Research Of Gait Recognition Based On Class Energy Image And Coupled Metric Learning

Gait recognition,as a unique perceptible biometric identification method at larger distances,has been a hot issue for its advantages of contactless,non-invasiveness,hide and fake with difficulties in the field of pattern recognition and machine vision in recent years.However,due to the non-rigid nature of body,the stability of the gait which is generated from the body could be influenced by many factors,such as the variations of clothing(including clothes,pants,shoes and hats),ages,walking speed,etc.Thus,the automatic gait recognition becomes a challenging task.Especially,the problems such as the changes of walking status,the characteristics of multiple view angles,which is brought by the changes of camera angle and human walking directions have become the bottlenecks of gait recognition in the video surveillance field.Accurate motion detection is a very important preliminary work in gait recognition.In this paper,the moving body detection method based on track area has been firstly carried out.Then,the research mainly focuses on the two key issues:the changes of walking states and multi-view problem in gait recognition.The recognition rates are decreased when walking states change.In order to solve this problem,two solutions are proposed from different perspectives:(1)We detailedly reviewe the current hot researches on the "Class Energy Image".The merits and demerits of each Class Energy Image method are compared through theoretical and experimental analysis.It can provide a useful reference for the selection of Class Energy Images under different walking status.(2)On the basis of the present Class Energy Image approaches,new and effective gait representation methods have been proposed,which have realized the expression of the gait information more complete to satisfy the gait feature representation of the variation of walking status;(3)For the problem of the multi-view gait recognition,we present a supervised coupled metric strategy mainly from the perspective of coupled distance metric.The supervised coupled metrics could be directly employed for feature extraction and classification of gait sequences from different views.The main contents of the paper are summarized as follows:1.Aiming at the motion detection problems of the complex scene,such as background disturbance,light changes,noise interference,etc.,specific scheme has been proposed.A codebook detection method based on track area was proposed,which can achieve fast and accurate detection of the moving body.The first frame of the track area is the complete moving body which is obtained by the codebook detection algorithm.The track area could be positioned accurately by using the online Adaboost fused with the Camshift target tracking algorithm.Furthermore,the codebook detection algorithm could be carried out in the track area.The proposed method only detected the track area,the other areas in monitoring window were all regarded as background.Thus,the redundancy detection which is unrelated to the target to be identified is avoided.It has certain significance for perfecting the real-time gait recognition system.Finally,aiming for the Class Energy Image approach based on the periodic gait images,a multi-view gait period detection method was proposed,and it could effectively extract the periodic gait sequences after moving body detection.2.Aiming at the problem of the common Class Energy Image approaches,which lack comprehensive and detailed commentaries,this paper provides the comparative analysis and research for the existing Class Energy Image methods.Based on the different ways of feature extraction and Class Energy Image generation,the present Class Energy Image methods could be divided into three categories:the gait information accumulation approach,the gait information introduction approach and the gait information fusion approach.The merits and demerits of the present Class Energy Image methods have been compared and analyzed through theoretical and experiments.It provides a useful reference for the selection of the Class Energy Image under different states.Moreover,the paper also pointed out the direction for the development of the gait feature representation method based on Class Energy Image:it is equally important for the recognition performance improvement to choose good fusion methods and fusion information.The fusion information could be the dynamic,static and temporal information from the gait information accumulation approach and the gait information introduction approach.3.Aiming at the problem that the gait information represents are incomplete for the existing Class Energy Image methods,three novel multi-channel colored Class Energy Image were further proposed,which were the Temporal Flow Energy Image(TFEI),the Active History Energy Image(AHEI)and the Enhanced Move Energy Image(EMEI).The proposed methods reflected the dynamic,static and temporal information from different aspects.Moreover,every channel of each colored Class Energy Image has different effects on the classification.Therefore,each channel has been regarded as a feature.Combined with the feature level fusion strategies,two effective multi-channel feature-level fusion methods were proposed.The proposed methods can better solve the difficulties of the gait recognition caused by walking status'changes.4.Aiming at the different viewing angle problem in gait recognition,this paper regarded it as a measurement problem between different data collections.The linear coupled metric learning method based on enhanced supervised sparsity preserving projection was proposed.Furthermore,the proposed method was extended to nonlinear situation,the kernel coupled metric learning method based on enhanced supervised sparsity preserving projection was proposed.Finally,the two proposed coupled metric learning methods were applied to multi-view gait recognition,respectively.The better recognition effect was obtained in the experiments.