A line-based progressive refinement of 3D rooftop models using airborne LiDAR data with single view imagery

Abstract In recent years, many mega-cities have provided 3D photorealistic virtual models, a digital replica of the geometrical structures of cities, for more effective decision support in public safety, urban planning, and engineering applications. Most research attempts at reconstructing geometric models of cities treat such urban systems as if they are in a static environment. However, cities are dynamic systems that continuously change over time. Accordingly, their virtual representations need to be regularly updated in a timely manner to allow for accurate analysis. The concept of progressive city modelling is to continuously reconstruct city models by accommodating changes recognized in the spatio-temporal domain, while preserving unchanged structures. This paper proposes a novel fusion method to progressively refine building rooftop models over time by integrating multi-sensor data. The proposed method integrates the line modelling cues of existing rooftop models produced by airborne laser scanning data with the new ones extracted from optical imagery. This modelling cue integration process is developed to progressively rectify geometric errors based on Hypothesize and Test optimization using Minimum Description Length. A stochastic method, Markov Chain Monte Carlo, coupled with simulated annealing, is employed to generate model hypotheses and perform a global optimization for finding the best solution. This fusion method is designed to offset the limitations of respective sensors and thus rectify various modelling errors (shape deformation, boundary displacement, and orientation errors) that are often involved in rooftop building models. The performance evaluation tested over the ISPRS show the proposed modelling method can achieve the improvements of 1.8%, 0.54°, 0.33 m, and 0.007 for the quality, orientation difference, Hausdorff distance, and turning function distance, respectively, compared with initial building models. In addition, the proposed methods show the highest performance in the quality measure among the state of the art methods, while demonstrates competitive performance in the completeness and correctness measures.