Feature Extraction for Low Overlap Point Cloud Registration

Abstract

In recent years, many achievements have been made in point cloud registration. However, when the overlap ratio of two input point clouds is low, the efficiency of most registration methods is greatly reduced. In order to solve this problem, we specially extract feature descriptors for overlapping region. We introduce a cross-attention based feature descriptor extraction model, which independently uses the kernel point convolution (KPConv) network to down sample two input point clouds and learn local geometric features, and then uses the cross-attention to make the feature descriptor also have co-contextual information features. Experiments show that this method can accurately learn the corresponding relationship between two point clouds in the low overlap region.