Multi-Branch Cooperative Network For Vehicle Re-Identification

The purpose of vehicle re-identification technology is to complete the task of vehicle identification in cross camera scenes.It is a new technology in the development of urban intelligence in China.It is widely used in the fields of traffic supervision,safety prevention and control,criminal investigation and so on.In order to deal with the problems of intra class difference and inter class similarity in vehicle images captured in complex scenes,a vehicle re-identification model based on multi-attribute branch cooperation and perspective similarity discrimination is designed and implemented in this thesis.Firstly,this thesis confirms the general idea of the model of joint learning using representation learning and measurement learning.In order to comprehensively utilize the appearance information of the vehicle in the picture,this thesis extends the attribute branch used in the classification task.Firstly,this thesis uses the color and model label in the data set to learn the multi label classification task of the model;At the same time,the vehicle image feature image output from the backbone network is segmented horizontally,and the stripe region of each feature image is pooled to obtain eight feature vectors containing the information of different regions in the image for classification tasks respectively;Finally,this thesis sets up the extraction branch of vehicle local component attribute features for the model.In the setting of local component branch,firstly,an object detector is pre-trained to detect the local components of the vehicle in the picture.In the improvement of the detector,this thesis uses the Faster R-CNN model as the benchmark,and adds the characteristic pyramid network structure of double tower structure to improve the detection ability of the detector to small vehicle components.In the setting of re identifying the attribute branch of local vehicle components in the network,it is proposed to use the target relationship module in the transformer model to represent the relationship between different vehicle components,and use the attention mechanism to express the appearance pixel features between each vehicle component and all other vehicle components,At the same time,a two-dimensional image relative position coding method is introduced,and the description of the relative position information between components is added to the local component attribute features.In order to solve the difficult problem of identification caused by the appearance change in the picture caused by different shooting angles,this thesis first uses the image segmentation technology to divide the vehicle in the picture into four areas,representing the front,rear,both sides and top areas respectively,and judge the shooting angle information of the picture according to the area proportion of each area in the body area.In the improvement of the metric learning loss function,the local region triple loss function using the multi region characteristics of the vehicle body is introduced into the vehicle re-identification model,which makes the model realize a more robust clustering method for multi view challenges.At the same time,in the selection of metric learning triples,a triplet screening scheme for sampling difficult samples according to the similarity of shooting angles is proposed,which makes the model pay more attention to the training of difficult samples in terms of angles.Finally,a comparative experiment using other global measures of learning loss is carried out to verify the rationality of the difficult sample triplet loss used in this thesis.This thesis sets up validation experiments for the above attribute branches and improved methods.Experiments show that the introduction of each attribute branch improves the overall model to varying degrees.After the introduction of all attribute branches,the average accuracy of the model increases by 13.4%,and after the introduction of perspective information,the average accuracy increases by 5.2%.Compared with other cutting-edge vehicle re recognition models,it has reached the current good level.The result proves the feasibility and progressiveness of each branch.