Using deep learning for sonar targets localization

International audienceThis paper addresses the problem of target localization in sonar signal in a 2D (range-azimuth) scene. The aim is to propose an approach based on an artificial neural network that outputs a binary occupancy grid. A dataset is generated using a sonar simulator and used to train and validate a deep neural network based on a U-net architecture. A pre-processing chain converts analog data to a form that can be passed through the neural network, in this case a (range-azimuth) 2D map with power received. Finally, the performances of the network are compared to those of an approach built around on a CFAR-based range estimation and a MUSIC-based direction of arrival estimation. The results show that the network is able to provide at least similar performances than the reference approach, without the algorithmic calibration currently required by the latter