Probabilistic normal distributions transform representation for accurate 3D point cloud registration

Abstract

This paper presents a probabilistic normal distributions transform (NDT) representation which improves the accuracy of point cloud registration by using the probabilities of point samples. Since conventional NDT does not generate distributions in cells having fewer point samples than the number threshold, it would be failed to represent the environment if the point cloud is divided by high-resolution cells. Also, it can lead to incorrect estimations of pose variations. To solve the problem, we define the probability of a point sample and compute the mean and covariance based on the probability. Besides, we show that the generalization property of the probabilistic NDT objective function. The probabilistic NDT has two advantages. First, it generates distributions in all of the occupied cells regardless of the resolution of cells. Second, it reduces the degeneration effect by using modified covariance. The experimental results show that all of the occupied cells have distributions even if the point cloud is divided by high-resolution cells and that the probabilistic NDT improves the accuracy of NDT-based registration.