The authors describe an approach to automatic city model construction, by means of a textured surface mesh, from airborne laser scans, and an additional ground-based data acquisition process that covers the building facades. The initial airborne surface mesh is filtered, to remove small objects, like trees or cars, that are impossible to reconstruct correctly at the given scanning resolution. In addition, rooftops are flattened, and the mesh is simplified to reduce the overall polygon count. Texture mapping of airborne aerial photos is done in a semi-automatic manner onto the surface mesh.

To get a detailed facade model of the buildings, the airborne model is complemented by a detailed ground-based model. The ground-based model acquisition is done via two two-dimensional (2D) laser scanners, and a camera mounted on an acquisition vehicle. Both laser scanners face the same side of the street; one is mounted horizontally, and the other vertically. The line-of-sight of the camera corresponds to the intersection between the horizontal and vertical scanning planes. Because of reliability problems with global positioning system (GPS) sensors in urban environments, the authors use a pre-processed 2D aerial image to determine the position of the acquisition vehicle, based on the results from the horizontal scanner, as described in Thrun et al. [1] for mobile robot applications.

Because of their higher resolution, it is reasonable to give preference to the ground-based facades wherever available, and to use the airborne mesh only for roofs and terrain shape. The approach favors photo-realistic appearances for walk- or drive-throughs, rather than the precision of computer-aided design (CAD) approaches.

As demonstrated on the basis of a complex example, the proposed approach works very well for urban areas, where the facades border the road. In this case, the ground data acquisition process is not impaired by trees, or similar objects between the facades and the camera. Consequently, in residential areas with lots of trees, the resulting model may contain artifacts. Future research will address this problem, as well as a fully automatic system that does not require user assistance during texture mapping.