三维激光扫描地图生成

Using Laser Range Data for 3D SLAM in Outdoor EnvironmentsDavid M. Cole and Paul M. Newman Oxford University Robotics Research Group Department of Engineering Science University of Oxford Parks Road, Oxford, OX1 3PJ, UK dmc, pnewmanrobots.ox.ac.ukAbstractTraditional Simultaneous Localization and Map- ping (SLAM) algorithms have been used to great effect inflat, indoor environments such as corridors and offices. We demonstrate that with a few augmentations, existing 2D SLAM technology can be extended to perform full 3D SLAM in less benign, outdoor, undulating environments. In particular, we willuse data acquired with a 3D laser range finder. We use a simple segmentation algorithm to separate the data stream into distinct point clouds, each referenced to a vehicle position. The SLAM technique we then adopt inherits much from 2D Delayed State (or scan-matching) SLAM in that the state vector is an ever growing stack of past vehicle positions and inter-scan registrations are used to form measurements between them. The registration algorithm used is a novel combination of previous techniques carefully balancing the need for maximally wide convergence basins, robustness and speed. In addition, we introduce a novelpost-registration classification technique to detect matches which have converged to incorrect local minima.Index TermsMobile Robotics, Outdoor 3D SLAM, 3D Laser Data, Delayed State EKF, Point Cloud SegmentationI. INTRODUCTION ANDPREVIOUSWORKThe SLAM problem has been thoroughly researched theo- retically, and has been demonstrated many times on mobilerobots in flat, 2D environments. Some successful implemen-tations working in corridors and offices include those in 1 and 2. Recently, some researchers have looked towards performing SLAM on mobile robots in fully 3D, outdoor environments, including work in 3 and 4. Perhaps the most successful work to date has been in 5 and 6, which present very compelling results. However, unlike this work they do not use a probabilistic framework which if adopted can offer principled behavior in terms of error distribution when loop closing. In this paper, we extend some of the methods used in 2D environments, and take advantage of their desirable qualities. We also offer several contributions to enable the 2D to 3D extension, which combine to form a system for SLAM in 3D,outdoor, non-flat terrain. At present, we use laser data acquired with a custombuilt 3D laser range finder, along with odometry. As the vehicle moves, we divide this data into 3D point clouds, eachThis work is supported by EPSRC Grant #GR/S62215/01Fig. 1.A 3D scanning laser range finder mounted on a research vehicle.referenced to a vehicle pose. This is achieved using a relatively straightforward segmentation algorithm, avoiding entirely the need to periodically stop and take data. As vehicle poses with attached 3D point clouds are formed, odometry provides dead- reckoned transformations between them. These are then used in augmenting a Delayed State Extended Kalman Filter (EKF) with new vehicle poses. Consecutive point clouds are registered together, or scan- matched (with the odometry derived transformation as aninitial estimate) using our modified registration algorithm to provide additional observations between poses. This processcan include invoking a classification technique based on a scan matchs nearest neighbor statistics to detect any matches that have converged to incorrect local minima. Whilst such a strategy requires supervised training, it proves to be a useful technique for identifying poor registrations which could otherwise progress undetected. Using existing 2D SLAM techniques has several distinct advantages, predominantly due to the probabilistic nature of the Delayed State EKF. Firstly, maintaining a state vector of poses and corresponding pose uncertainties can be used to detect potential loop closures; if for example a past vehicleProceedings of the 2006 IEEE International Conference on Robotics and Automation Orlando, Florida - May 20060-7803-9505-0/06/$20.00 ©2006 IEEE1556Fig. 2.The black line represents a vehicle trajectory, along which 3D laser data was taken (each scan with a different elevation angle). The grey dashes separate periodic potential point cloud regions of odometry. If a change in angle threshold is not broken along a potential point cloud region, OR is only broken after a point cloud has a certain number of points in it, a point cloud from that region is accepted. The point cloud generated is described relative to the vehicle pose at the beginning of the region.pose enters the current uncertainty bounds. Secondly, if aloop closure is detected, and we are able to find an accurate registration derived transformation between the two poses concerned (an observation), the Delayed State Architecture means that the entire state can be updated with a single observation. The errors will then be redistributed around the loop probabilistically. Additi