Simultaneous Localization and Map Building by Integrating a Cache of Features

Abstract

Localization is an important problem in autonomous mobile robots navigation. To solve this problem, robots must also be able to learn, maintain and update models of their environments. This paper describes a full implementation of a simultaneous localization and map building (SLAM) method. SLAM is the problem of an autonomous vehicle starting at an unknown position which then incrementally builds a world map and estimates the robot absolute pose according to the map. An extended Kalman filter (EKF) is used for estimation and data fusion. For perception, the method combines an adaptive break point detector, first and second order analysis, and the RANSAC algorithm for robust fitting of laser scan data in order to extract a model composed of line segments and their uncertainty. A dynamic cache is proposed and introduced in the world model in order to speedup the map search in the measurement prediction and feature matching phases of SLAM. Experimental results of simulation and real-robot experiments with a Nomad 200 are presented demonstrating the effectiveness of the SLAM methods and improvements attained with the cache of feature method.