Research And Implementation Of Vehicle Re-identification In Traffic Monitoring Video

With the improvement of transportation infrastructure,the number of vehicles in road traffic has soared,and a series of public safety problems have occurred frequently.The vehicle re-identification work in traffic monitoring video plays a key role in maintaining social public safety as a core module for large-scale vehicle identification,intelligent transportation and monitoring.Vehicle re-identification is designed to quickly retrieve,locate,and track target vehicles in the surveillance camera network so that the same vehicle can be identified in images taken by different cameras.In the vehicle re-recognition,the traditional method is to use the license plate information as the unique identifier of the vehicle identity to search,but the license plate information has the situation of forgery,occlusion,deck,etc.,in order to effectively solve the accuracy of using the license plate information for vehicle re-identification.The bottleneck problem,while considering the characteristics of the vehicle itself,enhances the discrimination of the same vehicle with very similar appearance.Based on deep learning,this paper proposes a vehicle re-recognition model based on Hard Triplet’s multi-task learning.Combining characterization learning with metric learning,feature learning from two dimensions,coarse and fine.The characterization learning is used to complete the coarse-grained feature extraction including the vehicle model,color and other attributes;in the metric learning,the grading feedback is used to learn the fine-grained features belonging to the vehicle after culling the attribute features learned by the learning.At the same time,using the improved Hard Triplet,it is difficult to distinguish between vehicles belonging to the same model but different IDs,which greatly improves the accuracy of vehicle re-identification.Experiments have shown that using Hard Triplet improves the metric learning loss function compared to the traditional simple triplet,the accuracy is improved by about 5%.In addition,in order to shorten the retrieval time when re-identifying the mass vehicle data set,the bucket retrieval method is designed.Compared with linear retrieval,bucket retrieval can achieve the purpose of faster retrieval speed in the case of massive data retrieval.Aiming at the problems of small data volume and poor image quality of vehicles in the current public vehicle dataset,this paper organizes and proposes a large-scale vehicle dataset BUPT-Vehicle based on traffic surveillance video images,with a total of 303,539 maps containing 24,419 vehicles.928 models,13 colors,and a pair of vehicle attribute information,to help follow-up research in the field of vehicle re-identification.The vehicle triple recognition model based on Hard Triplet’s multi-task learning not only can effectively solve the problem of vehicle re-identification under traffic monitoring video,but also can be used for fine-grained vehicle identification and color recognition.