时空序列联合约束下基于时空金字塔模型的套动作质量评估方法研究

Study on Quality Evaluation Method of Action Sets Based on Space-time Pyramid Model under Joint Constraints of Time-space Sequences

阐述基于多目视觉的动作空间和动作序列时间上的联合约束,采用列文伯格-马夸尔特方法进行多坐标系融合计算,将可见光视觉下无标记动作捕捉方法的姿态识别精度从厘米级别提升到毫米级别。同时针对连续动作姿态序列,提出基于空间分层结构和多时间尺度特征的时空金字塔网络建模方法,以及适用于多类套动作的套动作质量评估深度学习方法,并在KIMORE和UI-PRMD数据集上取得全面优于现有康复评估方法的效果。

This paper describes the unmarked motion capture technology based on multi-eye vision,greatly improves the accuracy of 3D reconstruction of human motion and continuous attitude measurement and quantitative analysis,and proposes a set of motion quality evaluation method based on spatio-temporal pyramid network model.Specifically,this paper presents for the first time the joint constraints of action space and action sequence time based on multi-eye vision,uses the LevenbergMarquardt method for multi-coordinate fusion calculations,and uses the unmarked motion capture method under visible light vision.The accuracy of gesture recognition has been improved from the centimeter level to the millimeter level,at the same time,for continuous action gesture sequences,a space-time pyramid network modeling method based on spatial hierarchical structure and multi-time scale features and a set of action quality evaluation suitable for multi-type sets of actions are proposed.Deep learning method,and achieved overall better results than existing rehabilitation evaluation methods on KIMORE and UI-PRMD datasets.