Gait Recognition Based On Joint Spatiotemporal Features

Based on the fact that biometrics is not easy to forge,it has the advantages of being unique and unchanging for life.Biometric identification technology has a natural advantage over traditional identification methods.Among them,gait recognition can even extract the gait biometrics of travelers in the environment of long distance,no contact with individuals and poor light,which has become one of the hotspots of biometrics research.However,the current mainstream gait recognition is based on gait contour map,which is easy to be interfered by covariates such as perspective,clothing,and carrying objects.Therefore,aiming at the problem of missing spatiotemporal information of contour map template and the influence of covariates,a joint spatiotemporal feature extraction model is established based on deep learning method.This project mainly carried out the following work:(1)In order to solve the problem that gait recognition is affected by clothing,carrying objects and other covariates,a gait recognition algorithm based on skeleton map and cyclic neural network is proposed.In order to verify the effectiveness of the algorithm,casia-b is selected as gait data set,which has large amount of data,including perspective,clothing,and changes in carrying objects.Using image algorithm to process the contour map,we can get the skeleton map with precise topological relationship,and transform the gait data from the contour map to the skeleton map,which solves the problem that the contour map is easy to be disturbed by covariates and distorted to some extent.Then,the feature vector of high-dimensional skeleton is reduced by PCA,which not only retains most of the original information,but also reduces the complexity of data space.In the aspect of feature learning,two kinds of models based on the cyclic neural network variant(LSTM,attention bilstm)are adopted.One cycle of reduced dimension static feature group is input into the cyclic neural network to extract the dynamic features in gait,and attention mechanism increases the ability of focusing attention in the time dimension.By comparing the experimental results of the two models with the existing gait recognition methods based on contour map,the results show that skeleton map is more robust than contour map in covariate interference,and the feature learning ability of gait model based on cyclic neural network combined with attention mechanism is stronger.(2)Aiming at the problem that the skeleton images are easy to overlap in the angle of 0 ° and 180 °,in order to select a gait template which is not affected by the angle of view and has less redundant features,a gait recognition algorithm based on human joint points and fusion of spatiotemporal feature network is proposed.Using attitude estimation algorithm to estimate the joint coordinates of RGB original image in casia-b dataset,28 low-dimensional gait static feature descriptors are obtained.In the first part of the network,attention bilstm extracts the gait dynamic features,in the second part,convolution neural network extracts the gait static features,and finally the two parts get the gait spatiotemporal features.In the same experiment on casia-b gait data set,compared with the experimental results of skeleton based gait method,the experimental results show that the gait recognition algorithm based on joint points and fusion of spatiotemporal features is more efficient.