Home // International Journal On Advances in Systems and Measurements, volume 11, numbers 3 and 4, 2018 // View article

The Application of a Radial Basis Function Network to Supervised Terrain Classification

Authors:
Tiny Du Toit
Hennie Kruger

Keywords: classification; inertial measurement unit; MLP; RBFN; sensor; terrain classification.

Abstract:
In this paper, inertial contact sensor-based terrain classification is performed with a Radial basis function network. Compared to the more popular Multilayer perceptrons, Radial basis function networks are also intelligent techniques and universal approximators, but with a much simpler structure and shorter training time. It has been shown that Radial basis function networks are efficient classifiers and, consequently may be used for terrain classification. For the experiments, a mobile robot platform recorded vibration training data with an inertial measurement unit while traversing five different terrains: asphalt, carpet, dirt, paving, and tiles. The composition of these terrains induces specific vibrations in the mobile platform, which are measured by the inertial measurement unit. The vibration signatures comprise the mobile robot’s linear acceleration, orientation, and the earth’s magnetic field. In contrast to most terrain classification techniques found in literature, no pre-processing of the data is performed. This reduces the computational overhead needed for real-time classification. A Radial basis function network is then trained using a hybrid conjugate gradient descent method and k-fold cross-validation. Identification of the terrain is performed in real time. The classification capability is empirically compared to that obtained by a Multilayer perceptron, a Naïve Bayes method and a Support Vector Machine, which have also been successfully applied to terrain classification in literature. It was found that the Radial basis function network outperformed the Support Vector Machine and Naïve Bayes techniques by a relatively large margin. The Multilayer perceptron, although performing slightly better than the Radial basis function network, has some disadvantages compared to the Radial basis function network. Consequently, the Radial basis function network, with no pre-processing of the input data, may be used successfully as an alternative contact sensor-based terrain classification method.

Pages: 396 to 406

Copyright: Copyright (c) to authors, 2018. Used with permission.

Publication date: December 30, 2018

Published in: journal

ISSN: 1942-261x