Research Of Appearance Based Person Re-Identification

With the development of video capturing technology,and the growing concerns of people about public safety,video surveillance system has been widely used in communications,security,transportation and other industries.Admittedly,the mass surveillance videos can bring convenience to people's lives,however,it also poses challenges to the traditional manual processing method concurrently.To obtain better service,people have put forward higher requirement for the video surveillance system,one specific requirement is that it should be intelligent.It can automatically identify objects,and alerts when abnormal condition occurs so as to assist the security personnel to deal with the crisis,as a powerful auxiliary tool.Compared with the traditional manual processing methods,computer processing methods have the advantages of economic,efficiency and high accuracy.Therefore,utilizing computer technology and multicamera networks to implement the intelligent video surveillance system is one direction for the future monitoring system.The mass surveillance videos together with the rapid development of image processing,pattern recognition and machine learning have made the intelligence of video surveillance become a possibility.Up till the present moment,intelligent video surveillance system is at the stage of preliminary study.Therefore,how to use computer vision techniques to realize intelligent video surveillance system becomes a urgent problem.It is under this background that the person re-identification technology emerges,which has become a hot topic in the field of computer vision,and plays an important role in maintaining the public safety.Person Re-identification aims to re-identify the same interest targets across different cameras.Due to the uncertainty of time information across different cameras,current methods of person re-identification are appearance based models.The key problem of person reidentification is the giant appearance change of the same targets due to illumination,view angle,gesture and background change.Meanwhile,person re-identification also suffers from limited samples and low resolution problems,so the accuracy of person re-identification is comparatively low.In a consequence,person re-identification is still an open problem and more research work should be done.Although many methods have been proposed in the past 5 years,current methods still suffer from two problems: 1)Methods tend to be effected by the large intra-class difference and iter-class similarity.There are large intra-class and inter-class similarity in the data.In certain circumstance,different targets may appear more similar than the same targets.2)Methods are dependant on the visual features.The appearance-based person re-identification methods extract low-level handcraft features,which effect the final results.However,these features are usually less representative and discriminative,containing large amount of redundency.To address the problems mentioned above,methods based on elastic projections and latent feature learning are proposed from the view of metric learning and feature learning,respectively.In the elastic projections based method,raw data are projected to different subspaces according to their different attributes.And then the discriminative subspaces are learnt by driving the same targets to be more similar while different targets to be more distinct.The latent feature based method assumes that there exists a latent space of the raw data,which can capture targets' inner attributes and correlations.In this paper,a matrix factorization method is proposed to obtain the latent feature representations of targets in different camera views.Due to the giant appearance change of the same targets,latent feature representations of the same targets across different cameras may have semantic representation gap.To cope with this problem,a projection matrix is utilized to bridge the gap of their corresponding latent features.To verify the performance of the proposed method in this paper,we test the proposed methods on three benchmark datasets and compare them with the state-of-the-art methods.Experiments results prove the robustness and effectiveness of the proposed methods.