Research On Cross-view Gait Recognition Based On Deep Learning

Gait is a kind of behavioral biometric feature,which is defined as the way a person walks.Compared with other biometric features such as face,fingerprint,iris and so on,the advantage of gait is that it can be recognized at a distance without the cooperation of subjects,which provides important guarantees for citizens’ information and social public security.However,gait recognition is easily affected by the camera’s viewpoint,dressing and carrying condition,leading to strong intra-class variation in the extracted gait features.To address the cross-view problem in gait recognition,this paper relies on deep learning to research gait recognition.The main work is as follows:The existing generative gait recognition methods focus on transforming gait templates to a specific view angle,which may decline the recognition rate in a large variation of view angles.To solve this problem,a gait recognition method based on generative adversarial networks with view transformation is proposed.A gait view transformation model consisting of generator,view discriminator,identity preserver and self-attention module is established with the gait energy images as input.In specific,the encoder-decoder structure of the generator is used to connect the input gait features and the target view indicator to achieve different views transformation,in the discriminant network,the view discriminator constrains the generated view to be consistent with the target view,the identity preserver with Tri-Hard loss maximizes the retention of the identity information of the input template.At the same time,adding self-attention module into the generative network and discriminant network can improve the quality of the generated images.Comparison,ablation and data enhancement experiments are conducted on the CASIA-B dataset.And comparison experiments are conducted on the OU-MVLP large-scale cross-view gait database.In order to extract temporal and spatial features of gait from gait sequences more efficiently and improve the flexibility of gait recognition,a gait recognition method based on multi-scale network of gait set is proposed to further improve the cross-view gait recognition rate.Firstly,a feature sub-network is used to extract low-level features of gait contours,then a multi-branch learning strategy is used to extract discriminative features at different scales.Furthermore,attentive temporal pooling is used to aggregate the frame-level features into set-level features,and a multi-layer global pipeline is added to better gather the shallow and deep gait set information.At the same time,a loss function scheme of classification-beforemetric is proposed,softmax loss is added to constrain the set-level features to learn the coarse information of classification,and the triple loss is applied to the feature vector after horizontal pooling to learn the fine information of metric.By applying the proposed methods to case studies on the CASIA-B and OU-MVLP datasets,experimental results show that the proposed gait view transformation model based on GEI not only can extract gait features that are robust to view variation,but also can improve the quality of the generated gait templates during gait view transformation.At the same time,the proposed multi-scale network of gait set can effectively extract the spatial and temporal information of gait features and improve recognition rates over existing methods in all three states as well as cross-view,which can be better applied to practical application scenarios with multiple covariates such as viewing angles,clothing and carrying conditions.