← Back to Projects
Exercise Classification with BlazePose
Video-based exercise classifier using pose estimation and GBM achieving 90%+ accuracy.
BlazePoseMediaPipePythonScikit-learnGBMOpenCV
Overview
Built an exercise classification system using MediaPipe BlazePose to extract framewise pose tracking data from workout videos. The pipeline generates a training dataset from skeletal keypoints, which is fed into a Gradient Boosting Machine (GBM) classifier. At inference time, the model predicts exercise type for each frame, and majority voting across all frames determines the final classification, achieving over 90% accuracy.
Key Features
- ✓Framewise pose tracking with BlazePose
- ✓Custom training dataset generation from videos
- ✓GBM classifier for exercise recognition
- ✓Majority voting for robust predictions
- ✓90%+ classification accuracy