Incomplete Point Cloud Registration Based On Unbalanced Optimal Transport

The goal of point set registration is to align two or more point clouds into a coherent coordinate frame by estimating their relative transformation.In the process of scanning,point cloud is often degraded by noise,outliers and missing points,which make registration more difficult.The thesis summarizes the core of rigid point cloud registration algorithms and states the reason for low precision in incomplete point cloud registration.For rigid point cloud registration,algorithms based on many-to-many correspondences have shown to be more robust than algorithms based on one-to-one correspondences.However,point clouds with severe outliers and missing data may lead to imprecise many-to-many correspondences,consequently inaccurate registration.Optimal tranport theroy is a more general many-to-many method,and the mass conservation laws in traditional optimal transport lead to inaccurate registration result.A general framework for point cloud registration based on unbalanced optimal transport is proposed.When the point clouds are with a large part of outliers or missing points,the proposed method behaves the effective and robust.The mainly research content is as followings:1.A point cloud registration algorithm based on unbalanced optimal transport theory is proposed.Point clouds are represented as probability measure,and the method establishes objective function according to optimal transport.Unbalanced optimal tranport,based on relaxing the mass conservation laws and constraining the range of total transport mass,is utilized to obtain the accurate partial mass optimal transport plan.The total mass constraint provides an explicit parameter to adjust the ratio of points that should be accurately matched and an efficient strategy to narrow the solution space,and helps avoid incorrect many-to-many correspondences in server missing points and outliers.2.The thesis utilizes a lot of different degraded artificial data and real scene point cloud data to evaluate efficiency,and the proposed algorithm compared with the classical one-to-one,many-to-many correspondence point cloud registration algorithm.The experiments demonstrate that the proposed approach elevates the robustness and accuracy of point cloud registration when point clouds include a large part of outliers and missing points.