Improving Model-Based Visual Self-Localization Using Gaussian Spheres

摘要

A new approach for global self-localization based on a world model and active vision using varying density gaussian spheres is introduced. The method simultaneously improves and extends our previous proposed approach [1] by contemplating the acquired uncertainty in the perception layer. The used appearance-based object recognition components deliver noisy percept subgraphs which are filtered and fused into an ego-centered frame. In subsequent stages, the required vision-tomodel associations are extracted by selecting ego-percept subsets in order to prune and match the corresponding world-model-graph. Ideally, these coupled subgraphs hold necessary information to obtain the model-to-world transformation, i.e. the pose of the robot. However, the estimation of such a pose is not robust due to the uncertainties introduced while recovering euclidean metric from images and the mapping from the camera to the ego-center. In this context, our approach models the uncertainty of the percepts with a radial normal distribution. This setup allows an optimization-solution in a closed-form, which not only obtains the maximal density position depicting the optimal ego-center, but it also ensures a solution even in situations where normal spheres might not meet at all. 1 Model-Based Visual Self-Localization The following resume of our previous approach offers a brief overview of the elements and their mechanisms involved in the vision-based self-localization, in this way a better understanding of the improvements and extensions in the following D. Gonzalez-Aguirre, T. Asfour, R. Dillmann Institute of Computer Science and Engineering, University of Karlsruhe, Haid-und-Neu-Strasse 7, Karlsruhe-Germany, e-mail: {gonzalez,asfour,dillmann}@ira.uka.de E. Bayro-Corrochano CINVESTAV, Av. Cientifica 1145, Zapopan-Jalisco, Mexico, e-mail: edb@gdl.cinvestav.mx