Analysis of Different Sensor Modalities for Movement Classification in Physical Therapy

This study investigates feature analysis and feature fusion from different sensor modalities for the task of identifying movement errors in physiotherapeutic exercises, using squats as a case study. Incorrectly performed exercises can lead to injuries, underscoring the need for accurate monitoring tools. In an experiment, ten participants performed squats in three variations: correct execution, forward lean, and lateral tilt. To identify movement patterns, we evaluated muscle activation through electromyography (EMG), kinematic data through Motion Capture (MoCap) and joint angles from video footage through MediaPipe Pose. Distinct movement patterns were identified for the erroneous variations: forward lean altered hip and knee angles, while lateral tilt caused asymmetries in posture. In the EMG signal, deviations in the activity of distinct muscles correlated clearly with specific erroneous movements. Activation in the Gluteus Maximus was higher for the forward lean, while activity in the Quadriceps was lower. For the lateral tilt, a clear difference between left and right muscle activation was visible. Signal processing techniques extracted key features, such as muscle activation peaks and joint angle deviations, that we used to discern between the different squat types with a decision tree model. MoCap-based features offered the highest precision when used on their own, but fusing different sensor modalities achieved the best results. Although the video-based classifications were less accurate, its cost-effectiveness and ease-of-use suggest potential for home rehabilitation. Future research should enhance marker-less technologies and enable real-time feedback for broader applications.