BUILDING DETECTION BY DEMPSTER-SHAFER METHOD USING MULTI-VIEW IMAGES

摘要

Building detection is important in updating GIS data layers. The major information contents in building detection are shape and spectrum when images are employed. Multiple image matching may generate 3D point clouds to form the 3D shape outline. Thus, we can obtain normalize digital surface model (nDSM) by excluding digital elevation model (DEM). On the other hand, multi-spectral images provide color information for building detection. The Normalized Difference Vegetation Index (NDVI) can separate vegetation from buildings. Besides, considering the building characteristic, we define a weighted matrix based on the building lines. We assume that around each building line there is a high possibility of building areas. Final, we employ the Dempster_Shafer method to perform classification of different data based on the evidence that each feature provides for each class hypotheses. Hence, these three features (nDSM, NDVI, weight) are used by Dempster-Shafer to detect buildings. The proposed method comprises six major steps: (1) determination of building lines, (2) multiple image matching, (3) generation of nDSM, (4) spectrum analysis, (5) weighted matrix formation, and (6) Dempster-Shafer classification. Extracting features by Canny operator and deciding the building lines by Hough transform from the reference image is the first step. Second, geometrically constrained cross-correlation algorithm and Central-Left-Right matching are selected in the multiple image matching to generate 3D point clouds. Third, the nDSM is derived by excluding terrain evaluation. Fourth, NDVI is obtained by spectrum analysis. Fifth, we form a weighted matrix based on the building lines. Finally, we use these three features (nDSM, NDVI, weight) to detect buildings by using Dempster-Shafer method. The test datasets include six aerial images with 9cm spatial resolution and the overlap and sidelap are both 60%. The experimental results show that this research has the high ability to locate building areas.