Research On High Precision Stitching Algorithm For 3D Point Cloud

Point cloud stitching is a key step in full surface reconstruction,and it has good application scenarios in medical imaging,industrial inspection,autonomous driving,and reverse engineering.In the real scene,due to the complexity of the target object and the limitations of the equipment,the sensor can only scan in a limited field of view to obtain local point cloud data.It is necessary to stitch the local point clouds to obtain complete object surface information.There are many challenges in the process of 3D local point cloud stitching due to its own characteristics such as disorder,rotation invariance,internal point pair relationship and external reasons such as only some points between two point clouds have corresponding points.The point cloud stitching algorithm studied in this paper is to align the coordinates of the point cloud which only partially overlaps due to the different viewpoint,so as to achieve high precision point cloud stitching.The main work of this paper are as follows:(1)A point cloud stitching algorithm based on multi-resolution dynamic graph convolution is proposed.Aiming at the three basic characteristics and processing difficulties of point cloud disorder,rotation invariance and internal point pair relationship,a point cloud feature learning module based on multi-resolution dynamic graph convolution is designed.This module takes coordinate information and normal vector information as input,uses neighborhood points to build a point-to-point relationship,and enriches the information of each point in the point cloud through a multi-layer perceptron and a maximum pooling layer.In order to make full use of the topological information of the point cloud,a multi-resolution dynamic graph convolutional network is proposed to learn the point features,so as to obtain the point cloud features.On the processed ModelNet40 data set,the point cloud feature learning module is integrated into the RPMNet framework.From both quantitative and qualitative perspectives,it is found that the stitching accuracy is better than the mainstream feature learning algorithms PointNet and DGCNN.Experiments have verified the effectiveness of the algorithm.(2)A point cloud registration algorithm based on key point selection and a local point cloud iterative stitching algorithm based on soft matching are proposed.Aiming at the problem that only some points in the local point cloud have corresponding points,a point cloud registration algorithm based on key point selection is designed.The registration algorithm uses the attention mechanism to select points that are more likely to have corresponding points as the key points to be registered,and calculate the similarity between the key points of the two point clouds to obtain the matching probability.Based on the idea of soft matching,virtual points are synthesized by matching probability to form a set of corresponding points.In order to obtain more accurate rotation and translation matrices,the matching probability is still used to perform weighted singular value decomposition to solve the transformation matrix.In order to optimize the results,a local point cloud iterative stitching algorithm based on soft matching is proposed.Iteratively uses the point cloud registration algorithm based on key point selection to gradually align the source point cloud to the target point cloud and optimize the stitching accuracy.Finally,the ModelNet40 data set is also used to conduct point cloud stitching comparison experiments with mainstream algorithms.The errors of the rotation angle and translation matrix obtained by the stitching algorithm proposed in this paper are smaller than the current mainstream stitching algorithm.The experimental results verify that the algorithm proposed in this paper effectively improves the stitching accuracy.