Exploration of computer vision fire detection and tracking in a realistic tunnel environment

Abstract Many tunnel fire detection devices and methods are adopted in practice, including vision-based and multispectral technologies. Vision-based methods are susceptible to interference from vehicle lights and other sources. Multispectral detectors sense unique infrared and ultraviolet signatures from fires, but they are sensitive to environmental parameters. This study explores the accuracy of vision-based fire detection and the likelihood of false alarms in realistic tunnel environments. By setting both static and moving fires in a real-scale tunnel, the capabilities, constraints, and sensitivities of five vision-based and multispectral fire detection systems are examined. All five sensors show satisfactory performance (accuracy above 0.748, false alarm rate below 0.192) when detecting a static 50 cm × 50 cm pool fire up to 35 m away. By lowering the detection threshold, fires up to 50 m away can be detected and a fire moving at 30 km·h−1 can be tracked (up to 20 continuous tracking frames), but the chance of false alarms significantly increases under tunnel conditions. Results also show that none of the tested sensors can detect and track a flame moving at speeds exceeding 30 km·h−1. Environmental sensitivity analysis reveals an asymmetric response of vision-based fire and smoke detection to illumination changes. This examination provides insight into the performance of vision-based fire detection and the potential for integrating multi-sensor fire detection systems for future tunnel safety.