Universal Physical Education: Training-free Action Quality Assessment Based on Dynamic Time Warping

Renguang Chen, Guolong Zheng, Xu Yang, Zhide chen, Xuechao Yang, Xun Yi

Keywords: Action Quality Assessment, Pose Estimation, Adaptive Constrained Dynamic Time Warping, Training-free

Abstract: The increasing integration of video-based evaluation into sports education underscores the need for efficient and adaptable methods to assess action quality. Existing approaches, often reliant on deep learning models and extensive labeled datasets, face challenges in generalizing to novel exercises or scenarios with limited training data. This paper introduces the Multi-Dimensional Exercise Distance Adaptive Constrained Dynamic Time Warping (MED-ACDTW) method, a training-free algorithm designed for action quality assessment. By leveraging features extracted from both 2D and 3D spatial dimensions and integrating multiple human body attributes, our method achieves a 2-3\% improvement in accuracy over single-dimensional pose estimation techniques. The proposed adaptive constraint scheme further enhances action quality discriminability, achieving a 30\% improvement in precision compared to conventional methods. Additionally, we introduce BGym, a novel dataset tailored for evaluating action quality in diverse and non-standardized perspectives. The MED-ACDTW method eliminates the dependency on labeled datasets, enabling direct comparison between template and test videos, thus offering a scalable and versatile solution for real-world applications.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 15729