Relative Motion Threshold for Rejection in ICP Registration

Abstract

Simultaneous Localization and Mapping (SLAM) iteratively builds a map of the environment by putting each new observation in relation with the current map. This relation is usually done by scan matching algorithms such as Iterative Closest Point (ICP) where two sets of features are paired. However as ICP is sensitive to outliers, methods have been proposed to reject them. In this article, we present a new rejection technique called Relative Motion Threshold (RMT). In combination with multiple pairing rejection, RMT identifies outliers based on error produced by paired points instead of a distance measurement, which makes it more applicable to point-to-plane error. The rejection threshold is calculated with a simulated annealing ratio which follows the convergence rate of the algorithm. Experiments demonstrate that RMT performs better than former techniques with outliers created by dynamical obstacles. Those results were achieved without reducing convergence speed of the overall ICP algorithm.

An Improved MbICP Algorithm for Mobile Robot Pose Estimation

“…The iterative closest point (ICP) algorithm is one of the most popular scan-matching techniques because of its efficiency and simplicity [18][19][20][21][22]. This algorithm iteratively establishes correspondences based on the criterion of the closest point and calculates a rigid transformation that minimizes the distance error until convergence.…”

“…Considerable research has been conducted to improve ICP and its variants in matching partially overlapping scans by enhancing the performance of correspondence establishment and rejecting incorrect correspondences [19,20,[22][23][24][25][26]. In terms of establishing correspondence, Chen proposed a point-to-plane approach, which determines the correspondence by finding the projection of the source point onto the tangent plane at a destination surface point that is the intersection of the normal vector of the source point [27], but this approach is only available for scan matching with three-dimension (3D) laser scans.…”

“…In terms of rejecting incorrect correspondences, Druon proposed a rejection strategy that was based on a statistical analysis of the correspondence distance, which adopted µ + σ as the rejection threshold [28][29][30]. Nevertheless, this strategy assumes that the distribution of distances is Gaussian, which is not a reasonable assumption in a dynamic environment [22]. Chetverikov proposed a trimmed ICP algorithm (TrICP) [26], which uses golden-section searching to estimate an overlapping ratio and uses a trimmed least-squares approach to discard incorrect correspondence pairs.…”

“…Chetverikov proposed a trimmed ICP algorithm (TrICP) [26], which uses golden-section searching to estimate an overlapping ratio and uses a trimmed least-squares approach to discard incorrect correspondence pairs. Despite the claimed robustness of this algorithm, TrICP is time consuming because it repeatedly searches for the best overlapping ratio [22,24]. This algorithm may also be misled by a large change in the overlap because of occlusions from a moving platform.…”

“…This algorithm may also be misled by a large change in the overlap because of occlusions from a moving platform. Pomerleau proposed an adaptive rejection technique called the relative motion threshold (RMT) [22], which calculates the rejection threshold with a simulated annealing ratio. This technique is flawed in that the rejection threshold heavily relies on the translation vector, so a large sensor rotation may produce an incorrect rejection [20].…”