Football Vision – End-to-End Broadcast Tracking & Analytics

Turn single-camera TV broadcasts into metric player/ball trajectories and football insights (distance, speed zones, heatmaps, possession proxy, formations, roles).
The pipeline runs frame-by-frame, with class-aware tracking, jersey OCR, and per-frame homographies from PnLCalib.

---

✨ What’s inside

• Detection: Domain-adapted YOLOv8 (players, goalkeeper, main ref, side ref, ball, others).
• Class-aware tracking: BoT-SORT for people (optional ReID on GPU), BoostTrack for the ball.
• Team clustering: Simple K-Means on jersey colors with periodic refits.
• Jersey OCR: PARSeq with a legibility gate and temporal voting for stable shirt numbers.
• Calibration: PnLCalib keypoint/line nets → per-frame homography (H_t) + stability gate.
• Mapping: Apply (H_t^{-1}) to foot points → metric pitch (105×68 m).
• Analytics: Distance & speed zones, positional heatmaps, possession measuring, formations, event detection.

---

🔧 Installation

pip install -r requirements.txt

Download: YOLOv8n Detection Model

Download: PARSeq_tiny Jersey OCR model

Download: Resnet34 Player crop Legibility Classifier (Origin-Repo)

Download: PnLCalib Keypoint Recognizer (Origin-Repo)

Download: PnLCalib Line Recognizer (Origin-Repo)

---

🚀 Quickstart

1) Detection + class-aware tracking (+ OCR)

python -m tracking_pipeline.main \
  path/to/input_video.mp4

Outputs

• Video preview with overlays: output/..._output_video.mp4

• Per-frame YOLO labels (optional): outputDet/<seq>/labels/000001.txt

• Tracking CSV (MOT-like w/ jersey):

  frame,id,x,y,w,h,conf,class,jersey

  • class is a string label (e.g., 'L'/'R' for teams, '2' ball, '3' ref, …)
  • jersey -1 means unknown

2) Camera calibration (PnLCalib → homographies)

cd PnLCalib
python inference_modified.py \
  --weights_kp weights/hrnetv2_w48_kp.pth \
  --weights_line weights/hrnetv2_w48_line.pth \
  --input_path path/to/input.mp4 \
  --input_type video \
  --device cuda:0 \
  --batch_size 8 \

Output

• Homographies CSV with rows:

  frame,H00,H01,H02,H10,H11,H12,H20,H21,H22

3) Map to 2D + overlay pitch

cd PnLCalib
python transformandplotsmooth_modified.py \
  --video path/to/input.mp4 \
  --tracking path/to/tracking_results.txt \
  --transforms_csv output/homographies.csv \
  --output_video output/overlay.mp4 \
  --output_csv output/coords.csv

Outputs

• Overlay video with a drawn pitch and 2D points

• 2D CSV:

  frame,id,model_x,model_y,class,jersey

  where model_x/y are in meters on a 105×68 m pitch.

4) Analytics (notebook)

Open Analysis.ipynb and point it to the 2D CSV to compute:

• distance & speed zones,
• positional heatmaps,
• possession/pass proxy (nearest-player rule to ball),
• formation & role clustering,
• simple event timeline.

---