Deep Learning for Injury Rehabilitation

This master's thesis presents the design, implementation, and evaluation of a deep learning system for automated injury rehabilitation exercise recognition. The system is framed within a human-centric paradigm, designed to augment rather than replace the expertise of physiotherapists by automating remote patient monitoring and adherence tracking. The methodology involves a data pipeline that transforms raw video into structured kinematic time-series data using Google's MediaPipe framework to extract 33 body landmarks. A novel classification model based on the Transformer architecture is developed to interpret these complex sequential movements. The model's architecture and hyperparameters were systematically optimized using the Keras Tuner library to ensure high performance. The final model achieved a robust overall accuracy of 85.25% and an Average Precision of 0.91 on a diverse test dataset. A detailed per-class analysis revealed excellent performance for exercises with distinct kinematic profiles (e.g., 'plank' with an F1-score of 0.98) but identified limitations with exercises dominated by sagittal-plane motion (e.g., 'romanian deadlift' with an F1-score of 0.74). This performance variance is attributed to the inherent depth ambiguity of the monocular camera input. The thesis concludes by affirming the viability of the Transformer architecture for this task and proposes future work focused on integrating true 3D sensors to enable quantitative form correction.