Research On Point Cloud Registration Method Based On Harris Features

In recent years,due to the rapid development of 3D laser scanning technology,point cloud registration has become one of the hot issues in the research of 3D laser scanning technology.Point cloud registration is a key step in point cloud data processing which affects subsequent data processing and modeling accuracy.Existing point cloud registration methods are mainly divided into feature-based registration and feature-free registration which can better realize point cloud registration from different perspectives.However,the existing algorithms still have some deficiencies.The feature-based registration method needs to extract the point cloud features,the registration accuracy depends on the feature extraction accuracy and it takes a lot of time to determine the corresponding relationship at the same time,so the registration efficiency is not high.The featureless methods mainly include the nearest point iteration algorithm,the normal distribution transformation algorithm and the four-point fast robust matching algorithm.The first two algorithms depend on the initial position of the point cloud and registration cannot be realized if the initial position is not good,thus the efficiency is relatively low.The third algorithm belongs to the global registration algorithm and does not depend on the initial position of the point cloud.However,when the point cloud itself has symmetry,erroneous registration results will easily occur.In this paper,aiming at some deficiencies in the above point cloud registration,,the main research contents and results are as follows1)To solve the problem that the bilateral filtering algorithm cannot filter large-scale noise,the bilateral filtering algorithm is optimized in this paper.The algorithm uses the mean and variance of the average distance of neighboring points to filter out the discrete noise points of the point cloud,reduces the influence of the discrete noise points on the bilateral filtering results and then uses the bilateral filtering algorithm to filter out the local noise of the point cloud.Experiments show that the algorithm in this paper can effectively remove noise and better preserve the point cloud features.On the basis of denoising,point cloud data is compressed by voxel grid compression method.The algorithm can effectively reduce the amount of point cloud data while ensuring the topological structure of point cloud and can effectively improve the efficiency of later data processing.2)Aiming at the problems of slow Harris corner extraction speed and manual threshold setting,Harris algorithm is optimized in this paper.The algorithm uses the curvature information of the point cloud to pre-screen corners,which increased computational efficiency of corner response values.Simultaneously the corner response threshold is replaced by the variance of point cloud curvature,which realizes the adaptation of the corner response threshold and ensures the real-time corner extraction,a new method for feature-based point cloud registration is provided.3)In order to solve the problem of low efficiency of the normal distribution transformation algorithm Heisenberg matrix,the quasi-Newton iteration method is used to optimize the normal distribution transformation algorithm.The algorithm avoids solving the second derivative and the inverse of the Heisenberg matrix.At the same time,the algorithm can ensure that the Heisenberg matrix is always the direction in which the objective function value decreases.The experimental results show that the registration efficiency is improved on the premise of ensuring the accuracy of the normal distribution transformation algorithm.4)In order to solve the registration problem of super four-point fast robust matching algorithm under symmetric point cloud,Harris feature is used to optimize the algorithm.The algorithm extracts Harris features from the source point cloud,to highlights the local features of the point cloud.The feature point cloud is used as the source point base to effectively improve the accuracy of matching points with the same name.Experiments show that algorithms in this paper can quickly and accurately realize the initial registration of point clouds and provide a good initial position for fine registration.