From Athlete to Algorithm: Transforming Canoe Technique Analysis with AI
summary
Pipeline performs foreground–background segmentation and waterline derivation, uses pose detection (trained neural net via transfer learning on ~30,000 labeled frames) plus PCA feature tracking and temporal filtering to locate paddle blade and compute paddle angle and angular velocity. It relies on OpenPose, a retrained MobileNet classifier and fallback YOLO detectors, plus hand/finger refinement and sigmoid state filtering. Initial accuracy is 37% on varied videos, training hits 95%, production yields ~70% exact, +20% one–two frames off (≈90% acceptable), cutting manual review from 20 to 2 minutes per video and producing annotated video, CSV and LaTeX reports.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 3:16Athlete to algorithm with AI
- 10:13Computer vision basics
- 16:56Finding paddle poses
- 24:23Initial results and challenges
- 29:43Improvements and accuracy
- 33:55Automation and coach feedback
- 34:57Future use cases and conclusion
related talks
44mFrom Athlete to Algorithm: Transforming Canoe Technique Analysis with AI
27mAnalysis of sports scenes using computer vision models
24mAI in maritime - Autonomous ships
35mFootball video analysis using Deep Learning
22mSemantic segmentation of tennis match videos
41mKEYNOTE: Image processing, AI, and autonomous systems
17mAI/ML demo derby
33mUsing machine learning to improve player safety in the NFL