Football video analysis using Deep Learning
summary
Combines classical CV and tailored CNNs (inspired by SSD) to detect ball and players from CCTV‑style long shots, trains on ~1,000 manually labeled frames and runs ~40 fps on 1920x1080. Reports ~90% ball and 94% player average precision, far faster than a general Faster R‑CNN baseline (≈8 fps). Builds a pipeline that adds postprocessing, camera calibration for pitch projection and 3D triangulation, and tracking‑by‑detection with RNN‑encoded appearance+motion and Hungarian assignment, and shows these components working in real game footage.
AI-written summary · about
chapters (15)
- 0:00Football video analysis intro
- 3:27Pipeline overview
- 4:28Ball/player detection challenges
- 7:22Real-time ball detection NN
- 10:13Top-down connections for detection
- 12:18Training with annotated data
- 13:46Detector results and metrics
- 16:39Post-processing false balls
- 17:46Player tracking with MOT
- 20:23Occlusions and similar appearances
- 22:33Player motion features
- 25:14Q&A on multi-camera 3D tracking
- 31:31Field/false positives filtering
- 33:20Jersey number recognition challenges
- 35:08Closing remarks and future
related talks
27mAnalysis of sports scenes using computer vision models
42mDeep Learning for Automatic Basketball Video Production
20mAI-Driven Real-Time Ball Tracking for Live Sports Streaming
42mHow Deep Learning is Revolutionizing Sports Analytics
38mDeep Learning plays Handball
51mVideo Analytics for Football games
01h 01mA practical guide to deep learning
10mLightning Talk: Ball Tracking and Detection in Soccer Videos - Comparison of...