Predicting Alignability of Point Cloud Pairs for Point Cloud Registration Using Features

Abstract

Point cloud registration is often used in fields like SLAM where the overlap of two consecutive point clouds is large. But in fields like multi-sensor fusion of point clouds and LiDAR-based localization, there is a high chance of registering non-overlapping point cloud pairs. Since in such cases, the result will always be a wrong transformation, it is useful to evaluate the alignability of the point cloud pairs prior to the registration. In this paper, an algorithm is presented that predicts the alignability of two point clouds based on the minimum distances of descriptors. It calculates statistical measures describing the minimum distances and classifies the point cloud pairs. The paper shows that it is possible to predict the alignability and evaluates the runtime compared to registration algorithms, as well as the ignoring of the largest minimum distances.