The Algorithmic Research On Person Re-identification Based On Attention And Attribute Learning

With the rapid development of science technology and the rapid change in social forms,people pay more and more attention to public security.A large number of video surveillance devices are placed in crowded public places,which plays an important role in the detection of criminal cases,tracking of missing persons,and urban public security management.Person re-identification is an important technology for video analysis,which aims to retrieve and match whether the detection target is the same person from different views.Due to the factors such as shooting angle,lighting,occlusion,and background in real scenes,the pedestrian images also have large difference,so person re-identification is a challenging research.The current research on person re-identification is mainly divided into two aspects: one is to extract the discriminative and robust feature vector,and the other is to design a reasonable and effective distance metric function to constrain the deep feature.In this paper,we design an effective person re-identification algorithm based on both aspects,at the same time,deep learning technology is considered.In order to describe the pedestrian image accurately,this paper designs a person re-identification algorithm model(SJ-AAN)based on attention and attribute learning.The model is a two branch structure,which can represent the pedestrian image from different levels.It is mainly summarized into two aspects:(1)Based on the global deep attention feature,this paper integrates the attention mechanism into the deep convolutional neural network to form the attention branch of the network.The attention branch is adaptively update by iterative training and focus on the saliency area of pedestrian in the image,so as to effectively reduce the noise caused by objective factors such as background and lighting.(2)Based on the local semantic attribute features,the attribute features contain the semantic information of the pedestrian image and show high robustness when the appearance and posture of the pedestrian changes.Therefore,In order to capture the semantic invariance of pedestrians,this paper uses the deep residual network to extract the attribute features of the image.In order to constrain the features of different branches in the feature space,this paper designs attention loss and attribute loss respectively.Specifically,attention loss constrains pedestrian identity,so that the same pedestrians aggregate with each other in the feature space,and different pedestrians disperse each other.Attribute loss constrains the local part of the pedestrian,which makes the probability that the pedestrian with more similar attributes is the same person greater,and the probability that the pedestrian with less similar attributes is the same person smaller.At present,most person re-identification research only cascades different features and ignores the correlation between different features.In view of the features of different branches,this paper creatively designs a person re-identification algorithm based on bilinear feature embedding(BEN)capturing the correspondence between different features.Bilinear feature embedding mainly includes two parts: bilinear feature aggregation and spatial pooling embedding.Bilinear feature aggregation first obtains the compact expression of features through affine transformation,then aggregates attention features and attributes by outer product,and finally the fusion feature is formed through spatial pooling.This greatly improves the effect of person re-identification.The current training of person re-identification networks relies on a large amount of labeled data,which severely limits the scalability of person re-identification in practical applications.In this paper,attribute related learning strategy is proposed to optimize the network model step by step.The network embeds the prediction result of the attribute branch as a "soft label" into the attention branch to help attention branch training iteratively,so as to carry out unsupervised person re-identification.Finally,in order to verify the effectiveness of the proposed algorithm,experiments are performed on two standard datasets: Market-1501 and DukeMTMC-reID.Experimental results show that the algorithm proposed in this paper can significantly improve the accuracy of person re-identification.