A Low-cost System for Markerless Body Part Interaction Detection in Physical Therapy

Yashar Talebirad, Simon Nakatsu, Brenda Luo

Keywords: computer vision, opencv, mediapipe, physical therapy

TL;DR: A low-cost webcam-based system achieves 96% accuracy in detecting self-touch interactions for tele-rehabilitation by using dynamic thresholding that adapts to body size and temporal filtering for tracking multiple simultaneous touches.

Abstract: This paper presents a low-cost markerless system for detecting body part interactions in tele-rehabilitation using only a single RGB camera. The system employs two key innovations: (1) dynamic thresholding normalized by inter-shoulder width to adapt to varying camera distances, and (2) temporal filtering with sliding window majority voting for robust multi-touch detection. A 2×2 experimental evaluation comparing dynamic vs. fixed thresholding with and without temporal filtering shows that dynamic thresholding is the primary performance driver, achieving 96% F1-score during contact phases compared to 87% for fixed thresholding. The temporal filtering mechanism proves essential for simultaneous multi-touch detection while maintaining high precision (98.5%).

Submission Number: 61