| The re-identification technology aims to identify the same target from different shooting scenes,and it is an important branch of surveillance research of smart cities and smart video.Vehicle re-identification is one of the problems that need to be solved in re-identification technology.It is a simple and effective method to achieve vehicle re-identification by license plate identification.However,vehicles in road surveillance video often there are no license plates or license plate is blocked.The occlusion situation brings a big challenge to the traffic law enforcement department.This requires the vehicle to be re-identified by extracting and comparing other information of the shooting vehicle.However,the pictures of vehicles captured in actual surveillance videos often have changes in illumination and the angle of view,which affect the texture,edges,and other characteristics of the vehicles in the pictures.In addition,the inconsistency of camera brands,parameters,etc.may also cause the same vehicle images exist big difference.Therefore,in the field of intelligent video surveillance,vehicle re-identification technologies is indispensable to monitoring vehicles.Therefore,the research of this paper mainly focuses on several problems encountered in the current vehicle re-identification field mentioned above.The main work is as follows:(1)There is one problem in the existing vehicle re-identification field is that the vehicle re-identification data set is lacking.So we made three vehicle re-identification data sets.One of the data sets is composition by 632 vehicle samples from the perspective of the frontal face of an ordinary expressway vehicle.The other data set is composition by 236 vehicle samples with large changes in viewing angle.The last one is composition by 500 vehicle data sets with large changes in lighting.(2)The extracted cross-view images of the same vehicle often there are large differences in the characteristics,and it is not reliable to directly measure the similarity of the vehicle’s image.This article refers to the method of pedestrian re-identification,and uses a coupled dictionary learning method for vehicle re-identification.The feature data is processed to realize the reconstruction of the vehicle image across the field of view.(3)Considering that intra-class commonalities and disparities between classes need to be emphasized when different images are captured by different cameras during vehicle re-identification.This paper designs a label matrix for label consistency constraints and label constraint information in the dictionary training model.The distance between similar training sample data can be well shortened,and the trained dictionary is more discriminative.In addition,the related information of the same sample is mapped to a public space through the mapping matrix,The experiments in the vehicle re-identification data set and pedestrian re-identification data set verify that the method has a good recognition rate. |
