Research On Human Multi-view Gait Recognition Algorithm Based On Deep Learning

With the rapid development of society,biometric-based identity recognition methods have been extensively studied for safeguarding the social security.Compared with other various biometric features,gait can be recognized in a long distance,even without the cooperation of the subject,which ignited the enthusiasm of the researchers.However,in practice,the performance of gait recognition is susceptible to some factors,such as multiview,walking speed,carrying status.Multi-view task is the focus of this thesis,and the main works are as follows:1.Firstly,some survey is conducted,including research status of multi-view gait recognition,and the achievements and challenges of existing methods,which lay a technical foundation for the methods proposed in this thesis.2.Secondly,an approach to multi-view gait recognition via view transformation and similarity learning is studied in this thesis.As we know,How to establish the relationship between different gait view angels is one of the challenges in multi-view gait recognition.In this thesis,the View Transform Model is adopted to establish the mapping relationship among different gait view angels,then the gait from the perspective with low distinctness is converted to the perspective with high distinctness.After that,an optimize Siamese Network is proposed,which is supervised by distance measurement and identity authentication.In the process of testing,a feedback verification algorithm is proposed,and finally the recognition results are obtained.3.Finally,a novel convolutional neural network(CNN)framework with spatial–temporal information is proposed in this thesis.Considering that human walking is a dynamic process,the proposed framework takes Chrono-Gait Image(CGI)as temporal information,and makes it more precise with step width.Then,both the spatial and temporal features are transformed to perspective with high distinctness.After that,a Spatial-Temporal convolutional neural network is trained,which fusion the information in the convolution layer.Then,fisher discriminative criterion is further applied to the network so that the subjects have smaller within-class scatter but larger between-class scatter.Experimental results show that the two algorithms proposed in this thesis can achieve an ideal recognition rate in multi-view gait recognition task.