Overview
This is peer-reviewed work, presented at the 4th IEEE International Conference on Artificial Intelligence and Signal Processing in October 2024. I am one of six authors on the paper, “Real-Time Human Pose Estimation Using MediaPipe and Artificial Intelligence Applications in Health and Fitness”.
The system detects a full 33-landmark skeleton, computes 3D joint angles and velocities from it, classifies activity with an LSTM, and counts exercise repetitions, with the whole chain holding real-time frame rates.
What the pipeline does
- Landmark detection of the full 33-point body skeleton through MediaPipe.
- Spatial reasoning on top of the raw landmarks: 3D joint angles and per-joint velocities, which is what turns a point cloud into something you can reason about.
- Activity classification with LSTM networks over the landmark time series.
- Repetition counting for exercise tracking, driven by the angle and velocity signals rather than by raw position.
Reported results
85% mAP on the COCO standard, 0.78 IoU, and 30+ FPS on GPU. These are the figures reported in the peer-reviewed paper and measured in that study, not numbers produced by this repository’s test suite. The repo ships benchmarking utilities so they can be re-run on other hardware and data, and the distinction between “published result” and “what this code asserts” is kept explicit in the README.