Repository files navigation

Anend-to-end, production-grade tennis video analysis system leveraging state-of-the-art computer vision and deep learning. This repository integratesUltralytics YOLOv8 for robust player detection and tracking, a fine-tuned model forhigh-velocity ball detection, and acustom ResNet50-based CNN for precise court keypoint extraction. The system transforms raw match video intophysics-aware metrics (e.g., shot speed, per-player distance covered, trajectory reconstructions) andtactical visualizations suitable for coaching, scouting, and advanced analytics workflows.

Tennis-Vision-AI solves the complex computer vision challenges inherent to court sports analytics: maintainingpersistent player identity during rapid interactions, detectinghigh-velocity, small-object balls amidst motion blur, and establishing ageometrically accurate mapping from video pixels to real-world court coordinates.

The system is engineered fordeterminism and auditability—from annotated training sets and serialized model weights to final exported telemetry—making it suitable for bothacademic research andapplied deployment in professional sports environments.

To ensure high-precision tennis analytics, this project relies on carefully curated datasets forball detection,player tracking, andcourt keypoint regression.

• Exported fromRoboflow with ~428 annotated images used for fine-tuning.
• Captures challenging conditions:motion blur, varyinglighting, andpartial occlusions.
• Optimized forhigh-velocity small-object detection.

• Based on standardperson/player annotations, adapted to tennis-specific viewpoints.
• Handlesclose-contact scenarios andocclusions typical of tennis matches.

• 14 labeled points per frame representing canonical court lines and intersections.
• Enablesaccurate homography-like mapping from video pixels → real-world court coordinates.

• Maintaintrain/validation/test splits, plus a dedicatedevaluation set withhigh-velocity rallies.
• Ensure dataset balance across:
  • Camera angles
  • Court surfaces
  • Lighting conditions
• This supports robustgeneralization for diverse match scenarios.