Human Gait Recognition Via Subspace Ensemble Learning

Gait recognition has been a hot research topic in the field of computer vision and biometrics.It aims at endowing computers with the ability to identify different human beings according to their walking patterns,termed gait.Such a research has both significant theoretic values and wide potential applications,closely related to many disciplines such as Computer Vision,Pattern Recognition,Video/Image Processing,Computer Graphics and Intelligent Human-Computer Interaction etc,and has great application prospect in human identification at a distance.For example,there is a need for an accurate and robust human identification technology at a distance for video monitoring systems in some security sensitive fields such as banks,subways,airports and railway stations.After more than 20 year's development,many gait recognition algorithms have been proposed.The current state-of-the-art can achieve the satisfying recognition rate under situations where the training and test data are captured under similar conditions,while under non-ideal conditions the recognition performance with the effect of a variety of covariate factors usually decline dramatically,which is difficult to meet practical application requirements.Therefore,gait recognition technologies are currently far from mature,and developing an accurate and robust gait recognition algorithm is still the subject of much of the current study.In this thesis,we mainly concentrate on designing efficient gait recognition methods based on subspace ensemble learning theory.The main work and contribution of this thesis includes:1.As the effect of a variety of covariate factors on gait,the performance of recognition algorithms declines.Against this problem,this thesis proposed a subspace ensemble learning via totally-corrective boosting based gait recognition approach,which innovatively applies totally-corrective boosting technology to build a general subspace ensemble learning framework to extract and optimally combine gait features in the multiple discriminant subspaces,in order to improve its robustness to gait covariate factors.2.To preserve temporal information which is crucial for gait recognition of gait image sequence,a multi-channel feature extraction technology is utilized to map temporal information of gait sequence to the channel weights.On the multiple channels gait contuer images located at different time regions are coded to become a preserving temporal information gait feature image,named gait gaussian energy image.To further extractnonlinear features of gait gaussian energy image,kernel subspace ensemble learning based gait recognition is proposed,which utilizes kernel method under the subspace ensemble learning framework via totally-corrective boosting to extract features of gait feature image in the nonlinear subspaces.Compared with features in the linear subspace,those nonlinear features are more distinguishable,which are good for gait recognition.3.In order to improve the robustness to gait covariate factors and classification ability of gait feature image,gabor wavelets with eminent characteristics in spatial local feature extraction and orientation selection are utilized to extract features of gait energy image with different scales and different directions,which is an efficient gait representation.Taking inherent spatial structural information and relevant information of gabor features with the positive effect on recognition into consideration,tensor subspace ensemble learning based gait recognition method and subspace ensemble learning using elementary multilinear projection based gait recognition method are proposed which utilize tensor to represent gait gabor features and respectively adopt tensor to tensor projection and tensor to vector projection of two different multilinear projection ways.According to tensor theory,these methods are extended from totally-correctively boosting based subspace ensemble learning framework,which have higher accurate rates of gait recognition compared with the tensor discriminant analysis method.4.Against the problem that the change of gait view causes significant reduction of recognition performance of algorithms,this thesis proposed a cross-view gait recognition method.Firstly,a sparse representation based gait view classification method is utilized to perform gait view matching,in order to determine the view angle.Then,CCA(Canonical Correlation Analysis)method is adopted to extract the most relevant features of gait with different views,in order to reduce the differences between them.Considering different impact of the view change on the different local regions of gait,a local patch based subspace ensemble learning method is extended from totally-corrective boosting based subspace ensemble learning framework by bringing the local patch selection of gait image.Finally,by building a reasonable triplet set of training sample set,CCA method and local patch based subspace ensemble learning method are combined in order to perform cross-view gait recognition.For verifying the effectiveness of gait recognition algorithms,a lot of comparative experiments have been performed in the USF Human ID gait database and CASIA-B gait database.The experimental results revealed that our algorithms have the higher recogni-tion rates and stronger robustness.