Pretreatment And Registration Technology Of 3D Scattered Point Cloud Data

Computer vision is one of the important challenging research fields and a comprehensive and multidisciplinary course. 3D reconstruction is a hot topic in the field of computer vision research currently. To get the 3D point cloud model of the target, the rotation and translation matrix of space rigid body transformation is solved for 3D point clouds under different perspectives, and point cloud with multiple perspectives can be merged into an integrated point cloud, namely point cloud registration technology. Point cloud registration technique is the core technology of 3D reconstruction, and is the research focus in the field of computer vision such as virtual reality, simulation design and the cultural relics digitalization. More and more commercial companies, such as Microsoft and Intel, provide cheap point cloud sampling equipment and develop disruptive human-computer interaction mode. A growing number of interactive design commercialization products has opened up a new era with great commercial value and practical significance. At the same time, the studies of this problem also can make rich and colorful the research field of computer vision with point cloud data of great theoretical significance. In light of the extraction problem of 3D model of scattered point cloud under the complicated environment, we aim at getting point cloud through the computer stereo matching technology and depth camera respectively; on the basis of the geometric information such as the normal vector and the mean curvature of point cloud, we conduct research on scattered point cloud filtering method based on the curvature and statistical methods; we study the initial point cloud registration based on different feature points setting the depth image feature point and the SIFT algorithm feature point as the research object; accurate registration algorithm of point cloud mainly includes the ICP algorithm and 3DNDT algorithm, and 3DNDT point cloud accurate registration algorithm is studied based on dynamically updating steps. In this paper, the related problems were studied and the main work and achievements are as follows:1. This paper thoroughly analyze the different acquisition methods of scattered point cloud under complex environment. The representative local matching algorithm and semi-global matching algorithm in stereo matching algorithms are studied. The SAD algorithm is adopted for local matching algorithm and SGBM algorithm for semi-global matching algorithm. The experiment proves that when SAD window is too small or too big, there are larger matching errors. The experimental results also show that this algorithm has good real-time performance and matching speed, but the precision is low. The matching cost calculation of SGBM algorithm takes the BT algorithm, and adds smooth constraints in energy formula. Experiments show that SGBM stereo matching algorithm is much better than SAD, and also has very good real-time performance. In conclusion, according to the low efficiency and accuracy of stereo matching algorithm, this paper introduces the method based on depth camera to get point cloud. Experimental results show the depth camera has faster detection speed and high precision.2. This paper analyze the defects of existing scattered point cloud filtering method under complex environment. Due to the density inhomogeneity, noise and outliers of the mass point cloud data, this paper introduced the calculation of mean curvature, and proposed CSF point cloud filtering based on curvature and statistical methods. Firstly, statistical analysis is conducted aiming at all the data points of scattered point cloud to calculate the mean value and variance of global distance and curvature of every point in scattered point cloud together with the average curvature; Then, a cube partition of scattered point cloud is done and curvature threshold judgment of all the data points inside the cube is made, in order to retain those data points of similar curvature within the cube, and partition is continued for those points in the cube which doesn’t meet the threshold and until meeting the threshold condition; all data points within the cube are represented by a center of gravity and point cloud data with uneven density can be homogenized through the grid filtering; finally, outliers are removed by the relationship between global distance threshold of the scattered point cloud, and the average distance between data point and its neighboring point. A large amount of the experiment results show that the algorithm is stable and reliable to quickly and efficiently to compress and filter the point cloud data, and greatly speed up the search point cloud.3. A point cloud initial configuration algorithm is proposed based on SIFT feature points, extending the SIFT operator of 2D images to a 3D point cloud space. First, the image and scale is localized, and the established feature points are selected using Gaussian difference formula which have invariance characteristic in rotation and scale-zooming. Second, specific location and scale is determined for all candidate points. Then the direction of key points is selected and the operation aiming at image data is replaced by position and scale of feature point keeping the invariance of operation. Finally, the gradient of key point within the scope is calculated to generate feature point descriptor. Through the calculated space differences between feature points and neighboring points of registration points cloud, a multi-dimensional histograms is formed to describe the k field geometric attributes of the feature points. Finally, the registration point is randomly selected from feature point histogram to calculate the relationship between point cloud rotation and translation.4. An initial consistent configuration algorithm for sampling point based on depth image is proposed. Edge detection is conducted for those places with depth mutation in neighboring regions of each depth image point, and depth image boundary is extracted for the four kinds of weights of each point in the depth image which represents the possibility of boundary in the up, down, left and right direction; The change of the surface of depth image and the direction of the boundary is calculated and main direction and curvature of the boundary point. Smoothing operation is done for the projection angle and weight of boundary point using Gaussian kernel function so as to extract feature points that we need; Through the calculated space differences between feature points and neighboring points of registering point cloud, a multi-dimensional histograms is formed to describe the k field geometric attributes of the feature points. Finally, the registration point is randomly selected from feature point histogram to calculate the relationship between point cloud rotation and translation.5. In view of the poor convergence of the 3D point cloud registration algorithm and easily local optimal, this paper presents an improved 3D norm distribution transform algorithm based on Newton’s method, hereinafter referred to as the NM-3DNDT. Scattered three-dimensional point cloud surface is described by first and second order derivative of piecewise smooth functions, then the point cloud space is divided into cube grid and the corresponding mean and covariance matrix is calculated. In order to decrease the complexity of the algorithm, the Gaussian function is introduced to approximate logarithmic likelihood function, the parameters in the probability density function is simplified for 3DNDT algorithm, calculating Hessian matrix and the gradient vector by Jacobi matrix and rotating translational equation. We also put forward an improved linear search to update the step length of Newton iteration algorithm and ensure the convergence of algorithm after a few iterations. Finally, the algorithm is compared with other algorithms and simulation experiments is conducted. The results show that the proposed algorithm can obtain good effect of registration with high registration precision and efficiency.