Research On Rigid Point Cloud Registration Method Based On Key Point Extract

Point cloud registration has a wide range of applications in areas such as autonomous driving,3D reconstruction,and robotics.The registration method based on the nearest point assumption usually requires a good initial position between point clouds,so its application scenarios are limited.To this end,some scholars have developed geometric feature extraction methods,which build correspondences through geometric feature matching,thus releasing the limitation of initial position.In this thesis,this type of method is called feature matching-based registration method.In this type of method,a set of point pairs containing a large number of outliers is usually obtained after feature matching.It is difficult to directly solve the rigid transformation through these point pairs,so it is necessary to The outlier removal operation is performed first.And the point cloud usually contains a large number of points,in order to improve the computational efficiency,the usual method is to random sample.However,in point clouds with low overlapping attributes,random sampling will sample useless points in non-overlapping regions,which leads to inefficient sampling and further increases the difficulty of registration.Aiming at the above problems,this thesis proposes a rigid point cloud registration method based on key point extraction.In the keypoint extraction stage,the extensive contextual information is first captured by the dynamic graph convolutional network,and then the cross-attention mechanism is used to perceive the overlapping area,and the probability of being located in the overlapping area point by point is output,so as to combine the feature matching probability sampling located in the overlapping area Key points with strong matching can improve the efficiency of sampling,thereby reducing the difficulty of registration.Aiming at the problem that the set of point pairs contains a large number of outliers,this thesis proposes a gradient non-convex optimization method,which converts the non-convex cost function into a convex changing surrogate function through Black-Rangarajan duality.When the surrogate function changes from convex to In the non-convex process,the weighted singular value decomposition algorithm is combined to complete the process of first reducing the error value,and then returning to the original cost function to converge,so as to obtain a better local optimal solution or global optimal solution.In the experiment,this thesis verifies the effectiveness of the key point extraction method and the point-to-point outlier removal method by comparing with the random sampling method and the commonly used outlier removal method.In the3 DMatch dataset,our method has obvious advantages over traditional methods and is also competitive with deep learning methods.Finally,the robustness of our method in lowoverlap registration is verified by comparative experiments with different overlap rates in the Model Net40 dataset.