Registration of Non-Uniform Density 3D Point Clouds using Approximate Surface Reconstruction (original) (raw)

3D laser scanners composed of a rotating 2D laser range scanner exhibit different point densities within and between individual scan lines. Such non-uniform point densities influence neighbor searches which in turn may negatively affect feature estimation and scan registration. To reliably register such scans, we extend a state-of-the-art registration algorithm to include topological information from approximate surface reconstructions. We show that our approach outperforms related approaches in both refining a good initial pose estimate and registering badly aligned point clouds if no such estimate is available. In an example application, we demonstrate local 3D mapping with a micro aerial vehicle by registering sequences of non-uniform density point clouds acquired in-flight with a continuously rotating lightweight 3D scanner. 1 Introduction 3D scanners provide robots with the ability to extract spatial information about their surroundings, detect obstacles and avoid collisions, build 3D maps, and localize. In the course of a larger project on mapping inaccessible areas with autonomous micro aerial vehicles (MAVs), we have developed a light-weight 3D scanner [13] specifically suited for the application on MAVs. It consists of a Hokuyo 2D laser scanner, a rotary actuator and a slip ring to allow continuous rotation. Just as with other rotated scanners, the acquired point clouds (aggregated over one full or half rotation) show the particular characteristic of having non-uniform point densities: usually a high density within each scan line and a certain angle between scan lines which depends on the rotation speed of the scanner (see Figure 1). In our setup, we aggregate individual scans of the continuously rotating laser scanner and form 3D point clouds over one half rotation (covering an omnidirectional field of view). To compensate for the MAV’s motion during aggregation, we use visual odometry [18] and inertial sensors as rough estimates, and transform the scans into a common coordinate frame. Since we use the laser scanner not only for mapping and localization but also for collision avoidance, we rotate the scanner fast resulting in a particularly low angular resolution of roughly 9. This reduces the point density in the aggregated point clouds but increases the frequency with which we perceive (omnidirectionally) the surroundings of the MAV (2 Hz). The resulting non-uniform point densities affect classic neighborhood searches in 3D and cause problems in local feature estimation and registration. To compensate for the non-uniform point densities, we extend the state-of-the-art registration algorithm Generalized-ICP [19] to include topological surface information instead of the 3D neighborhood of points. r u+1 u u-1 Figure 1: Classic neighbor searches in non-uniform density point clouds may only find points in the same scan line (red), whereas a topological neighborhood (green) can better reflect the underlying surface The remainder of this paper is organized as follows: after giving a brief overview on related work in Section 2, we present our approach in Section 3. In experiments, we demonstrate the superior performance of our approach and discuss the achieved results in Section 4. 2 RelatedWork 2.1 Laser Scanners for MAVs For mobile ground robots, 3D laser scanning sensors are widely used due to their accurate distance measurements even in bad lighting conditions and their large field-ofview. For instance, autonomous cars often perceive obstacles by means of a rotating laser scanner with a 360 horizontal field-of-view, allowing for the detection of obstacles in every direction [12]. Up to