Player Tracking System

A real-time player tracking and re-identification system for sports videos using computer vision and deep learning. This system can track multiple players across frames, handle occlusions, and re-identify players when they reappear after being lost.

Features

Core Tracking

• Multi-object tracking with advanced association algorithms
• Robust handling of occlusions and temporary disappearances
• Adaptive IoU thresholds based on detection confidence
• Motion prediction for improved tracking stability

Re-identification

• Deep feature extraction for player appearance modeling
• Gallery-based re-identification system with long-term memory
• Jersey number detection and verification
• Team classification based on uniform colors

Performance Optimizations

• Lightweight feature extraction mode for real-time processing
• Configurable confidence thresholds and tracking parameters
• Efficient FAISS indexing for large-scale gallery searches
• Smart track management with confidence decay

Installation

1. Clone the repository:

git clone <repository-url>
cd player-tracking

2. Create a virtual environment:

python -m venv venv
source venv/bin/activate

3. Install dependencies:

pip install -r requirements.txt

4. Put model.pt in the models folder

Quick Start

Basic Demo

Run the advanced tracking demo on a sample video:

python main_demo.py --input data/input/your_video.mp4 --output outputs/videos/tracked_output.mp4

Configuration

The system uses YAML configuration files. The main demo uses configs/config_demo.yaml:

python main_demo.py --config configs/config_demo.yaml --input your_video.mp4

System Architecture

Detection

• YOLO-based player detection with confidence filtering
• Support for multiple player classes and ball detection
• Boundary filtering to remove edge detections

Tracking

• Enhanced association using IoU, appearance, and motion cues
• Confidence-based track management with decay mechanisms
• Lost track buffer for handling temporary occlusions

Re-identification

• Lightweight feature extraction using color histograms and texture analysis
• Gallery system for storing and matching player appearances
• Jersey number OCR for additional verification

Visualization

• Real-time tracking visualization with player IDs
• Team color coding and confidence display
• Re-identification event highlighting

Configuration Options

Detection Settings

detection:
  conf_threshold: 0.4
  iou_threshold: 0.45
  max_det: 50

Tracking Parameters

tracking:
  track_high_thresh: 0.5
  track_low_thresh: 0.1
  track_buffer: 30
  appearance_thresh: 0.3

Re-identification

advanced_reid:
  enabled: true
  feature_extractor: "lightweight"
  feature_dim: 512
  similarity_threshold: 0.6

File Structure

player-tracking/
├── src/
│   ├── detection/     
│   ├── tracking/           
│   ├── features/    
│   ├── reidentification/ 
│   └── visualization
├── configs/         
├── data/                  
├── models/                
├── outputs/         
└── main_demo.py