Multi-Level 3D Building Modeling by Integration of Aerial and Ground View Images

Sung Chun Lee


Abstract

Accurate 3D building models of a city are useful for a variety of applications such as 2D and 3D GIS, fly-through rendering, and simulation for mission planning. Each application may require different aspects of the model. 2D GIS needs the roof boundary of buildings; fly-through rendering demands textures of building facades in addition to the 3D building model. The different levels of information are computed from different data sources. For example, information such as the roof boundary or height can be obtained by using aerial images or LIDAR data. Facade information such as its texture, or detailed 3D structure, can be computed from multiple ground view images.

In this thesis, I attempt to integrate the processes of obtaining the different level knowledge for building models from various data sources by deriving a multi-level representation of a 3D building model. To implement multi-level representations for 3D building models, I exploit the concept of Level Of Detail (LOD) from virtual reality literature to represent the different level of knowledge. The LOD is used for storing a number of representations of an object, where the complexity of each representation is varied so that an approximate model can be selected according to the complexity of the application. I define the multi-level representations of 3D building models as followings: Level 1: Structural information of building, Level 2: Facade texture information, and Level 3: Detailed geometry of building facade.

Using the multi-level building representation, the reconstruction process of more complex level models is aided by simpler level models. For level 1 model (simpler model), we use the knowledge that flat roof building is likely to have the roof parallel to the ground and walls perpendicular to the ground. The representation of level 1 model consists of the 2D roof boundary and its height. To obtain level 2 models, we exploit the knowledge from the level 1 model, namely the 3D information of wireframe building models such as its 3D vertices, boundary lines, and faces. With 3D building models, and automatically or interactively extracted vanishing points, we estimate the pose of the ground view camera to augment facade textures to the 3D building models. We use the knowledge that the orientation of the detailed structures of building facade (level 3 models) is perpendicular to the building facade that contains them. Then, we use the calibrated ground view camera information from level 2 and 3D model information from level 1 to reduce user interactions for creating level 3 models, which capture facade details such as entry ways, recesses, or window structures. The final result is a rich multi-level 3D model useful for a variety of applications depending on their needs.


Maintained by Philippos Mordohai