Object Detection Algorithms in Embedded Systems

My master's thesis about a modular and high-performance object detection platform using ONNX Runtime and OpenCV, designed for embedded platforms like NVIDIA Jetson. Supports image and video input with sandbox testing utilities.

Table of Contents

• Features
• Installation
• Usage
• Tests
• Notes
• Benchmarks
• Contributing
• License

Features

• Plug-and-play ONNX object detection
• Supports image and video sources
• Sandbox testing environment
• Easy configuration and modular design
• Optimized mainly for NVIDIA Jetson

Installation

Requirements

• Ubuntu 22.04.05 (WSL2 works)
• CMake >= 3.15
• C++17
• OpenCV
• ONNX Runtime GPU (or CPU)
• CUDA & cuDNN (for Jetson or GPU acceleration)

Dependencies

Name	Version
gcc	12.3
OpenCV	4.8
OpenCV contrib	4.8
CUDA	12.6
cuDNN	9.3
TensorRT	10.3
OnnxRuntime	1.22.1

Clone and build

git clone --recurse-submodules https://github.com/aleksander-piprek/Object-detection-algorithms-in-embedded-systems.git
cd Object-detection-algorithms-in-embedded-systems
mkdir build && cd build
cmake ..
make -j$(nproc)

Install

Few installation scripts are provided. Make use of them however you want.

Export YOLOV5 models

To export .onnx model you need to follow commands below

git clone https://github.com/ultralytics/yolov5.git
cd yolov5
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python export.py --weights yolov5s.pt --include onnx --dynamic --simplify
mkdir ~/Object-detection-algorithms-in-embedded-systems/bin
cp yolov5s.onnx ~/Object-detection-algorithms-in-embedded-systems/bin/

If Python environment has been already created already, just run

source .venv/bin/activate
python export.py --weights yolov5s.pt --include onnx --dynamic --simplify
cp yolov5s.onnx ~/Object-detection-algorithms-in-embedded-systems/bin/

TensorRT

Exporting models from .onnx format to TensorRT, can be done using example command below

/usr/src/tensorrt/bin/trtexec   --onnx=yolov5s.onnx   --saveEngine=yolov5s_fp16.engine   --fp16   --minShapes=images:1x3x640x640   --optShapes=images:1x3x640x640   --maxShapes=images:1x3x640x640

Usage

Code uses .cfg files for faster compilation. .vscode/settings.json is used for default argument call when using CMake extension. To run command line, run command below for all possible programs

./oda --help

Notes

• Do not use VS Code snap version! It uses its own packages which collides with the system's software (GLIBC, libpthread)
• Running detections on virtual machine works only on CPU (potential workaround is DDA or GPU passthrough)

Contributing

Feel free to fork the project and open pull requests. Bug reports and suggestions are also welcome via issues.

License

This project is licensed under the MIT License.