Physiotherapy Exercises Recognition for Scoliosis Patients with Raspberry Pi-Based Computer Vision

Abstract

Computer vision applications expand across business, education, agriculture, medicine, and numerous other fields. Researchers have successfully applied these techniques in physiotherapy to analyze movement patterns during exercises. This research implements an efficient classification system on a Raspberry Pi 4 to observe and classify three physiotherapy exercises: Supine Neck Lift, Quadruped Thoracic Rotations, and Lumbar Side Bends, utilizing 300 recorded videos for model development. The system integrates Optical Flow techniques with deep learning through a two-stream Convolutional Neural Network for comprehensive spatial and temporal feature analysis. The architecture incorporates Gated Recurrent Unit classification to enhance movement recognition accuracy. The model, deployed on the Raspberry Pi 4, enables fast inference for video classification. Experimental results demonstrate consistent performance across exercises: Lumbar Side Bends at 90.90%, Quadruped Thoracic Rotation on 85.42%, and Supine Neck Lift on 88.36%, achieving an overall model confidence level of 88.23% and an overall system accuracy of 81.25%. This research advances accessible, self-managed physiotherapy care through innovative computer vision solutions, demonstrating the potential for efficient exercise monitoring in clinical applications.