Padel Match Analytics Pipeline

A computer vision system that turns a single, ordinary broadcast camera into detailed padel match analytics, with no special sensors or multi-camera setup. From that one video feed it tracks all four players and the ball, sorts out the teams, reads each player's body pose, and recognizes shot types, including padel-specific shots like the bandeja, volea, and wall defense. The output is per-player stats: court coverage, shot selection, rally structure, and wall play. Fine-tuned on PadelTracker100: 2 full matches, ~100k labeled frames spanning ball, pose, and stroke annotations. Every stage reports held-out metrics (MOTA/IDF1 for tracking, per-frame ball recall) against both an in-distribution test match and an out-of-distribution clip, specifically to catch single-venue overfitting.

Dataset

PadelTracker100 (~100k labeled frames, 2 pro matches), PadelVic training frames, A1 broadcast clip held out for out-of-distribution eval

Models

YOLOv11, BoT-SORT with a custom court-space ID tracker, TrackNetV3 / WASB ball tracker, Keypoint R-CNN pose (RTMPose upgrade path), multi-modal cross-attention stroke classifier, rule-based event layer

Evaluation

Detection mAP and recall, tracking MOTA / IDF1, ball F1@4px, per-stroke F1, pose PCK, wall-rebound precision, each on held-out splits with an out-of-distribution clip alongside

Results & insight

Per-player coverage, shot types, rally structure, and wall-play from one fixed broadcast camera. The out-of-distribution eval caught a detector that scored 0.95 recall in-distribution but fell to 0.45 on unseen venues, recovered to 0.86 with multi-venue retraining. Player tracking generalizes; ball tracking and stroke classification are the remaining gaps.