Computer vision pipeline for analyzing padel tennis game videos. Processes video through court calibration, player detection and tracking, ball tracking, scoreboard OCR, and analytics to produce annotated videos and structured data exports.

• Court Calibration - DeepLSD line detection + template fitting for court registration
• Player Detection & Tracking - YOLO11 + ByteTrack for up to 4 players
• Ball Detection & Tracking - TrackNetV4 model with Kalman filtering and event detection
• Scoreboard OCR - PaddleOCR-based score extraction
• Analytics - Speed, acceleration, zone occupancy, formations, rally metrics
• Export - Annotated video, minimap overlay, JSON, and CSV outputs

• Python 3.10 or 3.11
• NVIDIA GPU with CUDA 12.6+ (or AMD GPU with ROCm)
• uv package manager

git clone --recurse-submodules https://github.com/<your-org>/padle-game-analysis.git
cd padle-game-analysis

If you already cloned without submodules:

git submodule update --init --recursive

uv sync --dev

This installs all Python dependencies including PyTorch with CUDA 12.6 support, ultralytics (YOLO11), PaddleOCR, and DeepLSD.

Create the models directory:

mkdir -p data/models

DeepLSD weights (court line detection):

wget -O data/models/deeplsd_md.tar https://cvg-data.inf.ethz.ch/DeepLSD/deeplsd_md.tar

TrackNetV4 model (ball detection):

1. Clone the TrackNetV4 repository:

git clone https://github.com/TrackNetV4/TrackNetV4.git third_party/TrackNetV4

2. Download TrackNetV4 .keras weights from Google Drive and set ball_tracking.tracknetv4.model_path in configs/default.yaml. The default config currently points at:

data/models/best_model_V1_NF_RIO_10u_e17.keras

Available weight variants:

• Variant 1
• Variant 2
• Variant 3

3. Install TensorFlow (required for TrackNetV4):

uv sync --extra tracknetv4

YOLO11 and PaddleOCR models are downloaded automatically on first run.

uv run python scripts/run_mvp.py --video <input.mp4> --config configs/default.yaml --out data/outputs/

Options:

data/outputs/
├── tracks.json          # Player tracks and ball detections
├── metrics.csv          # Per-frame kinematics
├── annotated.mp4        # Video with overlays
└── minimap.mp4          # Top-down court minimap

All pipeline parameters are in configs/default.yaml. Key sections:

• court - Court dimensions (standard padel: 10m x 20m)
• detection - YOLO model variant and confidence threshold
• tracking - ByteTrack hyperparameters
• calibration - DeepLSD settings and weights path
• ball_tracking - TrackNetV4 model path and Kalman filter tuning
• scoreboard - OCR engine and sampling interval
• export - Output formats and video options

The project includes dev container configurations for both NVIDIA (CUDA) and AMD (ROCm) GPUs in .devcontainer/. Open in VS Code or GitHub Codespaces for a pre-configured environment.

uv run pytest

uv run ruff check src/ scripts/ tests/

src/
├── calibration/       # Court registration (DeepLSD, template fitting)
├── detection/         # Player detection (YOLO11)
├── tracking/          # Player tracking (ByteTrack, identity assignment)
├── ball_tracking/     # Ball detection (TrackNetV4, Kalman filtering)
├── coordinates/       # Homography projection, smoothing
├── analytics/         # Kinematics, zone analysis, rally metrics
├── scoreboard/        # Scoreboard detection, OCR, state machine
├── export/            # JSON, CSV, video writers
├── visualization/     # Overlay annotations, minimap
├── video_io/          # Video I/O utilities
├── config/            # Configuration loader
└── schemas.py         # Pydantic data models
scripts/
├── run_mvp.py         # Full pipeline entry point
└── evaluate_*.py      # Component evaluation scripts
configs/
└── default.yaml       # Default pipeline configuration
third_party/TrackNetV4/ # TrackNetV4 ball detection (cloned separately)