Automated plane detection and extraction from airborne laser scanning data of dense urban areas

Popular science Digital 3D city models are nowadays used in many GIS applications. Airborne laser scanning is rapidly developing technology that can efficiently plot the 3D shapes of everything on the earth surface. Today, organizations that are responsible for, e.g. urban planning, railway and power line planning, apply such kind of the laser scanning data. For example they use this technique to evaluate if the designed building matches the surrounding landscape or if the trees are too close to the power line. However, those applications normally require that the raw ‘clouds’ of 3D laser scanning point are classified and modeled accordingly. Manual digitization of 3D buildings is rather tedious. In this study, a computer based method for automated extraction and estimation of building roof facets from urban laser scanning data are devised based on available methods and ideas. The method consists of three steps. In the first step, the ground and non-ground points are automatically classified using a method considering neighboring influence, which can effectively remove the points of cars and low shrubs from building points. In the second step, the roof points are extracted from the non-ground points by removing the points of vegetation and other unwanted points using a plane detection technique. And finally, the 3D polygons (or actual lines) of roof facets can be automatically generated from the roof points. In this method, urban laser scanning data is the only necessary input. It does not need any other auxiliary geographical data. This method can be used as a preliminary step of automatic creating of 3D city models. In this sense, the successive step is the assembly of those facet polygons into the complete buildings. In addition, the method can also provide a human-computer interaction method of manually drawing the digital 3D buildings. For example, users can obtain the roof facet polygon by clicking one of the points on the roof facet.Digital 3D city models are used in many GIS applications, such as urban planning. Manual digitization of 3D buildings is rather tedious, hence automated approaches are much preferred. In this study, a pipeline for automated extraction and estimation of building roof facets from LIDAR data are devised based on available methods and ideas, which is part of the entire frame work of automatic creating of 3D city models. In this pipeline, the ground and non-ground points are separated using a graph cuts based method combined with EM algorithm only from elevation data. The tree points and other undesirable clutters are further excluded using the criteria derived from principle component analysis. Finally, the model parameters of roof facets are estimated with a RANSAC based method. Extensive experiments have been conducted in the different steps of the pipeline. The splitting of data into subtiles of 25m by 25m can effectively reduce the influence of slightly varied topography of the test region. With a tuned spatial regularization parameter, graph cuts method can result in a more smooth classification, with some cars and low shrubs separated from buildings. The quality of extraction of roof points depends on both size of the neighboring radius and the choice of the threshold for the criteria. The kD-tree structure is used for searching neighboring points more efficiently. For the linear threshold classifier, 1m is an ideal radius size for filtering off non-plane points yet preventing the missing of smaller roof facets. In contrast, when applying graph cuts based classifier in detecting planes, even 2m radius does not result in missing points on the joint part between roof facets. The modified RANSAC algorithm can estimate each planar roof facet model from the data containing many potential models robustly and efficiently, however it cannot separate the facets that are fully coplanar. In this method, LIDAR data is the only necessary input. It does not need any other geographical data such as vector data or DEM