Vertical orientation correction of uav image-based point clouds using statistical modeling of gable roof geometry

Abstract

Coregistration of point clouds obtained from various sensors is an important part of workflows for automatic building reconstruction from remote sensing data. Many approaches assume a common Z axis between the coordinate systems, and perform coregistration in 2D. While this assumption is usually valid for laser scanning (LS) data, for photogrammetric point clouds the Z axis is in general different from the world Z axis, and requires correction e.g. by manually measured ground control points (GCP). In this paper, we propose a fully automatic, GCP-free procedure for finding the world Z axis in rural areas, based on the relationships of planar surfaces in building gable roofs. Instead of performing direct gable line detection, we derive these lines as theoretical intersections between adjacent roof planes from 3D shape fitting. Each gable roof then casts a vote for both the Z axis direction and sign based on roof convexity constraints, and the votes are aggregated through a non-parametric kernel density estimator model. Experiments on two real world UAV image-based point clouds show that the Z axis recovered by our method leads to high-accuracy planimetric coregistration, with a median distance over 89 as well as 149 matched linear feature pairs (respectively for dataset 1 and 2) lying below 1 cm. Our results indicate that a high-quality vertical orientation can be achieved without using any GNSS or IMU hardware, which enables the use of low-cost UAV platforms for suburban and rural mapping tasks.