AN IMPLICIT REGULARIZATION FOR 3D BUILDING ROOFTOP MODELING USING AIRBORNE LIDAR DATA
Keywords: 3D Rooftop Modeling, Binary Space Partitioning, Minimum Description Length, Regularization, LIDAR
Abstract. This paper proposes a new algorithm to generalize noisy polylines comprising a rooftop model by maximizing a shape regularity (orthogonality, symmetricity and directional simplications). The nature of remotely sensed data including airborne LiDAR often produce errors in localizing salient features (corners, lines and planes) due to weak contrast, occlusions, shadows and object complexity. A generalization or regularization process is well known algorithm for eliminating erroneous vertices while preserving significant information on rooftop shapes. Most of existing regularization methods achieves this goal base on a local process such as if-then rules due to lacking global objective functions or mainly focusing on minimising residuals between boundary observations and models. In this study, we implicitly derive rules to generate local hypothetical models. Those hypothesized models produce possible drawings of regular patterns that given rooftop vectors can possibly generate by combining global and local analysis of line directions and their connections. A final optimal model is globally selected through a gradient descent optimization. A BSP (Binary Tree Partitioning)-tree was used to produce initial rooftop vectors using ISPRS WGIII/4's benchmarking test sites in Veihngen. The proposed regularization algorithm was applied to reduce modelling errors produced by BSP-tree. An evaluation demonstrates the proposed algorithm is promising for updating of building database.