Fire Detection and Spatial Localization Approach for Autonomous Suppression Systems Based on Artificial Intelligence

The development of autonomous fire suppression systems is widely studied these days to ensure human safety during fire disasters. The evolution in convolutional neural networks (CNNs) has placed substantial emphasis on fire detection through images. Compared to other image processing methods, CNN architectures have proved their efficiency in very accurately detecting fires. However, delayed detection and computational complexity due to manual feature extraction are still concerns for researchers. Therefore, in this research, an autonomous 3D fire location prediction system is presented using YOLOv4 as a fire detector and a self-stereo vision camera setup to locate the position of the fire in real-world coordinates. The single-stage strategy of YOLOv4 has exceptionally increased the detection speed. The results show that YOLOv4 has achieved the fastest ever speed of 30 frames per second in very accurately detecting fires. Obtained by the YOLOv4 model, the pixel coordinates from the center of the bottom line of the bounding boxes around a fire are used in a self-stereo vision camera system to predict the 3D real-world coordinates of the source fire. The study reveals that the proposed system is economical and sufficiently efficient to be used as an early fire detector and location predictor in common surveillance systems.