Human pose estimation for physiotherapy applications

Physiotherapy is a pivotal component in rehabilitation for people who have been injured or underwent surgery. Conventionally, physiotherapists observe patients' exercise movements and supply real-time correction. Nonetheless, human observation is frequently constrained by availability and expense. This work introduces an automated system for estimating human pose based on deep learning models, aimed specifically at physiotherapy applications. The system uses pre-trained models including MoveNet Thunder, MoveNet Lightning, PoseNet, and Mediapipe's BlazePose, and compares their performance for real-time tracking of human pose. Moreover, a custom-trained model (YOLOv8-pose) is designed based on the University of Cordoba's UCO physical rehabilitation dataset, streamlined for physiotherapy-related tasks. The work embeds the most efficient model in an Android mobile app, which lets patients carry out exercises independently with real-time correction of posture. The work offers a description of current pose estimation systems, points out research shortcomings, as well as a thorough analysis of applied models. The results establish that a custom-trained YOLOv8-pose model outperforms pre-trained models significantly in physiotherapy-related exercises, and potential improvements consider increasing the dataset as well as 3D pose estimation.