Human Detection System

A human detection and tracking system using YOLOv12, improved re-ID, and camera recording capabilities.

Features

• ✅ YOLOv12 detection (latest model)
• ✅ ByteTrack for robust tracking
• ✅ Re-identification: maintains ID even when person leaves and re-enters the frame
• ✅ Real-time camera or video upload
• ✅ GUI with Tkinter
• ✅ CPU/GPU Selection: Choose device (Auto/CPU/GPU) for better performance
• ✅ Counters: current people & total people who passed
• ✅ Recording in camera mode with hotkeys (R/S)
• ✅ Device info display in GUI and video output
• ✅ Clean architecture

Installation

# Create virtual environment
python -m venv .venv

# Activate
# Windows:
.venv\Scripts\activate
# Linux/Mac:
source .venv/bin/activate

# Install dependencies
pip install -r requirements.txt

Usage

python demo.py

From the GUI you can select:

• Camera: Real-time detection from webcam
• Upload Video: Upload video file
• Open Records: Open the outputs folder

Project Structure

Human Detection/
├── .venv/                    # Virtual environment
├── data/                     # Video files
│   └── people_walking.mp4
├── models/                   # Saved models
├── outputs/                  # Output videos with detections
├── src/
│   ├── detection/
│   │   ├── __init__.py
│   │   └── detect.py        # YOLOv12 detector
│   ├── tracking/
│   │   ├── __init__.py
│   │   └── tracker.py       # ByteTrack tracker with re-ID
│   ├── utils/
│   │   ├── __init__.py
│   │   └── helpers.py       # Helper functions
│   └── demo.py              # GUI & main application
├── README.md
└── requirements.txt

Device Selection (CPU/GPU)

The system supports automatic detection and device selection:

• Auto-detect: Automatic selection (GPU if available, otherwise CPU)
• CPU: Force CPU usage
• GPU (CUDA): Force GPU usage (if available)

Device selection is done from the GUI before starting detection. The device is displayed in the video output.

How Re-identification Works

The system uses:

1. ByteTrack: For frame-to-frame tracking
2. Feature Memory: Stores features for each ID
3. Similarity Matching: When a new ID appears, it compares with old IDs
4. Threshold-based Re-ID: If similarity > 0.7, reuses the old ID

Tuning Parameters

In tracker.py you can adjust:

• reid_threshold: Threshold for re-identification (default: 0.7)
• max_time_lost: How many frames to remember an ID (default: 120)
• track_buffer: Buffer for ByteTrack (default: 30)

Requirements

• Python 3.8+
• PyTorch (required, supports both CPU and GPU)
• CUDA (optional, for GPU acceleration - installed via PyTorch)
• Webcam (for real-time mode)

Installing PyTorch with GPU Support

# CPU only (default)
pip install torch torchvision

# CUDA 11.8
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118

# CUDA 12.1
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121

Building to EXE

To create a standalone Windows executable:

Prerequisites

1. Install PyInstaller:

py -m pip install pyinstaller

Build Options

Option 1: Using the batch script (Recommended)

build_exe.bat

Option 2: Using the spec file

py -m PyInstaller Human_Detection_System.spec

Option 3: Direct command

py -m PyInstaller --name="Human_Detection_System" --onedir --windowed --add-data "models;models" --add-data "data;data" --hidden-import=ultralytics --collect-all=ultralytics --collect-all=torch demo.py

Output

The executable will be created in dist/Human_Detection_System/:

• Human_Detection_System.exe - Main executable
• _internal/ - All dependencies and libraries

Note: The entire dist/Human_Detection_System/ folder must be distributed together. The .exe requires the _internal/ folder to run.

File Size

The executable folder will be large (approximately 1-2 GB) due to PyTorch and Ultralytics dependencies. This is normal.

Distribution

To distribute the application:

1. Zip the entire dist/Human_Detection_System/ folder
2. Users can extract and run Human_Detection_System.exe directly
3. No Python installation required on the target machine

Troubleshooting

Issue: YOLOv12 doesn't download

• Solution: Manually download yolo12n.pt to the models/ folder

Issue: Slow performance

• Solution: Select GPU from GUI (if available) or use a smaller model (nano: yolo12n.pt)
• Note: GPU can provide 3-10x better performance than CPU

Issue: Camera doesn't open

• Solution: Check if the camera is being used by another application

Issue: pip command not found

• Solution: Use py -m pip instead of pip on Windows