Convolutional neural networks for robot vision: numerical studies and implementation on a sewer robot

Convolutional neural networks (CNNs) impose constraints on the weights, and the connectivity of a standard neural network, providing a framework well suited to the processing of spatially or temporally distributed data. Although CNNs have been applied to face and character recognition, they have still received relatively little attention. The present paper applies the CNN architecture to an artificial test problem and to an application in robot vision. Autonomous sewer robots must navigate independently the sewer pipe system using information from sensors. One required component is robust detection of pipe joints and inlets using data from the omnidirectional sensor. A simple CNN is shown to robustly classify 32x32 pixel normalized video frame data on a limited validation set. The study indicates that machine learning methods for robot vision are feasible in terms of classification accuracy and online implementation.