Automatic 3D Building Model Generation by Integrating LiDAR and Aerial Images Using a Hybrid Approach

Abstract

The development of sensor technologies and the increase in user requirements have resulted in many different approaches for efficient building model generation. Three-dimensional building models are important in various applications, such as disaster management and urban planning. Despite this importance, generation of these models lacks economical and reliable techniques which take advantage of the available multi-sensory data from single and multiple platforms. Therefore, this research develops a framework for fully-automated building model generation by integrating data-driven and model-driven methods as well as exploiting the advantages of images and LiDAR datasets. The building model generation starts by employing LiDAR data for building detection and approximate boundary determination. The generated building boundaries are then integrated into a model-based image processing strategy, because LiDAR derived planes show irregular boundaries due to the nature of LiDAR point acquisition. The focus of the research is generating models for the buildings with right-angled-corners, which can be described with a collection of rectangles (e.g., L-shape, T-shape, U-shape, gable roofs, and more complex building shapes which are combinations of the aforementioned shapes), under the assumption that the majority of the buildings in urban areas belong to this category. Therefore, by applying the Minimum Bounding Rectangle (MBR) algorithm recursively, the LiDAR boundaries are decomposed into sets of rectangles for further processing. At the same time the quality of the MBRs are examined to verify that the buildings, from which the boundaries are generated, are buildings with right-angled-corners. These rectangles are preliminary model primitives. The parameters that define the model primitives are adjusted using detected edges in the imagery through the least-squares adjustment procedure, i.e., model-based image fitting. The level of detail in the final Digital Building Model is based on the number of recursions during the MBR processing, which in turn are determined by the LiDAR point density. The model-based image fitting refines the search space and resolves the matching ambiguities in multiple images, which results in higher quality boundaries. This research thus develops an approach which not only automates the building model generation, but also improves the accuracy of the building model itself.