基于行走特征矢量图的步态识别方法

1)深圳大学机电与控制工程学院,深圳电磁控制重点实验室,广东深圳518060; 2)深圳大学生命科学与海洋学院,深圳市海洋生物资源与生态环境科学重点实验室,广东省海洋藻类开发与应用工程重点实验室,广东深圳518071

生物识别; 深度学习; 步态特征; 步态识别; 行走特征矢量图; 时空网络; 残差学习; 长短期记忆网络; softmax分类器

Gait recognition based on human walking feature vector diagram
PENG Xiaobo1, HUANG Haina1, YANG Huiyue1, LIU Junhong1, and HUANG Ying2

1)College of Mechatronics and Control Engineering, Shenzhen Key Laboratory of Electromagnetic Control, Shenzhen University, Shenzhen 518060, Guangdong Province, P.R.China;2)College of Life Sciences and Oceanography, Shenzhen Key Laboratory of Marine Bioresources and Ecology, Guangdong Technology Research Center for Marine Algal Bioengineering, Shenzhen University, Shenzhen 518071, Guangdong Province, P.R.China

biological recognition; deep learning; gait feature; gait recognition; walking feature vector diagram; spatiotemporal network; residual learning; long short-term memory network; softmax classifier

DOI: 10.3724/SP.J.1249.2021.05528

备注

为解决多视角下存在服饰和携带物等协变量时,基于轮廓或关节模型的步态识别方法识别效果不理想的问题,结合深度学习技术,提出基于人体行走特征矢量图(walking feature vector diagram, WFVD)的步态识别方法.首先,以视频流为输入,基于人体姿态估计模型与人体部位关联场,获取行走特征矢量图; 然后,设计以WFVD为输入的步态时空网络作为特征学习与识别网络,基于残差学习模块学习步态空间特征和长短期记忆(long short-term memory, LSTM)网络学习步态时间特征,在网络末端采用softmax分类器进行分类; 最后,基于中国科学院自动化研究所研发的CASIA-B数据集和自建数据集进行实验.实验结果表明,所提出的方法既充分利用了步态时空信息又避免了冗余信息的干扰,能较好地解决多视角下的服饰和携带物等协变量结合的真实环境中的步态识别难题.
In order to solve the problem that the current gait recognition methods using contour or joint models as gait features do not perform well in the presence of covariates such as clothing and carrying objects under multiple viewing angles, this paper proposes a gait recognition method based on human walking feature vector diagram(WFVD)and deep learning. Firstly, with the video stream as input, the human WFVD is obtained based on the human pose estimation model and the human part affinity fields. Then, a spatiotemporal gait network with the WFVD as the input is designed for gait feature learning and recognition. It learns the spatial and temporal gait features using a residual learning module and a long short-term memory(LSTM)network, respectively. The classification is performed in the softmax layer of the network. Finally, experiments are performed on the dataset CASIA-B developed by the Institution of Automation, Chinese Academy of Sciences(CASIA)and a self-built dataset. The experimental results show that the proposed method makes full use of spatiotemporal gait information while avoids the interference of redundant information, which allows the method to better solve the recognition difficulties due to covariates such as clothing and carrying objects in complex real environments.
·