YCFA-Net: A unified framework for vehicle detection and fire anomaly recognition in tunnel scenarios

Fire accidents in highway tunnels present a significant risk to urban transportation safety, often caused by vehicle crashes or spontaneous combustion. Accurately identifying vehicles on fire and their surrounding fire areas is crucial for tunnel fire detection. Most deep learning methods rely on extensive training data for detection, which is challenging to obtain in practical tunnel scenarios. To address this issue, we introduce the YOLO Coupled Hypersphere-Based Feature Adaptation Network (YCFA-Net), designed for simultaneous vehicle detection and unsupervised fire anomaly recognition. YCFA-Net employs an unsupervised approach to identify flames around vehicles, reducing the effects of limited training datasets. The designed framework achieves the highest AUCROC scores of 98.8% and 96.7% at the instance level and the pixel level, respectively, and the AUPRO score at the pixel level reaches 84.5%. Meanwhile, we introduce the Tunnel Fire-AD dataset, recorded through real-time monitoring of highway tunnels in Shanxi and Ningxia, China. Tunnel Fire-AD dataset broadens the experimental scenarios for detecting and unsupervised fire anomaly recognition. Validation on the proposed dataset demonstrates the framework's practical feasibility in addressing vehicle fire risks, providing an effective method for tunnel fire anomaly detection. Related Code and dataset have been publicly released at the following links: (https://github.com/RichardLio/Tunnel-Fire-AD);(https://github.com/RichardLio/UCFA).