Analysis of sports scenes using computer vision models
summary
Recorded a handball dataset of ~4 hours from three GoPro/action cameras with players aged 14–17 and train models end to end: Mask R-CNN and YOLO for player detection (transfer learning yields ~90% person accuracy, ball detection still fails), tracking pipelines that suffer ID switches under occlusion, 2D pose estimation reaching ~94% correct keypoints after retraining, and an action-localization system using an I3D two-stream (RGB+optical flow) model with frame-sampling strategies; system works on training data but live match robustness and small-ball tracking remain unsolved.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 3:19Handball dataset creation
- 6:21Player detection with deep learning
- 8:11Player tracking in games
- 10:18Player pose detection
- 12:47Action recognition
- 15:42Sensor-camera fusion
- 18:11Temporal action localization
- 19:14Challenges and future directions
- 21:24Live testing applications
- 23:57Model adaptation for sports
- 25:19Q&A
related talks
35mFootball video analysis using Deep Learning
42mDeep Learning for Automatic Basketball Video Production
20mAI-Driven Real-Time Ball Tracking for Live Sports Streaming
10mLightning Talk: Ball Tracking and Detection in Soccer Videos - Comparison of...
22mSemantic segmentation of tennis match videos
31mCV at the Dutch Tennis Federation: Utilizing YOLO to create insights for coaches
51mVideo Analytics for Football games
36m"Lights, camera, AI!" - Automated sports videography