Research On Person Re-identification Method Based On Deep Learning

Person re-identification refers to the problem of pedestrian object retrieval under the cross-camera view area.With the continuous increase in the coverage area of monitoring facilities and the increase in monitoring data,pe rson re-identification provides police officers with convenient target retrieval and identity authentication schemes.Due to the increase in the amount of data and the impact of various factors in the image capturing process,person re-identification tasks under different surveillance cameras face great challenges.How to extract discriminative person features under complex interference conditions and how to use appropriate similarit y measurement methods to improve the accuracy of re-recognition,these two problems are the current research difficulties of person re-identification tasks.Based on a large number of domestic and foreign research results,this paper uses deep learning methods to conduct research on person re-identification tasks.The specific work is as follows:1.Establish a person re-identification model using multiple feature learning methods.For pedestrian images are affected by many factors such as lig ht,occlusion,pose,angle,etc,pedestrian image cannot be fully characterized if a single feature extracted,so a new person re-identification model is proposed.The model uses the perspective information of the pedestrian image as a global feature,and uses the batch feature erasure method to learn the fine-grained features of the pedestrian image.Through joint learning of the two kinds of features,the model can learn the pedestrian image features more fully.The experimental results show that the proposed model can effectively improve the re-identification accuracy.2.Use multiple attention mechanisms to extract more discriminat ive pedestrian image features.Because the person re-identification datasets lack the attribute labels of pedestrian images,it makes it difficult to extract the attribute features of pedestrian images,so a multi-attention mechanism method is proposed.In the process of extracting person image perspective information,feature attention is introduced,which uses a view classifier to weight different view units,and then channel attention and spatial attention are also introduced to view units.Experiments show that the three different attention mechanisms work together can extract more discriminative perspective features,and can improve the accuracy of person re-identification.3.A new triplet loss function is proposed by constructing a triangle model.In the similarity measurement process of person re-identification,the closer the distance between positive sample pairs is,or the further the distance between negative sample pairs is,the model can gain better re-identification performance.Therefore,a new triplet loss function is proposed.This loss function establishes a triangle model with a anchor sample,a positive sample,and a negative sample.According to the correspondence between triangle angles and edges,the distance between positive sample pairs and negative sample pairs is redefined.The newly defined distance makes the distance between positive sample pairs closer and the distance between negative sample pairs further.According to the experimental results,it can be seen that the proposed loss function can improve the accuracy of person re-identification comparing with other loss functions,and has good generalization performance on three public person re-identification datasets.