|
|
|
3D Mapping in Unstructured EnvironmentsBackgroundThe project with the full title "Learning of 3-Dimensional Maps of Unstructured Environments on a Mobile Robot" was funded by the Deutsche Forschung Gemeinschaft (DFG). It deals with the generation of 3D maps with a mobile robot in unstructured environments like search and rescue scenarios. We are now trying to carry the approach over to the underwater domain in the EU project Cooperative Cognitive Control for Autonomous Underwater Vehicles (Co3-AUVs). The DFG project lead to a very fast and robust form of mapping approach in form of 3D Plane SLAM that has been successfully used in different application scenarios. The approach as well as results from various field experiments are presented below. 3D Plane SLAMThe mapping approach developed within this project consists of several
steps
The different steps are also illustrated in the image on the right (click to enlarge). We are using different 3D range sensors for the first step including time of flight (TOF) cameras like PMD camera and Swissranger, stereo cameras, and self-made actuated laser range finders (aLRF). Some results within the project deal with the preprocessing of these data sources, e.g., to achieve sub-pixel accurary in TOF cameras [1][2][3]. But most results are based on raw data from an aLRF in form of a SICK S300 that is rotation in a pitching motion by a simple servo. The main steps of the approach are short explained in the following. Plane ExtractionAfter a scan is acquired, a set of planar-patches is extracted from the point-cloud that is associated with the related robot frame. Apart from the planar patch's normal and distance to the origin, i.e., its defining parameters, the extraction procedure also gives their 4 x 4 covariance matrix. The plane extraction exploits the fact that a single 3D scan is based on 2.5D range measurements, the data can hence be treated as range images where an optimal fit in linear time is possible. In addition to the planes, their boundaries are extracted. The related polygons are not needed for the SLAM approach itself, the registration operates on the basic plane parameters. But the planes with their boundaries are the basis for the visualization and the actual usage of the 3D map. A detailed description of the plane extraction can be found in [4]; the formal basis and the computation of the proper uncertainties is described in [5]. Plane Based RegistrationThe plane based registration is the core of 3D Plane SLAM. It works by
The core of the algorithm and a formal derivation of its principles are introduced in a publication that is currently under review; please revisit this page occasionally, we will put an update here as soon as this submission is accepted. There, results in an indoor environment and a detailed comparison to ICP and NDT are also presented. The plane registration has several main advantages. First of all, it needs unlike any other approach to 3D SLAM no motion estimates from any other source like odometry, inertial navigation system, or similar. It hence still works under circumstances where all other known approaches fail. Second, it is very fast and significantly outperforms other approaches in direct comparisons with respect to computation time. As mentioned, the related detailed publication is currently under review. The experimental results presented below nevertheless already give an impression of the performance of the algorithm. Pose Graph SLAMThe plane registration presented in the previous section is in principle already suited for generating 3D maps by merging scans together. But though the errors are relatively small, they nevertheless accumulate and are hence in principle unbounded. It is therefore of interest to use proper Simultaneous Localization and Mapping (SLAM) for which the plane registration is fortunately particularly well suited for two main reasons. First of all, its theoretical basis allows for a proper determination of the uncertainties in the registration process. Second, theoretical analysis as well as practical experiment suggest that it is particularly robust with respect to rotations. The relaxation as result of loop closing in the SLAM can hence concentrate on translational errors for which an extremely fast closed form solution is possible. A detailed description of the pose graph embedding of the plane registration and the related SLAM algorithm can also be found in a publication that is currently under review. Again, please revisit this page occasionally, an update will be provided as soon as the submission is accepted. But below are some experimental results presented that give a first impression of the performance of the approach. Experiments and Results3D Plane SLAM, respectively core elements of it like the plane registration have already been very successfully applied in different application scenarios. Here an overview of an example in a disaster scenario, followed by several links to more detailed descriptions of the results including pictures and videos. Disaster City, NIST Response Robot Evaluation Exercise (RREE), 2008In the Collapsed Car Park experiment a 3D map is generated with:
Pictures from the mapped site:
Final 3D Map:
A second example from Disaster City is a map generated with full 3D-Plane-SLAM at the dwelling site, a scenario simulating a flooding disaster. This scenario was covered with 70 scans, which were turned into a 3D map without using odometry or similar motion estimates but only with plane based registration and relaxation. An accelerated movie from the robot's front camera gives an impression of the scenario. Below some pictures from the site and of the 3D map are shown.
front view: perspective views:
views inside the building:
Links to 3D Datasets and Experimental Results
ReferencesKaustubh Pathak,
Andreas Birk, Narunas Vaskevicius, and Jann Poppinga,
|
||||||||||||||||||||||||
|
|
|||||||||||||||||||||||||
