基于运动数据时空特征提取的人类运动片段分割方法

Segmentation of Human Motion Segments Based on Spatio-Temporal Feature Extraction of Motion Data

摘要: 动作示教方法是非专家用户对人形机器人进行控制的可靠形式，而对人类动作数据的运动分割与理解是其前提。利用现有方法对所捕获人类运动原始数据进行关键帧提取与运动分割时，由于数据特征不明确，导致难以准确定位运动起始帧、结束帧及分割帧。本文提出一种基于运动数据时空特征提取的运动分割方法。首先，对动态捕捉的人体运动数据进行预处理；然后，基于预处理数据提取人体质心速度、加速度以及简化倒立摆的摆角，从而揭示动态平衡时空特征；随后，提取步长、步高、步频、触地状态等步态参数，用于表现运动的周期性时空特征；从而构造出具有时空特征的运动数据集。最后，基于运动数据集的特征矩阵和运动数据，采用概率主成分分析(Probabilistic Principal Component Analysis, PPCA)方法进行运动分割，获得多种类型运动片段。为验证所提方法的有效性，进行了多种运动的捕捉与分割，并与基于原始数据的分割结果进行对比，结果表明本文方法具有更高的查准率和查全率。

Abstract: Motion demonstration is a reliable method for non-expert users to control humanoid robots, and the segmentation and understanding of human motion data are prerequisites for it. Existing al-gorithms struggle to extract key frames accurately because of multi-modal features of human motion data from motion capture are not intuitive. This paper proposes a motion segmentation method based on spatio-temporal feature extraction of motion data. Firstly, the dynamically captured human motion data are preprocessed; then, based on the preprocessed data, the human centre-of-mass velocity, acceleration, and the simplified pendulum angle of the inverted pendulum are extracted so as to reveal the dynamic equilibrium spatio-temporal features; subsequently, gait parameters such as stride length, stride height, stride frequency, and touchdown state are extracted for expressing the periodic spatio-temporal features of the motion; thus, the motion dataset with spatio-temporal features is constructed. Finally, based on the feature matrix and motion data of the motion dataset, Probabilistic Principal Component Analysis (PPCA) method is used for motion segmentation to obtain multiple types of motion segments. In order to verify the effectiveness of the proposed method, a variety of motion capture and segmentation are carried out and compared with the segmentation results based on the original data, and the results show that the method in this paper has a higher rate of checking accuracy and completeness.