结合倒置残差模块和可微分RANSAC算法的点云配准模型OA

In response to the issue of insufficient fusion of global and local features in complex and occluded point cloud data,as well as the challenge of integrating robust estimation algorithms into deep learning training pipelines,a point cloud registration network model based on an improved PointNet++and random sample consensus algorithm is proposed.Firstly,a PointNet++net-work fused with an inverted residual MLP is used to extract local point cloud features,generating feature descriptors that inte-grate both global and local feature information.Secondly,the local feature Transformer module is used to generate provisional point correspondence and confidence score.Then,a neural sampler is introduced to ensure the differentiability of the RANSAC sampling process,and a differentiable geometric solver is used to calculate the rigid transformation matrix between point cloud pairs.Finally,a trainable quality function is designed to optimize evaluation metrics during each iteration,integrating the robust estimation algorithm into the training pipeline,ultimately completing the point cloud registration.The results of multiple com-parative experiments on three public large-scale point cloud datasets,3DMatch,ETH,and KITTI,show that the feature match-ing recall rate of the method proposed in this paper on 3DMatch reaches 98.4%,which is 0.8 percentage points higher than that of the SpinNet network.On ETH and KITTI,the feature matching recall rate and accuracy reach 98.5% and 99.57% respec-tively,which are 5.7 percentage points and 0.5 percentage points higher than those of the SpinNet network.When dealing with multiple complex point cloud datasets with uneven density and occlusion,the proposed method performs better than existing ad-vanced methods and can effectively improve the registration accuracy.