Abstract
This paper proposes an innovative hazard identification and risk assessment mapping model for Urban Search and Rescue (USAR) environments, concentrating on a 3D mapping of the environment and performing grid-level semantic labeling to recognize all hazards types found in the scene and to distinguish their risk severity level. The introduced strategy employs a deep learning model to create semantic segments for hazard objects in 2D images and create semantically annotated point clouds that encapsulate occupancy and semantic annotations such as hazard type and risk severity level. After that, a 3D semantic map that provides situational awareness about the risk in the environment is built using the annotated point cloud. The proposed strategy is evaluated in a realistic simulated indoor environment, and the results show that the system successfully generates a risk assessment map. Further, an open-source package for the proposed approach is provided online for testing and reproducibility.