A Deep Learning CNN Approach Regarding Drone Surveillance in Fire-Fighting Scenarios

This article presents the development of a computer vision module for UAVs engaged in firefighting scenarios. The module features two deep neural networks trained on a customized database, which contains three classes of images: fire, smoke, and person. The aim is to give first responders details such as the number of human victims, their state and their positions, and the type of fuel that keeps the fire going, which helps firefighters to better prioritize their actions in a fire scenario and make the intervention safer for them as well. A faster RCNN and an SSD are used in order to detect these three classes, and the best model is then used to help first responders. The model achieves a precision of 0.58 for 50 IoU, 0.68 for the fire class, 0.68 for the person class, 0.50 for the smoke class because of smoke opacity. Even though the confidence score was high in detections, having false detections, especially with the smoke class, made a low precision overall.