Towards Vehicle Re-identification Based On:Unsupervised Learning

With the rapid development of the modern industry,vehicles have brought a convenient way to our life and improve our efficiency enormously.However,it gets into trouble because of traffic safety problem at the same time.In order to effectively restrain these illegal driving behaviors,one of the measures taken by the traffic control department is to punish illegal vehicle owners.Nevertheless,it is not possible to directly obtain the vehicle owners'information from the videos directly in some situation,which has made vehicle re-identification a hot research topic in recent years.Vehicle re-identification aims at searching for the same vehicle from non-overlapping cameras,which is an important research direction in intelligent transportation systems.In order to improve the generalization ability of vehicle re-identification model and apply it to the real-world scenarios,this research proposed a vehicle re-identification algorithm based on unsupervised learning.Firstly,there are some problems existing in vehicle re-identification.For instance,it is required to annotate a large number of videos and it is unrealistic to utilize the trained model to another different dataset.From the mentioned above,this study proposed a novel architecture for vehicle re-identification,which utilized the Attention-Based Feature Learning Network(AFLNet)to extract features.Unlabeled vehicle information was able to be learned by reducing dimension,clustering and selecting data iteratively.The unsupervised vehicle re-identification algorithm proposed in this paper could be applied to the unlabeled dataset.What's more,the generalization ability of the model could be improved.Secondly,in order to decrease the influence of the the vehicle images background on the dataset,a vehicle feature learning network based on attention mechanism is proposed,which can focus on the foreground of the vehicle images and improve the mAP.Thirdly,the unsupervised vehicle re-identification algorithm assumed that each class of vehicle images made the same contribution to the vehicle feature learning model training,therefore,they were supposed to be given the same weight in the loss calculation.However,after the feature information was clustered,the closer the distance between the feature and it's center was,the higher the credibility was.Therefore,this paper assigned different weight to each image according to the distance between it's feature and cluster center.Qualitative and quantitative experiments have been conducted on the VeRi-776 and VehicleID dataset,which has proved that the vehicle re-identification algorithm based on unsupervised learning proposed in this paper could be applied to the unlabeled dataset and the generalization ability of the vehicle re-identification model could be improved effectively.