Neural Network-Based Detection of Ultrasonic Targets with Respect to Noise and Number of Sampling Positions

A neural network-based approach for detection of sonar targets in air is presented in this contribution. Our approach may facilitate autonomous mobile systems to reliably detect and classify objects in their surrounding by using sonar information. This task might be extremely important in changing as well as unorganized environments. We perform target iden-tification with long short-term memory networks as classifiers. Such are capable of dealing with variable numbers of echoes from multiple positions per input sequence, which facilitates more flexible operation. The impact of the number of recording positions per sequence and of noise is investigated. Furthermore, we demonstrate the improvement in classification performance in comparison to previously obtained results from multi-layer-perceptrons.