APPEARANCE BASED ROBOT LOCALIZATION USING REGRESSION MODELS

Abstract

This paper proposes an appearance-based approach for mobile robot self-localization using least square regression models. In this approach we try to build a simple model that maps the observed images onto their locations directly, and use it for localization scheme. In the approach, we first obtain a pair of training data set from the environment, i.e., images and their corresponding locations. The subspace that describes relation between them can be derived from linear ridge regression. In localization, the coordinates that coincide with currently observed image is predicted in this subspace. The kernel method is also introduced to cope with the non-linearity of training data set. We have examined our method using experimental platform and indoor environment. Through the experiments, 3D localization performance and disturbance tolerance of our regression models is compared with that of conventional subspace method. The results shows that our method is robust against illumination change, additional noise and partial occlusion.