Evaluation of Neural Network Architectures for Classification of Sonar Echoes in Air

Abstract

In this contribution, we will classify sonar targets in air. The gained semantic environment information may be beneficial for autonomous mobile systems, such as robots and transport systems, for navigation in unknown as well as challenging settings with little or ambiguous optical and electromagnetic features. Examples for aforementioned environments may include food processing plants as well as medical buildings, in which obstacles are often comprised of transparent plastic or glass and may also have large shiny/reflecting surfaces. Targets are classified into three generic categories (flat, convex, concave), based on echoes from three subsequent recording positions. Multiple artificial neural networks with different architectures are designed, trained and evaluated as classifiers. The networks show promising prediction accuracy and an embedded implementation appears to be feasible. Especially, combinations of capsule networks and long short-term memory networks appear to be promising candidates for high classification performance.