Sports Analytics · Padel Vision v2
Padel match analytics
Padel Vision v2 turns a single game recording into broadcast-style analytics: the live match view, stable player tracking, and a real-time 2D tactical map drawn to the court's real 10 m × 20 m scale.
The challenge
One recording, many signals
The pipeline has to extract players, identities, positions, and tactical context from a single match video.
Real metrics, not pixels
Padel stats only become meaningful when image positions are projected into the fixed 10 m × 20 m court.
Glass-wall dynamics
Future ball and shot detection must handle wall bounces, where simple tennis-style velocity changes are not enough.
The approach
RF-DETR player segmentation
One Roboflow RF-DETR model returns both player boxes and per-player masks for the live match view.
ByteTrack stable IDs
Multi-player tracking with ByteTrack, smoothing, and stable player IDs keeps the analytics connected over time.
Homography tactical map
Court calibration projects players' feet into a 2D court map, converting pixels into real-world metres.
OpenCV rendering pipeline
OpenCV, NumPy, and Supervision render the broadcast view, tactical map, heatmaps, and live metrics.
Outcomes
Match + tactical map
Live footage paired with a scaled 2D court view.
Player analytics
Heatmaps, distance, average speed, and work rate.
Richer match intelligence
TrackNet-style ball trajectory, shot detection, pose keypoints, and a full tactical dashboard.