| Abstract [eng] |
This thesis investigates the analysis of back muscle electromyograms (EMG) and video data for assessing the quality of physical exercise performance. The research is conducted within the context of the international REMO project, focusing on remote rehabilitation solutions for nonspecific low back pain management. The aim of the study is to develop a mathematically grounded methodology integrating EMG signal analysis and biomechanical indicators for exercise quality assessment. A literature review of EMG signal processing and biomechanical analysis methods was conducted. EMG signals were processed using filtering, and continuous wavelet transform-based time-frequency analysis. Mean and median frequency indicators (MNF and MDF), intermuscular coherence, and phase synchronization metrics were analyzed. For video analysis, OpenPose, MediaPipe, and YOLO26-Pose human pose estimation models were evaluated. Biomechanical indicators and kinematic profiles were extracted from skeleton keypoints. The results showed that the YOLO26-Pose model provided the best balance between computational speed and pose estimation accuracy. EMG analysis revealed strong correlations between MNF and MDF indicators, as well as significant intermuscular coherence changes during different exercises. Phase synchronization indicators were found to be less sensitive to amplitude fluctuations and suitable for evaluating intermuscular interaction. The generated biomechanical profiles enabled motion phase identification and EMG signal segmentation according to movement phases. The obtained results demonstrate that multimodal EMG and video data analysis can be applied for automated physical exercise quality assessment in remote rehabilitation systems. |
