Research On Gait Recognition Method Based On Convolutional Neural Network

With the development of human society,the requirements for public safety have become higher and higher.And many fields have adopted biometric identification technology for authentication to ensure personal and public security.Gait recognition,as a relatively new biometric technology,aims to identify people by their walking posture.Compared with other biometric technologies,gait recognition has the characteristics of long-distance,non-contact and difficult to camouflage.For this reason,gait recognition has attracted wide attention of many researchers,and various gait recognition methods have been proposed.Because gait recognition is a relatively young research direction,some of the proposed methods still have some drawbacks.For example,they are susceptible to changes in perspective,clothing,and carrying bags.In order to better solve the problem of gait recognition,this paper studies the gait recognition based on the convolutional neural network method.The main work is as follows:1.Studying on gait recognition based on convolutional restricted Boltzmann machine.In this paper,a feature learning network based on convolutional restricted Boltzmann machine is proposed to train gait features from the GEI.For the obtained gait features,gait recognition using support vector machines,twin support vector machines,artificial neural networks,and K-nearest neighbor methods.Experimental comparisons were made with restricted Boltzmann machines,PCA and LDA feature extraction methods.The experimental results show that the convolutional restricted Boltzmann machine can obtain a higher recognition rate when using a KNN classifier,and the classification accuracy reaches 98.35%.2.Studying on gait recognition based on convolutional neural network with improved loss function.By analyzing the convolutional neural network,a convolutional neural network model for gait recognition is constructed;meanwhile,an improved loss function is introduced by combining the recognition loss and the center loss function.The CASIA-B dataset was used in the experiment to study the influence of the network parameter initializer,activation function and convolution kernel size on the recognition performance,and the performance under the improved loss function.The experimental results show that the model not only converges faster,but also obtain higher accuracy when the model is initialized by Xavier parameter initializer,using ELU activation function and with 5×5 convolution kernel.In the same walking state experiment,the improved loss function model improved the recognition accuracy of normal,backpack and coat walking by 0.46%,0.71%,and 0.96% respectively.Therefore,the improved loss function model makes the network converge faster and the recognition rate is improved to some extent.3.Studying on gait recognition based on 3D convolutional neural network.In order to take advantage of temporal information in gait silhouette sequence,a 3-Dimensional convolution gait recognition network model with an improved triplet loss function is proposed.In this way,spatiotemporal deep features are extracted from sequential gait images by 3-Dimensional convolutional neural network without any preprocessing,which is able to take advantages of periodical dynamic and motion pattern of human gait.Experiments on CASIA-B dataset and OU-MVLP dataset validate the proposed method,which better solve the effects of changes in perspective,carrying object and wearing on gait recognition performance.We compared with state-of-the-arts method and our method has obtain a higher accuracy on the side view.