Counting exercise repetition using live camera video
summary
Uses MediaPipe pose landmarks to read 33 body points, reduces them to 16 metrics (8 bones × X/Y) and runs PCA to find the dominant motion axis, then detects peaks on that projection to count reps. Runs entirely on-device (Flutter iOS/Android/web) at about 8 frames per second, keeps ~10 seconds of history and uses dual 3–4 second windows to detect when the exercise pattern changes. Source code and apps exist and the pipeline design is modular and working.
AI-written summary · about
chapters (9)
- 0:00Exercise tracking demo intro
- 1:15MediaPipe pose detection
- 2:20Calculating repetitions with principal component analysis
- 4:50Tracking multiple body points
- 6:59Peak and exercise change detection
- 10:29Finalizing repetition algorithm
- 12:43Key takeaways and apps
- 14:07Tracking system pipeline
- 16:16Q&A
related talks
18mFitness motion recognition with Android wear
30mPoseDance: Build a TikTok Trainer
40mPoseDance Build a TikTok Trainer
33mSee More, Be More
33mComputer Vision, Not Just for Breakfast Anymore
26mYoga Pose
06mIMUTube: Automatic extraction of virtual on-body accelerometry from video for human activity ...
13mUltraPoser: Pushing the Limits of IMU-based Full-Body Pose Estimation with Ultrasound Sensing on ...