Crossmodal point cloud registration in the Hough space for mobile laser scanning data

Abstract

In this paper we propose a general approach for registration of point clouds obtained by various mobile laser scanning technologies. Our method is able to robustly match measurements with significantly different density characteristic including the sparse and inhomogeneous instant 3D (I3D) data taken be self-driving cars, and the dense and regular point clouds captured by mobile mapping systems (MMS) for virtual city generation. The core steps of the algorithm are robust scan segmentation, abstract street object extraction, object based coarse transformation estimation in the Hough accumulator space, and point-level registration refinement. Experimental results are provided using three different sensors: Velodyne HDL64 and VLP16 I3D scanners, and a Riegl VMX450 MMS. Application examples are shown regarding self localization of autonomous cars through crossmodal I3D and MMS frame registration, IMUless SLAM and change detection based on I3D data.

Real-Time Point Cloud Alignment for Vehicle Localization in a High Resolution 3D Map

“…Registering point clouds with different modalities and density characteristics has a limited bibliography. A sequential technique for cross modal point cloud alignment has be proposed in [15], which extracts first abstract object patches in both point clouds, then it calculates a coarse alignment between the frames purely based on the estimated object centers, finally an NDT based point level refinement process is applied. As drawbacks, the object level matching may fail in case of several diverse object types, and the additional NDT steps induces significant computational overload.…”

SFM And Semantic Information Based Online Targetless Camera-LIDAR Self-Calibration

“…Most of the conventional point level iterative registration techniques, such as variants of ICP or NDT [10], may fail when the density characteristic is quite different between the point clouds, and in our case, they can also be misled by false correspondences on the ground caused by the typical ring patterns of RMB LIDAR data. To avoid such artefacts we proposed a robust object level alignment approach between sparse RMB LIDAR point clouds and a dense reference pont map in [11,12]. This technique extracts object blob centers in both point cloud frames, which are matched in the Hough domain, based on the idea of a fingerprint minutiae matching algorithm [13].…”

3D CNN-Based Semantic Labeling Approach for Mobile Laser Scanning Data

“…In dense urban areas, such as the downtown of Budapest, Hungary, one can experience that commercial Global Positioning Systems (GPS) can often only provide position information with large inaccuracies (1-10 m). In order to correct the error of the initial GPS based transformation we proposed a low-cost method [1], [27] which is able to precisely register the sparse and inhomogeneous RMB Lidar point clouds of the SDV to a dense geo-referenced point cloud which can be obtained by a MLS mapping system: To reduce the complexity of the algorithm first we fit a rectangular 2D grid onto the horizontal plane and we project all the 3D points to the corresponding 2D grid cell. In the next step the blobs of the estimated obstacles are separated both in the sparse RMB Lidar point clouds and in the dense MLS map, using an adaptive connected component extraction algorithm applying structure-based merge and split steps.…”

“…The core of the registration method is a quick object-level transformation estimation algorithm between the point clouds in the Hough domain, relying on several automatically extracted feature points [1]. Contrary to point level point cloud registration techniques such as the Iterative Closest Point (ICP) or the Normal Distributions Transform (NDT) [27], the proposed method [1] works in real-time, and it is able to manage arbitrary initial rotation differences between the two point clouds, as well as an initial translation error up to 10 − 15m. However, the above registration algorithm is based on the assumption that the reference landmark objects extracted from the MLS map correspond in majority to static and permanent scene elements (such as lamp posts, tree trunks, kiosks etc), while all the phantoms, and moving or movable objects of the MLS point cloud appear as noise factors during the estimation of the right transform.…”