The increasing need for rapid characterization and quantification of complex environments has created challenges for data analysis. This critical need comes from many important areas, including industrial automation, architecture, agriculture, and the construction or maintenance of tunnels and mines. Especially mobile systems with 3D laser scanners that automatically perform multiple steps such as scanning, gaging and autonomous driving have the potential to greatly advance the field of environment sensing. On the other hand, having 3D information available in real-time enables autonomous robots to navigate in unknown environments, e.g., in the field of autonomous mine inspection.
The problem of automatic environment sensing and modeling is complex, because a number of fundamental scientific issues are involved in this research. One issue is the control of an autonomous mobile robot and scanning the environment with a 3D sensor. Another question is how to create a volumetric consistent scene in a common coordinate system from multiple views. The latter problem is addressed here: The proposed algorithms allows to digitize large environments fast and reliably without any intervention and solve the simultaneous localization and mapping (SLAM) problem. Finally robot motion on natural outdoor surfaces has to cope with yaw, pitch and roll angles, turning pose estimation as well as scan matching or registration into a problem in six mathematical dimensions. This paper presents a new solution to the SLAM problem with six degrees of freedom. A fast variant of the iterative closest points (ICP) algorithm registers the 3D scans in a common coordinate system and relocalizes the robot. The computational requirements are reduced by two new methods: First we reduce the 3D data, i.e., we compute depth images that approximate the scanned 3D surface and contain only a small fraction of the 3D point cloud. Second a fast approximation of the closest point for the ICP algorithm is given. These extenstions of the ICP result in an algorithm for generating overall consistent 3D maps using global error minimization.
This paper describes an algorithm for acquiring volumetric maps of underground mines. Mapping underground mines is of enormous societal importance [16], as a lack of accurate maps of inactive, underground mines poses a serious threat to public safety. According to a recent article [3], ``tens of thousands, perhaps even hundreds of thousands, of abandoned mines exist today in the United States and worldwide. Not even the U.S. Bureau of Mines knows the exact number, because federal recording of mining claims was not required until 1976.'' The lack of accurate mine maps frequently causes accidents, such as a recent near-fatal accident in Quecreek, PA [18]. Even when accurate maps exist, they provide information only in 2D, which is usually insufficient to assess the structural soundness of abandoned mines. Accurate 3D models of such abandoned mines would be of great relevance to a number of problems that directly affect the people who live or work near them. One is subsidence: structural shifts can cause collapse on the surface above. Ground water contamination is another problem of great importance, and knowing the location, volume, and condition of an abandoned mine can be highly informative in planning and performing interventions. Accurate volumetric maps are also of great commercial interest. Knowing the volume of the material already removed from a mine is of critical interest when assessing the profitability of re-mining a previously mined mine.
Hazardous operating conditions and difficult access routes suggest that robotic mapping of abandoned mines may be a viable option to traditional manual mapping techniques. The idea of mapping mines with robots is not new. Past research has predominantly focused on acquiring maps for autonomous robot navigation in active mines. For example, Corke and colleagues [8] have built vehicles that acquire and utilize accurate 2D maps of mines. Similarly, Baily [2] reports 2D mapping results of an underground area using advanced mapping techniques. The experiments reported in this paper utilize data by CMU's mine mapping robot Groundhog [10,24], which relies on 2D mapping for explorating and mapping of abandoned mines. While Groundhog and a related bore-hole deployable device ``Ferret'' [16] utilizes local 3D scans for navigation and terrain assessment, none of these techniques integrates multiple 3D scans and generates full volumetric maps of abandoned mines.
The paper is organized as follows: The next two sections describe the state of the art in automatic 3D mapping and present the autonomous mobile robot. Section IV presents the registration algorithms and the solution to the SLAM problem. Furthermore we show how data reduction and access methods speed up computation and the methods become real-time capable. Section V describes the Mathies Mine experiment. Finally, section VI summerizes the results and concludes the paper. The paper is accompanied with a 3D map, given as video. The video is available for download at www.ais.fraunhofer.de/ARC/3D/mine/.