DeepStream-3D Multi-Modal V2XFusion Setup

The pipeline efficiently processes data from 4 cameras and 4 corresponding LiDARs, inferencing the 3D bboxes with a new model which combines the BEVFusion model and the BEVHeight model together to get more accurate detection result. This model is pre-trained by the DAIR-V2X dataset . The pipeline seamlessly employs ds3d::datatfiler for triton inference. Ultimately, users can inferencing 4 cameras’ videos/images and the corresponding LiDARs’ data to detect 3D bboxes of different objects in road. The results can be visualized together with videos/images and projected LiDAR point clouds in different ways by different rendering configurations.

Prerequisites

Follow the instructions in the apps/sample_apps/deepstream-app/README file to install the prerequisites for commeon DeepStream SDK modules.

Follow the instructions in the samples/configs/deepstream-app-triton/README file to prepare the Triton Server environment.

The following development packages must be installed for DS3D sample apps:

GStreamer-1.0

GStreamer-1.0 Base Plugins

GLES library

libyaml-cpp-dev

Getting Started

Start the deepstream-triton base container for V2XFusion tests. Skip this step if users have installed Triton dependencies manaully on Jetson host.

# running cmdline outside of the container
$ xhost +
# export DOCKER_GPU_ARG="--runtime nvidia --privileged" # for Jetson Orin
$ export DOCKER_GPU_ARG="--gpus all" # for x86

# start the container interactively
   $ docker run $DOCKER_GPU_ARG -it --rm --ipc=host --net=host -v /tmp/.X11-unix:/tmp/.X11-unix \
    -e DISPLAY=$DISPLAY \
    -w /opt/nvidia/deepstream/deepstream/sources/apps/sample_apps/deepstream-3d-lidar-sensor-fusion \
    nvcr.io/nvidia/deepstream:{xx.xx.xx}-triton-multiarch
# {xx.xx.xx} is deepstream sdk version number

To get started, navigate to the following directory:

$ cd /opt/nvidia/deepstream/deepstream/sources/apps/sample_apps/deepstream-3d-lidar-sensor-fusion
# sudo chmod -R a+rw . # Grant read/write permission if running on Jetson host

Download V2XFusion Models and Build TensorRT Engine Files

Download the pre-trained model from the V2XFusion pre-trained model download link and put it to the folder v2xfusion/models/v2xfusion.

Download the V2X-Seq-SPD (V2X-Seq Perception Dataset) example dataset from https://github.com/AIR-THU/DAIR-V2X and put the downloaded V2X-Seq-SPD-Example.zip in the v2xfusion/scripts folder.

$ cd /opt/nvidia/deepstream/deepstream/sources/apps/sample_apps/deepstream-3d-lidar-sensor-fusion/v2xfusion/scripts
$ ls V2X-Seq-SPD-Example.zip # verify the dataset is ready

Run the script to generate the model TensorRT engine file and ds3d required dataset.

$ pip install gdown python-lzf
$ cd /opt/nvidia/deepstream/deepstream/sources/apps/sample_apps/deepstream-3d-lidar-sensor-fusion/v2xfusion/scripts
$ ./prepare.sh

Note: The example dataset is provided by https://thudair.baai.ac.cn/coop-forecast. For each dataset an user elects to use, the user is responsible for checking if the dataset license is fit for the intended purpose.

Start DS3D V2XFusion Pipeline

$ cd /opt/nvidia/deepstream/deepstream/sources/apps/sample_apps/deepstream-3d-lidar-sensor-fusion
$ deepstream-3d-lidar-sensor-fusion -c ds3d_lidar_video_sensor_v2x_fusion.yml

Build from source:

To compile sample app deepstream-3d-lidar-sensor-fusion:

$ make
$ sudo make install (sudo not required in case of docker containers)

NOTE: To compile the sources, run make with “sudo” or root permission.

To compile the custom preprocessing and postprocessing libraries for the sample app deepstream-3d-lidar-sensor-fusion from source:

$ cd /opt/nvidia/deepstream/deepstream/sources/libs/ds3d/inference_custom_lib/ds3d_v2x_infer_custom_postprocess
$ make
$ cp libnvds_3d_v2x_infer_custom_postprocess.so /opt/nvidia/deepstream/deepstream/lib/
$ cd /opt/nvidia/deepstream/deepstream/sources/libs/ds3d/inference_custom_lib/ds3d_v2x_infer_custom_preprocess
$ make
$ cp libnvds_3d_v2x_infer_custom_preprocess.so /opt/nvidia/deepstream/deepstream/lib/