Dense LIDAR point clouds from room-scale scans

Abstract

LiDARs can capture distances with high accuracy and should be very useful to create point clouds that provide highly detailed representations of an environment. If these reconstructions are meant as baseline or ground truth for other algorithms, they must have a high density and accuracy. Currently available LiDARs do still face some limitations. Either they have a limited range, or they have a rather limited resolution in one or more dimensions. As a consequence, all of them have to undergo motion to capture a larger environment. While some systems follow extremely well-predictable motion paths such as satellite trajectories or robotic arms, others require more spontaneous and flexible motion. These systems use either visual simultaneous localization and mapping (vSLAM), GPS or IMU to achieve this, but they are generally designed in such a way that human intervention is required during the creation of high-quality point clouds. In this paper, we make use of a rotating LiDAR with an attached IMU to create dense point clouds of room-scale environments with the base accuracy of the LiDAR by compensating for the various inaccuracies that are introduced by the LiDAR's motion. The resulting dense scans are suitable as ground truths for other techniques because we retain the error distribution of the LiDAR itself through the densification. In contrast to other works, we do not aim at a visually pleasing or easily meshable result and we can therefore avoid potentially inaccurate assumptions about the flatness of surfaces. We take a two-step approach. First, we densify from a stationary position changing only the LiDAR's pitch. Second, we add free motion to expose obstructed views. We show that motion paths determined by repeated Iterative Closest Point (ICP) as well as image matching on height maps can be used to create feasible priors for densification using ICP.