Person Re-identification By Regularized Smoothing Cross-view Maximum Margin Metric Learning

In recent years,with the development of computer vision,person re-identification technology has become a hot research direction in the field.The main purpose of the technology is to determine whether the pedestrian is the same person at different periods of time under different cameras.This technology plays an important role in the supervision of the security department.But as of right now,this technology still faces many challenges:the resolution difference between cameras,the diversity of pedestrian posture,the occlusion of objects and the intensity of light.At present,the main research is broadly divided into two categories:Research on feature representation method based on pedestrians,seek and design a pedestrian representation that is robust to factors such as visual angle change,illumination change and so on;Research based on metric learning method,using pedestrian sample to train a more discriminative metric decision function,the similarity between people is far greater than the similarity of different pedestrians.When calculating the similarity of pedestrian eigenvector,the Euclidean distance,cosine distance and other functions are usually used.However,these classical distance functions usually do not take into account the sample characteristics,which makes the recognition rate of the algorithm is not very high.For the past few years,some new discriminated measurement learning methods are based on training samples with labels,and think over the sample characteristics,and have a significant improvement in recognition rate.The Mahalanobis distance metric learning has been widely used in some algorithms,such as the large margin nearest neighbor classification distance metric learning algorithm(LMNN),keep it simple and straight(KISS)distance metric learning algorithm,and cross-view quadratic discriminant analysis(XQDA).However,in the process of projecting the high-dimensional data into the new low-dimensional feature space,some distance metric learning algorithm without considering the small sample problem;at the same time also does not consider the difference between positive and negative samples.To solve these problems,this paper put forward the relevant solutions.The contents of this paper are summarized as follows:(1)In order to classify and identify pedestrians more accurately,and seek better projection subspace to reduce the dimension of the sample,we make improvements in the XQDA algorithm to learn a low dimensional discriminant subspace by using the maximum margin criterion,which named the cross-view largest distance metric learning algorithm(XMM).the algorithm we proposed can solve the small sample problems that caused by samples with training labels.The experimental results show that the XMM algorithm has a certain improvement compared with the XQDA algorithm.(2)At the same time,some distance metric learning does not consider the difference between positive and negative samples,thus affecting the robustness of the sample covariance matrix.In this paper,the covariance matrix is dealt with by regularized smoothing in XMM algorithm,which named the regularized smoothing cross-view maximum distance metric learning algorithm(RSXMM).The experimental results show that the efficiency of the RSXMM algorithm is superior to the XMM algorithm.(3)The algorithm of person re-identification is based on robust person feature representation,and for this reason,we present a robust pedestrian representation(Saliency weighted local maximum occurrence representation,SWLOMO).Learning weight by saliency detection in various parts of the body,then,when extracting color features,we first multiplied the pixel value of each channel by the weight of the position,which can make better use of the color information of pedestrian appearance.Experimental results show that,this feature has obvious enhancement effect in different algorithms.