Previous Abstract Return to Session D1 Next Abstract

Session D1: Robotic and Indoor Navigation

Laser-Camera Based 3D Reconstruction of Indoor Environments
Jamal Atman, Institute of Systems Optimization (ITE), Karlsruhe Institute of Technology (KIT), Germany; Gert F. Trommer, ITE, KIT, Germany and ITMO University, Russia
Location: Spyglass

Obtaining relevant information as soon and as detailed as possible is a crucial issue in a rescue scenario. Hence, micro air vehicles (MAVs) are becoming more and more attractive due to their flexibility in terms of the place of use and their perception abilities using various sensors. The MAV is able to quickly reach locations which are hard to access by pedestrians or ground vehicles. After exploring the area, the sensor data can be processed, such that the operator is able to easily interpret the situation. Thus, this paper deals with data processing of particularly a 2D laser rangefinder and a monocular camera in order to reconstruct the explored indoor environment.
Our previous publications have shown that a deep integration of laser rangefinder and camera information results in accurate navigation solution. Now, based on a very accurate navigation solution at all times an accurate map of the unknown area is generated. This is done by using a minimal setup of sensors. The laser rangefinder is mainly used to provide the depth information in order to regain the geometry of the MAV’s environment and the camera provides the realistic texture. Combining both sensors and being able to fuse both sensor information by knowing their exact relative pose, it is possible to reconstruct the explored area to a three-dimensional map with realistic texture. This resulting map simplifies the planning of the ongoing mission.
The first step in order to build such maps is to generate a laser point cloud by accumulating the laser data over time. This means that the relative measurements are assigned to a global frame. Even though the laser rangefinder scans only in a plane, the resulting point cloud is three-dimensional due to the MAV’S motion in six-degree-of-freedom. However, not all parts of a certain environment, especially in vertical direction, is necessarily scanned. Therefore, images are obtained in order to enrich the spare laser rangefinder information.
In order to achieve this goal a second step, which comprises a segmentation of the generated point cloud, is necessary. As vertical and horizontal planes are dominant in indoor environments, it is checked if points represent such planes. In this context, it is shown how to apply plane fitting methods to a local area in the point cloud. The local selection reduces the outlier ratio and thus also the robustness of plane detection. This assures geometric accuracy. Moreover, it is described how to combine neighboring planes with similar model representation. The resulting planes allow for a more comprehensive geometric representation.
Based on the segmentation, the texturing of the 3D model is presented. It is described how to calculate the homography analytically, which establishes a relationship between image and world plane. Furthermore, the assignment of the relevant texture of the image to the corresponding plane is described. Using the camera information the height of the planes can be extracted, so that the sparse information of the point cloud and even the lack of the laser’s depth information is enriched and augmented, respectively.
Experiments with real data show that the methodology of combining navigation solution, 2D laser rangefinder, and camera information is very promising. In this context, MAV’s sensor data were recorded in a typical indoor scene consisting of corridor and offices. The processed map is compared to an existing floor plan, which indicates its geometric accuracy. It is shown that in presence of challenging camera perspective floor, ceiling, and walls are textured in a high grade. As a result, it can be concluded that indoor environments can be comprehensively represented, such that the operator can easily interpret the situation by using the processed raw data.



Previous Abstract Return to Session D1 Next Abstract