基于动作标准序列的3D视频人体动作识别

Human action recognition for 3D video based on action standard sequence

基于3D视频的人体动作识别近年来受到越来越广泛的关注。基于动态时间规整的算法考虑了动作的时序信息,并能较好地解决人体运动在时间上的不确定性,但是随着训练样本增加,效率会变得较低。本文提出了一种基于动作标准序列的动作识别方法。通过特征提取将3D动作视频样本构建为动作序列,在动态时间规整度量下将动作标准序列学习建模成一个序列平均的优化问题,并使用动态时间规整重心平均算法(DBA)求解。对于动作类别类中存在显著差异的场景,研究了多重动作标准序列学习,并针对无监督学习的情况,提出了DBA-K-means聚类算法。实验结果表明,该方法可进一步提高动作识别的效率和准确率。

Human action recognition for 3D videos has taken more and more attention in recent years.Approaches based on Dynamic Time Warping(DTW)method consider the information of sequential order and can deal with the temporal uncertainty of action.But with the increase of training action samples,the efficiency of action recognition decreases.In this work,a new framework is designed for action recognition based on the action standard sequence.First,the action sequences is constructed from 3D action video samples by feature extraction.Then,the learning of action standard sequence is modeled as an optimization problem of sequence averaging under DTW measure,and the problem is solved by DTW Barycenter Averaging(DBA)algorithm.Furthermore,the learning of multiple action standard sequences is studied for the situation where there is large intra-class variation within one action category,and DBA-K-means algorithm is proposed for the unsupervised learning of multiple standard sequences.The experiment results show that both accuracy and efficiency can be improved by the proposed approach.