Adaptive Moving Least-squares Surfaces For Multiple Point Clouds Registration

Multi-view point clouds registration is a key technology in the multi-sensor basedclosed loop measurement system, and in improving the surface reconstruction quality inReverse Engineering. Existing methods for multi-view point clouds registration are oftenaffected by the complex shape of components and measurement noise, thus theregistration accuracy is limited. In this paper, we systematically analyzed the methods formulti-view point clouds registration, the main research contents and results can besummarized as follows:(1) By analyzing the main factors that affect the accuracy of multi-view point cloudsregistration, we identify the key issue as to define an underlying smooth surface of anytopology which could faithfully resembles the original surface. On this basis, we introducethe surface reconstruction directly from measuring point cloud using MovingLeast-Squares (MLS) surface, and further point out that the Gaussian Kernel h is the mainfactor that affects the surface reconstruction accuracy.(2) We define a novel Adaptive Moving Least-Squares (AMLS) surface whichenables the Gaussian Kernel h to be adaptively adjusted according to local principlecurvature. For models of complex shape, the AMLS surface guarantees the smoothnessand continuousness in flat area while avoids the over smooth and distortion for highcurvature areas and sharp features.(3) In order to overcome the shortcomings of traditional curvature calculationmethods in a differential way which is sensitive to noise, we propose a multi-scaleintegration method based on Integral Invariants for local curvature estimation.(4) We propose an Adaptive Moving Least-Squares (AMLS) surface based approachfor surface reconstruction and multi-view point clouds registration. To validate thepresented approach, two synthetic and two actual mechanical examples are demonstratedto show the improvement on registration accuracy, and the robustness to point clouddensity, noise level, and initial position.In this paper, we propose an Adaptive Moving Least-Squares (AMLS) surface formulti-view point clouds registration, and verify the presented approach by experiments todemonstrate its accuracy and robustness.