DanielHfnr/tensorrt_yolo_ros

C++ ROS Node for TensorRT inference of YOLO object detection network

C++ ROS node for TensorRT inference of YoloV7 (yolov7-tiny) object detection. Node load a yolov7.onnx model and create a TensorRT engine from that. Gets camera input data, runs inference and postprocesses network outputs.

Publishes detected bounding boxes as well as rendered input image with bounding box overlay.

Tested with:

• Ubuntu 20.04
• Nvidia Driver 520
• CUDA 11.6
• TensorRT 8.4.3
• ROS Noetic

Prerequisites

• Installed ROS: http://wiki.ros.org/noetic/Installation/Ubuntu
• Install CUDA: https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html
• Install TensorRT: https://docs.nvidia.com/deeplearning/tensorrt/install-guide/index.html

Follow the instructions to install ROS, CUDA and TensorRT.

Install ROS vision messages:

sudo apt-get install ros-distro-vision-msgs

Optional:

Install clang-format for code formatting:

sudo apt-get install clang-format

Usage

Clone the repository into your catkin workspace src folder:

git clone https://github.com/DanielHfnr/tensorrt_yolo_ros.git

Clone the TensorRT Base package which is a catkin package library that serves as a base class for this project:

git clone https://github.com/DanielHfnr/tensorrt_base.git

Build your workspace using catkin_make or catkin build.

Launch the ROS node using the provided launch file (adjust topic names where needed):

roslaunch tensorrt_yolo tensorrt_yolo.launch

<launch>
  <arg name="output" default="screen"/>
  
  <node name="tensorrt_yolo" pkg="tensorrt_yolo" type="tensorrt_yolo_node" output="$(arg output)">
      <param name="yolo_onnx_model" value="$(find tensorrt_yolo)/models/yolov7-tiny.onnx" type="str" />
      <param name="class_labels_file" value="$(find tensorrt_yolo)/models/coco_classes.txt" type="str" />

      <param name="device" value="GPU" type="str" />
      <param name="precision" value="FP32" type="str" />
      <param name="allow_gpu_fallback" value="true" type="bool" />


      <param name="image_topic_in" value="/camera/image_raw" type="str" />
      <param name="image_topic_out" value="/camera/image_detections" type="str" />
      <param name="bounding_boxes_topic_out" value="/object_detection/bounding_boxes_2d" type="str" />

      <param name="show_output_image" value="true" type="bool" />
  </node>
</launch>

Usage with CARLA simulator pretrained model and video

To test if everything works correctly get camera input data. For that you can for examples use the following ros package: http://wiki.ros.org/video_stream_opencv

Clone the repository in your catkin workspace and build it:

git clone https://github.com/ros-drivers/video_stream_opencv.git
catkin build video_stream_opencv

Afterwards you can start the video streamer:

roslaunch tensorrt_yolo video_file_stream.launch

Then start the object detector pretrained on the CARLA simulator object detection dataset:

roslaunch tensorrt_yolo tensorrt_yolo_carla.launch