Long-distance Terrain Labeling Based on Bayesian Kernel Principal Component Analysis

Abstract

When adopting predetermined appearance features in autonomous navigation of an outdoor mobile robot,terrain labeling may not be effectively done because of the dynamic uncertainty and the mapping deviation of appearance feature distribution in unstructured outdoor scenes.To solve this problem,a long-distance terrain labeling method based on Bayesian kernel principal component analysis(BKPCA) is proposed.The method incorporates the posterior probability of cluster centers using Bayesian formula and employs a self-defined kernel function.The method can implement the maintenance of the original data structure in the low dimensional space and can extract suitable appearance features for current scene labeling. The experimental result validates that the BKPCA model improves accuracy of long-distance terrain labeling.