Research On Point Cloud Matching Based On Deep Learning

Point cloud matching has important significance in computational graphics and vision.In order to change the current problems of low resolution,noise interference and only part of 3D data in some 3D point cloud data sets,and inspired by the recent success of neural networks,The main research of this article is point cloud matching based on deep learning.Its main core processing can be divided into three steps: key point acquisition,key point description and matching with descriptors.Aiming at these three steps,this paper proposes a method of obtaining key points to obtain a deep learning data set,so as to obtain key points;then use the convolutional neural network for deep learning training to obtain the matching descriptor.Finally,the RANSAC algorithm is used to complete the point cloud matching,which makes the point cloud matching more accurate.The main work and results of this article are as follows:(1)Aiming at the problems of low resolution,noisy interference and incompleteness of the current 3D data,a method for obtaining key point pairs in a point cloud is proposed,and a data set is established based on the key point pairs.First,reconstruct two RGB-D images into a three-dimensional point cloud,randomly sample the key points from them,and map the 3D position of the key points in the reconstruction to the two RGB-D frames that are in the camera's viewing cone and are not blocked In,directly use the corresponding position of the key point in the two RGB-D frames as the key point pair.The experimental results show that the key points obtained by the method in this paper are time-saving and labor-saving,and have high accuracy,which can be used as a data set for deep learning training.(2)Aiming at the problem that most of the descriptors are manually defined by humans,resulting in low accuracy and accuracy,a method for generating key point descriptors is proposed.First,this paper uses the deep learning framework Marvin to train the key points on the set and obtain the weight of the network;then,the key points are input into the network,so as to calculate the descriptor corresponding to the key points.The experimental results show that: the descriptor obtained by this method has a small error,which is much better than the descriptor obtained by the traditional method.(3)To solve the problems of low point cloud matching accuracy and long matching time,this paper uses RANSAC algorithm to achieve point cloud matching.Obtain a key point in the first point cloud,map it to the position in the RGB-D frame,and then reconstruct and merge the RGB-D frame into the second point cloud,then the key point in the first point cloud and the key point at this position in the second point cloud form a more accurate matching pair,Thus,the time of eliminating mismatches is reduced.After multiple iterations,the rigid transformation between point clouds is estimated,and the point cloud matching is accomplished quickly through the transformation matrix.The experimental results show that: using RANSAC algorithm for point cloud matching,the matching accuracy and running time have been significantly improved.