Object-based building instance segmentation from airborne LiDAR point clouds

Abstract

ABSTRACT Building instance segmentation is of very importance to parallel reconstruction, management and analysis of building instance. Previous studies of building instance segmentation mainly focused on the building scenes where the building spacing is much larger than the point spacing, while the accuracy of building instance segmentation for complex buildings scenes and the building point clouds where the space between buildings is similar with point spacing is low. To improve the accuracy of building instance segmentation for complex building scenes, we propose a novel object-based building instance segmentation (OBBIS) method from airborne light detection and ranging (LiDAR) point clouds. Firstly, our proposed method divides building point clouds into objects, and then the objects are classified according to the characteristics of building roof plane objects, roof accessory objects and building facade objects. Secondly, we use node to represent object and then a fix-size feature vector is inferred for each node. Thirdly, vertical cylinder neighbour node graph is constructed. Finally, the energy function is constructed according to the relationship between the nodes, and then the objects are merged according to the energy minimum (that is, objects are merged with a minimum energy to obtain the building instances). Comprehensive experiments on benchmark datasets demonstrate that the proposed OBBIS method performs better than eight state-of-the-art building instance segmentation methods.