A Scalable Real-Time Multi-Camera Vehicle Tracking System for Urban Environments
Oriol Alàs, Ricard Cervera, Robert Sanfeliu · IEEE Access · 2026
Real-time Multi-Camera Vehicle Tracking (MCVT) systems are challenging to design and deploy due to the need for real-time processing at the edge, as well as issues such as high noise, low image resolution, and reidentification. With the significant advances of computer vision techniques and transformer-based models for image classification, new opportunities have emerged for MCVT. Moreover, the growth of urban populations and the increasing demand for sustainable mobility have intensified the need for scalable traffic monitoring solutions. Despite recent progress, several critical challenges remain unresolved, including unreliable re-identification under severe occlusions and viewpoint changes, limited temporal synchronisation between heterogeneous cameras, and the lack of architectures that operate efficiently on fog and edge infrastructures. In this paper, we propose a scalable, low-latency MCVT pipeline designed for heterogeneous CCTV networks deployed across the edge–fog continuum. The system integrates lightweight vehicle detection, transformer-based appearance embeddings, and topology- and travel-time-constrained cross-camera association within a modular deployment-aware architecture. Experimental validation on both public benchmarks and a real-world logistics hub demonstrates stable identity preservation, correctly identifying more than 80% of vehicle matches under deployment conditions while maintaining strict computational constraints. The system sustains real-time performance (20–29 FPS per stream) under CPU-based detection across 27 CCTV cameras and seven edge nodes, demonstrating scalable multi-camera tracking under production-level constraints.