Introduction

Automatic and precise reconstruction of indoor environments is an important task in robotics and architecture. Autonomous mobile robots equipped with a 3D laser range finder are well suited for gaging the 3D data. Due to odometry errors the self localization of the robot is an unprecise measurement and therefore can only be used as a starting point for registration of the 3D scans in a common coordinate system. Furthermore the merging of the views as well as the scanning process itself is noisy and small errors may occur. We overcome these problems by extending the reconstruction process with a new knowledge based approach for the automatic model refinement. Since architectural shapes of environments follow standard conventions arising from tradition or utility (9) we can exploit knowledge for reconstruction of indoor environments. The used knowledge describes general attributes of the domain, i.e., architectural features as plane walls, ceilings and floors. For various domains different knowledge is needed, e.g., for reverse engineering of CAD parts (20). We show that applying general knowledge for recovering specific knowledge improves reverse engineering. In mobile robotics one important task is to learn the environment to fulfill specific jobs. 3D maps are needed for plan execution and obstacle avoidance (23). Volumetric maps, i.e., 3D point clouds are often large and difficult to use directly in control tasks. Therefore some groups have attempted to generate compact flat 3D models (12,15) or compact bounding box models (24). This paper presents algorithms for building compact and precise 3D models and generates a coarse semantic interpretation, thus creates coarse semantic maps. The proposed algorithm consists of three steps: First we extract features, i.e., planes from registered unmeshed range data. The planes are found by an algorithm which is a mixture of the RANSAC (Random Sample Consensus) algorithm and the ICP (Iterative Closest Point) algorithm (5,1). Second the computed planes are labeled based on their relative orientation. A predefined semantic net implementing general knowledge about indoor environments is employed to define these orientations. Finally architectural constraints like parallelism and orthogonality are enforced with respect to the gaged 3D data by numerical methods. The paper is organized as follows. After discussing the state of the art in the following part we present the 3D laser range finder and the autonomous mobile robot. The second section presents the range image registration, followed by a description of the feature extraction algorithm. The algorithms for semantic interpretation of the data is given in section four. In section 5 the model refinement is described. Section 6 summarizes the results and concludes the paper.