Research On Gait Recognition Algorithm Based On Convolutional Neural Network

Today,biometrics plays an important role in personal identification systems,but there are also many problems.For example,face recognition is susceptible to illumination,makeup,age distance,etc.Fingerprints are easy to forge.Gait recognition is identification by human walking posture.The gait features are characterized by long-distance,non-contact and non-disguise,which can overcome the shortcomings of face and fingerprint features.However,gait recognition is susceptible to covariates such as clothing,backpacks,and perspectives.In order to reduce the influence of covariates on gait recognition,this paper conducts an in-depth study on gait recognition algorithms.The main work is as follows:1.Gait image preprocessing.The advantages and disadvantages of algorithms such as inter-frame difference method,background subtraction method and optical flow method are analyzed.The background subtraction method was used to extract the gait contour map from the gait video image sequence and perform image preprocessing.The gait energy map is calculated by detecting the gait cycle using a change in the ratio of the width to the height of the gait profile.2.Select the optimal network model.The LeNet network model is selected for the gait data set CASIA-B to analyze the influence of parameter initialization mode,activation function and convolution kernel size on the performance of the network model.It is proved by experiments that when the standard normal distribution is selected as the parameter initialization mode,ReLU is the activation function,and the convolution kernel size is 5><5,which made the performance of the network model better.Under the same initialization and activation functions,LeNet converges faster and has a higher recognition rate than AlexNet.For the residual network model,not only can the network convergence speed be accelerated,but also the recognition rate can be improved.Therefore,the idea of the residual network is combined into the LeNet network model to improve the LeNet network model.3.Verify the performance of the improved LeNet network model.In the same walking state experiment,the improved LeNet network model improved the recognition rate of normal,backpack and coat walking by 0.3%,0.5%,and 0.8%respectively.In the experiment under cross-view,the improved LeNet network model improved the recognition rate of normal,backpack and coat walking by 4%,2.415%,and 6.245%respectively.In the experiment from the same perspective,the recognition rate of the improved LeNet network model increased by 0.27%and 2.95%for normal and wearing coats respectively.Therefore,the improved LeNet network model makes the network converge faster and the recognition rate is improved to some extent.