Segmentation of Sloped Roofs from Airborne LiDAR Point Clouds Using Ridge-Based Hierarchical Decomposition

Abstract

This paper presents a new approach for roof facet segmentation based on ridge detection and hierarchical decomposition along ridges. The proposed approach exploits the fact that every roof can be composed of a set of gabled roofs and single facets which are separated by the gabled roofs. In this work, firstly, building footprints stored in OpenStreetMap are used to extract 3D points on roofs. Then, roofs are segmented into roof facets. The algorithm starts with detecting roof ridges using RANSAC since they are parallel to the horizon and situated on the top of the roof. The roof ridges are utilized to indicate the location and direction of the gabled roof. Thus, points on the two roof facets along a roof ridge can be identified based on their connectivity and coplanarity. The results of the segmentation benefit the further process of roof reconstruction because many parameters, including the position, angle and size of the gabled roof can be calculated and used as priori knowledge for the model-driven approach, and topologies among the point segments are made known for the data-driven approach. The algorithm has been validated in the test sites of two towns next to Bavaria Forest national park. The experimental results show that building roofs can be segmented with both high correctness and completeness simultaneously.

A Hybrid Approach for Three-Dimensional Building Reconstruction in Indianapolis from LiDAR Data

“…First of all, compared with other existing hybrid approaches discussed above [41][42][43][44][45], the proposed hybrid approach can reconstruct very complicated buildings, such as those having irregular footprints and roof structures with flat, shed and tiled sub-structures mixed together (Building 4-7 in Figure 12). Roof structures with a round-shaped footprint can also be identified by the MBC threshold after edge-based segmentation, which is also not considered in those existing hybrid approaches.…”

“…As the author stated, the video camera cannot catch all of the building surfaces and frames at one time, which would lead to missing that information. Fan et al [45] presented a building reconstruction method using a hierarchical ridge-based decomposition algorithm to obtained ridge position, as well as the angle and size of roof planes for subsequent model-driven reconstruction, based on high-density (6-9 pts/m 2 ) LiDAR point cloud data. They achieved a 95.4% completeness for the reconstruction of seven buildings, but as they stated in the conclusion, the limitation of their method is that it has strict requirements for the density of LiDAR point clouds.…”

“…However, even more complete planes can be generated by the hybrid approaches' accuracy of plane boundaries, which still can be limited by the precision of algorithms used in data-driven approaches, which detect the edges and ridges. Some studies [42,45,46] rely on the high resolution LiDAR point cloud data to achieve a high 3D building reconstruction quality. Using the high-density LiDAR point clouds can produce high segmentation accuracy, but it requires very large storage memory in computers and large amounts of time to process when reconstructing Remote Sens.…”

“…Many researchers tried to utilize the hybrid approach [41][42][43][44][45][46], which integrates data-driven and model-driven approaches, as they think it might produce higher quality (more complete and detailed) building models. The principle is to use data-driven approaches to detect the features, such as ridges and edges, as prior knowledge for a subsequent model-driven approach.…”

“…Existing studies of the reconstruction of 3D city models from LiDAR-derived nDSM or high-density point cloud data with or without additional data sources included three major categories of strategies: data-driven (bottom-up) [5,7,8,[20][21][22][23][24][25][26][27][28][29][30][31][32], model-driven (top-down) [1,2,6,9,13,14,16,[33][34][35][36][37][38][39][40] and hybrid approaches [41][42][43][44][45][46].…”