Semi-automatic building segmentation from 1m resolution aerial images

Abstract

Since the emergence of aerial imagery, extraction of man-made objects from imagery has been one of the main issues in Computer Vision and buildings have been main targets among man-made objects. Furthermore, if 1-m resolution spaceborne imagery is available in the near future, building detection from such imagery will be more important in digital mapping and remote sensing. However, it has been addressed that most of the proposed building detection systems are for high resolution(sub-meter) aerial imagery and they need modification if they are to be applied to 1 m resolution imagery. In this thesis the problems of conventional building detection systems are discussed and the development of a new building detection system using balloons is described. In the proposed system, 1) building boundaries are extracted by balloons with human interactions incorporated, and 2) corners are detected from the boundaries to acquire 2-D building descriptions. Most buildings can be efficiently expressed with polygonal models using theses detected corners. After the initial establishment of semi-automatic building detection system using balloons, the system is improved by using line information which makes balloons apt to extract building boundaries and building hypotheses which help to extract undetected buildings by the initial system. For this purpose an enhanced line extraction algorithm for 1-m resolution imagery which finds globally optimal lines by planar fitting is developed. Qualitative evaluations of the proposed system are performed and the results support that the system can be used for operational building segmentation from 1-m resolution imagery. The proposed method is new and timely as 1-m resolution spaceborne imagery will be available in the near future.