Yong Xiao

Center for Earth Observation and Digital Earth,Chinese Academy of Sciences

China

Description of my method

The building models are reconstructed thorough a data-driven method. The method is composed of three major steps: building detection, roof segmentation and model generation.


1.     Building Detection:  After ground points are removed by the progressive TIN filter algorithm, building points are extracted by performing connected component analysis.

2.     Roof Segmentation: For each building, building roofs are separated by a new segmentation method based on the Expectation-Maximization algorithm. (We are now planning to utilize this dataset to evaluate the performance of this new segmentation method.)

3.     Model Generation: 3D building models are reconstructed thorough a data-driven method. The building outlines are acquired by a raster-based approach. Firstly, roof points of each building are rasterized. (If a ridge is detected, the points are rotated according to the angle between the ridge and X axis.)  Then building boundary points are extracted by the 2D alpha shape method and long segments are identified by the RANSAC algorithm.  Finally, after the building outline loop is obtained by connecting these long segments, parallel segments are merged and some segments are adjusted to make them orthogonal to each other. Meanwhile, roof intersection segments are computed based on roof plane equations. At last, building models are derived by combining regularized building outlines and roof intersection segments.