Registration of 3D Point Cloud with Rotation Invariant Features Learning for Partial Overlapping

Abstract

Point cloud registration attempts to register a pair of point cloud through rigid transformation to form a complete point cloud. Traditional methods rely on searching closest neighbors in the feature space and filtering the outliers through RANSAC to find the point-to-point correspondence. Recently, learning-based algorithms incorporate learning to local feature descriptors and perform better than the traditional methods. However, they continue to adopt the feature-based matching and point-level correspondence approach for pose estimation. In this work, attention mechanism is incorporated to learning transformation-invariant features and an end-to-end solution is used to predict the correspondences. A two-stream feature extraction architecture consisting primarily of feature extractor and transformer learns the rotation-invariant features through transformer layer. The network learns to predict the point coordinates mapped to the other point cloud and the probability between the correspondences in the overlap region. The rigid relative pose transformation can be solved from the predicted correspondences in a closed form. The registration method is applied to pose estimation in the robotic grasping and sorting task and achieves favorable result.