LAUSR

Human pose estimation (HPE) technology, a vital tool for assessing exercise posture by extracting the three-dimensional coordinates of each joint, has been applied in many studies. It is effective for determining static postures, such as yoga; however, it still presents challenges in assessing dynamic exercise that involves considering each joint’s velocity. Although the HPE technology enables the derivation of position and velocity from each joint, assessing exercise posture using multiple time-series data requires time consumption and expert knowledge. Therefore, this study addressed this challenge by introducing a method for determining the velocity-based exercise posture, which combines the coordinates of significant-extracted joints using HPE with the relative-phase method. The relative phase angle ( $\Delta \phi _{Angle}$ ) is valuable for assessing the combination of position and velocity. This study added the relative phase distance ( $\Delta \phi _{Distance}$ ). An experiment was conducted to compare the exercise postures of experts and beginners during barbell back squats using a constructed dataset of time-series data for the positions and velocities of each joint and the relative phase Angle and Distance. Training and prediction were performed using a one-dimensional deep learning model. The results demonstrated the effectiveness of the proposed index in velocity-based exercise assessment with over 95% accuracy and confirmed the robustness of the method without requiring expert knowledge in real time. This study has significant implications for practical application in sports science and biomechanics. It has the potential to revolutionize the assessment and improvement of exercise posture.