Keywords
Pose-estimation, Posture, Templates, Feedback, Modular

Technologies
Python, MediaPipe-Pose-Landmarker, OpenCV, Streamlit

This project set out to build a real-time, reusable framework for analysing a single user's pose from a standard camera and turning noisy 2D pose landmarks into clear, interpretable feedback. The main objective was to avoid hard-coded, domain-specific logic by using configurable templates that define rules, thresholds, and feedback, allowing the same core pipeline to be reused across different applications.

The implemented system uses MediaPipe Pose Landmarker to detect body landmarks and derives geometric features such as joint angles, orientations, distances, bearings, and key reference points. These features are evaluated by declarative rules to support live on-screen feedback, and the framework is delivered through a browser-based interface with schema validation, hot reload support, and safe template handling.

To make live feedback more stable, the system uses low-latency smoothing together with dwell and cooldown timing so brief jitter does not cause flicker or double counting. The result is a modular template-driven pipeline that supports posture monitoring, exercise feedback, rehabilitation, and ergonomics scenarios without requiring changes to the source code.