Enhanced flame detection in virtual tunnels using DEV-YOLOv8 and digital twin systems

Highway tunnel fires pose significant safety challenges requiring reliable detection and response strategies. Traditional testing methods face limitations due to insufficient sample diversity and high costs, thus making virtual environments a promising alternative for algorithm development. This study proposes DEV-YOLOv8, an enhanced YOLOv8 model integrated with digital twin technology for flame detection in simulated tunnel environments. The model incorporates three key enhancements: deformable convolutions for adaptive feature extraction, an efficient multi-scale attention mechanism for stable parameter updates, and a cross-stage partial connections module for efficient multi-scale fusion. A Unity-ROS2-based digital twin system enables cost-effective simulation and dataset generation. Experiments demonstrate DEV-YOLOv8’s superiority with 2.2%/1.5%/1.2% improvements in accuracy/recall/[email protected] respectively, while reducing computational cost by 0.8 GFLOPs. The integrated system achieves 98.21% simulation accuracy and 96.19% physical deployment accuracy, with latency differences $<$ 3.15 ms and RRT delays below 9.91 ms. This work establishes an efficient framework combining deep learning and digital twins for adaptive fire detection in tunnel environments.