Real-time semi-occluded fire detection and evacuation route generation: Leveraging instance segmentation for damage estimation

Abstract

Disasters, such as building fires, pose substantial hazards to both human existence and ecological integrity. Existing methods sometimes fail to detect fires, especially semi-occluded fires. Simultaneously, there is a lack of methodologies to identify optimal evacuation routes and exit points during fire emergencies. Existing smoke alarms trigger too late, causing significant damage. Current methods also struggle to accurately depict instances of fire damage, requiring manual on-site assessments to estimate the extent of the destruction caused by fires. By using YOLOv8 to identify the fire and the Double Verification algorithm to assess the colour percentages within potential fire regions, the results show the effectiveness of our approach for semi-occluded fire detection. The YOLOv8 model, with a high confidence threshold (>50 %), could not detect the semi-occluded fires. The same YOLOv8 model with the lower threshold (>25 %) had improved detection but increased false positives. Our model, with lower confidence (>25 %) and double verification, outperformed, detecting semi-occluded fires with high precision and fewer false positives. Our model emerges as the effective choice, demonstrating superior accuracy (0.73) and F1 score (0.81) compared to the first two YOLOv8 models. In conclusion, the integration of our double verification algorithm with YOLOv8 notably improves detecting semi-occluded fires, achieving an 81 % detection rate. The safest path algorithm plays a vital role in generating a path that aids in the safe evacuation of individuals by identifying the safest exit routes. Employing instance segmentation techniques enables us to identify objects in a picture vulnerable to fire damage, facilitating a cost estimation process for further application.