CN-122023480-A - Terrain change area point cloud registration method driven by spatial similarity and multi-threshold optimization

CN122023480ACN 122023480 ACN122023480 ACN 122023480ACN-122023480-A

Abstract

The invention belongs to the technical field of point cloud registration, and discloses a point cloud registration method for a terrain change area driven by spatial similarity and multi-threshold optimization. The method comprises the steps of firstly carrying out initial registration and voxel segmentation on two-stage ground point cloud data, secondly calculating the space geometrical characteristics of each voxel, constructing a point cloud space similarity measurement model based on GMM, primarily classifying all voxels into deformed voxels and stable voxels, recursively subdividing the deformed voxels to screen the stable voxels therein, defining all marked stable voxels as initial stable voxels, and then further screening a stable region by adopting a self-adaptive dynamic threshold distance constraint strategy according to the distance distribution characteristics from the point contained in the initial stable voxels to a fitting plane. And finally, performing fine registration based on the extracted stable region point cloud. The method effectively solves the problem of point cloud registration dilemma caused by unknown position of the terrain stabilization area in the surface deformation monitoring, and provides reliable technical support for geological disaster monitoring.

Inventors

LI YANYAN
Guo Fangjia
CHEN CHUANFA
LUO MINGLIANG

Assignees

山东科技大学
西华师范大学

Dates

Publication Date: 20260512
Application Date: 20260114

Claims (8)

1. A method for registering point clouds of a terrain change area driven by spatial similarity and multi-threshold optimization is characterized by comprising the following steps: I. performing initial registration and voxel segmentation on two-stage ground point cloud data of a source point cloud and a target point cloud; II, calculating the space geometric characteristics of each voxel, calculating the characteristic similarity of the voxels in the source point cloud and the voxels in the corresponding target point cloud, and estimating a similarity threshold by combining a Gaussian Mixture Model (GMM); Then, preliminarily classifying all voxels into deformed voxels and stabilized voxels based on a similarity threshold; recursively subdividing the deformed voxels to screen stable voxels in the deformed voxels; Finally, defining all marked stabilization voxels as initial stabilization voxels; step III, based on the initial stabilizing voxels extracted in the step II, firstly performing plane fitting on each initial stabilizing voxel in the target point cloud to obtain a voxel fitting plane corresponding to each initial stabilizing voxel in the target point cloud; Then calculating the distance from each point in each initial stabilizing voxel in the source point cloud to a voxel fitting plane determined by the corresponding initial stabilizing voxel in the target point cloud, and further screening a stabilizing region by adopting a self-adaptive dynamic threshold distance constraint strategy; And IV, based on the stable region point cloud extracted in the step III, performing fine registration from the source point cloud to the target point cloud.
2. The method for registering a point cloud of a terrain variation area driven by spatial similarity and multi-threshold optimization according to claim 1, wherein, The step II specifically comprises the following steps: II.1, calculating feature similarity between voxels in a source point cloud and voxels in a corresponding target point cloud; Firstly, using the distance from the point in the voxel to the centroid as the weight, establishing a weighted covariance matrix, and calculating the eigenvalues lambda 1 、λ 2 and lambda 3 , and the eigenvectors n 1 、n 2 and n 3 corresponding to lambda 1 、λ 2 and lambda 3 respectively The spatial feature descriptors F, F based on the feature values and the feature vectors are designed according to the expression formula: In the formula, For linearity, measuring whether the point cloud in the voxel is in obvious linear distribution; The voxel local planarity is marked for the flatness; reflecting the degree of dispersion of the point cloud spatial distribution in the voxels as the degree of dispersion; The total variance represents the discrete degree of the point cloud in the voxel in each direction; the anisotropy is used for reflecting the non-uniformity of the point cloud in the voxels in different directions; describing uncertainty of local structures in the voxels for feature entropy; for trace, quantifying the geometric structure intensity within the voxel; the verticality represents the deviation degree of the voxel local normal vector and the vertical direction; And The maximum curvature and the minimum curvature respectively represent the maximum and minimum bending degrees of the shape of the inner surface of the voxel; then, cosine similarity is adopted Feature similarity between voxels in the source point cloud and voxels in the corresponding target point cloud is measured, and a calculation formula is as follows: In the formula, And Respectively representing the spatial characteristics of voxels corresponding to the source point cloud and the target point cloud; II.2. adaptive estimation of feature similarity threshold with GMM When (when) When the corresponding earth surface is not deformed, otherwise, the corresponding earth surface is deformed, and all voxels are further divided into two types, namely deformed voxels and stabilized voxels; performing further recursion subdivision on the deformed voxels obtained in the step II.2, firstly subdividing the octree of deformed voxels into sub-voxels, and then performing feature similarity estimation and stable voxel identification on each sub-voxel again according to the steps II.1 to II.2; And (3) repeating the step II.1 and the step II.2 until the number of point clouds in the deformed voxels is smaller than a preset threshold value or reaches a preset layer number, and finally, defining all marked stabilizing voxels as initial stabilizing voxels.
3. The method for registering a point cloud of a terrain variation area driven by spatial similarity and multi-threshold optimization according to claim 2, wherein, In the step II.2, the GMM is combined to adaptively estimate the characteristic similarity threshold value The method comprises the following steps: First, a probability density function is used The distribution of the data is described and, The expression is as follows: Wherein x is cosine similarity K is the number of mixed components; gaussian mixture weights for the kth gaussian distribution; is the kth Gaussian distribution with the average value of Covariance matrix is ; Then, adopting an expectation maximization algorithm to carry out iterative estimation, wherein the expectation maximization algorithm comprises an E step and an M step; wherein, the step E is that under the condition of a given parameter theta, the posterior probability of each data point x i belonging to each Gaussian class k is calculated , The formula is: In the formula, Represents the probability density function value of data point x i at the kth gaussian component, A probability density function value for data point x i at the j-th gaussian component; m step is based on the posterior probability obtained in E step Updating the GMM parameters; Repeating the step E and the step M until convergence to obtain optimized parameters and obtain a Gaussian mixture model , The formula is: In the formula, Represents the kth gaussian distribution; When the GMM fit is two components, the similarity threshold Taking the probability density function intersection points of two Gaussian distributions: solving the equation to obtain an intersection point x value which is the optimal similarity threshold value 。
4. The method for registering a point cloud of a terrain variation area driven by spatial similarity and multi-threshold optimization according to claim 1, wherein, The step III specifically comprises the following steps: Based on the initial stabilizing voxels extracted in the step II, fitting a point plane contained in each initial stabilizing voxel in the target point cloud, calculating the distance d i between each point in each initial stabilizing voxel in the source point cloud and the fitting plane of the initial stabilizing voxel in the corresponding target point cloud, and then obtaining an optimal distance threshold t s according to a self-adaptive dynamic threshold distance constraint strategy; If the distance d i is smaller than or equal to the optimal distance threshold t s , the point is marked as a stable point, finally, the proportion of stable points in each initial stable voxel is counted, and if the proportion is not smaller than the set proportion, the initial stable voxel is judged as a stable voxel.
5. The method for point cloud registration of a terrain variation area driven by spatial similarity and multi-threshold optimization as set forth in claim 4, The self-adaptive dynamic threshold distance constraint strategy specifically comprises the following steps: firstly, defining an objective function; The root mean square error is selected as an optimization objective function: Where n represents the number of corresponding points that can be successfully matched, And Respectively representing coordinates of a kth corresponding point between a source point cloud and a target point cloud, wherein R is a rotation matrix, and t is a translation vector; then, the initial distance threshold t s0 is set as the average value of the distances from each point in each initial stable voxel in the source point cloud to the initial stable voxel fitting plane in the corresponding target point cloud: In the formula, Representing all point coordinates in the source point cloud voxel, Representing parameters of the target point cloud after the voxel fitting plane is corresponding to the target point cloud, wherein N is the number of stabilizing voxels; finally, setting iteration times, and carrying out iterative updating on the distance threshold t s ; And (3) obtaining a corresponding stable region result every time the distance threshold t s is updated in an iteration way, carrying out registration calculation on root mean square errors of source point clouds and target point clouds on the obtained stable region result, taking a distance threshold t s corresponding to the minimum root mean square error as an optimal distance threshold t s , and obtaining a stable region refinement result by utilizing the optimal distance threshold t s .
6. The method for point cloud registration of a terrain variation area driven by spatial similarity and multi-threshold optimization according to claim 5, wherein, The iterative updating of the distance threshold t s is specifically as follows: Constructing a descending gradient by using the difference between the stable proportions of two adjacent iterations, simultaneously introducing an attenuation coefficient, calculating the adjustment quantity of the distance threshold t s on the basis of the attenuation coefficient, and then updating the distance threshold t s by using gradient descent: Wherein lambda is a proportionality coefficient for enhancing the influence of the stable proportion on the updating of the distance threshold t s ; The proportion of stable voxels in each iteration process, gamma is the attenuation coefficient, epsilon is a constant term preventing denominator from being 0.
7. The method for registering a point cloud of a terrain variation area driven by spatial similarity and multi-threshold optimization according to claim 1, wherein, The step IV specifically comprises the following steps: And (3) finishing the fine registration of the two-stage ground point cloud data based on the stable region point cloud extracted in the step (III), and applying the solved conversion parameters to the whole unregistered point cloud for realizing the coordinate conversion from the source point cloud to the target point cloud.
8. The utility model provides a topography change district point cloud registration system of space similarity and multi-threshold optimization drive which characterized in that includes following module: The preprocessing module is used for carrying out initial registration and voxel segmentation on the two-period ground point cloud data of the source point cloud and the target point cloud; the initial stable region screening module is used for calculating the space geometric characteristics of each voxel, calculating the characteristic similarity of the voxels in the source point cloud and the voxels in the corresponding target point cloud, and estimating a similarity threshold by combining a Gaussian Mixture Model (GMM); Then, preliminarily classifying all voxels into deformed voxels and stabilized voxels based on a similarity threshold; recursively subdividing the deformed voxels to screen stable voxels in the deformed voxels; Finally, defining all marked stabilization voxels as initial stabilization voxels; The stable region selection module is used for carrying out plane fitting on each initial stable voxel in the target point cloud to obtain a voxel fitting plane corresponding to each initial stable voxel in the target point cloud; Then calculating the distance from each point in each initial stabilizing voxel in the source point cloud to a voxel fitting plane determined by the corresponding initial stabilizing voxel in the target point cloud, and further screening a stabilizing region by adopting a self-adaptive dynamic threshold distance constraint strategy; And the fine registration module is used for carrying out fine registration from the source point cloud to the target point cloud on the stable region point cloud extracted by the stable region fine selection module.

Description

Terrain change area point cloud registration method driven by spatial similarity and multi-threshold optimization Technical Field The invention belongs to the technical field of point cloud registration, and particularly relates to a point cloud registration method for a terrain change area driven by spatial similarity and multi-threshold optimization. Background The three-dimensional laser point cloud is a core data source for monitoring the surface deformation with the observation advantages of full elements, high precision, abundant details and the like. However, due to the constraints of device performance difference, data acquisition errors, local independent coordinate systems and the like, the multi-time phase point cloud often has the problems of different references, system deviation and the like. Therefore, the multi-time phase point cloud needs to be strictly registered before the monitoring of the terrain change. In recent years, scholars at home and abroad develop a great deal of research around point cloud registration, and the research is divided into coarse registration and fine registration according to registration procedures. The coarse registration (also called global registration) mainly starts from any initial gesture, and calculates initial conversion parameters between two-period point clouds, so as to provide reasonable initial values for subsequent fine registration. The prior coarse registration method mainly comprises registration based on geometric primitives such as points, lines, planes and the like, 4PCS and the like. Fine registration (also called local registration) is based on coarse registration, and conversion parameters are further optimized to obtain high-precision registration results, and representative methods include ICP, NDT, various variants thereof, and the like. However, in various geological disaster scenarios, the position of the terrain stability area is unknown in advance, so that the point cloud registration algorithm neglecting the surface deformation is easy to sink into local optimization. Thus, accurate identification of terrain stability is critical. The current strategy is that firstly, a source point cloud and a target point cloud are respectively divided into a plurality of voxels, and then a stable region is identified according to the information entropy difference, the centroid distance, the point-plane distance and other indexes of the source point cloud and the target point cloud among the corresponding voxels. However, when the method is used for processing non-uniform surface deformation, local deformation characteristics of the surface in the voxels are ignored, and part of deformation areas are easily misjudged as stable areas. In summary, aiming at the challenge that the existing research is difficult to capture the surface stable region and the registration accuracy is poor, a new point cloud registration method for the terrain variation region is needed to be proposed. Disclosure of Invention The invention aims to provide a point cloud registration method for a terrain change area driven by spatial similarity and multi-threshold optimization, which can accurately capture a ground surface stable area to achieve accurate registration. In order to achieve the above purpose, the present invention adopts the following technical scheme: a method for registering point clouds of a terrain change area driven by spatial similarity and multi-threshold optimization comprises the following steps: I. performing initial registration and voxel segmentation on two-stage ground point cloud data of a source point cloud and a target point cloud; II, calculating the space geometric characteristics of each voxel, calculating the characteristic similarity of the voxels in the source point cloud and the voxels in the corresponding target point cloud, and estimating a similarity threshold by combining a Gaussian Mixture Model (GMM); Then, preliminarily classifying all voxels into deformed voxels and stabilized voxels based on a similarity threshold; recursively subdividing the deformed voxels to screen stable voxels in the deformed voxels; Finally, defining all marked stabilization voxels as initial stabilization voxels; step III, based on the initial stabilizing voxels extracted in the step II, firstly performing plane fitting on each initial stabilizing voxel in the target point cloud to obtain a voxel fitting plane corresponding to each initial stabilizing voxel in the target point cloud; Then calculating the distance from each point in each initial stabilizing voxel in the source point cloud to a voxel fitting plane determined by the corresponding initial stabilizing voxel in the target point cloud, and further screening a stabilizing region by adopting a self-adaptive dynamic threshold distance constraint strategy; And IV, based on the stable region point cloud extracted in the step III, performing fine registration from the source point cloud to the targe