Target Geometry Estimation Using Deep Neural Networks in Sonar Sensing

Abstract: Accurate imaging of target shape is a crucial aspect of wideband FM biosonar in echolocating bats, for which we have developed new algorithms that provide a solution for the shape of complicated targets in the computational domain. We use recurrent neural networks and convolutional neural networks to determine the number of glints (i.e., major reflecting surfaces) making up the target's structure and the distances between the glints (target shape in sonar). Echoes are dechirped relative to broadcasts, and the dechirped spectrograms are scanned in short time segments to find local spectral ripple patterns arising from different interglint delay separations. By proceeding in successive time-window slices, we mimic time-frequency neural processing in the bat's auditory system as a novel means of real-time target discrimination for sonar sensing in robotics.