Pi Auto-Tracking Camera

Real-time person detection and auto-tracking camera system for Raspberry Pi 5 with Hailo AI accelerator.

Features

• Person Detection: YOLOv8s model running on Hailo-8L NPU at ~30 FPS
• Auto-Tracking Zoom: Smooth camera zoom that follows detected persons
• Multi-Person Tracking: Automatically frames all detected people
• Web Streaming: Live MJPEG stream accessible from any browser
• Zoom Hysteresis: Delayed zoom-out prevents jittery tracking
• Virtual Frame Mode: Preview tracking region without zooming
• Configurable: All tracking parameters adjustable via command line

Requirements

• Raspberry Pi 5
• Hailo-8L AI Kit
• Pi Camera Module (wide-angle supported)
• Python 3.x

Dependencies

sudo apt install python3-gi python3-gi-cairo gir1.2-gst-1.0 python3-flask python3-opencv python3-numpy
pip install -r requirements.txt

Project Structure

├── main.py                   # Main entry point
├── scripts/
│   ├── tracking.py           # Zoom tracking state and calculations
│   ├── pipeline.py           # GStreamer pipeline setup
│   └── streaming.py          # Flask web streaming
├── requirements.txt
└── README.md

Usage

# Run with camera
python main.py

# Run with video file
python main.py -i video.mp4

# Virtual frame mode (shows tracking region without zooming)
python main.py --frame-mode

# Hide detection boxes
python main.py --no-boxes

# Custom settings
python main.py -p 8000 -s 0.05 -c 0.6

Access the stream at http://<pi-ip>:8080

Command Line Options

Option	Default	Description
-i, --input	camera	Video file path
-p, --port	8080	Web server port
-s, --smooth	0.1	Smooth factor (lower = smoother)
-d, --delay	30	Frames to wait before zooming out
-c, --confidence	0.5	Minimum detection confidence
--padding	0.3	Padding around detected person
--no-boxes	false	Hide detection boxes
--frame-mode	false	Show virtual frame instead of zooming
--fps	false	Show FPS counter

How It Works

1. Camera captures 640x480 frames
2. Hailo NPU runs YOLOv8s inference
3. All persons above confidence threshold are detected
4. Bounding box containing all persons is calculated
5. Target crop region is calculated with padding
6. Current crop smoothly interpolates toward target
7. Cropped/zoomed frame is streamed via Flask