Football Player Tracking & Re‑Identification

A Python-based pipeline for detecting, tracking, and re-identifying football players (or other person-like objects) in video streams. It combines YOLO object detection, ByteTrack multi-object tracking, and OSNet person re-identification to assign consistent IDs to players across frames.

---

🔍 Features

• Accurate object detection using YOLOv11 (ultralytics).
• Robust multi-object tracking with ByteTrack algorithm.
• Person re-identification (ReID) via OSNet model (torchreid) to maintain consistent identities across occlusions.
• Customizable gallery: catalog feature embeddings and match new detections with a threshold.
• Color-coded bounding boxes for different classes (players, goalkeeper, ball, referees).
• Export annotated video with labeled bounding boxes and IDs.

---

🛠️ Requirements

• Python 3.11.4
• CUDA-enabled GPU (recommended) or CPU fallback

---

📥 Installation

1. Clone the repository:

   git clone https://github.com/KunalNath04/Re-id_initial.git
   cd Re-id_initial

2. Create and activate a virtual environment (optional but recommended):

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

3. Install dependencies:

   pip install -r requirements.txt

4. Download weights:

   • Place your YOLO model weights (best.pt) in models/
   • Place your input video (15sec_input_720p.mp4) in assets/
   • TorchReID will automatically download OSNet weights on first run.

---

🚀 Usage

Run the pipeline on a video file and generate an output with annotated tracks:

python object_tracking_reid.py

• Output is saved to the path specified by OUTPUT_FILE.
• Progress and final confirmation printed in console.

---

⚙️ Configuration

• Paths & parameters are set at the top of object_tracking_reid.py:

  YOLO_WEIGHTS   = "models/best.pt"
  VIDEO_FILE     = "assets/15sec_input_720p.mp4"
  TRACKER_CONFIG = "bytetrack.yaml"
  OUTPUT_FILE    = "output_reid_fixed.mp4"
  
  REID_THRESH    = 0.8           # gallery matching threshold
  CROP_SIZE      = (128, 256)    # width, height for ReID crops
  REID_CLASSES   = {"player"}    # classes to apply ReID

• ByteTrack settings are in bytetrack.yaml:

  tracker_type: bytetrack
  track_high_thresh: 0.25
  track_low_thresh: 0.1
  new_track_thresh: 0.6
  track_buffer: 30
  match_thresh: 0.8
  fuse_score: False

• Color scheme in object_tracking_reid.py:

  COLORS = {
      'ball':       (0, 255, 255),   # cyan
      'goalkeeper': (255,   0,   0),  # red
      'player':     (  0, 255,   0),  # green
      'referee':    (  0,   0, 255),  # blue
  }

---

📂 File Structure

├── assets/                      # Input videos, sample clips
│   └── 15sec_input_720p.mp4
│
├── models/                      # YOLO model weights
│   └── best.pt
│
├── configs/                     # Tracking configs
│   └── bytetrack.yaml
│
├── gallery.py                   # PersonGallery: stores & matches embeddings
├── reid_model.py                # ReidModel: feature extractor (OSNet)
├── object_tracking_reid.py      # Main tracking & ReID script
├── test.py                      # Quick test for torchreid imports
└── requirements.txt

📝 How It Works

1. Detection & Tracking

   • Load YOLOv11 detector and ByteTrack tracker.
   • For each frame: detect objects, associate detections across frames, assign ByteTrack IDs.

2. Crop & Embedding

   • For detections of class player, crop the bounding box region.
   • Resize to CROP_SIZE and feed to OSNet to extract feature embeddings.

3. Gallery Matching

   • Maintain a gallery of registered embeddings with unique UIDs.
   • For each new embedding, compute cosine similarity to gallery means.
   • If similarity ≥ REID_THRESH, reuse UID; else, register new UID.

4. Visualization

   • Draw colored bounding boxes and labels (<class> <UID>).
   • Non-player classes show <class>_<trackID>.

---

🔧 Customization

• Thresholds & sizes: adjust REID_THRESH and CROP_SIZE for different domains.
• Classes for ReID: modify REID_CLASSES to include additional person-like classes.
• Color scheme: change COLORS mapping for your preferences.
• Tracker parameters: tune bytetrack.yaml for detection/tracking sensitivity.

---

Happy tracking!