一种视角无关的时空关联深度视频行为识别方法  

A Perspective-independent Method for Behavior Recognition in Depth Video via Temporal-spatial Correlating

作　　者：吴培良[1,2,3] 杨霄 毛秉毅 孔令富[1,3] 侯增广 WU Peiliang;YANG Xiao;MAO Bingyi;KONG Lingfu;HOU Zengguang(School of Information Science and Technology,Yanshan University,Qinhuangdao 066004,China;State Key Laboratory of Management and Control for Complex Systems,Institute of Automation,Chinese Academy of Sciences,Beijing 100190,China;The Key Laboratory for Computer Virtual Technology and System Integration of Hebei Province,Qinhuangdao 066004,China)

机构地区：[1]燕山大学信息科学与工程学院,秦皇岛066004 [2]中国科学院自动化研究所复杂系统管理与控制国家重点实验室,北京100190 [3]河北省计算机虚拟技术与系统集成重点实验室,秦皇岛066004

出　　处：《电子与信息学报》2019年第4期904-910,共7页Journal of Electronics & Information Technology

基　　金：国家自然科学基金(61305113);河北省自然科学基金(F2016203358);中国博士后基金(2018M631620);燕山大学博士基金(BL18007)~~

摘　　要：当前行为识别方法在不同视角下的识别准确率较低,该文提出一种视角无关的时空关联深度视频行为识别方法。首先,运用深度卷积神经网络的全连接层将不同视角下的人体姿态映射到与视角无关的高维空间,以构建空间域下深度行为视频的人体姿态模型(HPM);其次,考虑视频序列帧之间的时空相关性,在每个神经元激活的时间序列中分段应用时间等级池化(RP)函数,实现对视频时间子序列的编码;然后,将傅里叶时间金字塔(FTP)算法作用于每一个池化后的时间序列,并加以连接产生最终的时空特征表示;最后,在不同数据集上,基于不同方法进行了行为识别分类测试。实验结果表明,该文方法(HPM+RP+FTP)提高了不同视角下深度视频识别准确率,在UWA3DII数据集中,比现有最好方法高出18%。此外,该文方法具有较好的泛化性能,在MSR Daily Activity3D数据集上得到82.5%的准确率。Considering the low recognition accuracy of behavior recognition from different perspectives at present,this paper presents a perspective-independent method for depth videos.Firstly,the fully connected layer of depth Convolution Neural Network(CNN)is creatively used to map human posture in different perspectives to high-dimensional space that is independent with perspective to achieve the Human Posture Modeling(HPM)of deep-performance video in spatial domain.Secondly,considering temporal-spatial correlation between video sequence frames,the Rank Pooling(RP)function is applied to the series of each neuron activated time to encode the video time sub-sequence,and then the Fourier Time Pyramid(FTP)is used to each pooled time series to produce the final spatio-temporal feature representation.Finally,different methods of behavior recognition classification are tested on several datasets.Experimental results show that the proposed method improves the accuracy of depth video recognition in different perspectives.In the UWA3DII datasets,the proposed method is 18%higher than the most recent method.The proposed method(HPM+RP+FTP)has a good generalization performance,achieving a 82.5%accuracy on dataset of MSR Daily Activity3D.

关 键 词：视频行为识别 深度视频 视角无关 卷积神经网络 时空关联 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]