Research on Cluttered Scene Grasping Based on Rapid Point Cloud Registration Algorithm

Robot grasping in stacked environments relies on object information acquisition and recognition. Currently, most recognition methods based on point cloud rely on Iterative Closest Point(ICP) registration, which still performs poorly in terms of real-time performance and convergence. This paper proposes a rapid point cloud registration algorithm based on key-point extraction and builds a grasping system to achieve real-time grasping of stacked objects. Firstly, point cloud segmentation of stacked objects is implemented based on Local Convex Connection Blocks (LCCP), and then the 3D Hough voting algorithm is used to identify the template library point cloud corresponding to the single object point cloud after segmentation. Secondly, the Sample Consensus Initial Alignment(SAC-IA) method that integrates Intrinsic Shape Signature(ISS) key-point extraction is adopted to achieve rapid coarse registration of point clouds. Then, an improved ICP based on Normal-based Adaptive Weighted Principal Component Analysis(NW-PCA) is proposed to achieve rapid and accurate registration of point clouds. Experiments have shown that the registration algorithm significantly reduces computation time while achieving optimal registration error in The Stanford 3D Scanning Repository. Using UR5 to grasp the stacked scenes of randomly placed objects with Double-layer and Triple-layer, the average grasp success rates were <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 2. 2 4 \%}$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{8 8. 6 2 \%}$</tex>, respectively.