CCTV Monitor with YOLO and ONNX Runtime

This project provides a Python script to monitor a CCTV RTSP stream using a YOLO object detection model accelerated with ONNX Runtime. It supports GPU acceleration on NVIDIA, Intel, and AMD hardware.

The CCTV stream is from V380 Pro IP camera, like this:



I also tested this using the DroidCam Webcam app.

Sample detection:



CPU and GPU Load:

Features

• Real-time object detection on RTSP or webcam streams
• Uses Ultralytics YOLO models (.pt), automatically exports to ONNX (.onnx)
• Hardware acceleration via DirectML on Windows OS. Fallback to CPU.
• Displays detected objects with class names and confidence
• Shows real-time FPS and can skip frames to maintain target frame rate
• Draw custom monitoring line to raise alarm message whenever a detected object intersects it

Requirements

• Python 3.8+
• Windows (DirectML/AMD GPU support), Linux (CUDA/NVIDIA GPU support), or Intel OpenVINO
• RTSP-compatible CCTV camera or webcam

Python Packages

Install dependencies with:

pip install -r requirements.txt

Contents of requirements.txt:

onnxruntime-directml
opencv-python
ultralytics

For NVIDIA GPU support, install onnxruntime-gpu instead of onnxruntime-directml. For Intel GPU/CPU support, install onnxruntime-openvino.

Usage

1. Clone the repository:

   git clone https://github.com/afdp777/cctv_monitor.git
   cd cctv_monitor
   

2. Create and activate a virtual environment:

   python -m venv venv
   venv\Scripts\activate   # On Windows
   # or
   source venv/bin/activate  # On Linux/macOS
   

3. Install dependencies:

   pip install -r requirements.txt
   

4. Configure the stream URL:

   • Edit watch_cctv_gpu.py and set stream_url to your RTSP stream.

5. Run the script:

   python watch_cctv_gpu.py
   

Hardware Acceleration

The script tries to use the best available provider in this order:

1. CUDAExecutionProvider (NVIDIA GPU)
2. OpenVINOExecutionProvider (Intel GPU/CPU)
3. DmlExecutionProvider (AMD GPU via DirectML)
4. Falls back to CPU if no GPU provider is available

You can check which provider is active in the console output.

Notes

• The script will download and export the YOLO .pt model to ONNX format on first run.
• Class names are extracted from the Ultralytics model before export.
• For best performance, ensure your GPU drivers are up to date.
• The maximum value for the displayed fps is the actual frame rate of the video source. The displayed fps gets lower depending on your GPU performance. Set the target_fps at or below this lower frame rate.

License

MIT License

---