Building Edge Detection Using Small-footprint Airborne Full-waveform Lidar Data
The full-waveform lidar technology allows a complete access to the information related to the emitted and backscattered laser signals. Although most of the common applications of full-waveform lidar are for forest studies, some recent works have demonstrated the relevance of such methods in urban areas, mainly for land-cover classification. As a basis for many applications such as strip registration, classification, building detection, building database change detection or updating, we aim to focus on building edge detection using airborne and small-footprint (0.8 m) datasets. To our knowledge, such subject has never been investigated on real data.
The goal of our work is first to coarsely extract the edges of buildings, and then to use a physical model based on the lidar equation to retrieve a more accurate position (1) of the estimated segments than the size of the footprint lidar and (2) of the echoes that can be extracted from the waveforms.
The proposed algorithm is divided into five parts.
First, since we directly work on a full dense urban area (ca. 3.5 km?), a focusing step on areas with building edges is carried out. This allows to approximately retrieve where buildings are located. This step is not necessary but improves the algorithm efficiency since this permits to remove large vegetated areas. Satisfactory results are achieved since no buildings are omitted and each building is at least in one focus area.
Secondly, for each focus area, waveforms lying on building edges are segmented. This allows to coarsely estimate the gutter height. Then, all the 3D points retrieved from all waveforms of the area (using a standard Gaussian decomposition) and whose altitude is close to the estimated gutter height, are labelled as ??border points??. In case of multiple building in one region of interest, this process is iteratively performed, starting from the highest building.
Thirdly, segments are extracted from these border points using the robust Ransac algorithm. The Ransac procedure is improved using slope and length criteria in order to improve the merge of aligned but distinct building outlines.
Then, since a part of the extracted segments may lie on vegetated areas, such false detections are filtered using physical characteristics of the lidar backscatter over vegetation: multiple echoes resulting in a scattered 3D point as well as a particular spred of the laser pulse.
Finally, the remaining segments are qualified and their location is improved by inverting a physical model based on the lidar equation. For each ??border?? waveform, the accurate edge location within the footprint is retrieved by comparing the backscattered laser power on the roof with the laser power that we should received if all the footprint is on the roof. Then, new segments with adjusted 3D points are computed. Accuracy is improved up to 20 cm.
The proposed method provide very satisfactory results, and thus reliable primitives for a large body of applications. The main limitation concerns the fine adjustment of the segments using the lidar equation. This step assumes a Lambertian behaviour of the roofs in order to provide a simple computational problem. Our experiments show that this is not always sufficient and may not always lead to better 3D segments.
Keywords : Building, outline, detection, lidar, waveform, primitive, segmentation, qualification.