Terrain Segmentation with On-Line Mixtures of Experts for Autonomous Robot Navigation

Abstract

We describe an on-line machine learning ensemble technique, based on an adaptation of the mixture of experts (ME) model, for predicting terrain in autonomous outdoor robot navigation. Binary linear models, trained on-line on images seen by the robot at different points in time, are added to a model library as the robot navigates. To predict terrain in a given image, each model in the library is applied to feature data from that image, and the models’ predictions are combined according to a single-layer (flat) ME approach. Although these simple linear models have excellent discrimination in their local area in feature space, they do not generalize well to other types of terrain, and must be applied carefully. We use the distribution of training data as the source of the a priori pointwise mixture coefficients that form the soft gating network in the ME model. Single-class Gaussian models are learned during training, then later used to perform density estimation of incoming data points, resulting in pointwise estimates of model applicability. The combined output given by ME thus permits models to abstain from making predictions for certain parts of the image. We show that this method outperforms a less sophisticated, non-local baseline method in a statistically significant evaluation using natural datasets taken from the domain.