A Real-Time Vehicle Re-Identification System Based on Big Data Streaming Video Processing

Intelligent Transportation Systems require robust methods for associating vehicle identities across non-overlapping surveillance cameras, yet large viewpoint variation, illumination changes, occlusion, and limited image quality continue to impede reliable cross-camera matching. This paper presents a deep learning-based vehicle re-identification system that combines a lightweight, high-fidelity feature extraction network with an efficient, scalable similarity-search pipeline built for large-scale matching. The proposed feature extractor is designed to produce discriminative appearance embeddings while maintaining low computational and memory footprint, enabling deployment on resource-constrained inference nodes. The similarity-search component exploits spatio-temporal constraints, trajectory-level aggregation, and an index-based retrieval architecture to enable accelerated nearest-neighbor search across many cameras and long time horizons. The full pipeline is engineered for real-time operation and integration within big-data environments. We evaluate the proposed system on a dedicated real-world Vietnamese vehicle Re-ID dataset that captures the diversity of local traffic-dense flows, mixed vehicle types (motorcycles, passenger cars, light commercial and heavy vehicles), variable weather and lighting, and common imaging artifacts such as compression noise and motion blur. Experimental results demonstrate state-of-the-art performance, achieving a Rank-1 accuracy of 90.0 % and a mean Average Precision (mAP) of 94.01 %, while maintaining stable throughput and latency under varying loads, confirming both recognition robustness and runtime scalability. The principal contributions are: (1) a real-world Vietnamese vehicle Re-ID dataset that encapsulates the visual and operational variability of local traffic and enables realistic benchmarking for cross-camera recognition and tracking, (2) an end-to-end, scalable similarity-search pipeline for large-scale, low-latency matching that combines appearance and spatio-temporal cues with trajectory-aware feature aggregation, and (3) a deploymentoriented software architecture based on Apache Kafka and Spark Streaming, together with empirical validation demonstrating high recognition accuracy and real-time scalability, confirming the system's practical suitability for ITS deployments.