Biomimetic solutions to finding passageways in foliage

Abstract

Many bats species live in densely vegetated habitats. Hence, they must have evolved the ability to detect passageways through the foliage. To identify echo features that bats have at their disposal to accomplish this sensing task, a biomimetic sonar head was used to ensonify artificial foliages in the laboratory. These foliages consisted of plastic vines that were arranged to create gaps of defined width and height. The conventional sonar approach to detecting gaps between scatterers is based on the drop in energy that occurs when the sonar beam is aimed at a gap. However, the performance of such an energy-detector decrease as the sonar-beam width increases relative to the angle subtended by the gap edges. Hence, reliable detection of a narrow gap with a wide beam and/or from a distance is not feasible with this approach. Here, a machine-learning approach based on convolutional neural networks was employed to identify features in energy-normalized spectrogram that could support passageway-detection independently of echo energy. The results indicate that features other than echo energy exist in the spectrogram that by themselves support much better passageway detection performance than the energy-based reference. Work to understand the features learned by the convolutional neural network is currently underway.