Study On Gait Recognition Based On Convolutional Neural Network

Biometric-based recognition technology is one of the methods to achieve identity authentication by using the unique physiological or behavioral characteristics of human.Compared with the traditional identity authentication technologies,the biometric-based identification has higher recognition efficiency and security level.Gait recognition,as an emerging biometric identification method,is used for identification by using the way of human walking.It has the advantages of long-distance recognition and difficulty in camouflage.With the development of convolutional neural network and its successful application in the field of computer vision,more and more people have introduced it into the research on gait recognition.Gait is affected by many variables easily,such as view,clothing and carrying.Gait recognition in complex scenes still faces great difficulties.In this paper,we study on the method of gait recognition based on deep learning,and use convolutional neural network to learn distinguishing deep features in gait images,which can make classification more effective.Specifically,a method of gait recognition based on feedback weight convolutional neural network is proposed in this paper.The feedback mechanism is introduced to combine deep features and handcrafted features,and a deep model is trained to complete classification for gait recognition.In addition,in order to address the problem of the inconspicuous inter-class variations and the large intra-class variations in gait features because of the change of view,a method of gait recognition based on multi-loss convolutional neural network is proposed to increase the difference between intra-class variations and inter-class variations while achieving classification.More precisely,the scheme designs a module to combine the classification loss function and the triplet loss function to optimize the mapping relationship between gait images and deep features in the classification model.At the same time,the batch normalization is used in the proposed method to avoid over-fitting.Therefore,the deep features learned in the classification model are not only class-dividable,but also compact within the class.The accuracy of gait recognition is further improved in this model.Finally,the proposed methods are evaluated on two gait data sets(i.e.,CASIA-B,OU-ISIR),and the experimental results show the effectiveness of our methods.