The Research Of Person Re-identification Based On Local Feature And Metric Learning

With the establishment of a large-scale surveillance system in the cities,a lot of video information is generated every day,which helps the police to solve the crimes.Checking these videos with human eyes not only misses some important clues and fails to catch the opportunity to break the case,but also is very labor consuming.Person re-identification refers to matching the object person across non-overlapping cameras.As an important part in intelligent surveillance,it has drawn intensive attention.However,it still faces enormous challenges due to the illumination,occlusion,pose and viewpoint.Current person re-identification methods focus on feature representation and metric learning.The feature representation aims to obtain the invariant pedestrian descriptor,while the metric learning makes the intra-class distance less than inter-class distance.This paper makes the following contributions:In the methods of feature representation,global feature usually discriminates different person from the overall appearance.However,individual global feature is not sufficient to express local detail information,so that it does not meet the reality.In this paper,we propose a deep convolutional neural network based on mid-level attribute recognition and high-level identity classification.In our model,we create a local branch and use the horizontal average pooling to obtain the local feature.Besides,attribute recognition can catch the salient feature,so we utilize mid-level attribute recognition to enhance the high-level identity feature.We jointly train the network to obtain complementary global feature and local feature.From the perspective of metric learning,we propose a novel batch-contrastive loss function to solve some current problems.In our loss function,we make full use of the distance between all pairs.Because the feature learned from hard pairs can better reflect the subtle difference,we have a process of online hard examples mining for each sample.In order to avoid choosing image pairs,we introduce the center loss to force the samples close to their centers,so that it indirectly reduces the distance between intra-class samples.Furthermore,for each sample,the inter-class distance and the intra-class distance are relatively balanced.We verify the efficiency of our feature representation model and the proposed metric learning loss function in two large-scale person re-identification datasets and the performance of our methods is superior to most current algorithms.