Welcome to the on-board part of CAD2CAV project. For a detailed project documentation, see the main repo..

• Nvidia Jetson Xavier NX Development Kit (hardware device)
• Linux Ubuntu 20.04 LTS/18.04 LTS
• ROS Noetic/Melodic
• GCC 7+/Clang 7+

• Boost 1.71
• OpenCV 4.5.1
• TensorRT 8.0.1

First run and install the following packages from Ubuntu repository:

sudo apt install libboost-all-dev

Then install OpenCV 4.5.1 on Xavier NX board by running

cd scripts
sudo ./Install_OpenCV-4-5-1.sh

Note. It might take ~2 hrs to fully install OpenCV 4.5.1 on the device.

Then install other dependencies manually, by either building from source or following the official documentation.

The entire process is ready for installation on an F1/10 Autonomous Racing Car, which uses an Nvidia Jetson Xavier Board. Users must also install the system-level drivers for F1/10 cars from this repository.

mkdir -p ~/catkin_ws/src
cd ~/catkin_ws
catkin_init_workspace
cd src
git clone https://github.com/shineyruan/cad2cav_onboard.git
git clone https://github.com/f1tenth/f1tenth_system.git
git clone --recursive https://github.com/shineyruan/cad2cav_common.git
cd ..

To build the project on the car (ROS Melodic), you have to manually specify Python 3 path for cv_bridge building compliance. Run:

catkin_make_isolated --use-ninja -DPYTHON_EXECUTABLE=/usr/bin/python3 -DPYTHON_INCLUDE_DIR=/usr/include/python3.6m -DPYTHON_LIBRARY=/usr/lib/aarch64-linux-gnu/libpython3.6m.so
source devel_isolated/setup.bash

For object detection with YOLO in darknet architecture, one could also experiment with Shumin326/darknet_ros. The package has been adapted from the upstream repo and is verified to run successfully on the F1Tenth car.

cd ~/catkin_ws/src
git clone --recursive https://github.com/Shumin326/darknet_ros
cd ..

1. For object detection, run

rosrun object_detection object_detection_node --model=darknet

You can also specify a pre-recorded video to visualize by adding --video=<PATH_TO_VIDEO> as a command line argument.

The object detection part of project uses a pre-trained model of Tiny-YOLO v4 directly from the author's website.