Real-Time Vehicle Re-Identification Method in Intelligent Transportation Environments Based on Siamese Network

Abstract

An intelligent transportation system's real-time vehicle re-identification feature must be capable of cross-camera tracking, traffic flow analysis, infraction tracing, and vehicle route reconstruction. However, changes in viewpoint, lighting, motion blur, partial occlusion, and very similar vehicle models can readily modify a vehicle's appearance. This research proposes a Siamese network-based real-time vehicle re-identification approach to improve recognition reliability and inference speed. A multi-granularity representation module is devised to incorporate compact semantic signals, local discriminative areas, and global body appearance, while a dual-branch shared-weight feature extractor is built to learn identity-sensitive vehicle embeddings from paired photos. The distance between identities is then increased while maintaining the compactness of intra-identities using a distance-aware similarity learning technique. The suggested approach has a Rank-1 accuracy of 94.3%, a mAP of 86.7%, and an average matching latency of 21.4 ms instead of 38.6 ms when compared to a conventional deep re-identification baseline, according to experimental study of multi-camera traffic scenes. The findings suggest that Siamese metric learning can be utilized to develop a practical and efficient solution for real-time vehicle identity association in intelligent traffic monitoring.