle tracking use-case as well.

People Tracking

NVIDIA® pre-trained PeopleNet detects person, bag, and face classes. The pre-trained model with ResNet-34 backbone is on NVIDIA NGC catalog. It can be used in PGIE module in conjunction with various low level trackers for people tracking.

Setup

Here deepstream-app pipeline is used as an example, which can be naturally extended to other applications. Steps are:

• Download the detector model files here, and place the files under /opt/nvidia/deepstream/deepstream/samples/models/peoplenet

• Copy detector config file in below sections config_infer_primary_PeopleNet.txt for the specific low level tracker (for example PeopleNet + NvDCF) into work directory: cp config_infer_primary_PeopleNet.txt /opt/nvidia/deepstream/deepstream/samples/configs/deepstream-app

• Assuming deepstream-app config file is /opt/nvidia/deepstream/deepstream/samples/configs/deepstream-app/deepstream_app_config.txt, set PGIE and tracker config files in it:

deepstream_app_config.txt

## Other groups

[primary-gie]
## Use PeopleNet as PGIE
config-file=config_infer_primary_PeopleNet.txt
## Other [primary-gie] configs

[tracker]
## Specify the low level tracker (for example NvSORT)
# ll-config-file=config_tracker_IOU.yml
ll-config-file=config_tracker_NvSORT.yml
# ll-config-file=config_tracker_NvDCF_perf.yml
# ll-config-file=config_tracker_NvDCF_accuracy.yml
# ll-config-file=config_tracker_NvDeepSORT.yml
## Other [tracker] configs

/opt/nvidia/deepstream is the default DeepStream installation directory. The paths will be different if the user sets up in different directory.

PeopleNet + NvSORT

This pipeline performs high performance people tracking with reasonable accuracy. Such a deepstream-app pipeline is constructed with the following components:

• Detector: PeopleNet v2.6.2 (w/ ResNet-34 as backbone)

• Post-processing algorithm for object detection: Hybrid clustering (i.e., DBSCAN + NMS)

• Tracker: NvSORT with config_tracker_NvSORT.yml configuration in DeepStream release

A set of recommended detector config params for PeopleNet v2.6.2 to be used with NvSORT tracker is:

config_infer_primary_PeopleNet.txt

[property]
## model-specific params. The paths will be different if the user sets up in different directory.
int8-calib-file=../../models/peoplenet/resnet34_peoplenet_int8.txt
labelfile-path=../../models/peoplenet/labels.txt
onnx-file=../../models/peoplenet/resnet34_peoplenet.onnx
tlt-model-key=tlt_encode

gpu-id=0
net-scale-factor=0.00392156862745098
input-dims=3;544;960
process-mode=1
model-color-format=0
## 0=FP32, 1=INT8, 2=FP16 mode
network-mode=1
num-detected-classes=3
interval=0
gie-unique-id=1
## 1=DBSCAN, 2=NMS, 3= DBSCAN+NMS Hybrid, 4 = None(No clustering)
cluster-mode=3
maintain-aspect-ratio=1

[class-attrs-all]
pre-cluster-threshold=0.1555
nms-iou-threshold=0.3386
minBoxes=2
dbscan-min-score=1.9224
eps=0.3596
detected-min-w=20
detected-min-h=20

PeopleNet + NvDeepSORT

This pipeline enables people Re-ID capability during tracking. Such a deepstream-app pipeline is constructed with the following components:

• Detector: PeopleNet v2.6.2 (w/ ResNet-34 as backbone)

• Post-processing algorithm for object detection: Hybrid clustering (i.e., DBSCAN + NMS)

• Tracker: NvDeepSORT with config_tracker_NvDeepSORT.yml configuration in DeepStream release

A set of recommended detector config params for PeopleNet v2.6.2 to be used with NvDeepSORT tracker is:

config_infer_primary_PeopleNet.txt

[property]
## model-specific params. The paths will be different if the user sets up in different directory.
int8-calib-file=../../models/peoplenet/resnet34_peoplenet_int8.txt
labelfile-path=../../models/peoplenet/labels.txt
onnx-file=../../models/peoplenet/resnet34_peoplenet.onnx
tlt-model-key=tlt_encode

gpu-id=0
net-scale-factor=0.00392156862745098
input-dims=3;544;960
process-mode=1
model-color-format=0
## 0=FP32, 1=INT8, 2=FP16 mode
network-mode=1
num-detected-classes=3
interval=0
gie-unique-id=1
## 1=DBSCAN, 2=NMS, 3= DBSCAN+NMS Hybrid, 4 = None(No clustering)
cluster-mode=3
maintain-aspect-ratio=1

[class-attrs-all]
pre-cluster-threshold=0.1653
nms-iou-threshold=0.5242
minBoxes=2
dbscan-min-score=1.7550
eps=0.1702
detected-min-w=20
detected-min-h=20

PeopleNet + NvDCF

This pipeline performs more accurate people tracking. For the output visualization, a deepstream-app pipeline is first constructed with the following components:

• Detector: PeopleNet v2.6.2 (w/ ResNet-34 as backbone)

• Post-processing algorithm for object detection: Hybrid clustering (i.e., DBSCAN + NMS)

• Tracker: NvDCF with config_tracker_NvDCF_accuracy.yml configuration in DeepStream release

For better visualization, the following changes were also made:

• featureImgSizeLevel: 5 is set under VisualTracker section in config_tracker_NvDCF_accuracy.yml

• tracker-height=960 and tracker-width=544 under [tracker] section in the deepstream-app config file

A set of recommended detector config params for PeopleNet v2.6.2 to be used with NvDCF_accuracy tracker is:

config_infer_primary_PeopleNet.txt

[property]
## model-specific params. The paths will be different if the user sets up in different directory.
int8-calib-file=../../models/peoplenet/resnet34_peoplenet_int8.txt
labelfile-path=../../models/peoplenet/labels.txt
onnx-file=../../models/peoplenet/resnet34_peoplenet.onnx
tlt-model-key=tlt_encode

gpu-id=0
net-scale-factor=0.00392156862745098
input-dims=3;544;960
process-mode=1
model-color-format=0
## 0=FP32, 1=INT8, 2=FP16 mode
network-mode=1
num-detected-classes=3
interval=0
gie-unique-id=1
## 1=DBSCAN, 2=NMS, 3= DBSCAN+NMS Hybrid, 4 = None(No clustering)
cluster-mode=3
maintain-aspect-ratio=1

[class-attrs-all]
pre-cluster-threshold=0.1429
nms-iou-threshold=0.4688
minBoxes=3
dbscan-min-score=0.7726
eps=0.2538
detected-min-w=20
detected-min-h=20

The resulting output video of the aforementioned pipeline with (PeopleNet + Hybrid clustering + NvDCF) is shown below, but please note that only ‘Person’-class objects are detected and shown in the video:

While the video above shows the per-stream output, each animated figure below shows (1) the cropped & scaled image patch used for each target on the left side and (2) the corresponding correlation response map for the target on the right side. As mentioned earlier, the yellow + mark shows the peak location of the correlation response map generated by using the learned correlation filter, while the puple x marks show the the center of nearby detector objects.

Person 1 (w/ Blue hat + gray backpack)	Person 6 (w/ Red jacket + gray backpack)
Person 4 (w/ Green jacket)	Person 5 (w/ Cyan jacket)

The figures above show how the correlation responses progress over time for the cases of no occlusion, partial occlusion, and full occlusions happening. It can be seen that even when a target undergoes a full occlusion for a prolonged period, the NvDCF tracker is able to keep track of the targets in many cases.

If featureImgSizeLevel: 3 is used instead for better performance, the resolution of the image patch used for each target would get lower like shown in the figure below.

Person 1 (w/ Blue hat + gray backpack)

Person 6 (w/ Red jacket + gray backpack)

Vehicle Tracking

To perform tracking of vehicle and other types of targets (such as pedestrians, bicycles and road signs), a NVIDIA® pre-trained detector model is available in NGC:

• TrafficCamNet: A newer model based on ResNet-18 backbone with higher detection accuracy on NVIDIA NGC.

The following samples demonstrate using TrafficCamNet and various trackers with different detection intervals for performance and accuracy tradeoff.

Setup

Here deepstream-app pipeline is used as an example, which can be naturally extended to other applications. Steps to setup a pipeline are:

• For TrafficCamNet, download the detector model files here, and place the files under /opt/nvidia/deepstream/deepstream/samples/models/trafficcamnet

• Copy detector and tracker config files in below sections for the specific tracker type (for example TrafficCamNet + NvDCF) into work directory: cp config_infer_primary_TrafficCamNet.txt /opt/nvidia/deepstream/deepstream/samples/configs/deepstream-app

• Assuming deepstream-app config file is /opt/nvidia/deepstream/deepstream/samples/configs/deepstream-app/deepstream_app_config.txt, set detector and tracker config files in it like the following:

deepstream_app_config.txt

## Other groups

[primary-gie]
## Use TrafficCamNet as PGIE
config-file=config_infer_primary_TrafficCamNet.txt
## Other [primary-gie] configs

[tracker]
## Specify the low level tracker (for example NvDCF_accuracy)
# ll-config-file=config_tracker_IOU.yml
# ll-config-file=config_tracker_NvSORT.yml
# ll-config-file=config_tracker_NvDCF_perf.yml
ll-config-file=config_tracker_NvDCF_accuracy.yml
# ll-config-file=config_tracker_NvDeepSORT.yml
## Other [tracker] configs

/opt/nvidia/deepstream is the default DeepStream installation directory. The paths will be different if the user sets up in different directory.

TrafficCamNet + NvSORT

This pipeline performs high performance vehicle tracking with reasonable accuracy. Such a deepstream-app pipeline is constructed with the following components:

• Detector: TrafficCamNet v1.0.3 (w/ ResNet-18 as backbone)

• Post-processing algorithm for object detection: Hybrid clustering (i.e., DBSCAN + NMS)

• Tracker: NvSORT with configuration as below

The detector and tracker config files:

config_infer_primary_TrafficCamNet.txt

[property]
## model-specific params. The paths will be different if the user sets up in different directory.
int8-calib-file=../../models/trafficcamnet/resnet18_trafficcamnet_pruned_int8.txt
labelfile-path=../../models/trafficcamnet/labels.txt
onnx-file=../../models/trafficcamnet/resnet18_trafficcamnet_pruned.onnx
tlt-model-key=tlt_encode

gpu-id=0
net-scale-factor=0.00392156862745098
input-dims=3;544;960
process-mode=1
model-color-format=0
## 0=FP32, 1=INT8, 2=FP16 mode
network-mode=1
num-detected-classes=4
interval=0
gie-unique-id=1
## 1=DBSCAN, 2=NMS, 3= DBSCAN+NMS Hybrid, 4 = None(No clustering)
cluster-mode=3
maintain-aspect-ratio=0

[class-attrs-all]
pre-cluster-threshold=0.2327
nms-iou-threshold=0.1760
minBoxes=2
dbscan-min-score=0.7062
eps=0.4807
detected-min-w=20
detected-min-h=20

config_tracker_NvSORT.yml

%YAML:1.0

BaseConfig:
  minDetectorConfidence: 0.0415
TargetManagement:
  enableBboxUnClipping: 0
  maxTargetsPerStream: 300
  minIouDiff4NewTarget: 0.6974
  minTrackerConfidence: 0.8049
  probationAge: 5
  maxShadowTrackingAge: 42
  earlyTerminationAge: 1
TrajectoryManagement:
  useUniqueID: 0
DataAssociator:
  dataAssociatorType: 0
  associationMatcherType: 1
  checkClassMatch: 1
minMatchingScore4Overall: 0.2042
  minMatchingScore4SizeSimilarity: 0.2607
  minMatchingScore4Iou: 0.3708
  matchingScoreWeight4SizeSimilarity: 0.2639
  matchingScoreWeight4Iou: 0.4384
  tentativeDetectorConfidence: 0.1054
  minMatchingScore4TentativeIou: 0.4953
  usePrediction4Assoc: 1
StateEstimator:
  stateEstimatorType: 2
  noiseWeightVar4Loc: 0.0853
  noiseWeightVar4Vel: 0.0061
  useAspectRatio: 1

TrafficCamNet + NvDeepSORT

This pipeline enables vehicle Re-ID capability during tracking. Such a deepstream-app pipeline is constructed with the following components:

• Detector: TrafficCamNet v1.0.3 (w/ ResNet-18 as backbone)

• Post-processing algorithm for object detection: Hybrid clustering (i.e., DBSCAN + NMS)

• Tracker: NvDeepSORT with configuration as below. No vehicle Re-ID models are included in DeepStream so far, so the user needs to follow Customize Re-ID Model to setup a vehicle Re-ID model and change ReID section in tracker config.

The detector and tracker config files:

config_infer_primary_TrafficCamNet.txt

[property]
## model-specific params. The paths will be different if the user sets up in different directory.
int8-calib-file=../../models/trafficcamnet/resnet18_trafficcamnet_pruned_int8.txt
labelfile-path=../../models/trafficcamnet/labels.txt
onnx-file=../../models/trafficcamnet/resnet18_trafficcamnet_pruned.onnx
tlt-model-key=tlt_encode

gpu-id=0
net-scale-factor=0.00392156862745098
input-dims=3;544;960
process-mode=1
model-color-format=0
## 0=FP32, 1=INT8, 2=FP16 mode
network-mode=1
num-detected-classes=4
interval=0
gie-unique-id=1
## 1=DBSCAN, 2=NMS, 3= DBSCAN+NMS Hybrid, 4 = None(No clustering)
cluster-mode=3
maintain-aspect-ratio=0

[class-attrs-all]
pre-cluster-threshold=0.1654
nms-iou-threshold=0.7614
minBoxes=3
dbscan-min-score=2.4240
eps=0.3615
detected-min-w=20
detected-min-h=20

config_tracker_NvDeepSORT.yml

%YAML:1.0

BaseConfig:
  minDetectorConfidence: 0.0451
TargetManagement:
  preserveStreamUpdateOrder: 0
  maxTargetsPerStream: 150
  minIouDiff4NewTarget: 0.0602
  minTrackerConfidence: 0.7312
  probationAge: 9
  maxShadowTrackingAge: 59
  earlyTerminationAge: 1
TrajectoryManagement:
  useUniqueID: 0
  DataAssociator:
  dataAssociatorType: 0
  associationMatcherType: 1
checkClassMatch: 1
  thresholdMahalanobis: 34.3052
  minMatchingScore4Overall: 0.0231
  minMatchingScore4SizeSimilarity: 0.3104
  minMatchingScore4Iou: 0.3280
  minMatchingScore4ReidSimilarity: 0.6805
  matchingScoreWeight4SizeSimilarity: 0.7103
  matchingScoreWeight4Iou: 0.5429
  matchingScoreWeight4ReidSimilarity: 0.6408
  tentativeDetectorConfidence: 0.0483
  minMatchingScore4TentativeIou: 0.5093
StateEstimator:
  stateEstimatorType: 2
  noiseWeightVar4Loc: 0.0739
  noiseWeightVar4Vel: 0.0097
  useAspectRatio: 1
ReID: # need customization
  reidType: 1
  batchSize: 100
  workspaceSize: 1000
  reidFeatureSize: 128
  reidHistorySize: 100
  inferDims: [128, 64, 3]
  networkMode: 0
  inputOrder: 1
  colorFormat: 0
  offsets: [0.0, 0.0, 0.0]
  netScaleFactor: 1.0000
  keepAspc: 1
  # custom Re-ID model path

TrafficCamNet + NvDCF

This pipeline performs more accurate vehicle tracking. For the output visualization, a deepstream-app pipeline is first constructed with the following components:

• Detector: TrafficCamNet v1.0.3 (w/ ResNet-18 as backbone)

• Post-processing algorithm for object detection: Hybrid clustering (i.e., DBSCAN + NMS)

• Tracker: NvDCF with configuration as below. No vehicle Re-ID models are included in DeepStream so far, so the user needs to follow Customize Re-ID Model to setup a vehicle Re-ID model and change ReID section in tracker config.

The detector and tracker config files:

config_infer_primary_TrafficCamNet.txt

[property]
## model-specific params. The paths will be different if the user sets up in different directory.
int8-calib-file=../../models/trafficcamnet/trafficcamnet_int8.txt
labelfile-path=../../models/trafficcamnet/labels.txt
tlt-encoded-model=../../models/trafficcamnet/resnet18_trafficcamnet_pruned.etlt
tlt-model-key=tlt_encode

gpu-id=0
net-scale-factor=0.00392156862745098
input-dims=3;544;960
uff-input-blob-name=input_1
process-mode=1
model-color-format=0
## 0=FP32, 1=INT8, 2=FP16 mode
network-mode=1
num-detected-classes=4
interval=0
gie-unique-id=1
output-blob-names=output_cov/Sigmoid;output_bbox/BiasAdd
## 1=DBSCAN, 2=NMS, 3= DBSCAN+NMS Hybrid, 4 = None(No clustering)
cluster-mode=3
maintain-aspect-ratio=0

[class-attrs-all]
pre-cluster-threshold=0.3034
nms-iou-threshold=0.5002
minBoxes=3
dbscan-min-score=1.2998
eps=0.1508
detected-min-w=20
detected-min-h=20

config_tracker_NvDCF_accuracy.yml

%YAML:1.0

BaseConfig:
  minDetectorConfidence: 0.0382
TargetManagement:
  enableBboxUnClipping: 0
  preserveStreamUpdateOrder: 0
  maxTargetsPerStream: 150
  minIouDiff4NewTarget: 0.1356
  minTrackerConfidence: 0.2136
  probationAge: 1
  maxShadowTrackingAge: 49
  earlyTerminationAge: 1
TrajectoryManagement:
  useUniqueID: 0
  enableReAssoc: 1
  minMatchingScore4Overall: 0.0324
  minTrackletMatchingScore: 0.2979
  minMatchingScore4ReidSimilarity: 0.4329
  matchingScoreWeight4TrackletSimilarity: 0.5117
  matchingScoreWeight4ReidSimilarity: 0.8356
  minTrajectoryLength4Projection: 14
  prepLength4TrajectoryProjection: 50
  trajectoryProjectionLength: 116
  maxAngle4TrackletMatching: 180
  minSpeedSimilarity4TrackletMatching: 0
  minBboxSizeSimilarity4TrackletMatching: 0.2154
  maxTrackletMatchingTimeSearchRange: 16
  trajectoryProjectionProcessNoiseScale: 0.0100
  trajectoryProjectionMeasurementNoiseScale: 100
  trackletSpacialSearchRegionScale: 0.0742
  reidExtractionInterval: 2
  enableVanishingTrackletReconstruction: 0
  minInclusionRatio4DuplicateTrackletRemoval: 0.5705
  minIou4DuplicateTrackletRemoval: 0.5260
  minMatchRatio4ValidTrackletDetermination: 0.4385
  minVisibility4VanishingTrackletReconstruction: 0.3485
  visibilityThreshold4VanishingTrackletDetection: 0.5817
 DataAssociator:
  dataAssociatorType: 0
  associationMatcherType: 1
  checkClassMatch: 1
  minMatchingScore4Overall: 0.1103
  minMatchingScore4SizeSimilarity: 0.0392
  minMatchingScore4Iou: 0.0980
  minMatchingScore4VisualSimilarity: 0.3234
  matchingScoreWeight4VisualSimilarity: 0.4223
  matchingScoreWeight4SizeSimilarity: 0.8416
  matchingScoreWeight4Iou: 0.6517
  tentativeDetectorConfidence: 0.0198
  minMatchingScore4TentativeIou: 0.1844
StateEstimator:
  stateEstimatorType: 1
  processNoiseVar4Loc: 374.6508
  processNoiseVar4Size: 3364.1350
  processNoiseVar4Vel: 3.6082
  measurementNoiseVar4Detector: 164.4517
  measurementNoiseVar4Tracker: 3439.5683
VisualTracker:
  visualTrackerType: 1
  useColorNames: 1
  useHog: 1
  featureImgSizeLevel: 4
  featureFocusOffsetFactor_y: 0.0652
  filterLr: 0.0993
  filterChannelWeightsLr: 0.0549
  gaussianSigma: 0.9047
ReID: # need customization
  reidType: 2
  batchSize: 100
  workspaceSize: 1000
  reidFeatureSize: 128
  reidHistorySize: 148
  inferDims: [128, 64, 3]
  networkMode: 0
  inputOrder: 1
  colorFormat: 0
  offsets: [0.0, 0.0, 0.0]
  netScaleFactor: 1.0000
  keepAspc: 1
  # onnxFile: customize_onnx_path

Below is a side-by-side comparison of TrafficCamNet with different trackers on a challenging scene with lots of occlusions. From the top-left corner in the clockwise direction are detection only, NvSORT, NvDeepSORT, and NvDCF results. NvDCF has the highest tracking accuracy and robustness to occlusion.

Tracker Accuracy Tuning

Tracker sample pipelines demonstrated in above section include lots of detector and tracker parameters optimized for people tracking. When users deploy such pipelines for other scenarios, such as traffic, animals, etc., a pain point is how to find the optimal parameters with the highest accuracy KPI for each use case. Manual parameter tuning requires in-depth knowledge on the tracker algorithm and how each parameter would affect the functionality. Given the large number of parameters, the complexity of such process would increase exponentially.

Starting from DeepStream 7.0, a new tool PipeTuner is released to allow automatic accuracy tuning. It efficiently explores the (potentially very high-dimensional) parameter space and automatically finds the optimal parameters for the pipelines, which yields the highest KPI on the dataset. Base on their use case, users can tune the tracker on public multi-object tracking datasets, such as MOT Challenge, KITTI, or create their own dataset with sample videos and ground truth (bounding box and object ID). Users can select common tracking accuracy metrics including HOTA, MOTA and IDF1 as the KPI. Visit Accuracy Tuning Tools on how to setup PipeTuner and start DeepStream tracker accuracy tuning.

How to Implement a Custom Low-Level Tracker Library

To write a custom low-level tracker library, users are expected to implement the API defined in sources/includes/nvdstracker.h, which is covered in an earlier section on NvDsTracker API , and parts of the API refer to sources/includes/nvbufsurface.h. Thus, the users would need to include nvdstracker.h to implement the API:

#include "nvdstracker.h"

Below is a sample implementation of each API. Assuming that the low-level tracker library defines and implements a custom class (e.g., NvMOTContext class in the sample code below) to perform actual operations corresponding to each API call. Below is a sample code for initialization and de-initialization APIs:

Note

The sample code below have some skeletons only. Users are expected to add proper error handling and additional codes as needed

NvMOTStatus NvMOT_Init(NvMOTConfig *pConfigIn,
                         NvMOTContextHandle *pContextHandle,
                         NvMOTConfigResponse *pConfigResponse)
{
     if(pContextHandle != nullptr)
     {
          NvMOT_DeInit(*pContextHandle);
     }

     /// User-defined class for the context
     NvMOTContext *pContext = nullptr;

     /// Instantiate the user-defined context
     pContext = new NvMOTContext(*pConfigIn, *pConfigResponse);

     /// Pass the pointer as the context handle
     *pContextHandle = pContext;

     /**
      * return NvMOTStatus_Error if something is wrong
      * return NvMOTStatus_OK if everything went well
      */
}

/**
 * This is a sample code for the constructor of `NvMOTContext`
 * to show what may need to happen when NvMOTContext is instantiated in the above code for `NvMOT_Init` API
 */
NvMOTContext::NvMOTContext(const NvMOTConfig &config, NvMOTConfigResponse& configResponse)
{
     // Set CUDA device as needed
     cudaSetDevice(m_Config.miscConfig.gpuId)

     // Instantiate an appropriate localizer/tracker implementation
     // Load and parse the config file for the low-level tracker using the path to a config file
     m_pLocalizer = LocalizerFactory::getInstance().makeLocalizer(config.customConfigFilePath);

     // Set max # of streams to be supported
     // ex) uint32_t maxStreams = config.maxStreams;

     // Use the video frame info
     for(uint i=0; i<m_Config.numTransforms; i++)
     {
          // Use the expected color format from the input source images
          NvBufSurfaceColorFormat configColorFormat = (NvBufSurfaceColorFormat)m_Config.perTransformBatchConfig[i].colorFormat;

          // Use the frame width, height, and pitch as needed
          uint32_t frameHeight = m_Config.perTransformBatchConfig[i].maxHeight;
          uint32_t frameWidth = m_Config.perTransformBatchConfig[i].maxWidth;
          uint32_t framePitch = m_Config.perTransformBatchConfig[i].maxPitch;

          /* Add here to pass the frame info to the low-level tracker */
     }

     // Set if everything goes well
     configResponse.summaryStatus = NvMOTConfigStatus_OK;

}

void NvMOT_DeInit(NvMOTContextHandle contextHandle)
{
     /// Destroy the context handle
     delete contextHandle;
}

During the initialization stage (when NvMOT_Init() is called), the context for the low-level tracker is expected to be instantiated, and its pointer is passed as the context handle (i.e., pContextHandle) as the output as well as the output status in pConfigResponse. Users may allocate memories based on the information about the video frames (e.g., width, height, pitch, and colorFormat) and streams (e.g., max # of streams) from the input NvMOTConfig *pConfigIn, where the definition of the struct NvMOTConfig can be found in nvdstracker.h. The path to the config file for the low-level tracker library in pConfigIn->customConfigFilePath can be also used to parse the config file to initialize the low-level tracker library.

After initialization is finished, the tracker plugin queries parameters needed from low-level tracker library. The query function needs to be implemented like below:

NvMOTStatus NvMOT_Query(uint16_t customConfigFilePathSize,
                         char* pCustomConfigFilePath,
                         NvMOTQuery *pQuery)
{
     /**
      * Users can parse the low-level config file in pCustomConfigFilePath to check
      * the low-level tracker's requirements
      */

     /** An optional function queryParams(NvMOTQuery&) can be implemented in context handle to fill query params. */
     /*
     if (pQuery->contextHandle)
     {
          pQuery->contextHandle->queryParams(*pQuery);
     }
     */

     /**  Required configs for all custom trackers. */
     pQuery->computeConfig = NVMOTCOMP_GPU;       // among {NVMOTCOMP_GPU, NVMOTCOMP_CPU}
     pQuery->numTransforms = 1;                   // 0 for IOU and NvSORT tracker, 1 for NvDCF or NvDeepSORT tracker as they require the video frames
     pQuery->supportPastFrame = true;             // Set true only if the low-level tracker supports the past-frame data
     pQuery->batchMode = NvMOTBatchMode_Batch;    // batchMode must be set as NvMOTBatchMode_Batch
     pQuery->colorFormats[0] = NVBUF_COLOR_FORMAT_NV12; // among {NVBUF_COLOR_FORMAT_NV12, NVBUF_COLOR_FORMAT_RGBA}

     #ifdef __aarch64__
          pQuery->memType = NVBUF_MEM_DEFAULT;
     #else
          pQuery->memType = NVBUF_MEM_CUDA_DEVICE;
     #endif
     // among {NVBUF_MEM_DEFAULT, NVBUF_MEM_CUDA_DEVICE, NVBUF_MEM_CUDA_UNIFIED, NVBUF_MEM_CUDA_PINNED, ... }

     pQuery->maxTargetsPerStream = 150;           // Max number of targets stored for each stream

     /** Optional configs to set for additional features. */
     pQuery->maxShadowTrackingAge = 30;           // Maximum length of shadow tracking, required if supportPastFrame is true
     pQuery->outputReidTensor = true;             // Set true only if the low-level tracker supports outputting reid feature
     pQuery->reidFeatureSize = 256;               // Size of Re-ID feature, required if outputReidTensor is true

     /**
      * return NvMOTStatus_Error if something is wrong
      * return NvMOTStatus_OK if everything went well
      */
}

Once the low-level tracker library creates the tracker context and performs query during the initialization stage, it needs to implement a function to process each frame batch, which is NvMOT_Process(). Make sure to set the stream ID properly in the output so that pParams->frameList[i].streamID matches with pTrackedObjectsBatch->list[j].streamID if they are for the same stream, regardless of i and j. The method NvMOTContext::processFrame() in the sample code below is expected to perform the required multi-object tracking operations with the input data of the video frames and the detector object information, while reporting the tracking outputs in NvMOTTrackedObjBatch *pTrackedObjectsBatch.

Users can refer to Accessing NvBufSurface memory in OpenCV to know more about how to access the pixel data in the video frames.

NvMOTStatus NvMOT_Process(NvMOTContextHandle contextHandle,
                         NvMOTProcessParams *pParams,
                         NvMOTTrackedObjBatch *pTrackedObjectsBatch)
{
     /// Process the given video frame using the user-defined method in the context, and generate outputs
     contextHandle->processFrame(pParams, pTrackedObjectsBatch);

     /**
      * return NvMOTStatus_Error if something is wrong
      * return NvMOTStatus_OK if everything went well
      */
}

/**
 * This is a sample code for the method of `NvMOTContext::processFrame()`
 * to show what may need to happen when it is called in the above code for `NvMOT_Process` API
 */
NvMOTStatus NvMOTContext::processFrame(const NvMOTProcessParams *params,
                                        NvMOTTrackedObjBatch *pTrackedObjectsBatch)
{
     // Make sure the input frame is valid according to the MOT Config used to create this context
     for(uint streamInd=0; streamInd<params->numFrames; streamInd++)
     {
          NvMOTFrame *motFrame = &params->frameList[streamInd];
          for(uint i=0; i<motFrame->numBuffers; i++)
          {
               /* Add something here to check the validity of the input using the following info*/
               motFrame->bufferList[i]->width
               motFrame->bufferList[i]->height
               motFrame->bufferList[i]->pitch
               motFrame->bufferList[i]->colorFormat
          }
     }

     // Construct the mot input frames
     std::map<NvMOTStreamId, NvMOTFrame*> nvFramesInBatch;
     for(NvMOTStreamId streamInd=0; streamInd<params->numFrames; streamInd++)
     {
          NvMOTFrame *motFrame = &params->frameList[streamInd];
          nvFramesInBatch[motFrame->streamID] = motFrame;
     }

     if(nvFramesInBatch.size() > 0)
     {
          // Perform update and construct the output data inside
          m_pLocalizer->update(nvFramesInBatch, pTrackedObjectsBatch);

          /**
           * The call m_pLocalizer->update() is expected to properly populate the ouput (i.e., `pTrackedObjectsBatch`).
           *
           * One thing to not forget is to fill `pTrackedObjectsBatch->list[i].list[j].associatedObjectIn`, where
           * `i` and `j` are indices for stream and targets in the list, respectively.
           * If the `j`th target was associated/matched with a detector object,
           * then `associatedObjectIn` is supposed to have the pointer to the associated detector object.
           * Otherwise, `associatedObjectIn` shall be set NULL.
           */
     }
}

The low-level tracker can send miscellaneous data to the tracker plugin by NvMOT_RetrieveMiscData() API. Here past frame data is used as an example. Other types of customized miscellaneous data can be added in NvMOTTrackerMiscData struct and also be outputted in retrieveMiscData.

NvMOTStatus NvMOT_RetrieveMiscData(NvMOTContextHandle contextHandle,
                              NvMOTProcessParams *pParams,
                              NvMOTTrackerMiscData *pTrackerMiscData)
{
     /// Retrieve the past-frame data if there are
     contextHandle->retrieveMiscData(pParams, pTrackerMiscData);

     /**
      * return NvMOTStatus_Error if something is wrong
      * return NvMOTStatus_OK if everything went well
      */
}

/**
 * This is a sample code for the method of `NvMOTContext::processFramePast()`
 * to show what may need to happen when it is called in the above code for `NvMOT_ProcessPast` API
 */
NvMOTStatus NvMOTContext::retrieveMiscData(const NvMOTProcessParams *params,
                                             NvMOTTrackerMiscData *pTrackerMiscData)
{
     std::set<NvMOTStreamId> videoStreamIdList;

     ///\ Indiate what streams we want to fetch past-frame data
     for(NvMOTStreamId streamInd=0; streamInd<params->numFrames; streamInd++)
     {
          videoStreamIdList.insert(params->frameList[streamInd].streamID);
     }

     ///\ Retrieve past frame data
     if (pTrackerMiscData && pTrackerMiscData->pPastFrameObjBatch)
           {
                m_pLocalizer->outputPastFrameObjs(videoStreamIdList, pTrackerMiscData->pPastFrameObjBatch);
           }

     /**
      * Add other types of miscellaneous data here
      */
}

For the cases where the video stream sources are dynamically removed and added, the API call NvMOT_RemoveStreams() can be implemented to clean-up the resources no longer needed.

NvMOTStatus NvMOT_RemoveStreams(NvMOTContextHandle contextHandle,
                                   NvMOTStreamId streamIdMask)
{
     /// Remove the specified video stream from the low-level tracker context
     contextHandle->removeStream(streamIdMask);

     /**
      * return NvMOTStatus_Error if something is wrong
      * return NvMOTStatus_OK if everything went well
      */
}

/**
 * This is a sample code for the method of `NvMOTContext::removeStream()`
 * to show what may need to happen when it is called in the above code for `NvMOT_RemoveStreams` API
 */
NvMOTStatus NvMOTContext::removeStream(const NvMOTStreamId streamIdMask)
{
     m_pLocalizer->deleteRemovedStreamTrackers(streamIdMask);
}

In sum, to work with the NvDsTracker APIs, users may want to define class NvMOTContext like below to implement the methods in the code above. The actual implementation of each method may differ depending on the tracking algorithm the user choose to implement.

/**
* @brief Context for input video streams
*
* The stream context holds all necessary state to perform multi-object tracking
* within the stream.
*
*/
class NvMOTContext
{
public:
     NvMOTContext(const NvMOTConfig &configIn, NvMOTConfigResponse& configResponse);
     ~NvMOTContext();

     /**
      * @brief Process a batch of frames
      *
      * Internal implementation of NvMOT_Process()
      *
      * @param [in] pParam Pointer to parameters for the frame to be processed
      * @param [out] pTrackedObjectsBatch Pointer to object tracks output
      */
     NvMOTStatus processFrame(const NvMOTProcessParams *params,
                                   NvMOTTrackedObjBatch *pTrackedObjectsBatch);
     /**
      * @brief Output the miscellaneous data if there are
      *
      *  Internal implementation of retrieveMiscData()
      *
      * @param [in] pParam Pointer to parameters for the frame to be processed
      * @param [out] pTrackerMiscData Pointer to miscellaneous data output
      */
     NvMOTStatus retrieveMiscData(const NvMOTProcessParams *params,
                                   NvMOTTrackerMiscData *pTrackerMiscData);
     /**
      * @brief Terminate trackers and release resources for a stream when the stream is removed
      *
      *  Internal implementation of NvMOT_RemoveStreams()
      *
      * @param [in] streamIdMask removed stream ID
      */
     NvMOTStatus removeStream(const NvMOTStreamId streamIdMask);

protected:

     /**
      * Users can include an actual tracker implementation here as a member
      * `IMultiObjectTracker` can be assumed to an user-defined interface class
      */
      std::shared_ptr<IMultiObjectTracker> m_pLocalizer;

};