Development of a Markerless Computer Vision Prototype for Range-of-Motion Evaluation in Rehabilitation

Abstract

The work develops a markerless computer vision software prototype for real-time joint range of motion (ROM) estimation during rehabilitation exercises, using only a standard webcam (< USD 50), Python 3.10, OpenCV 4.8, and MediaPipe Pose.

The system detects 33 anatomical landmarks per frame, computes joint angles through 2D vector geometry, and simultaneously monitors eight bilateral joints: elbows, shoulders, knees, and hips. It incorporates a 5-frame sliding window filter to reduce frame-to-frame noise and a visibility threshold of 0.6 to discard unreliable estimates.

In preliminary testing with three healthy volunteers, the system operated at 28.4 FPS without GPU acceleration, achieved a mean absolute error of 5.3° against manual goniometry (below the 10° minimum clinically important difference), and reduced frame-to-frame angle noise by 68%. All eight functional test cases met the acceptance criteria.

The authors position it as an early-stage prototype with motivational relevance to ALS — where ROM tracking could complement the ALSFRS-R scale — but without clinical validation to date. Future work proposes extending validation to diverse populations, non-controlled environments, and ethics-approved clinical studies with ALS patients.