DeepStream 3D Lidar Inference App

The deepstream-lidar-inference-app sample application is provided at app/sample_apps/deepstream-lidar-inference-app/ for your reference. The deepstream_lidar_inference_app provides an end-to-end inference sample for lidar pointcloud data. The sample application reads the point cloud data from dataset files and send the data to Triton Inferencing filter with PointPillarNet model, the inferencing result is the group of 3D bounding boxes of the objects. The sample application loads different pipelines based on different application config files. There are 2 pipelines configured in the sample app.

1. Lidar Triton inference for 3D objects detection and file dump.

2. Lidar Triton inference and 3D objects detection and GLES 3D rendering.

Inside these sample configurations, the inference model is a 3D TAO model based on PointPillar. For more details about PointPillar, see https://arxiv.org/abs/1812.05784.

Note

TensorRT 8.5 has a bug for FP16 mode when converting this specific TAO model to TensorRT engine file. DeepStream fallback to FP32 mode for this release.

This is a snapshot for deepstream-lidar-inference-app running with lidar data inference objects detection and GLES 3D rendering with 3D bounding box display on screen.

Prerequisites

You must have the following development packages installed:

• GStreamer-1.0

• GStreamer-1.0 Base Plugins

• X11 client-side library

• libyaml-cpp-dev

To install these packages, execute the following command:

sudo apt-get install libgstreamer-plugins-base1.0-dev libgstreamer1.0-dev \
libgstrtspserver-1.0-dev libx11-dev libyaml-cpp-dev

Lidar Point Cloud to 3D Point Cloud Processing and Rendering

The application can be configured as different pipelines according to the application configuration file.

1. Lidar data inference and 3D bounding box dump pipeline:

This pipeline is from lidar point cloud, 3D point cloud inferencing, to the 3D objects data dump.

This pipeline is setup by the config_lidar_triton_infer.yaml. It has 3 components: ds3d::dataloader for lidar pointcloud data file reading, ds3d::datafilter for point-cloud Triton inferencing, and ds3d::datarender for 3D Bounding Box file dump.

• ds3d::dataloader loads custom lib libnvds_lidarfileread.so and creates a dataloader through the createLidarFileLoader function. This specific loader is configured by lidar dataset file list data_config_file. Gst-appsrc connects the dataloader into the deepstream pipeline.

  name: lidarsource
  type: ds3d::dataloader
  out_caps: ds3d/datamap
  custom_lib_path: libnvds_lidarfileread.so
  custom_create_function: createLidarFileLoader
  

• ds3d::datafilter loads custom lib libnvds_tritoninferfilter.so and creates a lidar point cloud Triton inferencing filter through the createLidarInferenceFilter function. For this specific configuration, The Lidar Triton filter inferences the point cloud data with TAO model PointPillarNet model and return the 3D bounding boxes around each object.

name: lidarfilter
type: ds3d::datafilter
in_caps: ds3d/datamap
out_caps: ds3d/datamap
custom_lib_path: libnvds_tritoninferfilter.so
custom_create_function: createLidarInferenceFilter

ds3d::datafilter is loaded by the nvds3dfilter Gst-plugin which accepts in_caps as sink_caps and out_caps as src_caps. It creates a custom ds3d::datafilter instance and processess data as ds3d/datamap.

• ds3d::datarender loads custom lib libnvds_lidarfilewrite.so to dump the detected 3D bounding boxes to a file.

name: lidarrender
type: ds3d::datarender
in_caps: ds3d/datamap
custom_lib_path: libnvds_lidarfilewrite.so
custom_create_function: createLidarFileDataRender

2. Lidar data inference and Lidar data rendering with 3D bounding box display pipeline:

This pipeline is from lidar point cloud data file, 3D point cloud inferencing, to the 3D point cloud rendering with colors.

This pipeline is setup by the config_lidar_source_triton_render.yaml. It has 3 components: ds3d::dataloader for lidar pointcloud data file reading, ds3d::datafilter for point-cloud Triton inferencing, and ds3d::datarender for Lidar 3D data LidarXYZI and 3D Bounding Box rendering.

• ds3d::dataloader loads custom lib libnvds_lidarfileread.so and creates a dataloader through the createLidarFileLoader function. This specific loader is configured by lidar dataset file list data_config_file. Gst-appsrc connects the dataloader into the deepstream pipeline.

  name: lidarsource
  type: ds3d::dataloader
  out_caps: ds3d/datamap
  custom_lib_path: libnvds_lidarfileread.so
  custom_create_function: createLidarFileLoader
  

• ds3d::datafilter loads custom lib libnvds_tritoninferfilter.so and creates a lidar point cloud Triton inferencing filter through the createLidarInferenceFilter function. For this specific configuration, The Lidar Triton filter inferences the point cloud data with TAO model PointPillarNet model and return the 3D bounding boxes around each object.

  name: lidarfilter
  type: ds3d::datafilter
  in_caps: ds3d/datamap
  out_caps: ds3d/datamap
  custom_lib_path: libnvds_tritoninferfilter.so
  custom_create_function: createLidarInferenceFilter
  

ds3d::datafilter is loaded by the nvds3dfilter Gst-plugin which accepts in_caps as sink_caps and out_caps as src_caps. It creates a custom ds3d::datafilter instance and processess data as ds3d/datamap.

• ds3d::datarender loads custom lib libnvds_3d_gl_datarender.so and creates GLES Lidar point cloud rendering context to display XYZI or XYZ lidar data and 3D bounding boxes with custom_create_function: createLidarDataRender.

  name: lidarrender
  type: ds3d::datarender
  in_caps: ds3d/datamap
  custom_lib_path: libnvds_3d_gl_datarender.so
  custom_create_function: createLidarDataRender
  

Getting Started

Run Lidar Point Cloud Data File reader, Point Cloud Inferencing filter, and Point Cloud 3D rendering and data dump Examples

1. Prepare PointPillarNet model and Triton environment, this app will use Triton to do inference, for more details of Triton Inferencing Server, refer to https://docs.nvidia.com/metropolis/deepstream/dev-guide/text/DS_plugin_gst-nvinferserver.html.

   Follow instructions in deepstream-lidar-inference-app/README to prepare testing resources and Triton CAPI and gRPC environments.

2. Run the lidar point cloud data inference pipeline in 2 modes.

1. Run lidar data reader, point cloud 3D objects detection inference and 3D data GLES rendering pipeline:

   $ deepstream-lidar-inference-app -c configs/config_lidar_source_triton_render.yaml
   

2. Run lidar data reader, point cloud 3D objects inference and 3D objects file dump pipeline:

   $ deepstream-lidar-inference-app -c configs/config_lidar_triton_infer.yaml
   

3. This part sets up a lidar point cloud loader dataloader. It then streams ds3d/datamap to the downstream datafilter component lidarfilter.

   name: lidarsource
   type: ds3d::dataloader
   out_caps: ds3d/datamap
   custom_lib_path: libnvds_lidarfileread.so
   custom_create_function: createLidarFileLoader
   

4. It streams ds3d/datamap to the nvds3dfilter Gst-plugin which loads lidarfilter to do Triton inferencing on point cloud. For more details on nvds3dfilter Gst-plugin, See Gst-nvds3dfilter.

   name: lidarfilter
   type: ds3d::datafilter
   in_caps: ds3d/datamap
   out_caps: ds3d/datamap
   custom_lib_path: libnvds_tritoninferfilter.so
   custom_create_function: createLidarInferenceFilter
   

• Field of model_inputs is the description of the model input layers. Includes layer name, layer data type and layer dimensions.

  model_inputs:
  - name: points      # name of the 1st layer
    datatype: FP32    # data type of the 1st layer
    shape: [1, 204800, 4] # data dimension of the 1st layer
  - name: num_points  # name of the 2nd layer
    datatype: INT32   # data type of the 2nd layer
    shape: [1]        # data dimension of the 2nd layer
  

• Field of model_outputs is the description of the model output layers. Includes layer name, layer data type and layer dimensions.

  model_outputs:
  - name: output_boxes # name of the 1st layer
    datatype: FP32     # data type of the 1st layer
    shape: [1, 393216, 9] # data dimension of the 1st layer
  - name: num_boxes    # name of the 2nd layer
    datatype: INT32    # data type of the 2nd layer
    shape: [1]         # data dimension of the 2nd layer
  

• Field of labels is the label list of the point cloud inferencing model.

  labels:          # YAML list for object labels
  - Vehicle
  - Pedestrian
  - Cyclist
  

5. Finally the data stream as ds3d/datamap is delivered to the render component.

   1. 3D detection file dump component lidarfiledump.

   name: lidarfiledump
   type: ds3d::datarender
   in_caps: ds3d/datamap
   custom_lib_path: libnvds_lidarfilewrite.so
   custom_create_function: createLidarFileDataRender
   

   2. GLES lidar data rendering component lidardatarender.

   name: lidardatarender
   type: ds3d::datarender
   in_caps: ds3d/datamap
   custom_lib_path: libnvds_3d_gl_datarender.so
   custom_create_function: createLidarDataRender
   

Check the /opt/nvidia/deepstream/deepstream/sources/apps/sample_apps/deepstream-lidar-inference-app/README file for more details.

DeepStream Lidar Inference App Configuration Specifications

deepstream-lidar-inference-app [ds3d::userapp] group settings

The table below demonstrates the group settings for config_lidar_triton_infer.yaml and config_lidar_source_triton_render.yaml as the examples.

point cloud inferencing app user debug supported settings

Group	Property	Meaning	Type and Range	Example
LidarFileLoader	data_config_file	lidar data list file path	string	data_config_file: lidar_data_list.yaml
LidarFileLoader	points_num	number of the points in pointcloud file	fixed value	points_num: 204800
LidarFileLoader	lidar_datatype	data type of the dataset	String:FP32 FP16 INT8 INT32	lidar_datatype: FP32
LidarFileLoader	mem_type	memory type of process data:just support cpu now	String:gpu cpu	mem_type: cpu
LidarFileLoader	mem_pool_size	Size of the data read pool	Integer: >0	mem_pool_size: 4
LidarFileLoader	output_datamap_key	datamap key in lidarsource	string	output_datamap_key: DS3D::LidarXYZI
LidarFileLoader	file_loop	flag for file reading loop	boolean	file_loop: False
LidarInferenceFilter	in_streams	which data type will be processed	fixed value	in_streams: [lidar]
LidarInferenceFilter	mem_pool_size	Size of the input tensor pool	Integer	mem_pool_size: 8
LidarInferenceFilter	model_inputs	model ‘s input layers	Array	refer to config_lidar_triton_infer.yaml
LidarInferenceFilter	model_outputs	model ‘s output layers	Array	refer to config_lidar_triton_infer.yaml
LidarInferenceFilter	input_tensor_mem_type	input tensor memory type after preprocess	String:GpuCuda CpuCuda	input_tensor_mem_type: GpuCuda
LidarInferenceFilter	custom_preprocess_lib_path	preprocessing library path	String	custom_preprocess_lib_path: /opt/nvidia/deepstream/deepstream/lib/libnvds_lidar_custom_preprocess_impl.so
LidarInferenceFilter	custom_preprocess_func_name	customized preprocessing function name	String	custom_preprocess_func_name: CreateInferServerCustomPreprocess
LidarInferenceFilter	labels	label list for the detection model	Array	refer to config_lidar_triton_infer.yaml
LidarInferenceFilter	postprocess_nms_iou_thresh	NMS IOU threshold	Float	postprocess_nms_iou_thresh: 0.01
LidarInferenceFilter	postprocess_pre_nms_top_n	number of TOPs of NMS	Integer	postprocess_nms_top_n: 4096
LidarInferenceFilter	config_file	nvinferserver configuration file	String	config_file: triton_mode_CAPI.txt
LidarInferenceFilter	gpu_id	GPU id for the tensor memory(for native Triton Server Inferencing)	Integer	gpu_id: 0
LidarInferenceFilter	filter_input_datamap_key	input datamap key from lidarsource	String	filter_input_datamap_key: DS3D::LidarXYZI
LidarFileDataRender	frames_save_path	the path of the dump file	String	frames_save_path: ../data/
LidarFileDataRender	input_datamap_key	input key from the custom_postprocess for inferencing objects	String	input_datamap_key: DS3D::Lidar3DBboxRawData
LidarDataRender	title	the title of the render	String	title: ds3d-lidar-render
LidarDataRender	streams	the stream key(s) for the input to be rendered	List	streams: [lidardata]
LidarDataRender	width	render area width	Integer	width: 1280
LidarDataRender	height	render area height	Integer	height: 720
LidarDataRender	block	the flag of enabling block function	Boolean	block: True
LidarDataRender	view_position	the view position for lookat vector	List	view_position: [0, 0, 80]
LidarDataRender	view_target	the view target for lookat vector	List	view_target: [0, 0, 0]
LidarDataRender	view_up	the up vector of the visualizer	List	view_up: [1, 0, 0]
LidarDataRender	near	the near z-plane of the visualizer constance	Float	near: 0.3
LidarDataRender	far	the far z-plane of the visualizer constance	Float	far: 100
LidarDataRender	fov	degree for field of view	Integer	fov: 45
LidarDataRender	lidar_color	the RGB color description of the lidar data	List,lidar_color: [0, 255, 0]
LidarDataRender	lidar_data_key	the lidar data key name in datamap	String	lidar_data_key: DS3D::LidarXYZI
LidarDataRender	element_size	the lidar data element size (4, XYZI or 3, XYZ)	Integer	element_size: 4
LidarDataRender	lidar_bbox_key	the 3D bbox data key name in datamap	String	DS3D::Lidar3DBboxRawData
LidarDataRender	enable_label	Indicate flag to enable labels rendering	Boolean	enable_label: True

DS3D Custom Components Configuration Specifications

See more details in the DS_3D supported custom components specifications section in the DeepStream-3D Custom Apps and Libs Tutorials.

Build application From Source

1. Go to the folder sources/apps/sample_apps/deepstream-lidar-inference-app.

2. Run the following commands:

   $ sudo make
   $ sudo make install
   

Note

Check the source code for more details on how to load dataloader/datarender through Gst-appsrc and Gst-appsink. datafilter is loaded by the nvds3dfilter Gst-plugin.