Outdoor Scene Registration of 3D Range Data using Coarse GPS and Modified ICP with Cubic Filtering
Date
2015-03-24
Authors
Tonia, Lado
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
In the process of modeling an outdoor scene using 3D range data, the input data must be registered to a common origin. Registration is the process of determining the precise translation and rotation of the sensor between captures and is used to transform all data into a shared coordinate system. In the case of large and complex outdoor scenes data sets can include thousands of captures. Precise registration through an automated system is necessary for creating a consolidated model of the scene. Complexities in outdoor environments, the propagation of minute errors, and the nondeterministic motion of the sensor render a number of commonly used registration techniques ineffective. In this thesis we survey some of the existing methods of registration and propose a synthetic process that can robustly determine these transformations to a higher degree of accuracy than any of its component processes. The process begins with coarse GPS data and implements a method of optimizing the coincidence of overlapping binary cubes. This is followed by multiple iterations of the iterative closest point (ICP) algorithm, with and without a novel method of cubic filtering. The results of the registration for a single outdoor scene in a closed loop around a small building are presented and compared to a collection of alternative registration methods over multiple LIDAR captures of the same scene. The results show that the proposed process is the most robust to variations in sensor speed. It also provides accurate enough registration in all captures to successfully display 3D scenes. After registration, the point at which the sensor returned to the beginning of the loop is determined to within 2 meters of the true location, which is the cumulative error over roughly 400 registrations. The results found justify the development of fully automated or semi-automated 3D modeling of complex outdoor scenes from LIDAR data and low quality GPS. The proposed process has been created as a basis for an automated system of creating 3D models of outdoor scenes using a Velodyne HDL-64E sensor mounted on a golf cart using GPS data from a modern cellphone.
Description
Keywords
Registration, LIDAR, Binary cube, 3D, Point cloud, ICP