Vehicle Re-identification method based on Swin-Transformer network

Abstract

Vehicle Re-identification is to find out the exact vehicle captured by other cameras given a vehicle image. In natural traffic surveillance systems, vehicle re-identification can play a role in target vehicles localization, supervision, and criminal investigation. While current convolutional neural network-based approaches have been very successful, most studies consider learning representations from individual images, ignoring potential interactions between them. The Swin Transformer method currently shows excellent advantages in image classification. It can handle features at different scales like convolutional neural networks and make elements more closely related to each other, enhancing the learning of robust features and improving feature robustness and discrimination compared to convolutional neural nets. Therefore, this paper adopts the Swin Transformer method for vehicle re-identification and conducts experimental validation on two large vehicle datasets, VeRi-776, and VehicleID. The experimental results show that the recognition effect of this method is significantly better than existing vehicle re-identification methods, such as VAM, FDA-Net, TransReID, etc.