The Research Of Clothing-invariant Gait Recognition Based On Deep Learning

As a dynamic biometric feature,gait can be identified without interaction of subjects and can be captured from long distance,so gait recognition attracts a lot of researchers' attention.However,there are many challenges for gait recognition such as walking speed,the wearing condition,viewpoint of the cameras and illumination conditions.The variation of wearing condition is one of the biggest issue.The deep learning methods can not only the high level representation automatically,but also simpler and less computation costing.However,the deep learning method usually uses the single identity information of the target identity,so the extracted features are single.At the same time,it has the problem of the network structure itself,the extracted features are not strong enough.Therefore,there is still space to improve.Therefore,this paper propose a deep learning method to deal with the variation of clothing in gait recognition.And making some improvement for the existing problems.(1)To strengthen the robustness and discrimination of CNN features,we put forward attention mechanism to strengthen the robustness of CNN features,and combine the Contrastive Loss with Softmax Loss which constraints the CNN features and minimizes the distance of features from the same subject and maximizes the distance of features from different subjects(namely strengthen the discrimination of CNN features).(2)The trajectory of limbs in gait recognition is an important feature,we take it as prior knowledge and combine with CNN features for better performance.(3)To increase effective features strengthening the discrimination of features,latent semantic analysis is proposed to extracting latent semantic attributes which compensate the shortcomings of one single label.And it enhances the power of CNN features.(4)Based on those approaches aforementioned,an end-to-end method combining with those features and mechanisms is proposed.This method makes good use of the edges,and improve the performance of gait recognition greatly.Compared our method with various advanced methods on the CASIA B and OU-ISIR Treadmill dataset B,the proposed algorithm can eventually get the accuracy of 92% and 90% respectively,and fusion method outperforms most of the state-of-the-art methods.