A Vision-Based Exergame for Real-Time Posture Evaluation Using Pose Estimation and Adaptive Feedback in Upper Back Rehabilitation

Abstract

Upper back pain is one of the most common musculoskeletal complaints in sedentary populations, but supervised physiotherapy is costly and difficult to access outside urban centers. This paper proposes a vision-based exergame that evaluates rehabilitation posture in real time from a single consumer webcam. Skeletal landmarks are extracted from a pre-trained pose estimation model, joint angles are calculated using vector geometry, and each angle is classified against an exercise-specific reference range for shoulder rolls, scapular retractions, and cervical extensions. Instead of counting repetitions, the score is based on the quality of execution. A posture is held within its range of motion, and points are awarded according to the time. Fifty participants attended a structured session on standard laptops. We collected and report separately two classes of measure: objective system measures (execution success, key point detection accuracy, end-to-end latency and sustained frame rate) and subjective acceptance measures (System Usability Scale, perceived effectiveness, engagement). The pipeline ran to completion in every trial, key point detection accuracy was 94.3%, mean end-to-end latency was 31 ms, and the System Usability Scale score of 87.5 sits above the 70–80 band commonly reported for digital rehabilitation tools. The system is further benchmarked against traditional physiotherapy, AI-based mobile applications and IoT sensor platforms on hardware cost, feedback timing and engagement support. Clinical outcomes such as range-of-motion gain or pain reduction were not measured, so the evidence reported here concerns system performance and user acceptance rather than therapeutic efficacy.