Multi-view Gait Recognition Based On Deep Learning

Nowadays,fingerprint,face,iris and other biometrics are widely used in the field of identification.The development of biometrics technology provides an important guarantee for the stable and safe operation of society.Gait features,as one of the biological features,overcomes the shortcomings of traditional biological features by virtue of its characteristics of large differences between individuals,difficult to disguise and can be captured in uncontrolled state remotely,therefore,it gradually plays an important role in various places and fields.However,gait recognition still faces many problems in practical applications,such as the change of camera angle,target clothing,physical and psychological status and carrying objects,etc.In this paper,the deep learning technology is used to study the multi-view problem in gait recognition.The main work of this paper is as follows:1)Preprocessing of gait sequence.CASIA-B data set and DHU-Gait infrared gait data set collected in the laboratory are preprocessed respectively.Firstly,this paper chooses background subtraction based on the Gaussian Mixture Model by comparing some current mainstream background subtraction.After getting a better background image from the gait sequence,the preliminary gait binary images are obtained through difference operation.Finally,the postprocessing operation is done on the extracted gait image to get the normalized gait binary image,which greatly improves the quality of the follow-up experimental foreground target image.2)Aiming at the multi-view problem in gait recognition,the gait recognition method based on feature fusion network of view transformation is adopted.In multi-view gait recognition,gait features of different views contain different amounts of information.Therefore,VTM-GAN is used to transform gait features of different views into images with most sufficient information.While breaking the limits of multi-view,the gait sample pairs for training can be greatly expanded by the transformed features.Secondly,although most gait features used for gait recognition can retain the spatial information of gait,they often lack the use of temporal information.Therefore,on the premise of fully retaining the spatial information of gait,Chrono-Gait Image(CGI for short)is fused to retain the temporal information of gait sequence,so as to improve the effectiveness of recognition.Meanwhile,comparing with the experiment which only retains the spatial information of gait features,the feature fusion method is verified to be effective.Using Siamese neural network based on sequence information fusion for gait recognition.The difference image between two raw adjacent silhouettes is the really change part of the target motion in gait recognition,which contains a lot of motion and temporal information.Thus after fusing the information of frame difference sequences and the raw silhouettes sequences,this paper uses the Siamese neural network to predict the similarity between two fixed feature maps.Meanwhile,this paper verifies the effectiveness of the fusion of the frame difference sequences through the contrast experiments.