Multi-feature Fusion and Non-Local Operation for Vehicle Re-identification

As one of the most important tasks in the computer vision, vehicle re-identification aims to retrieve and identify the same vehicle under different surveillance cameras, which plays a key role in urban road traffic safety and intelligent traffic management system. However, the large intra-class difference and high inter-class similarity are still main challenges, as well as the diversity in lighting conditions, camera's shooting angle, and occlusion degrees. In order to further improve the average accuracy and algorithm performance, this paper proposes a vehicle re-identification algorithm based on multi-feature fusion and non-local operation. We embed non-local operation into the ResNet50 network, and employ feature slicing and reorganization to obtain multiple feature branches. Besides, learning rate warm-up and cosine annealing scheduler are also used. The experimental results show that our proposed method achieves higher accuracy on two commonly used datasets VeRi-776 and VehicleID.