Automatic reconstruction of 3D building model using shape knowledge

PhD project

3D Building reconstruction from point clouds: building laser points (left), 3D building models (right).

This research is to find an automatic way to reconstruct 3D architecture models from point clouds acquired by airborne laser scan technology or image dense matching technology. Though building reconstruction techniques have been dramatically improved recently, there are still many problems preventing it from comprehensive utilization. Those problems, including data insufficiency, complex building shapes and complex boundary, stem from the characters of point cloud as well as the complexity of building and environment.

Challenges in 3D building reconstruction: data insufficiency (1st figure), complex building shapes (2nd figure) and complex boundaries (3rd and 4th figure).

This research will focus on representation, learning, and utilization of knowledge on building shapes for 3D building model acquirement. Here building knowledge can be any regularity, pattern or repetition of building shapes. This research will first try to find the grouping type and dependences between geometric elements and regularities behind building structure. Building knowledge is searched by matching comparing topology graph and be used in surface refinement and surface hypotheses.

Work flow of building reconstruction.

The near future research is to statistic building type information in given city and area, and to construct mutual relationship between object models (from manually collection) and data models (from point clouds). A Gaussian network for the mutual relationship will be constructed and used to find best matched target model for each building in framework of global optimization.