Research On Multi-view Gait Recognition Method Based On Deep Learning

Gait recognition is an important branch of biometric recognition,which is to identify individuals according to the subtle differences in walking posture,it has wide application prospects in the fields of security and access control.Compared with other biometric technologies,gait recognition has the advantages of long distance,uncontrolled and not easy to camouflage.However,in the real recognition scene,dressing,shielding and carrying objects will change people's walking posture and physical characteristics,and the change of view angle will also influence the overall contour of the human body,which will have a certain impact on the recognition results.In this thesis,while improving the influence of covariates,aiming at the problem of the change of the gait view and combined with the method of deep learning,the main work is as follows:1.A network loss function model of joint learning is designed.Recognition tasks are mostly based on the data of distance measurement,the traditional triplet loss function only requires that the intra-class distance between sample pairs is smaller than the inter-class distance,without considering similarities or differences between intra-class and inter-class samples,which may lead to misjudgment in the process of recognition and affect the final experimental results.At the same time,this thesis considering the distance between classes within the class,using enhanced constraints Triplet loss joint Center loss for learning while adding Softmax loss to accelerate network convergence.Experimental results show that this method can get better recognition results.2.Based on convolution neural network,this thesis makes a deep research on gait recognition.Firstly,the CASIA-B dataset was preprocessed such as period detection,normalization and morphology processing,and then the silhouettes processed in the period were divided into multiple channels by the size of the stride,the silhouettes in each channel are processed according to the average gait,and the multi-channel gait template data set containing spatiotemporal information are obtained.Finally,the multi-channel gait templates are regarded as the image set as the input of convolution neural network,and the network itself is used to extract the temporal and spatial relationship between the gait templates.At the same time,STN network is added to the feature extraction and aggregation part to maintain the feature invariance of the gait template.The experimental results show that the algorithm in this thesis has achieved good experimental results in various gait view.