Vehicle re-identification and multi-camera tracking in Vietnamese urban traffic: a new dataset and distributed framework

Abstract

Urbanization has accelerated the deployment of Intelligent Transportation Systems (ITS), yet maintaining a vehicle’s identity across sprawling, non-overlapping camera networks remains a formidable challenge. This paper presents a distributed, event-driven framework designed to bridge the gap between isolated Vehicle Re-Identification (Re-ID) models and real-world Multi-Camera Tracking (MCT). Our approach leverages a big data infrastructure—comprising Apache Kafka for message buffering, Apache Spark for distributed inference, and Apache Airflow for orchestration. The pipeline integrates YOLOv8 for detection and ByteTrack for localized tracking, coupled with a deep embedding model for cross-camera association. Experimental results on a custom-collected multi-camera dataset show a Localization Accuracy (LocA) of 74.9% and a HOTA of 33.4%. While the system scales horizontally to handle high-definition streams, we identify drastic viewpoint variations as the primary bottleneck for identity consistency. These findings offer a pragmatic blueprint for deploying scalable vision analytics in smart city environments.