Point Cloud Alignment Algorithm Based On Feature Detection And Feature Description

In recent years,with the continuous development and progress of science and technology,3D reconstruction technology has been increasingly widely used in all walks of life,and has gradually penetrated into reverse engineering,virtual reality and biomedical fields.Since 3D scanning equipment can only obtain the point cloud data of the object under a fixed perspective,it is necessary to scan the object under multiple perspectives to obtain the complete data of the object surface.The process of converting the point cloud data scanned from multiple perspectives to the same coordinate system is point cloud alignment,which can be divided into coarse alignment and fine alignment.The coarse alignment is to summarily overlap two far separated point clouds.Fine alignment is based on coarse alignment to further optimize alignment posture and reduce errors between models.This paper mainly studies the coarse alignment algorithm of point cloud.Starting from two aspects of feature detection and feature description,it proposes some improved methods to the existing coarse alignment algorithm.The main work includes:(1)An improvement of the traditional Super4 PCS algorithm.To solve the problem of low alignment efficiency of the Super4 PCS algorithm,a Super4 PCS point cloud alignment algorithm combined with ISS feature detection is adopted.The number of feature points extracted by ISS feature detection is small and the repeatability is high.Taking the feature point set as the search range of the consistent four-point set in the Super4 PCS algorithm can not only improve the alignment efficiency,but also ensure the alignment accuracy.(2)An improvement to the handmade FPFH descriptor.For single scale FPFH descriptor is sensitive to noise data,using a multi-scale FPFH character description of point cloud alignment algorithm,by using a different search radius build descriptor weighted joining together,form a multi-scale descriptors,can not only improve the difference between sex and descriptive descriptor,also increases the robustness to noise data.(3)An improvement of 3DMatch algorithm based on deep learning.(1)In 3DMatch algorithm,random selection of points leads to poor repeatability of points to be matched.ISS feature detection is used instead of random sampling to select points to be matched,which not only ensures the alignment accuracy of point cloud,but also greatly reduces the number of sampling points;(2)In3DMatch algorithm,big training sample data quantity lead to high computational overhead,this paper proposes a 3D-2D mapping strategy combined with the idea of PCA,which can keep space information at the same time as reduce the dimension of the training sample,reduce the cost of the follow-up,and simplifies the siamese network,further improve the training efficiency of the network.