Disaster Area Recognition from Aerial Images with Complex-Shape Class Detection

This paper presents a convolutional neural network (CNN) model for event detection from Unmanned Aerial Vehicles (UAV) in disaster environments. The model leverages the YOLOv5 network, specifically adapted for aerial images and optimized for detecting Search and Rescue (SAR) related classes for disaster area recognition. These SAR-related classes are person, vehicle, debris, fire, smoke, and flooded areas. Among these, the latter four classes lead to unique challenges due to their lack of discernible edges and/or shapes in aerial imagery, making their accurate detection and performance evaluation metrics particularly intricate. The methodology for the model training involves the adaptation of the pre-trained model for aerial images and its subsequent optimization for SAR scenarios. These stages have been conducted using public datasets, with the required image labeling in the case of SAR-related classes. An analysis of the obtained results demonstrates the model's performance while discussing the intricacies related to complex-shape classes. The model and the SAR datasets are publicly available.