Research On Vehicle Re-identification Algorithm Based On Deep Learning

In recent years,with the rapid development of artificial intelligence,it has become possible to use deep learning technology to process criminal investigation data.When polices are pursuing criminal vehicles,they can quickly and accurately retrieve suspect vehicles from different cameras through vehicle re-identification technology,which can improve the efficiency of criminal investigation cases.Each vehicle has a unique identification,which is the license plate.Vehicle re-identification through the license plate is the most direct and effective way.However,when faced with a concealed or forged license plate,etc,it is necessary to consider using the vehicle appearance information to complete the re-identification task.The deep learning technology is used to extract and analyze the global and local features of the vehicle to vehicle re-identification.However,at this stage,there are still many problems in vehicle re-identification through deep learning technology.First of all,there will be interference from factors such as complex background,illumination,viewing angle,etc in the vehicle image,which will cause the network to not be able to extract the characteristic vehicle features well.Secondly,the differences between the same models are small and difficult to distinguish.This paper focuses on the existing problems of vehicle re-identification,and the main research contents are as follows:(1)Aiming at the interference of complex background,illumination,viewing angle and other factors,this paper proposes a vehicle re-identification algorithm based on the attention mechanism.The attention mechanism is added to the convolutional neural network to extract the global characteristics of vehicles,so that the network gives more important channels and space higher weight.And through the improved batch hard triplet loss and cross-entropy loss joint learning method,the image distance of the same vehicle under different viewing angles is reduced,and the image distance of different vehicles with similar appearances is increased.After using the above method,a higher accuracy rate is obtained.(2)Aiming at the problem of small differences between the same models and difficult to distinguish,this paper proposes a vehicle re-identification algorithm based on object relationship modeling,and uses a object detector to detect the local region of interest objects of the vehicle such as annual inspection signs,special displays,windows,lights and logos.And then the relationship between the local objects is modeled through the object relationship module,so that each object feature output also contains the relationship features with other objects,and combining multiple local object features into a uniquely discriminative vehicle local feature,which improves the discrimination of the same type of vehicle.(3)The vehicle global features obtained by using the attention mechanism and joint learning and the vehicle local features obtained by using object relationship modeling are fused together,and the vehicle re-identification task is carried out through the final fusion feature.And use reordering in the vehicle re-identification algorithm to build a more robust and accurate vehicle re-identification model.Experimental results show that the accuracy of using this method is better than that of using global or local branches alone.