Design And Implementation Of Vehicle Re-Identification System Integrating Multi-Attribute And Local Features

With the rapid development of the Internet of Things,a massive amount of surveillance videos are produced in real-time.Traditional manual retrieval and tracking of illegal and criminal vehicles are time-consuming and require a large amount of manpower.Therefore,the study of vehicle re-identification systems is of great significance.Vehicle re-identification is essentially an image retrieval problem that retrieves images of the same vehicle from a gallery set captured by other cameras,given a user input image of a vehicle.Vehicle re-identification is an important research direction in the field of computer vision and has a wide range of applications in traffic safety and vehicle tracking.However,there are still issues with the task of vehicle re-identification,such as significant viewpoint differences across different cameras,insufficient local feature extraction,and lengthy retrieval times.In this paper,we conducted research on the re-identification problem based on deep learning.The main research contents are as follows:(1)To alleviate the viewpoint differences of vehicles in different perspectives,we used a synthetic dataset with rich attribute annotations to assist in the vehicle re-identification task.However,the synthetic data and real data differ,which limits the model’s generalization ability.Therefore,we introduced domain-adaptive adversarial training to fully utilize the labels of the synthetic dataset and improve the attribute recognition ability of the real dataset.We also studied the importance of different attributes for re-identification and used an attention mechanism to construct the weight relationship between attributes and the final recognition effect to improve the vehicle re-identification effect.(2)To address the problem of insufficient local feature extraction in current vehicle re-identification,we proposed a directional feature dropout branch combined with the direction information unique to vehicles.This branch generates a corresponding mask matrix by judging the different directions of the vehicle,forcing the network to extract more extensive local features for re-identification and effectively improving the vehicle reidentification effect.(3)We designed and implemented a vehicle re-identification system that can output vehicle-related attribute information based on a user input image of the vehicle and retrieve other images of the target vehicle from a gallery set.To improve the retrieval efficiency of massive data,the system adopts vector indexing technology.Finally,we conducted comprehensive functional testing and non-functional testing of the system and demonstrated the operation interface of relevant modules.The test results show that the system’s functions can meet the requirements and can effectively help investigators conduct rapid vehicle retrieval and reduce manual costs.