Volume IV-4/W4
ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci., IV-4/W4, 89-96, 2017
https://doi.org/10.5194/isprs-annals-IV-4-W4-89-2017
© Author(s) 2017. This work is distributed under
the Creative Commons Attribution 4.0 License.
ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci., IV-4/W4, 89-96, 2017
https://doi.org/10.5194/isprs-annals-IV-4-W4-89-2017
© Author(s) 2017. This work is distributed under
the Creative Commons Attribution 4.0 License.

  13 Nov 2017

13 Nov 2017

A NEW MASK FOR AUTOMATIC BUILDING DETECTION FROM HIGH DENSITY POINT CLOUD DATA AND MULTISPECTRAL IMAGERY

M. Awrangjeb and F. U. Siddiqui M. Awrangjeb and F. U. Siddiqui
  • Institute for Integrated and Intelligent Systems, Griffith University, Nathan, QLD 4111, Australia

Keywords: Building extraction, building mask, building detection, point cloud data, LIDAR, multispectral imagery

Abstract. In complex urban and residential areas, there are buildings which are not only connected with and/or close to one another but also partially occluded by their surrounding vegetation. Moreover, there may be buildings whose roofs are made of transparent materials. In transparent buildings, there are point returns from both the ground (or materials inside the buildings) and the rooftop. These issues confuse the previously proposed building masks which are generated from either ground points or non-ground points. The normalised digital surface model (nDSM) is generated from the non-ground points and usually it is hard to find individual buildings and trees using the nDSM. In contrast, the primary building mask is produced using the ground points, thereby it misses the transparent rooftops. This paper proposes a new building mask based on the non-ground points. The dominant directions of non-ground lines extracted from the multispectral imagery are estimated. A dummy grid with the target mask resolution is rotated at each dominant direction to obtain the corresponding height values from the non-ground points. Three sub-masks are then generated from the height grid by estimating the gradient function. Two of these sub-masks capture planar surfaces whose height remain constant in along and across the dominant direction, respectively. The third sub-mask contains only the flat surfaces where the height (ideally) remains constant in all directions. All the sub-masks generated in all estimated dominant directions are combined to produce the candidate building mask. Although the application of the gradient function helps in removal of most of the vegetation, the final building mask is obtained through removal of planar vegetation, if any, and tiny isolated false candidates. Experimental results on three Australian data sets show that the proposed method can successfully remove vegetation, thereby separate buildings from occluding vegetation and detect buildings with transparent roof materials. While compared to existing building detection techniques, the proposed technique offers higher objectbased completeness, correctness and quality, specially in complex scenes with aforementioned issues. It is not only capable of detecting transparent buildings, but also small garden sheds which are sometimes as small as 5 m2 in area.