3D point cloud registration method based on overlapped region mask prediction

Abstract

The invention discloses a 3D point cloud registration method based on overlapped region mask prediction. An iterative convolutional neural network is used to learn point cloud rigid body transformation from coarse to fine. The 3D rigid body transformation is represented as a quaternion and translation distances in the X-axis direction, the Y-axis direction and the Z-axis direction, and the quaternion can avoid the situation of a universal lock. In addition, we learn masks with attention mechanisms in the network. The mask not only highlights the overlapped part of the source point cloud and the target point cloud, but also can filter out non-overlapped areas and noise interference therein. After each iteration process, the mask predicted by the iteration can be used in the point cloud feature extraction process of the next iteration so as to ensure that the prediction of the point cloud overlapped region and the regression of the 3D rigid body transformation are mutually promoted.

Description

technical field [0001] The invention relates to the fields of computational graphics and computer vision, in particular to a 3D point cloud registration method based on overlapping region mask prediction. Background technique [0002] 3D point cloud registration is the process of matching and superimposing two or more point clouds acquired at different times, different viewing angles, and different sensors. This technology is widely used in 3D scene reconstruction (3D Scene Reconstruction), simultaneous positioning and maps Applications and fields such as Simultaneous Localization And Mapping, Augmented Reality and Autopilot. [0003] Among the various existing point cloud registration methods, the algorithm based on Iterative Closest Point has been widely used due to its simplicity and efficiency. This method iteratively calculates the nearest neighbor points of coordinates in Euclidean space to determine matching point pairs, and then uses Singular Value Decomposition (Si...

AI Technical Summary

Problems solved by technology

However, when the overlapping area between the source point cloud and the target point cloud is small, it is easy to cause mis-matching of corresponding points.

In order to allow the algorithm to have a stronger registration ability for partially overlapping point clouds, a method based on outlier filtering is proposed, but because it still needs to match the corresponding point pairs, it can only use the information of some sparse points in the point cloud. It is difficult to improve the registration accuracy

Images