Research On Cross-view Gait Recognition Algorithms Based On Deep Feature Fusion

With the technological advancements of hardware and software,traditional biometric technologies such as face recognition and fingerprint recognition have been widely used in production and life,which greatly facilitates people’s life and makes intelligent life possible.Gait recognition,as a relatively new biometric technology,is a kind of pedestrian identification technology based on the characteristics of pedestrian’s walking posture.It has the characteristics of long-distance and non-contact,and can obtain the gait information of pedestrians in a non-cooperative manner.Gait features have excellent distinction,and the acquisition of them has low requirements for ambient light,sampling distance and other conditions.Therefore,gait recognition technology has gradually attracted the attention of more and more researchers.Gait recognition technology has broad application prospects in the field of security and danger warning.However,in real life,the installation height,the perspective of cameras and other equipment responsible for collecting gait information are not fixed,which leads to differences in perspective among probe samples and gallery samples.Therefore,improving the accuracy of cross-view gait recognition is the key to promoting the implementation of gait recognition technology in real life.Aiming at cross-view gait recognition,which is the most difficult problem in gait recognition at present stage,we study the cross-view gait recognition technology based on depth feature fusion,and make full use of the information contained in the sequence by fusing spatio-temporal features and block features.Specifically,the main contributions of this paper are as follows:In Chapter 3,a cross view gait recognition algorithm based on spatio-temporal feature fusion is proposed.The algorithm mainly consists of three parts:spatial feature extraction,Staged Horizontal Pyramid(SHP)and temporal feature extraction.Specifically,in order to extract spatial feature from gait sequences,the Backbone network is used to extract both low-level features and high-level features,After that,Staged Horizontal Pyramid mapping is used in multiple stages of Backbone network and the mapped features are spliced together as spatial feature.In order to extract temporal sequence information,the output of Backbone network is further extracted through feature smooth convolution layer,and encoded into the Recurrent Neural Network(RNN).Subsequently,the maximum pooling of the fitted sequence features is carried out to complete the extraction of temporal feature.Finally,the spatial feature extracted from Staged Horizontal Pyramid mapping and temporal feature extracted by Recurrent Neural Network are fused as the final spatio-temporal feature.Extensive experiments on CASIA-B,OU-ISIR and OUMVLP,as well as comparison with other similar methods,fully demonstrate the effectiveness and advancement of the proposed algorithm.In Chapter 4,a cross-view gait recognition algorithm based on layered framework and block feature fusion is proposed,which mainly includes layered gait recognition framework,Partial Feature Blending Mask(PFBM),Improved Staged Horizontal Pyramid(ISHP)and Block Feature Fusion Module(BFFM).Specifically,in order to break the strict order of feature extraction followed by feature mapping in the general gait recognition framework,a layered gait recognition framework is proposed,which logically divides the model into feature extraction layer,feature connection layer and feature mapping layer.In order to premix the information of each block in the images at the input stage,the sequence first passes through a Partial Feature Blending Mask,and the information of each block is mixed with the information of other blocks.In order to fully extract gait features with low computational cost,a simple Backbone network is used to extract sequence features.Then,an Improved Staged Horizontal Pyramid is used to further extract features and smooth the output feature channels.In order to redistribute the extracted gait features according to blocks,a Block Feature Fusion Module is used for feature mapping,and the feature of each block is fused with the features of other adjacent and non-adjacent blocks.Extensive experiments on CASIA-B,OU-ISIR and OUMVLP,as well as comparison with other similar methods,fully demonstrate the effectiveness and advancement of the proposed algorithm.