Preprint: Human-Size Interactive Mirror for Real-Time Movement Feedback: A Technical Framework for Physiotherapy Applications

Abstract

We present a modular framework for a life-size interactive mirror based on stereo RGB cameras, markerless human pose estimation, and real-time 3D avatar rendering. The system reconstructs the user’s three-dimensional body pose from two camera views, maps the resulting keypoints to a humanoid VRM avatar, and renders the avatar as a perspective-aware virtual reflection on a large display. We evaluate multiple HPE configurations from MediaPipe and RTMLib with respect to processing performance and robustness under varying subject distance, partial occlusions, additional people, and temporal stability. MediaPipe Full provided the best overall balance between pose quality and processing speed and was therefore selected for the prototype. A proof-of-concept interactive application demonstrates real-time movement-based interaction with the virtual avatar. The framework provides a technical foundation for future movement-feedback applications, including physiotherapy, where exercise-specific analysis and visual feedback can be investigated.