Classification of sonar data for a mobile robot using neural networks

We study an innovative architecture of an ultrasonic sensor, in conjunction with a neural network-based classification algorithm, in order to recognize some geometric obstacles encountered by a mobile robot. The ultrasonic sensor is made of the association of an array of ultrasonic transducers, building an acoustic antenna, and providing acoustic scans with a fine resolution. The neural network is a multilayer perceptron which was trained with a set of features extracted from the sonar data. Results show that, by selecting appropriate features, the network can be trained to classify some geometric shapes, like wall corners and edges.