Improving architectural 3D reconstruction by constrained modelling

Abstract

This doctoral thesis presents new techniques for improving the structural quality of automatically-acquired architectural 3D models. Common architectural properties such as parallelism and orthogonality of walls and linear structures are exploited. The locations of features such as planes and 3D lines are extracted from the model by using a probabilistic technique (RANSAC). The relationships between the planes and lines are inferred automatically using a knowledge-based architectural model. A numerical algorithm is then used to optimise the position and orientations of the features taking constraints into account. Small irregularities in the model are removed by projecting the irregularities onto the features. Planes and lines in the resulting model are therefore aligned properly to each other, and so the appearance of the resulting model is improved. Our approach is demonstrated using noisy data from both synthetic and real scenes.