基于AFAM机制的鲁棒点云配准方法OA

In addressing the shortcomings of the existing methodologies for point cloud registration,including inadequate robustness,inaccurate feature representation and sensitivity to initial registration,particularly in the context of noisy point clouds,a novel point cloud registration network structure,RPM-AFAMNet,is proposed.This structure is constructed by integrating an adaptive feature aggregation module(AFAM)within the RPM-Net network.The AFAM is comprised of two modules:the batch attention mechanism transformer(BatchFormer)and the dynamic re-aggregated feature representation(DRFR)module.The purpose of the AFAM is twofold:firstly,to enhance the robustness of the network,and secondly,to optimise the representation of point cloud features.The BatchFormer is designed to mitigate the interference of noisy points by batch-weighted learning of point cloud features.The DRFR module,meanwhile,enhances the understanding of point cloud spatial relationships through a dynamic feature reorganisation strategy.This,in turn,improves the accuracy of point cloud alignment.Finally,the network is tested on the ModelNet40 dataset and significant improvements are achieved in six evaluation metrics,such as rotational and translational mean square error.Compared with RPM-Net,the following reductions are achieved:32.05%for rotational mean square error,75.86%for rotational mean absolute error,33.33%for translational mean square error,77.78%for translational mean absolute error,33.03%for rotation angle error,and 36.36%for translation distance error.The experimental results demonstrate the efficacy of the proposed method in enhancing the registration performance of noise-containing point clouds.