Research On Person Re-identification Representation Learning Based On Semantic Information And Attention

The awareness of public safety in today's society is getting stronger and the demand is more urgent.Pedestrian detection and person re-identification based on video data collected by security systems have great application value in maintaining public safety,etc.Compared with traditional manual methods,the person re-identification algorithm has great advantages in saving manpower and improving efficiency in retrieving specific individual from massive data.In view of this,research on person re-identification has important value in both theories and practicability.In recent years,research on person re-identification based on convolutional neural networks has made great progress.On the one hand,large data sets such as Market1501,DukeMTMC-reID,and MSMT17 have been proposed successively.On the other hand,various ingenious and effective works in the directions of representation learning and metric learning are in full swing.However,as a retrieval problem,person re-identification itself is difficult,and it is also bound by external conditions such as pose,lighting,and occlusion.Therefore,research on person re-identification is still far from practical applications.As far as representation learning is concerned,researchers often fail to take into account the fine-grained information lying in pedestrian area and the hierarchical information of the network while resort to global and local features.In addition,in the field of person re-identification,the attention model has also been proven to be effective for separating pedestrians from the background of the whole frame,but often the attention masks in the model are obtained more generally and do not encode the weights against the location.In view of the above problems,this thesis makes an exploratory study on the basis of previous studies.For exploring person semantic information,this article makes use of the pedestrian attributes and hierarchical information of convolutional layers.In addition,an attention module is proposed which takes advantage of the relative positional relationship between the spatial positions.All the network structures in this thesis have been conducted sufficient experiments on multiple data sets to demonstrate their feasibility and effectiveness.And t this work can be divided into the following two aspects:1.Considering that person attributes are robust and informative to identify pedestrians.This thesis proposes a multi-branch model,namely Part-based Attribute-Aware network(PAAN),to not only utilizes person ID label visible to the whole image but also utilizes pedestrian attribute information.One attribute branch can be used for pedestrian attribute recognition,and the other branch is used to predict pedestrian ID labels.The overall objective function is linearly coupled by the cross-entropy loss of the two branches through the coefficients.In order to enhance the pedestrian representation of the shared convolutional layers before the branches,a semantic bridge is designed to fuse middle-level and high-level semantic information in backbone.2.In order to make the learned pedestrian representation more accurate,a weak spatial attention module(WSAM)is proposed in this thesis.This module mainly obtains the relative importance between each spatial position of the module's input through the Gram matrix,and this mask acts as the input's weight.The resulting output is linearly combined into the output of the module through one learnable parameter.The visual analysis of the module revealing that the output of the module is more accurate than the input in locating pedestrians,which confirms the original intention of the module design.This module can be easily integrated into each convolutional layer of the basic network to improve the retrieval performance of person re-identification.