Research On Distance Metric Learning For Person Re-identification

With the reduction of the cost of monitoring equipment,video surveillance system has been widely used in many fields.The work of observing and processing massive monitoring images by artificial way becomes heavier,so computer vision technology came into being.Through the computer processing monitor images to replace the artificial way,making people's work efficiency has been greatly improved.Where,person re-identification is a very important technology in field of computer vision.Person re-identification is to match target person images observed from different camera views of non-overlapping multi-camera surveillance systems.The technology has important applications in the field of video surveillance such as providing useful clues for the police,helping for finding children who are separated from their families in public places and multi-camera person tracking.However,the same person observed in different camera views undergoes significant variations in illumination,viewpoints,and poses,it is likely to form a lot of differences in appearance and that causes interference in person re-identification.Facing these challenges,how to build a peculiar and robust representative feature to describe the person appearance in a variety of changing environments,and how to get an effective distance metric learning algorithm.In this paper,the problem of person re-identification based on distance metric learning is discussed.The main research results include the following aspects:(1)Aiming for these dominant algorithms to learn a similarity is the metric learning that learns a Mahalanobis Similarity Function(MSF)to estimate the similarity of a pair of persons.However,the MSF only projects a pair of persons into feature difference space and ignores the appearance of each individual.This paper proposed to learn a Bidirectional Relationship Similarity Function(BRSF)to calculate the similarity between a pair of person images.BRSF not only describes the cross-correlation relationship of a pair of person images features,but also correlates the autocorrelation relationship.The proposed method use the ideal of the KISSME(Keep It Simple Straightforward Metric)algorithm to learn a similarity function.Specifically,the auto correlation relationship and cross correlation relationship of a pair of sample features are expressed by Gaussian distribution.Finally,by converting the ratio of the final Gaussian distribution into the form of BRSF,the proposed method gets a similarity function which is robust to the change of background,viewpoint and posture.The experimental results show that the proposed method have a higher person re-identification rate than the existing methods.(2)Due to person images are easily affected by illumination changes,different viewpoints,varying poses,and the person image feature is in the linear original feature space can not be separated,the similarity and difference between samples can not be distinguished by the metric matrix obtained,which leads to poor person re-identification effect.This paper proposed dense horizontal stripes and kernel space mapping for person re-identification.First,the each horizontal stripe of person images is extracted from color features and a texture feature by using the top-down sliding horizontal stripe.Then,fusing multi-features of person images and mapping the obtained features to kernel space.Finally,the proposed algorithm gets a similarity function which is robust to the change of background,viewpoint and posture by learning in kernel space.Pedestrians are re-identified by comparing rankings of similarities.Experimental results show that compared with similar algorithms,the proposed algorithm obtains more distinguishable person image features and a more efficient similarity function,and has a higher person re-identification rate.