Multi-Constraints Guided Single-View Point Cloud Registration for Adaptive Robotic Manipulation

Abstract

In model-based adaptive industrial robotic manipulation, the target pose uncertainty and workspace restriction are prevalent, where single-view point cloud registration is an effective step for real-time pose estimation. However, single-view 3D registration suffers from challenges of small target occupancy, significant rotation deviations, high presence of outliers and noises, and limiting the effectiveness of current approaches. To address these challenges, we propose a novel single-view point cloud registration method multi-constraints guided single-view point cloud registration (MCSVR), which aims to leverage multiple constraints of single-view imaging to guide a coarse-to-fine registration mechanism, thereby achieving more accurate and versatile pose estimation for targets with complex structures of varying sizes and orientations. First, a region-level matching based on Gaussian mixture models (GMMs) is proposed to screen target regions. Subsequently, in the point-level matching stage, we introduce a multiconstraint-guided hybrid compatibility to obtain more reliable correspondence consensus. Finally, we devise a dynamic registration strategy based on single-view constraints to achieve precise registration. Experimental evaluations and practical applications demonstrate the superior performance of MCSVR.