多聚点子空间下的时空信息融合及其在行为识别中的应用  

Recognizing Action Using Multi-center Subspace Learning-based Spatial-temporal Information Fusion

作　　者：杨天金 侯振杰[1,2] 李兴 梁久祯 宦娟[1] 郑纪翔[1] YANG Tian-Jin;HOU Zhen-Jie;LI Xing;LIANG Jiu-Zhen;HUAN Juan;ZHENG Ji-Xiang(School of Computer Science and Artificial Intelligence Aliyun School of Big Data School of Software,Changzhou 213164;Jiangsu Key Laboratory of Internet of Things Mobile Internet Technology Engineering,Huai'an 223003)

机构地区：[1]常州大学计算机与人工智能学院阿里云大数据学院软件学院,常州213164 [2]江苏省物联网移动互联技术工程重点实验室,淮安223003

出　　处：《自动化学报》2022年第11期2823-2835,共13页Acta Automatica Sinica

基　　金：国家自然科学基金(61803050,61063021);江苏省物联网移动互联技术工程重点实验室开放课题基金(JSWLW-2017-013);浙江省公益技术研究社会发展项目(2017C33223)资助。

摘　　要：基于深度序列的人体行为识别,一般通过提取特征图来提高识别精度,但这类特征图通常存在时序信息缺失的问题.针对上述问题,本文提出了一种新的深度图序列表示方式,即深度时空图(Depth space time maps,DSTM).DSTM降低了特征图的冗余度,弥补了时序信息缺失的问题.本文通过融合空间信息占优的深度运动图(Depth motion maps,DMM)与时序信息占优的DSTM,进行高精度的人体行为研究,并提出了多聚点子空间学习(Multi-center subspace learning,MCSL)的多模态数据融合算法.该算法为各类数据构建多个投影聚点,以此增大样本的类间距离,降低了投影目标区域维度.本文在MSR-Action3D数据集和UTD-MHAD数据集上进行人体行为识别.最后实验结果表明,本文方法相较于现有人体行为识别方法有着较高的识别率.Human action recognitions from depth map sequences improve the recognition accuracy by extracting feature maps.A new representation of depth map sequences called depth space time map(DSTM)is proposed in this paper for overcoming the lack of temporal information in the feature maps.DSTM reduces the redundancy of action features.We conduct high-precision human action recognitions by fusing depth motion maps(DMM)and DSTM based on a new multi-modal data fusion algorithm called multi-center subspace learning(MCSL).The algorithm constructs multiple projection centers for each class data to expand the samples inter-class distance and reduce the projection target area dimension.Experiments conducted on MSR-Action3D and UTD-MHAD depth database show the effectiveness of the proposed method.

关 键 词：行为识别 信息融合 深度时空图 多聚点子空间学习 

分 类 号：TP311.13[自动化与计算机技术—计算机软件与理论]