Computer Vision-based Fire Detection and Localization Inside Urban Rail Transit Stations

To efficiently address the occurrence of in-station fire incidents in rail transit, this paper proposes a computer vision-based model for fire detection and precise fire localization within the rail stations, which is referred to as Fire-Detect. First, this study created the Fire-Rail dataset using the Unity simulation and collecting internet images, which established the dataset to train the fire detection and precise localization algorithms. Then, a fire detection algorithm was developed to integrate convolutional neural networks, residual structures, and channel attention mechanisms. This algorithm classifies each frame of surveillance video within the station as either "normal" or "suspected fire" status. In the "suspected fire" status, the model activates the precise localization algorithm. It processes the "suspected fire" image along with subsequent frames, providing real-time, detailed fire localization information to station attendants. Experimental results on the Fire-Rail dataset demonstrated a fire detection accuracy of 95.12% on the test set. Furthermore, hierarchical experiments with convolutional neural network layers balance the resource consumption and accuracy. Ablation experiments confirmed the effectiveness of individual components, and robustness experiments indicated the algorithm's ability to handle most noise. The overall model achieves an average fire localization detection accuracy (mAP) of 77.3% and is suitable for deployment in video surveillance equipment within rail transit stations. © 2024 Science Press. All rights reserved.