Abstract
Effective management and safe operation of tunnel infrastructure are critically challenged by the persistent vehicle tracking across non-overlapping camera domains and the lack of real-time situational awareness for emergency incidents. This work proposes a dynamic digital twin framework to create a high-fidelity virtual representation that synchronously updates with physical tunnel environments by leveraging existing Closed-Circuit Television (CCTV) networks. The core is a novel Tunnel Re-identification Model (TRIM) that can create discriminative feature embeddings for seamless vehicle tracking and re-identification across multi-camera domains. Comparative validations unequivocally demonstrate that the TRIM significantly outperforms Convolutional Neural Networks (CNNs) and standard Vision Transformer backbones, achieving a state-of-the-art mAP@10 of 0.684. Analysis of attention heatmaps reveals its ability to generate a comprehensive and granular feature representation, focusing on preserving high-quality characteristics for specific vehicle identity. Furthermore, the emergency response ability of the digital twin is augmented with a vision-based fire detection module to real-time localize and visualize emergent tunnel fire events in the digital twin. By consolidating and distilling vision data from CCTV cameras into an intuitive representation, the proposed framework enhances situational judgment for operators and significantly reinforces the safety resilience of tunnel infrastructure without requiring extensive hardware redevelopment.