Gait Recognition Research Based On Deep Transfer Learning

Gait recognition refers to the identification of a person's identity or identity attributes(such as gender)through the manner of person walking style.It is a new biometric recognition technology which attracting more and more researchers in recent years.Compared with other biometric technologies,gait recognition has the advantages of long-distance recognition,difficult camouflage,strong security and non-cooperate recognition,etc.These advantages make the gait recognition technology has great potential in security,intelligent video surveillance and other fields.However,in real life,human gait is easily affected by covariates such as view,carry-on,and clothing,resulting in low recognition rate.If the performance of the gait recognition method can be further improved,it will help the gait recognition to be put into commercial applications and bring more convenience to users.On the basis of summarizing and analyzing previous studies,this paper proposes a gait recognition method based on Dense Net deep transfer learning,named as DNDTL.The method first extracts the binary contour map of the human body from the original gait video sequence frame,calculates the gait energy image.Then use the Dense Net as a pre-training model for feature extraction,and input the gait energy image into the model for fine-tuning.Next,use the fine-tuned trained model as a feature extractor,extract the output of the last layer as a deep gait feature,and recognize the identity of the person through the nearest neighbor classifier.Or directly use DNDTL as a binary classifier to the gender of the person Classification and evaluation using K-fold crossvalidation.At the same time,in view of the problem that gait is greatly affected by view,and the recognition rate is low in cross-view gait recognition.Therefore,this paper proposes a cross-view method of DNANL deep transfer learning.This method combines variance analysis and linear discriminant analysis to optimize the intra-class and inter-class variance of gait recognition,and extract the deep gait features more suitable for cross-view gait recognition.This method is evaluated on multiple tasks including gait identity recognition based on the same-view,cross-view,and multi-state,as well as gender classification under various views on the data set CASIA-B published by the institute of automation of the Chinese academy of sciences,and compared with other mainstream gait recognition methods.The results show that the proposed method has high recognition rate and good robustness to the three covariates of view,carry-on,and clothing.