Classification of ultrasonic signals using multistage neural networks

This paper describes an architecture and training method for a novel neural network which is designed to classify reflector shapes from ultrasonic echoes for use in mobile robot navigation. Experimental results are presented for the classification of 10 different shapes of reflectors at 21 distinct ranges and azimuths within the active volume of a linear bistatic array. Each stage of the network can include up to 11 separate neural networks, each of which is designed to classify either a reflector shape or the case where no reflector is present. Later stages of the network take as inputs the outputs from all previous stages as well as the features themselves. Although feature selection is introduced to restrict the number of network inputs, as successive stages are added to the network, it is apparent that a design decision must be taken which determines a suitable compromise between network complexity and improved performance. Both multilayer perceptron (MLP) and radial-basis function (RBF) versions of the multi-stage network are proposed and, to assess their performance, comparison are made with results obtained from standard MLP neural networks and nearest neighbour classifiers. In comparison with the standard MLP neural network, the results show an improvement in the number of reflector types which can be correctly classified, and demonstrate that the training time can be reduced. The results also compare the classification performance of the nearest neighbour approach whose training time is shorter, but which is likely to be unsuitable in applications where real-time classification is required.