Abstract: In the field of point cloud registration,effectively capturing detailed features and enhancing registration accuracy,particularly when handling point clouds with low overlap rates,are two primary challenges.Traditional feature extraction methods have achieved some success but remain insufficient in mining geometric information,leading to limited feature discrimination.Current techniques primarily rely on position encoding and geometric embedding strategies,which enhance geometric understanding but still struggle with accuracy in high-outlier scenes.To address these issues,this paper introduces ROPNet,a novel registration network that integrates adaptive optimization and multi-dimensional focusing.ROPNet’s design includes multi-dimensional focusing,adaptive modulation kernels,and a dynamic optimization selector.These components enable the network to capture both global features and local details,accurately identify spatial positions and correspondences,and better understand the intrinsic structure of point cloud data.Additionally,ROPNet’s robust design significantly improves the identification of inliers,thereby significantly improving registration accuracy.Experimental results demonstrate ROPNet’s superior performance.On the 3DMatch dataset,ROPNet achieves a 92.4% registration recall rate and a 71.3% inlier ratio.On the KITTI dataset,it attains 99.8% registration accuracy,with relative rotation and translation errors reduces to 0.24 degrees and 6.6 cm,respectively.

Key words: Point cloud registration, Multi-dimensional focusing, Deep learning, Feature extraction, Adaptive optimization