Deep Learning for Victims Detection from Virtual and Real Search and Rescue Environments

Robotic interventions in post-disaster environments to carry out search and rescue explorations allow optimizing time for identifying victims and safeguarding the rescuer's integrity. The rising in the neural networks field and their application in image detection algorithms have made it possible to facilitate the early detection of victims in these first phases of exploration. This article analyses the effectiveness of applying neural network models obtained from training with different datasets of both: images captured in real environments and synthetic images from recreated virtual post-disaster environments for detecting victims. For this development, tests have been carried out in environments at the ETSII-Universidad Politecnica de Madrid, generating models from images obtained with the ARTU-R robot (A1 Rescue Task UPM Robot) and have been validated with real search and rescue exercises at the University of Malaga. The main results show that the models obtained from virtual environments apply to real ones.