Research On Vehicle Re-identification Technology Based On Deep Learning

The purpose of vehicle re-identification is to search,locate and track the target vehicle quickly through the surveillance camera network.Simply,the vehicle re-identification can judge whether the vehicles under different cameras belong to the same vehicle.Nowadays,as one of key technologies,vehicle re-identification,with great significance to build a smart city and maintain social security,play a key role of among modern intelligent transportation system.Most of the previous vehicle related research focuses on vehicle detection,classification and tracking.In contrast,vehicle re-identification is a relatively new research hotspots meanwhile as a far from solved topic during computer vision.There are two main challenges for accurate vehicle re-identification:(1)inter-class similarity,i.e.different vehicles have similar appearance;(2)intra-class variance,i.e.images belonging to the same vehicle usually have large differences due to the change of perspective and light.Aiming at the two main challenges of vehicle re-identification,this article studies the mining of discriminative vehicle features based on deep learning technology.The research of this article is as follows:(1)Vehicle re-identification based on convolutional neural network.Firstly,summarizes the basic models used in vehicle re-identification in recent years and mainly analyze several commonly used loss functions.Second,the performance of various basic models and loss functions are analyzed by experiments.In this way,it paves the way for further research.(2)Vehicle re-identification based on local and multichannel convolutional neural network.In order to extract features with high discrimination,in this article we combines local and global features of vehicles to extract multi-channel features of vehicles.The experimental results on open datasets show that our method can archive advanced performance.(3)Vehicle re-identification based on attention mechanism and multi task learning.In order to further improve the accuracy of vehicle re-identification,in this article we build a multi task model,which using vehicle classification task to improve the accuracy of vehicle re-identification task.Furthermore,channel attention module is added to the backbone of the model.The experimental results on the public vehicle re-identification data set show that the model can improve the accuracy of vehicle re-identification.