An Efficient Real-Time Vehicle Re-Identification Scheme Using Urban Surveillance Videos

With the explosive use of surveillance and onboard cameras, vision-based Vehicle Re-identification (VReID) has attracted widespread attention. The goal of VReID is to search and identify the target vehicle from a large number of images. An efficient VReID model can help the police make fast decisions and improve regional security. The major challenge of the VReID is to distinguish the subtle visual difference between different vehicles. In this work, we propose a Compact Attention Unit (CAU) that relies on a single attention map to extract the discriminative local features of the vehicle. We add two CAUs to the truncated ResNet to construct a small but efficient VReID model, ResNetT-CAU. The feature representation of the vehicle image is the concatenation of the extracted global and local features. Compared with the original ResNet, the model size of ResNetT-CAU is reduced by 60% and has excellent VReID performance. We conduct experiments on two benchmark datasets, VeRi and VehicleID. The results show the proposed model stably achieves excellent VReID performance with very fast processing speed.