GAN-Siamese Network for Cross-Domain Vehicle Re-Identification in Intelligent Transport Systems

• Published in: IEEE transactions on network science and engineering Vol. 10; no. 5; pp. 2779 - 2790
• Main Authors: Zhou, Zhili, Li, Yujiang, Li, Jin, Yu, Keping, Kou, Guang, Wang, Meimin, Gupta, Brij Bhooshan
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.09.2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Artificial neural networks; Cross-domain matching; Deep learning; Distance metric learning; Domain transformer; Feature extraction; GAN-Siamese network; Generative adversarial networks; Intelligent transportation systems; Matching; Measurement; Task analysis; Transformers; Transforms; Vehicle Re-ID; Vehicles; Visualization
• ISSN: 2327-4697; 2334-329X
• DOI: 10.1109/TNSE.2022.3199919

The vehicle re-identification (Re-ID) has become one of most important techniques for tracking vehicles in intelligent transport system. Vehicle Re-ID aims at matching identical vehicle images captured by different surveillance cameras. Recent vehicle Re-ID approaches explored deep learning-based features or distance metric learning methods for vehicle matching. However, most of the existing approaches focus on the vehicle Re-ID in the same domain, but ignore the challenging cross-domain problem, i.e. , identifying the identical vehicles in different domains including the day-time and night-time domain. To tackle this problem, we propose a GAN-Siamese network structure for vehicle Re-ID. In this network structure, a generative adversarial network (GAN)-based domain transformer is employed to transform the domains of two input vehicle images to another domains, and then a four-branch Siamese network is designed to learn two distance metrics between the images in the two domains, respectively. Finally, the two distances are fused to measure the final similarity between the two input images for vehicle Re-ID. Experimental results demonstrate the proposed GAN-Siamese network structure achieves the state-of-the-art performances on four large-scale vehicle datasets, i.e. , VehicleID, VERI-Wild, VERI-Wild 2.0, and VeRi776.