CIX Multi-Channel Video Recognition

This is a multi-channel video stream YOLO object detection inference system based on Radxa O6/O6N.

Environment Setup

The relevant environment needs to be configured in advance.

• Environment Setup
• AI Model Hub

Features

• Multi-channel video stream parallel processing
• Thread/process dual-mode inference
• NPU hardware acceleration
• Real-time FPS display
• Multi-stream combined display, fullscreen switching

Project Structure

cix-multistream-yolo/
├── main.py                      # Program entry point
├── src/
│   ├── capture/                # Video capture
│   │   ├── video_reader.py     # Video frame reading
│   │   └── video_reader_pipeline.py  # Read pipeline
│   ├── processing/              # Inference processing
│   │   ├── inference.py         # NPU inference engine
│   │   ├── inference_pipeline.py # Inference pipeline
│   │   └── post_processing.py  # Post-processing (NMS)
│   └── utils/
│       ├── manager.py          # Main coordinator
│       ├── tools.py            # Utility functions
│       ├── displaying.py       # Display module
│       └── download_model.sh   # Model download script
├── models/                     # Model files
├── data/                       # Test videos
│   ├── test_videos_360P/
│   └── test_videos_720P/
└── test/                       # Test code

Dependencies

• Python 3.11+
• OpenCV
• FFmpeg
• NumPy
• libnoe (NPU library)

Usage

Environment Dependencies

O6 / O6N

pip install opencv-python numpy

Download Project

O6 / O6N

git clone https://github.com/Ronin-1124/cix-multistream-yolo.git
cd cix-multistream-yolo

Basic Usage

python main.py

Specify Parameters

# Specify video source
python main.py -i video.mp4
python main.py -i video1.mp4 video2.mp4
python main.py -i data/test_videos_360P/

# Specify model
python main.py -m models/yolov8s.cix

# Select inference mode (thread or process)
python main.py -t process

Demo

As shown, with 8-channel inference, it can achieve up to 173 FPS real-time throughput, with an average of about 20 FPS per channel.

Parameter Description

Parameter	Description	Default
-i, --input	Video file, directory or multiple paths	data/test_videos_360P
-m, --model	Model file path	models/yolov8n.cix
-t, --type	Inference mode: thread/t or process/p	thread

Keyboard Shortcuts

Key	Function
q	Quit program
f	Toggle fullscreen mode

Supported Models

• yolov8n - YOLOv8 Nano
• yolov8s - YOLOv8 Small

Model files are automatically downloaded from ModelScope if not present.

Architecture Design

Video Reading (VideoReader)
       ↓
   Frame Queue (Queue)
       ↓
  Preprocessing (pre_processing)
       ↓
 NPU Inference (InferenceEngine)
       ↓
 Post-processing (NMS)
       ↓
 Result Queue
       ↓
 Display Module (Display)

Testing

pytest test/