一种使用3D骨架片段表示的人体动作识别方法

Using the Representation of 3D Skeleton Snippet for Human Action Recognition

提出一种新的使用3D骨架片段表示的人体动作识别方法.人体动作可以表示为一系列3D骨架坐标(即一系列高维空间中的点).通过传统的滑动窗口方法,将一个人体动作划分为3D骨架片段序列.对于每一个骨架片段,计算该片段所包含点(3D骨架坐标)的均值及协方差矩阵,其中均值代表了人体动作在这一时间段的主要姿势,提取协方差矩阵的前T个特征向量则代表了人体动作在这一时间段的主要动作趋势,两者相结合即是一个全面的对人体动作在该时间段的表示.方法定义了3D骨架片段与片段之间的距离,该距离的定义同时结合了主要姿势和主要动作趋势两方面特征.为了描述动作的全局时间关系,用动态时间规整算法来度量两个不同动作之间的距离,则测试动作的标签即为与它距离最近的训练动作的标签.本文方法在两个公开发布的人体动作数据集(KARD数据集和CAD60数据集)上进行实验,实验结果表明所提出的使用3D骨架片段表示的方法适用于人体动作识别.

In this paper, we propose a new approach to recognize human action, which uses the representation of 3D skeleton sequence- snippet. A human action is represented by 3D joint locations sequence,which can be seen as a series of high-dimensional points. We divide an action sequence into skeleton sequence-snippets through traditional sliding window method. For each skeleton sequence-snip- pet, we compute the poims mean as Major Posture Feature { MPF } and extract the top-K eigenvectors of corresponding covariance ma- trix as Main Tendency Feature{ MTF）. The representation of the snippet is the combination of MPF and MTF, and we define a 3D skeleton sequence snippet-snippet distance measurement method which takes both MPF and MTF into consideration. To encode global temporal relationship,Dynamic Time Warping （DTW） algorithm is used to measure the distance between different human actions which is composed of a series of sequence-snippets. We assign the test action to be the train action label that has the closest distance. Our approach is applied in the publicly available KARD and CAD60 datasets, and experiment results show that the proposed approach is effective for human action recognition task.