Unsupervised Vehicle Re-identification Via Spatio-Temporal Inference

Traffic monitoring system is an important means for transport management departments to carry out traffic control.The traditional manual way is becoming more and more difficult to cope with the rapid growth of traffic big data due to the growing scale of the road network.The construction of intelligent traffic monitoring system has become an urgent need.Among them,vehicle re-identification is a key technology in the field of intelligent video surveillance,which means to analyze and infer the running state of vehicles in the road network by identifying different vehicles in the surveillance camera network without overlap.The vehicle re-identification technology realizes the intelligent analysis of large-scale traffic surveillance data and effectively saves the manpower cost of processing traffic surveillance video.It provides an application basis for multi-camera cooperative tracking of target vehicles,abnormal diagnosis of vehicle driving state and intelligent control of road traffic.Due to the development of economy and the update of technology,the speed of vehicle replacement is getting more and more faster,and the existing datasets are more and more difficult to adapt to the real traffic scene.Due to the low efficiency and high cost of manual tags,it is difficult to obtain new tagged datasets,which seriously affects the research of supervised vehicle re-identification.Around the above interference factors,this paper designs an unsupervised vehicle re-identification method with stronger cross-scene generalization ability from two aspects.The main research work is as follows:(1)In order to make better use of the unlabeled dataset training model,an Contour Guide Masked Autoencoder for Unsupervised Vehicle Re-Identification is designed in this paper.Under the guidance of the vehicle contour,the key areas of the image are mined.By encoding and decoding the key areas,the original image of the vehicle is reconstructed,and the encoder with good feature expression ability is trained by self-supervision.The application of encoder to unsupervised vehicle re-identification effectively improves the performance of unsupervised vehicle re-identification model.(2)In order to make effective use of the spatio-temporal information without manual labeling,a Time-Spot Contrast for Unsupervised Vehicle Re-Identification is designed in this paper.By fully mining the position of the surveillance camera and the capture time of the vehicle image,the fusion features combined with the depth visual features of the vehicle image are constructed,and the fusion features are clustered to form a feature dictionary.By comparing the clustering noise to estimate the loss optimization model,the accuracy of the re-identification model is improved.In addition,in order to further verify the performance of the model in the actual traffic scene,in this paper,we uses the monitoring data collected from the real highway scene in China to label a vehicle re-identification dataset containing 784 vehicles and a total of 12236 pictures,and specially labeled spatio-temporal information labels.Comparative experiments are carried out on this dataset,and the experimental results show that the use of spatio-temporal information can effectively improve the performance of the unsupervised vehicle re-identification model in the real traffic scene.