Research On The Construction Of Point Cloud Local Coordinate System Based On Deep Learning

3D point cloud is one of the important 3D data expression methods,which has extensive and important applications in robot vision,automatic driving,3D reconstruction and 3D face recognition.Its related research is one of the hot spots in the field of computer vision.3D point cloud surface matching algorithm is one of the core tasks of 3D computer vision,and also an important basis of 3D registration and 3D reconstruction algorithm.3D point cloud surface matching algorithm can be roughly divided into two categories,which are based on local feature and global feature respectively.Under the same conditions,the local feature algorithm is more robust to occlusion,deletion and Gaussian noise than the global feature algorithm.The key of local feature algorithm is to construct a repeatable and robust point cloud local reference coordinate system(LRF).In order to achieve this goal,many calculation methods based on the construction of local reference coordinate system of point cloud have been proposed by researchers in the last decade,which can be mainly divided into two types:analysis of covariance(CA)and point spatial distribution(PSD),but their repeatability and robustness are limited.This paper mainly studies the traditional PSD method and deep learning method to construct point cloud local reference coordinate system,and applies the constructed point cloud local reference coordinate system to specific point cloud core applications.The main research contents and achievements of this paper are as follows:Firstly,the SES-Net network model is proposed for building local reference coordinate system of point cloud by studying the construction method of local reference coordinate system of point cloud,twin network Siamese-Net,T-Net network and SE attention mechanism.Secondly,based on the 2D feature and discriminative 3D feature of local block of point cloud,this paper studies and implements the construction method of local reference coordinate system of point cloud by combining SES-Net network model and PSD method.The repeatability and robustness of the constructed point cloud local reference coordinate system are evaluated.The experimental results show that the point cloud local reference coordinate system constructed by the method used in this paper is more repeatable and robust than the point cloud local reference coordinate system constructed by most traditional PSD methods.Thirdly,the 3D reconstruction system of point cloud is designed and implemented based on the local reference coordinate system of point cloud constructed by the 3D feature of point cloud and SES-Net network model.