Self-feature-based point cloud registration method with a novel convolutional Siamese point net for optical measurement of blade profile

Abstract

Currently, non-contact optical measurement technologies have shown prominent power in measuring blade profiles. One of the critical issues is how to align measured point clouds in different viewpoints into an entire profile. However, extracting efficient features for registration is difficult due to the thin-walled and twisted spatial freeform surfaces of blades. Besides, various densities and small overlaps between two adjacent point cloud data also increase the difficulty of registration. To overcome these issues, in this paper, laying on the foundation of a developed multi-axis motion system, we present a novel and general framework titled convolutional Siamese point net (CSPN) for partial-to-partial point cloud registration. Our network mainly consists of three parts: feature extraction, matching matrix computation and singular value decomposition (SVD). To make a balance between low-level and high-level semantics, we design a pyramidal architecture to learn the multiscale features of each point. In matching matrix computation, both feature space and coordinate space are utilized. Also, a novel attention mechanism is proposed to deal with the conflicts between matching matrices obtained by feature space and coordinate space. Experimental results demonstrate the feasibility and good practical application prospect of this method.