AI-Driven Real-Time Ball Tracking for Live Sports Streaming
summary
An end-to-end system detects and tracks the ball in live locked‑off streams using CNNs and YOLO-SORT integrated into an optimised GPU cloud architecture, producing smooth crops at 30 fps at 1920x1080 with ~15–20 ms processing latency. Per‑frame detection feeds a multiobject tracker and a Kalman filter that interpolates through occlusions, rule‑based centroid fusion decides framing, and deployed tests across multiple sports report robust operation on modest hardware including A10 GPU setups with five CPU cores.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 1:04AI in video processing
- 2:26Live ball tracking demo
- 3:32Real-time challenges
- 5:22Applications beyond sports
- 6:25Architecture constraints
- 8:12Object detection pipeline
- 10:02Key event detection
- 11:51Window framing smoothing
- 13:58Centroid modeling
- 15:01Live pipeline with Zeus
- 15:44Portability across sports
- 16:43Football tracking demo
- 17:03Video format conversion
- 17:47Tracking for analysis
- 18:51Longtail sports systems
- 19:34Deployment and future
related talks
42mDeep Learning for Automatic Basketball Video Production
51mVideo Analytics for Football games
35mFootball video analysis using Deep Learning
27mAnalysis of sports scenes using computer vision models
10mLightning Talk: Ball Tracking and Detection in Soccer Videos - Comparison of...
29mAI in Sports - Basketball Rule Violation Detection
31mCV at the Dutch Tennis Federation: Utilizing YOLO to create insights for coaches
30mWhen performance matters: running a real-time computer vision pipeline