Skip to main content

Object Detection with OpenCV and CUDA

This tutorial covers setting up object detection capabilities on your FEI JetRacer using OpenCV with CUDA acceleration and YOLOv4-Tiny for real-time performance.

Overview​

We'll use:

  • OpenCV with CUDA: For accelerated computer vision operations
  • YOLOv4-Tiny: A lightweight object detection model optimized for embedded devices
  • cuDNN: NVIDIA's deep learning acceleration library

Note​

If you are using our pre-configured JetRacer image, some of this setup is already complete!

You can skip the OpenCV with CUDA installation steps below and go directly to Installation Verification.

Manual Installation​

Step 1: Install OpenCV with CUDA Support​

We'll use the Qengineering installation script which automates the complex process of building OpenCV with CUDA and cuDNN support.

Download the Installation Script​

cd ~
git clone https://github.com/Qengineering/Install-OpenCV-Jetson-Nano.git
cd Install-OpenCV-Jetson-Nano

Prepare Your System​

Before running the installation script, ensure your system is ready:

sudo apt update && sudo apt upgrade -y
sudo apt autoremove -y

Free up disk space and increase swap:

sudo systemctl disable nvzramconfig
sudo fallocate -l 4G /var/swapfile
sudo chmod 600 /var/swapfile
sudo mkswap /var/swapfile
sudo swapon /var/swapfile
sudo swapon -s

Run the Installation Script​

Warning: This process can take 2-4 hours to complete. Ensure stable power supply and don't interrupt the process.

chmod +x ./OpenCV-4-11-0.sh
./OpenCV-4-11-0.sh

The script will:

  • Install all necessary dependencies
  • Download OpenCV source code
  • Configure build with CUDA support
  • Compile OpenCV (this takes the longest time)
  • Install the compiled libraries

Verify Installation​

After installation completes, verify OpenCV with CUDA support:

python3 -c "import cv2; print('OpenCV version:', cv2.__version__); print('CUDA devices:', cv2.cuda.getCudaEnabledDeviceCount())"

You should see output similar to:

OpenCV version: 4.11.1
CUDA devices: 1

Step 2: Configure Environment​

Ensure the OpenCV installation is properly configured in your DonkeyCar environment:

source ~/env/bin/activate
cd ~/mycar

Test OpenCV in your virtual environment:

python3 -c "import cv2; print('OpenCV loaded successfully with version:', cv2.__version__)"

Step 3: Setup YOLOv4-Tiny​

Now we'll add YOLOv4-Tiny object detection model to your DonkeyCar setup.

Create Models Directory​

cd ~/mycar
mkdir -p models
cd models

Download YOLOv4-Tiny Files​

Download the pre-trained YOLOv4-Tiny model files:

wget https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v3_optimal/yolov4-tiny.weights
wget https://raw.githubusercontent.com/AlexeyAB/darknet/master/cfg/yolov4-tiny.cfg
wget https://raw.githubusercontent.com/AlexeyAB/darknet/master/data/coco.names

Verify Model Files​

Check that all required files are present:

ls -la ~/mycar/models/

You should see:

  • yolov4-tiny.weights (approximately 23MB)
  • yolov4-tiny.cfg (configuration file)
  • coco.names (class labels file)

Test Object Detection​

Test by running the DonkeyCar web interface:

cd ~/mycar
python manage.py drive

You should see the object detection in the camera feed in the web interface.

Performance Tips​

Optimize for Real-time Performance​

  1. Resolution Tuning: Balance between accuracy and speed by adjusting input resolution

  2. Detection Frequency: Run detection every few frames rather than every frame

  3. ROI Processing: Only process regions of interest in the image

Troubleshooting​

OpenCV Installation Issues​

If the installation script fails:

  1. Insufficient storage: Ensure at least 16GB free space
  2. Memory issues: Increase swap file size
  3. Power issues: Use adequate power supply

CUDA Not Working​

If CUDA acceleration isn't working:

# Check CUDA installation
nvcc --version
# Verify cuDNN
cat /usr/local/cuda/include/cudnn_version.h | grep CUDNN_MAJOR -A 2

Model Loading Errors​

If YOLO models fail to load:

  1. Verify file integrity:

    md5sum ~/mycar/models/yolov4-tiny.weights

  2. Check file permissions:

    chmod 644 ~/mycar/models/*

References​