Formula Student Driverless inference with ZED and ONNX

This project uses a EfficientDet-D0 model exported to ONNX and a ZED stereo camera to detect and range traffic cones. The model is a self trained EfficientDet-D0 architecture with no other optimizations but anchor sizes. It is only for showcasing and debugging.

Performance

The performance depends heavily on the execution provider and the used hardware. With a RTX2060 the inference reaches roughly 55 ± 5Hz while a Jetson NX Xavier only reaches about 30Hz. This can probably be improved with the ONNX-TensorRT provider or through discarding ONNX completely and switching to TensorRT. Here is surely some room for optimization.

Installation & Setup

To run this code one obviously needs ONNX and the ZED SDK but also OpenCV, CUDA & CUDNN. In the following my installation process is briefly described. Others may work too or mine may not, depending on your existing environment.

An example .svo file can be downloaded from here: https://drive.google.com/file/d/1oHrZGJ2r6h4mJaQN11YjG5cwEa9o6X1p/view?usp=sharing

Build Opencv:

• Install the following packages if missing:
  • sudo apt install g++ cmake make git libgtk2.0-dev pkg-config
• Download OpenCV source code:
  • https://github.com/opencv/opencv/releases
• Create build directory...
  • mkdir -p build && cd build
• .. and build:
  • cmake ../opencv
  • make -j16 (or your core count)
  • sudo make install

Install CUDA & CUDNN:

Make sure to use the right versions for your preferred version of the ZED SDK.

• https://developer.nvidia.com/cuda-toolkit
• https://developer.nvidia.com/cudnn

Install ZED SDK:

• Install Cuda (and CUDNN)
• https://www.stereolabs.com/developers/release/

Build ONNX Runtime:

• Install CUDA and CUDNN
• Clone onnxruntime https://github.com/microsoft/onnxruntime
  • Important: clone from release commit. E.g. https://github.com/microsoft/onnxruntime/tree/v1.13.1
• build with:
• ./build.sh --use_cuda --cudnn_home <CUDNN HOME PATH> --cuda_home <CUDA PATH> --parallel --build_shared_lib --config=Release
  • <CUDNN HOME PATH> is probably something like: /usr/lib/x86_64-linux-gnu
  • <CUDA PATH> is probably something like: /usr/local/cuda-11.7
  • parallel makes compiling faster. If OOM occurs exclude it.
• Set the ONNXRUNTIME_ROOT_PATH path in CMakeLists.txt to your build directory.

For further information you can look here: https://onnxruntime.ai/docs/build/eps.html#cuda