Research On Cross-view Gait Recognition Based On Feature Subspace

With the development of modern society and the raising safety consciousness of people,human identification is indispensable to the safety of more and more important public places such as stations,banks,government departments,military bases and so on.The traditional biometrics such as face,fingerprint and iris usually require proximal sensing or physical contact.Gait,which is recognized by one's walking pattern,is considered to be the most potential biometric in the application of intelligent visual surveillance at a distance.However,in the actual scenarios,there are many factors affecting correct match of gait features.In this paper,based on feature subspace learning,we project gait features of different views onto one public subspace to recognize cross-view gait.The main contributions of this paper are as follows:1.A new method of cross-view gait recognition based on coupled locality preserving projections is proposed.Gait features of two different views are projected onto a common subspace by Coupled Locality Preserving Projections(CLPP).In the projected subspace,gait can be measured directly.Based on the objective function of coupled metric learning,we add a constraint term of preserving local structure of samples to keep neighboring relationship in the learnt subspace.Experimental results show the effect of preserving neighboring relationship.The greater viewing angle changes,the more improvement of gait recognition rate.2.A new method of cross-view gait recognition based on Multiview Max-Margin Subspace Learning(MMMSL)is proposed.Multiview Max-Margin Subspace Learning seeks for a group of transform matrices for multiple views.With the transform matrices,gait features under different viewing angles are projected onto a common subspace so that they can be measured directly.In the common subspace,same-class samples from all views cluster together,and each different-class cluster is kept away from its nearest neighbors as far as possible.Experimental results demonstrate that the proposed method outperforms other classical multiview subspace learning methods over all views.3.A new method of cross-view gait recognition based on Large Margin Deep Distance Metric Learning(LMDDML)is proposed.A short and simple CNN structure is designed firstly for gait feature extraction.Experiments of gait classification show that the welldesigned CNN has excellent ability of feature expression.Based on our CNN,a Large Margin Deep Distance Metric Learning method is proposed to seek a more discriminant common feature subspace.With the constraint of large margin in the feature subspace,distance of same-class samples is reduced and the gap between different-class samples is kept in a large margin.Cross-view gait experimental results show that the proposed method perform best in all subspace learning methods.