Experimental Results

The proposed methods have been tested in several experiments with our autonomous mobile robot in the GMD Robobench. Figure 4 shows an example 3D point cloud (single 3D scan with 184576 points) and the semantic interpretation. The corresponding original and refined model is given in figure 5 (top: Original model, $E(P)$ = 14.57 +
$\gamma$ 173.09, bottom: Refined model, $E(P)$ = 26.68 + $\gamma$ 2.35,
$\gamma$ was set to 100.0). The figure shows the reduction of the jitters at the floor and ceiling (circled). The orientation of the model in the bottom image is transformed along the axis of the coordinate system and the meshing algorithm produces flat walls. The total computation time for the optimization is about one minute (Pentium-IV-2400). An octree-based algorithm (19) generates the mesh (cube width: 5cm) to visualize the differences between the images. Starting from a cuboid surrounding the whole scene the mesh generation recursively divides the scene into 8 smaller cubes. Empty nodes of the resulting octree are pruned.

Figure 6: Example of a compact flat surface model reconstructed by an autonomous mobile robot (eight merged 3D scans). The persons in the scene are filtered out through the plane detection. Top left: Photo of the corridor scene. Top middle and right: Extracted surfaces with their semantic interpretation. Bottom left: Rendered scene with reflectance values. Bottom middle: Unconstrained mesh. Bottom right: Constrained mesh.

The second example in figure example 6 consists of eight merged scans acquired by the autonomous mobile robot driving in the GMD Robobench. The scene consists of a 32 meter corridor connecting 15 offices. Two persons are standing inside at the beginning. Figure 6 top, left shows the 3D data and reflectance values. The next two pictures (top middle and right) show the extracted and labeled planes. The two persons and other non-flat objects, e.g., dynamic objects, are not explained by the semantic net and therefore filtered from the plane model. The door behind the right person becomes visible. Figure 6 bottom shows the original (left) and refined (right) octree model with marked differences. The images contains the silhouette of two persons, because all points not assigned to planes are unchanged and included.