ASPRS 2012 Annual Conference Sacramento, California March 19-23, 2012 AUTOMATIC 3D BUILDING MODEL GENERATION USING A HYBRID APPROACH Eunju Kwak a , Mohannad Al-Durgham b , Ayman Habib a a Dept of Geomatics Engineering, University of Calgary 2500 University Dr. NW, Calgary, AB, T2N 1N4, Canada b Dept of Civil Engineering, University of Toronto 35 St. George Street, Toronto, ON, M5S 1A4, Canada ekwak@ucalgary.ca, mohannad.al.durgham@utoronto.ca, ahabib@ucalgary.ca ABSTRACT Accurate and up-to-date 3D building models are quite valuable for several applications such as city planning, disaster management, and military simulations. As location-based services and personal navigation become more accessible to the public, automated and efficiently generated 3D models are required more urgently than ever. Considering the importance of 3D building models, they still lack economic and reliable techniques for their generation while taking advantage of the available multi-sensory data from single and multiple platforms. The research conducted on 3D building model generation may fall into the following three categories: data sources used (single or multi-source approaches), the processing strategy (data-driven or model-driven), and the amount of user interaction (semiautomatic or fully automatic). The objectives of this research is to propose fully-automatic building generation approach by integrating data-driven and model-driven methods while making use of multiple images and LiDAR datasets. The focus of reconstruction is on complex structures, which comprise a collection of rectangular primitives. The proposed methodology generates building hypotheses and initial-boundaries from LiDAR data (i.e., data-driven method) and this information is used to restrict the search space and to resolve the matching ambiguities in the images’ space (i.e., model-based image fitting). To reduce the number of involved models, rectangular primitives are used and model parameters, height and slopes are determined from LiDAR data. An automatic algorithm for decomposing the initial LiDAR boundaries into several rectangular primitives is introduced. KEYWORDS: LiDAR, Photogrammetry, Digital Building Model, Reconstruction INTRODUCTION Accurate and up-to-date 3D building models are key components in urban planning, disaster mitigation planning and aftermath evaluation, and military simulations. With the availability of current personal navigational technologies such as Google Earth and Microsoft’s Virtual Earth, the accessibility of building models to the general public has increased. Due to the importance of 3D building models, there have been many research activities on the automatic and efficient building model generation. Existing building reconstruction approaches fall into the following categories: the data sources they use (single or multi-source approaches), the processing strategy (data-driven or model-driven), and the amount of user interaction (semiautomatic or fully automatic) (Vosselman and Mass, 2010). One of the common data sources traditionally used has been aerial imagery. While it produces reliable results, an automatic matching is still a challenging task especially over urban area resulting in low degree of automation. Light Detection And Ranging (LiDAR) data, as a popular alternative source of 3D building model generation, eliminate the matching problem with the availability of direct 3D position acquisition. However, regularization of the boundaries and lower horizontal resolution remain as an issue to be addressed. Due to the limitations of single data source, integration of multi-sources for building model generation has been recommended (Rottensteiner et al., 2005; Cheng et al., 2008; Demir et al., 2009; Habib et al., 2010). Therefore, how to integrate the two data sources in a way that their weakness can be compensated effectively is an active issue of current research (Awrangjeb et al. 2010). In terms of processing strategy of the building reconstruction, it can be processed using three different approaches: data-driven method (bottom-up), model-driven method (top-down), or hybrid method (Faig and Widmer, 2000; and Vosselman and Maas, 2010). These approaches differ on how much relevant information regarding buildings is incorporated during each process. Data-driven approaches can model any shape of buildings, but the little knowledge of the models makes its implementation complex leaving the question of how to set the rules
