基于视角转换的多视角步态识别方法  

Multi-view Gait Recognition Method Based on View Transformation

作　　者：瞿斌杰 孙韶媛[1,2] Samah A.F.Manssor 赵国顺 QU Binjie;SUN Shaoyuan;Samah A.F.Manssor;ZHAO Guoshun(College of Information Science and Technology,Donghua University,Shanghai 201620,China;Engineering Research Center of Digitized Textile and Fashion Technology of Ministry of Education,Donghua University,Shanghai 201620,China)

机构地区：[1]东华大学信息科学与技术学院,上海201620 [2]东华大学数字化纺织服装技术教育部工程研究中心,上海201620

出　　处：《计算机工程》2021年第6期210-216,共7页Computer Engineering

基　　金：上海市科委基础研究基金(15JC1400600)。

摘　　要：针对步态识别中步态视角变化、步态数据样本量少及较少利用步态时间信息等问题,提出一种基于视角转换的步态识别方法。通过VTM-GAN网络,将不同视角下的步态能量图及含有步态时间信息的彩色步态能量图,统一映射到保留步态信息最丰富的侧视图视角,以此突破步态识别中多视角的限制,在视角转换的基础上,通过构建侧视图下的步态正负样本对来扩充用于网络训练的数据,并采用基于距离度量的时空双流卷积神经网络作为步态识别网络。在CASIA-B数据集上的实验结果表明,该方法在各状态、各角度下的平均识别准确率达到92.5%,优于3DCNN、SST-MSCI等步态识别方法。The existing gait recognition methods are limited by multiple factors,including the changes of view,small gait sample size,and under-utilization of the temporal information of gaits.To address the problems and improve the gait recognition performance,this paper proposes a gait recognition method based on view transformation.Through VTM-GAN,Gait Energy Images(GEI)and Chrono Gait Images(CGI)with temporal gait information of different views are mapped to the side view that contains the most abundant gait information in order to break the limitation of views in gait recognition.On the basis of view transformation,the positive and negative sample pairs of gait data in the side view are constructed to extend the volume of network training data.The Spatial-temporal double flow convolutional neural network based on distance measurement is taken as the gait recognition network.Experimental results on the CASIA-B dataset show that the average recognition accuracy of this method in all states and views reaches 92.5%,higher than that of 3DCNN,SST-MSCI and other gait recognition methods.

关 键 词：步态识别 视角转换 VTM-GAN网络 时空双流卷积神经网络 CASIA-B数据集 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]