Research On Re-Identification Algorithm In Non-Overlapping Domains

In recent years,person re-identification has attracted more and more research interest.It is a task of recognizing a given individual when he is viewed across a non-overlapping views(through different images)by measuring the similarity between two person-centered bounding boxes.It is difficult as each view may be taken from a different angle and distance,featuring different backgrounds with different lighting conditions,or low resolution.In order to tackle this problem,current research efforts primarily focus on two aspects: developing feature representations and developing machine learning methods.This paper focuses on the former by proposing two improved image representations algorithm.One is a method to extract multi-features with the framework of PS model(Pictorial Structures model)for person re-identification.A re-identification mechanism takes in bounding boxes which containing segmented images of individuals as input.Prior to feature extraction,re-identification requires first detecting a person with attempting to segment the pixels of a person in the bounding box from background which is also included in the bounding box.In order to perform feature extraction more pure or more reliable,PS model has been exploited in this paper before feature extraction,as it has outlined person more clearly and has eliminated contamination by background.Color feature is more robust to viewpoint changes than other features,but single color feature do not perform best to some specific objects,for example,person with apparent textural characteristic.So,multi-features have been extracted and obtain competitive performance.Specially,several texture features have been compared and the most efficient one has been combined with color features to get a richer descriptor.The other method is based on multiply color space and covariancematrix.The feature is extracted firstly through the use of Gabor filters from different color space.Then they are encoded by the covariance descriptor which have extracted statistical information about its strength,texture and shape.The Gabor filters and the covariance descriptor improve the robustness to the illumination variation,measuring the similarity of neighboring scales limits the influence of the background.At the same time,by superimposing a plurality of color space characteristics,and specially designed covariance matrix,the algorithm performance greatly improved.All the schemes in this paper are realized with MATLAB.Our experiments show that the proposed scheme has a improved performance.