Method and device for automatic registration of forestry unmanned aerial vehicle point cloud and ground point cloud without target

Abstract

The application discloses a forestry unmanned aerial vehicle point cloud and ground point cloud non-target automatic registration method and device. The method comprises the following steps: S1, extracting two kinds of data source breast height position point cloud slice data, respectively using DBSCAN clustering algorithm to extract single tree coordinate information; S2, determining the corresponding relationship of the two based on the starburst diagram composed of the single tree coordinates in the two kinds of data sources, and solving the initial transformation matrix; S3, using ICP algorithm to perform fine matching on the two kinds of data, and obtaining the final transformation matrix, S4, and then performing space coordinate system conversion on the ground laser scanning point cloud based on the unmanned aerial vehicle laser point cloud. In the case that complete forest structure information cannot be obtained from a single point cloud data source, the initial transformation matrix is obtained by extracting single tree coordinates and using the starburst diagram composed of the geometric relationship between the single trees, the transformation matrix is calculated by combining ICP fine registration, and finally the transformation from the ground point cloud coordinate system to the unmanned aerial vehicle point cloud coordinate system is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing, and in particular to a forestry unmanned aerial vehicle point cloud and ground-based point cloud no-target automatic registration method. BACKGROUND

[0002] Forests are the main body of the terrestrial ecosystem on earth, with rich vegetation types and complex spatial structures. Forest parameters are important indicators for describing the spatial structure of forests, which can directly reflect the growth conditions of forests. Complete forest structure reconstruction is the key prerequisite for accurate extraction of forest parameters, which helps to improve the estimation accuracy of forest parameters such as biomass and carbon storage, and has important significance for studying forest ecological processes, biodiversity, and sustainable forest management.

[0003] Light Detection and Ranging (LiDAR) as an active remote sensing technology has incomparable advantages in accurate reconstruction of forest structure. Among them, ground-based and unmanned aerial vehicle LiDAR are two kinds of LiDAR systems widely used in forest surveys. Ground-based LiDAR can obtain millimeter-level precision structure information of the scene, but due to tree shading, the information of the tree canopy in the sample plot is difficult to obtain from the ground-based LiDAR scan. Unmanned aerial vehicle LiDAR can quickly realize large-scale forest observation from the air and accurately obtain canopy structure information (such as crown width, tree height, etc.). However, due to the shading of the forest canopy, a single platform LiDAR system cannot obtain complete forest structure information. Ground-based LiDAR has limitations in collecting canopy structure information, and unmanned aerial vehicle LiDAR cannot observe the complete understory structure, which affects the accurate extraction of forest structure parameters. In order to change the current situation of incomplete information acquisition of a single data source, it is of great significance to develop an automatic splicing method for multi-source LiDAR point clouds of forests. Fusing ground-based and unmanned aerial vehicle (air-ground) LiDAR data can realize the complementary advantages of the two platform systems, which plays an important role in reconstructing the complete forest sample plot structure and realizing accurate forest surveys. SUMMARY

[0004] The present application aims to provide a forestry unmanned aerial vehicle point cloud and ground-based point cloud no-target automatic registration method, which is mainly applicable to multi-source point cloud data in forestry, solves the problem of coordinate unification of unmanned aerial vehicle point cloud and ground-based point cloud multi-source data, and improves the efficiency and stability of registration.

[0005] According to a first aspect, the embodiments of the present application provide a forestry unmanned aerial vehicle point cloud and ground-based point cloud no-target automatic registration method, which comprises the following steps:

[0006] S1, for the forest point cloud from two data sources of a ground-based laser scanning system and a UAV laser scanning system, extract the point cloud slice of a preset height position of each data source, process the point cloud slice by using a density-based noise spatial clustering algorithm (Density-Based Spatial Clustering of Applications with Noise, referred to as DBSCAN), and extract the single tree coordinate information of each data source.

[0007] S2, based on the starburst diagram composed of the single tree coordinate information of each data source, determine the point pairs having a corresponding relationship between the two data sources, and solve an initial transformation matrix based on the point pairs.

[0008] S3, based on the initial transformation matrix, match the point pairs by using an iterative closest point (Iterative Closest Point, referred to as ICP) algorithm, and obtain a final transformation matrix.

[0009] S4, according to the final transformation matrix, perform spatial coordinate system conversion on the ground-based point cloud based on the UAV point cloud.

[0010] In one embodiment, step S1 specifically includes:

[0011] S11, obtain forest point cloud data by using a UAV laser scanning system and a ground-based laser scanning system.

[0012] S12, divide the point clouds of the two data sources into ground point clouds and non-ground point clouds by using a cloth simulation filtering algorithm (Cloth Simulation Filter, referred to as CSF), and obtain normalized point clouds based on the respective ground point clouds.

[0013] S13, obtain slice point clouds by intercepting normalized point clouds according to a single tree preset height position.

[0014] S14, extract single tree coordinate information from the slice point clouds of the two data sources respectively by using a DBSCAN clustering algorithm.

[0015] Exemplarily, step S14 specifically includes:

[0016] S141, process the point cloud slice of each data source by using a DBSCAN clustering method to obtain a point cloud cluster of each single tree.

[0017] S142, take the average value of the point cloud plane coordinates in the point cloud cluster as the coordinate position of each single tree.

[0018] In one embodiment, step S2 specifically includes:

[0019] S21, randomly select the coordinate information of a single tree from the single tree coordinate information of two data sources, and connect each single tree in the homologous data in turn to form a starburst diagram.

[0020] S22, repeat the step S21 until the starburst diagram of all single trees in each data source is constructed.

[0021] S23, determine the correspondence of the starburst diagram through the shape of the starburst diagram in two data sources.

[0022] S24, based on the correspondence, solve the initial transformation matrix by using the least square fitting method. For example, the step S23 specifically includes:

[0023] S231, convert the starburst diagram of each tree in the unmanned aerial vehicle point cloud into a polar coordinate system with each tree as the center and the longest side in the starburst diagram as the polar axis, and also convert the starburst diagram of each tree in the ground-based point cloud into the corresponding polar coordinate.

[0024] S232, randomly select a starburst diagram from the starburst diagram of the ground-based point cloud.

[0025] S233, match the starburst diagram of the unmanned aerial vehicle point cloud with the most coordinates from the starburst diagram of the ground-based point cloud, and count the maximum value N of the number of coordinates.

[0026] S234, repeat the steps S232 and S233 until all ground-based starburst diagrams are traversed.

[0027] S235, select the best three results from the matching results of all starburst diagrams as the point pairs for determining the correspondence between the ground-based point cloud and the unmanned aerial vehicle point cloud.

[0028] In one embodiment, the step S3 specifically includes:

[0029] S31, take the feature set of the unmanned aerial vehicle point cloud as the initial point cloud feature set, and construct the corresponding Kd tree.

[0030] S32, search the nearest neighbor point in the initial point cloud feature set based on the Kd tree; the target point cloud feature set is the feature set of the coarse matching ground-based point cloud determined based on the ground-based point cloud and the initial transformation matrix.

[0031] S33, calculate the similarity of the nearest neighbor point pair, select a threshold t, and if the similarity is less than the threshold, record the matching point pair to obtain a matching point pair set.

[0032] S34, calculating the rotation matrix and the translation vector by singular value decomposition method according to the matching point pair set, calculating the coordinates of the initial point cloud feature set after the first transformation by the rotation matrix and the translation vector, setting the iteration number n and the convergence threshold, until the iteration number is n or the iteration converges, then obtaining the point set of the initial point cloud after the final transformation, and the convergence condition is the distance variance of the matching point set after two transformations.

[0033] Further, the judgment criterion of the nearest neighbor point in step S32 is the Euclidean distance.

[0034] Further, the similarity judgment criterion in step S33 includes the minimum eigenvalue of the covariance matrix of the point set in the respective neighborhood of the nearest neighbor point pair and the length of the triangle formed by the respective neighborhood of the nearest neighbor point pair, and the point pairs with most matching errors can be removed. Optionally, the default threshold t is set to 0.05 and 0.2 respectively.

[0035] Further, the translation vector is calculated by the rotation matrix, and the formula is R = VU T . The iteration number n is set to 20 by default, the convergence threshold is set to 1e -6 , and the iteration convergence condition is that the root mean square of the distance of the matching point set after two transformations is lower than the convergence threshold.

[0036] Further, the rotation matrix calculation method is actually a least squares optimization problem. A new transformation matrix is repeatedly calculated until the iteration number reaches the threshold or the iteration converges. In the experiment, the time consumption of ICP registration can be controlled by setting a smaller iteration number threshold.

[0037] According to the second aspect, a device for automatic registration of forestry unmanned aerial vehicle point cloud and ground point cloud without target is provided, and the device comprises: an extraction module configured to extract a point cloud slice at a preset height position of each data source from point clouds of a forest area from two data sources of a ground-based laser scanning system and an unmanned aerial vehicle laser scanning system, process the point cloud slice by using a density-based spatial clustering algorithm for noise DBSCAN, and extract single tree coordinate information of each data source; an initial transformation matrix determination module configured to determine a point pair having a corresponding relationship between the two data sources based on a starburst diagram formed by the single tree coordinate information of each data source, and solve an initial transformation matrix based on the point pair; a final transformation matrix acquisition module configured to match the point pair by using an ICP algorithm based on the initial transformation matrix, and acquire a final transformation matrix; and a spatial coordinate system conversion module configured to perform spatial coordinate system conversion on the ground-based point cloud based on the unmanned aerial vehicle point cloud according to the final transformation matrix.

[0038] In one embodiment, the extraction module specifically comprises: a scanning unit for acquiring forest area point cloud data for a UAV laser scanning system and a ground-based laser scanning system; a normalization unit for using cloth simulation filtering algorithm CSF to respectively divide the point clouds of the two data sources into ground point clouds and non-ground point clouds, and obtaining normalized point clouds based on the respective ground point clouds; a slicing unit for intercepting the normalized point clouds according to single-tree preset height positions to obtain sliced point clouds; and an extraction unit for using DBSCAN clustering algorithm to extract single-tree coordinate information from the sliced point clouds of the two data sources respectively.

[0039] Further, the extraction unit is specifically configured to: process the point cloud slices of each data source by using DBSCAN clustering method to obtain point cloud clusters of each single tree; and take the average value of the point cloud plane coordinates in the point cloud cluster as the coordinate position of each single tree.

[0040] In one embodiment, the initial transformation matrix determination module specifically comprises: a starburst diagram construction unit for randomly selecting coordinate information of a single tree from the single-tree coordinate information of the two data sources respectively, sequentially connecting each single tree in the homologous data to form a starburst diagram, and constructing a starburst diagram for all single trees in each data source until the construction is completed; a correspondence relationship determination unit for determining the correspondence relationship of the starburst diagrams through the shapes of the starburst diagrams in the two data sources; and a matrix solving unit for solving the initial transformation matrix by using least square fitting method based on the correspondence relationship.

[0041] Exemplarily, the correspondence relationship determination unit specifically comprises: converting each-tree starburst diagram in the UAV point cloud into a polar coordinate system with each tree as the center and the longest side in the starburst diagram as the polar axis, and simultaneously converting each-tree starburst diagram in the ground-based point cloud into a corresponding polar coordinate; randomly selecting a starburst diagram from the starburst diagrams in the ground-based point cloud; matching the starburst diagram with the starburst diagram in the UAV point cloud having the largest number of coordinates in common from all the starburst diagrams in the UAV point cloud, and counting the maximum value N of the number of coordinates in common; repeating the above two operations: randomly selecting a starburst diagram from the starburst diagrams in the ground-based point cloud, matching the starburst diagram with the starburst diagram in the UAV point cloud having the largest number of coordinates in common, and counting the number N of coordinates in common in the best matching, until all the starburst diagrams in the ground-based point cloud are traversed; and selecting the best three results from the matching results of all the starburst diagrams as the point pairs for determining the correspondence relationship of the ground-based point cloud and the UAV point cloud.

[0042] In one embodiment, the final transformation matrix acquisition module is specifically configured to: on the basis of the initial transformation matrix, match the point pairs by using iterative closest point ICP algorithm to obtain the final transformation matrix.

[0043] Exemplarily, the final transformation matrix obtaining module specifically comprises: a Kd tree constructing unit, which takes the feature set of the unmanned aerial vehicle point cloud as an initial point cloud feature set and constructs a corresponding Kd tree in a k-dimensional space; a nearest neighbor searching unit, which searches for the nearest neighbors in the initial point cloud feature set based on the Kd tree; the target point cloud feature set is a feature set of a coarse matching ground point cloud determined based on the initial transformation matrix; a matching point pair set obtaining unit, which calculates the similarity of the nearest neighbor pairs, selects a threshold t, and if the similarity is less than the threshold, records the matching point pairs and obtains a matching point pair set; and a final transformation matrix obtaining unit, which calculates a rotation matrix and a translation vector based on the matching point pair set by using a singular value decomposition method, and then obtains the coordinates of the initial point cloud feature set after one transformation; the number of iterations n and the convergence threshold are set until the number of calculations is n or the iteration converges, and then the point set of the initial point cloud after the final transformation is obtained, and the convergence condition is the root mean square distance of the matching point sets after two transformations.

[0044] Further, in the nearest neighbor searching unit, the judgment criterion of the nearest neighbor is the Euclidean distance.

[0045] Further, the similarity judgment criteria in the matching point pair set obtaining unit are the minimum eigenvalue of the covariance matrix and the length of the triangle side in the neighborhood of the nearest neighbor, respectively.

[0046] Further, the translation vector is calculated based on the rotation matrix, and the formula is R = VU T . The number of iterations n is 20 by default, the convergence threshold is 1e -6 , and the iteration convergence condition is that the root mean square distance of the matching point sets after two transformations is lower than the convergence threshold.

[0047] Further, the rotation matrix calculation method is actually a least squares optimization problem. A new transformation matrix is repeatedly calculated until the number of iterations reaches the threshold or the iteration converges. In the experiment, the time consumption of ICP registration can be controlled by setting a smaller iteration threshold.

[0048] In one embodiment, the spatial coordinate system conversion module is specifically configured to: convert the ground point cloud based on the unmanned aerial vehicle point cloud according to the final transformation matrix.

[0049] In summary, the forestry unmanned aerial vehicle point cloud and ground point cloud automatic registration method and device disclosed in the embodiments of the present specification are adopted. First, the breast height position point cloud slice data of the two data sources is extracted, and the single tree coordinate information is extracted by using the DBSCAN clustering algorithm respectively; then the corresponding relationship of the two is determined based on the starburst diagram composed of the single tree coordinates in the two data sources, and the initial transformation matrix is solved; then the ICP algorithm is used for fine matching of the two data to obtain the final transformation matrix, and then the ground laser scanning point cloud is converted in the spatial coordinate system based on the unmanned aerial vehicle laser point cloud. In the case that the complete forest structure information cannot be obtained from a single point cloud data source, the initial transformation matrix is obtained by extracting the single tree coordinates and using the starburst diagram composed of the geometric relationship between the single trees, and the transformation matrix is calculated by combining the ICP fine registration, so that the ground point cloud coordinate system is finally transformed into the unmanned aerial vehicle point cloud coordinate system, and the algorithm efficiency and registration accuracy of the ground point cloud and unmanned aerial vehicle point cloud fine registration are improved. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 is a flowchart of a forestry unmanned aerial vehicle point cloud and ground point cloud automatic registration method provided by the embodiments of the present application;

[0051] Figure 2 is a registration flowchart of a forestry unmanned aerial vehicle point cloud and ground point cloud automatic registration method provided by the embodiments of the present application;

[0052] Figure 3 is a flowchart of the construction of a starburst diagram and initial transformation provided by the embodiments of the present application;

[0053] Figure 4 is a flowchart of the fine registration of feature point clouds by using the ICP matching method in the embodiments of the present application;

[0054] Figure 5 is a structural schematic diagram of a forestry unmanned aerial vehicle point cloud and ground point cloud automatic registration device provided by the embodiments of the present application. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical scheme and advantages of the present application more clear, the technical scheme in the embodiments of the present application will be described below with reference to the drawings.

[0056] In the description of the embodiments of the present application, the words such as "exemplary", "for example", "for instance" or the like are used to mean example, illustration or description. Any embodiment or design scheme described as "exemplary", "for example" or "for instance" in the embodiments of the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. In fact, the words such as "exemplary", "for example", "for instance" or the like are intended to present the relevant concept in a specific manner.

[0057] In the description of the embodiments of the present application, the term "and / or", merely describes an association relationship of associated objects, indicates that there can be three relationships, for example, A and / or B, which can represent three cases of A alone, B alone, and A and B existing at the same time. In addition, unless otherwise specified, the term "multiple" means two or more. For example, multiple systems refer to two or more systems, and multiple screen terminals refer to two or more screen terminals.

[0058] In addition, the terms "first", "second" are only for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more features. The terms "include", "contain", "have" and their variants mean "include but not limited to", unless otherwise specifically emphasized.

[0059] The embodiments of the present application provide a forestry unmanned aerial vehicle point cloud and ground point cloud no-target automatic registration method, which can realize the transformation of the ground point cloud coordinate system to the unmanned aerial vehicle point cloud coordinate system, and improve the algorithm efficiency and registration accuracy of the ground point cloud and the unmanned aerial vehicle point cloud.

[0060] Figure 1 is a flowchart of a forestry unmanned aerial vehicle point cloud and ground point cloud no-target automatic registration method provided by the embodiments of the present application. Figure 2 is a registration flowchart of a forestry unmanned aerial vehicle point cloud and ground point cloud no-target automatic registration method provided by the embodiments of the present application, and each step shown in Figure 1 and Figure 2 , are described in detail. Figure 1

[0061] In step S1, for the forest area point cloud from two data sources of a ground laser scanning system and an unmanned aerial vehicle laser scanning system, the point cloud slice of the preset height position of each data source is extracted, the point cloud slice is processed by using a density-based noise point spatial clustering algorithm DBSCAN, and the single tree coordinate information of each data source is extracted. The sub-steps of this step are described as follows:

[0062] S11, acquiring forest area point cloud data by using an unmanned aerial vehicle laser scanning system and a ground laser scanning system.

[0063] S12, using cloth simulation filtering algorithm CSF to divide the point cloud of the two data sources into ground point cloud and non-ground point cloud respectively, and obtaining normalized points based on the ground point cloud of each.

[0064] S13, intercepting normalized point cloud according to the preset height position of single tree to obtain slice point cloud. ​

[0065] In some embodiments, the preset height position of a single log can be breast height.

[0066] S14, use the DBSCAN clustering algorithm to extract individual tree coordinate information from the slice point clouds of the two data sources respectively.

[0067] In some embodiments, the DBSCAN clustering method is used to process the point cloud slices of each data source to obtain a point cloud cluster for each individual tree; the average value of the point cloud plane coordinates in the point cloud cluster is used as the coordinate position of each individual tree.

[0068] Based on the above, the coordinate information of individual trees can be extracted from the point clouds of forest areas from both ground-based laser scanning systems and UAV laser scanning systems.

[0069] In step S2, based on the star map composed of the single-tree position coordinate information from each different data source, corresponding point pairs with corresponding relationships between the two data sources are determined, and the initial transformation matrix is ​​solved based on the point pairs. Figure 3 This is a schematic diagram illustrating the construction and initial transformation of a starburst diagram according to an embodiment of this application. The following is a flowchart illustrating the process. Figure 3 To describe the sub-steps of this step, such as Figure 3 As shown:

[0070] S21: Randomly select the coordinate information of a single tree from the single tree coordinate information of the two data sources, and connect them sequentially to each single tree in the same source data to form a star map.

[0071] S22, repeat steps S21 until the star charts of all individual trees in each data source are constructed.

[0072] S23, determine the correspondence of star charts by the shapes of star charts from the two data sources.

[0073] S24. Based on the aforementioned correspondence, the initial transformation matrix is ​​solved using the least squares fitting method.

[0074] In some embodiments, step S23 specifically comprises: S231, converting the starburst diagram of each tree in the unmanned aerial vehicle point cloud into a polar coordinate system with each tree as the center and the longest side in the starburst diagram as the polar axis, while also converting the starburst diagram of each tree in the ground point cloud into the corresponding polar coordinate; S232, randomly selecting a starburst diagram from the starburst diagrams of the ground point cloud; S233, matching the starburst diagram of the ground point cloud with the starburst diagram of the unmanned aerial vehicle point cloud with the largest number of coordinates from all the starburst diagrams of the unmanned aerial vehicle point cloud, and counting the maximum value N of the number of coordinates; S234, repeating the steps of S232 and S233 until all the starburst diagrams of the ground are traversed; S235, selecting the best three results from the matching results of all the starburst diagrams as the point pairs for determining the correspondence between the ground point cloud and the unmanned aerial vehicle point cloud. Based on the correspondence, the least square fitting method is used to solve the initial transformation matrix.

[0075] From the above, the initial transformation matrix can be obtained.

[0076] In step S3, based on the initial transformation matrix, the ICP algorithm is used to match the point pairs to obtain the final transformation matrix, Figure 4 A flowchart for precisely matching feature point clouds using the ICP matching method is provided for the embodiments of the present application. The following describes the sub-steps of this step according to the flowchart: Figure 4 The sub-steps of this step are described as follows:

[0077] S31, taking the feature set of the unmanned aerial vehicle point cloud as the initial point cloud feature set, and constructing the corresponding k-dimensional space Kd tree.

[0078] It can be understood that the Kd tree is a data structure for organizing points in a three-dimensional Euclidean space, which can realize fast search and nearest neighbor search. The Kd tree of the target point cloud is constructed to search for the nearest points in the initial point cloud corresponding to the target point cloud.

[0079] S32, searching for the nearest points in the initial point cloud feature set corresponding to the target point cloud feature set based on the Kd tree; the target point cloud feature set is the feature set of the coarsely matched ground point cloud determined based on the initial transformation matrix.

[0080] In some embodiments, the criterion for judging the nearest points is the Euclidean distance. It can be understood that the Euclidean distance, also known as the Euclidean distance, is the most common distance measure, which measures the absolute distance between two points in a multi-dimensional space.

[0081] S33, calculating the similarity of the nearest point pairs, selecting a threshold t, and if the similarity is less than the threshold, recording the matching point pairs to obtain a set of matching point pairs.

[0082] In some embodiments, the similarity judgment criteria are the minimum eigenvalue of the covariance matrix and the length of the nearest neighbor triangle edge of the nearest points, respectively. Most of the matching error point pairs can be removed to improve the algorithm accuracy and efficiency in ICP precise matching.

[0083] In some embodiments, the default threshold values are set to 0.05 and 0.2, respectively.

[0084] In step S34, according to the set of matching points, the rotation matrix and the translation vector are calculated by singular value decomposition method, and then the coordinates of the initial point cloud feature set after the first transformation are obtained; the number of iterations n and the convergence threshold are set, and the point set after the final transformation of the initial point cloud is obtained until the number of calculations is n or the iteration converges, and the convergence condition is the root mean square distance of the matching point set of the two transformations.

[0085] In some embodiments, the translation vector is calculated from the rotation matrix, and the formula is R = VU T The coordinates of the initial point cloud feature set after the first transformation are calculated from the rotation matrix and the translation vector, the number of iterations n (default is 20) and the convergence threshold (default is 1e -6 ) are set, and the point set after the final transformation of the initial point cloud is obtained until the number of calculations is n or the iteration converges. The iteration convergence condition is that the root mean square distance of the matching point set of the two transformations is lower than the convergence threshold.

[0086] For example, the above rotation matrix calculation method is actually a least squares optimization problem. A new transformation matrix is repeatedly calculated until the number of iterations reaches a threshold or the iteration converges. In the experiment, by setting a smaller iteration threshold, the time consumption of ICP registration can be controlled.

[0087] From the above, the final transformation matrix can be obtained.

[0088] In step S4, according to the final transformation matrix, the ground point cloud is converted in the spatial coordinate system based on the unmanned aerial vehicle point cloud. In the case that the complete forest structure information cannot be obtained from a single point cloud data source, the present application extracts the coordinate information of single trees in the two data sources by preprocessing the ground point cloud and the unmanned aerial vehicle point cloud through CSF filtering, slicing, DBSCAN clustering, etc. Secondly, the single tree coordinates are used to form a starburst diagram of each tree, and the starburst diagram is converted into a polar coordinate system for fast searching of a matching starburst diagram. The number of single trees with consistent coordinates in the starburst diagrams of the two data sources is compared to determine the starburst diagrams that match each other. The least squares fitting method is used to solve the initial transformation matrix. Finally, the ICP matching algorithm based on kd-tree optimized matching point search is used to speed up the search for matching points and improve the algorithm efficiency and registration accuracy of the ground point cloud and the unmanned aerial vehicle point cloud.

[0089] Corresponding to the above method, the present application also discloses an automatic registration device. Figure 5 is a structural schematic diagram of a forestry unmanned aerial vehicle point cloud and ground point cloud target-free automatic registration device provided by the present application, as shown in Figure 5As shown, the device 1400 comprises the following modules: an extraction module 1410 configured to, for forest area point clouds from both a ground-based laser scanning system and a UAV laser scanning system, extract a point cloud slice of a preset height position of each data source, process the point cloud slice by using a density-based spatial clustering algorithm for noise DBSCAN, and extract single tree coordinate information of each data source; an initial transformation matrix determination module 1420 configured to determine a point pair having a corresponding relationship between the two data sources based on a starburst diagram formed by the single tree coordinate information of each data source, and solve an initial transformation matrix based on the point pair; a final transformation matrix acquisition module 1430 configured to, based on the initial transformation matrix, match the point pair by using an ICP algorithm to obtain a final transformation matrix; and a spatial coordinate system conversion module 1440 configured to perform spatial coordinate system conversion on the ground-based point cloud based on the UAV point cloud according to the final transformation matrix.

[0090] In one embodiment, the extraction module 1410 specifically comprises: a scanning unit configured to obtain forest area point cloud data for a UAV laser scanning system and a ground-based laser scanning system; a normalization unit configured to respectively divide the point clouds of the two data sources into ground point clouds and non-ground point clouds by using a cloth simulation filtering algorithm CSF, and obtain normalized point clouds based on the respective ground point clouds; a slicing unit configured to obtain slice point clouds by cutting the normalized point clouds according to single tree preset height positions; and an extraction unit configured to respectively extract single tree coordinate information from the slice point clouds of the two data sources by using a DBSCAN clustering algorithm.

[0091] Further, the extraction unit 1410 is specifically configured to: process the point cloud slice of each data source by using a DBSCAN clustering method to obtain a point cloud cluster of each single tree; and take an average value of point cloud plane coordinates in the point cloud cluster as a coordinate position of each single tree.

[0092] In one embodiment, the initial transformation matrix determination module 1420 specifically comprises: a starburst diagram construction unit configured to randomly select coordinate information of a single tree from single tree coordinate information of the two data sources, sequentially connect each single tree in the homologous data to form a starburst diagram, and construct a starburst diagram of all single trees in each data source until the construction is completed; a corresponding relationship determination unit configured to determine a corresponding relationship of the starburst diagrams by shapes of the starburst diagrams in the two data sources; and a matrix solving unit configured to solve an initial transformation matrix by using a least square fitting method based on the corresponding relationship.

[0093] Exemplarily, the correspondence determining unit specifically comprises: converting the starburst diagram of each tree in the unmanned aerial vehicle point cloud into a polar coordinate system with each tree as the center and the longest side in the starburst diagram as the polar axis, and also converting the starburst diagram of each tree in the ground point cloud into the corresponding polar coordinate; randomly selecting a starburst diagram from the starburst diagram of the ground point cloud; matching the starburst diagram of the ground point cloud with the starburst diagram of the unmanned aerial vehicle point cloud with the largest number of coordinates, and counting the maximum value N of the number of coordinates; repeating the above steps: selecting a starburst diagram from the starburst diagram of the ground point cloud, matching the starburst diagram of the ground point cloud with the starburst diagram of the unmanned aerial vehicle point cloud with the largest number of coordinates, and counting the number N of coordinates in the best matching, until all the starburst diagrams of the ground are traversed; selecting the best three results from the matching results of all the starburst diagrams as the point pairs for determining the correspondence between the ground point cloud and the unmanned aerial vehicle point cloud.

[0094] In an embodiment, the final transformation matrix obtaining module 1430 is specifically configured to: on the basis of the initial transformation matrix, match the point pairs by using an ICP algorithm to obtain a final transformation matrix.

[0095] Exemplarily, the final transformation matrix obtaining module 1430 specifically comprises: a Kd tree constructing unit configured to: take the feature set of the unmanned aerial vehicle point cloud as an initial point cloud feature set, and construct a corresponding Kd tree; a nearest point searching unit configured to: search for the nearest points in the initial point cloud feature set based on the Kd tree; the target point cloud feature set is a feature set of a coarsely matched ground point cloud determined based on the initial transformation matrix; a matching point pair set obtaining unit configured to: calculate the similarity of the nearest point pairs, select a threshold t, if the similarity is less than the threshold, record the matching point pairs, and obtain a matching point pair set; and a final transformation matrix obtaining unit configured to: calculate a rotation matrix and a translation vector by using a singular value decomposition method according to the matching point pair set, and then obtain the coordinates of the initial point cloud feature set after one transformation; set an iteration number n and a convergence threshold, until the iteration number is n or the iteration converges, and then obtain the point set of the initial point cloud after the last transformation, and the convergence condition is the root mean square distance of the matching point sets after two transformations.

[0096] Further, in the nearest point searching unit, the judgment criterion of the nearest points is the Euclidean distance.

[0097] Further, the similarity judgment criteria in the matching point pair set obtaining unit are the minimum eigenvalue of the covariance matrix and the length of the triangle side of the neighborhood of the nearest points, respectively.

[0098] Further, the translation vector is calculated from the rotation matrix, and the formula is R = VU T . The iteration number n is 20 by default, the convergence threshold is 1e -6 , and the iteration convergence condition is that the root mean square distance of the matching point sets after two transformations is lower than the convergence threshold.

[0099] In an embodiment, the spatial coordinate system conversion module 1440 is specifically configured to: perform spatial coordinate system conversion on the ground-based point cloud based on the UAV point cloud according to the final transformation matrix.

[0100] In the above embodiments, the entire or part of the embodiments can be implemented through software, hardware, firmware or any combination thereof. When implemented through software, the entire or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded on a computer, the computer executes the computer program instructions, and thus the entire or part of the flow or function according to the embodiments of the present application is generated. The computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable apparatus. The computer instructions can be stored in or transmitted from a computer readable storage medium. The computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through a wired manner, such as a coaxial cable, an optical fiber, a digital subscriber line (DSL), or a wireless manner, such as infrared, wireless, microwave, etc. The computer readable storage medium can be any available medium that can be accessed by a computer or a data storage computing device, such as a server, a data center, etc., which includes one or more available media sets. The available media can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape, etc.), an optical medium (such as a DVD, etc.), or a semiconductor medium, such as a solid state disk (SSD), etc.

[0101] It can be understood that various numerical numbers involved in the embodiments of the present application are only distinguished for convenience of description, and are not used to limit the scope of the embodiments of the present application.

Claims

1. A method for automatic registration of forestry drone point cloud and ground point cloud without target, characterized in that, The method comprises the following steps: S1, for forest area point clouds from two data sources of a ground-based laser scanning system and an unmanned aerial vehicle laser scanning system, extracting point cloud slices at preset height positions of each data source, processing the point cloud slices by using a density-based spatial clustering algorithm for noise DBSCAN, and extracting single tree coordinate information of each data source; S2, determining point pairs having a corresponding relationship between the two data sources based on starburst diagrams constituted by the single tree coordinate information of each data source, and solving an initial transformation matrix based on the point pairs; S2 specifically comprises: S21, randomly selecting coordinate information of a single tree from single tree coordinate information of the two data sources respectively, sequentially connecting each single tree in the homologous data to constitute a starburst diagram; S22, repeating the step S21 until the starburst diagrams of all single trees in each data source are constructed; S23, determining the corresponding relationship of the starburst diagrams through the shapes of the starburst diagrams in the two data sources; S24, solving the initial transformation matrix by using a least square fitting method based on the corresponding relationship; S23 specifically comprises: S231, converting the starburst diagram of each tree in the unmanned aerial vehicle point cloud into a polar coordinate system with each tree as the center and the longest side in the starburst diagram as the polar axis, and also converting the starburst diagram of each tree in the ground-based point cloud into the corresponding polar coordinate; S232, randomly selecting a starburst diagram from the starburst diagrams of the ground-based point cloud; S233, matching the starburst diagram of the unmanned aerial vehicle point cloud with the same number of coordinates most from all the starburst diagrams of the unmanned aerial vehicle point cloud, and counting the maximum value N of the number of the same coordinates; S234, repeating the steps S232 and S233 until all the starburst diagrams of the ground-based point cloud are traversed; S235, selecting the best three results from the matching results of all the starburst diagrams as the point pairs for determining the corresponding relationship of the ground-based point cloud and the unmanned aerial vehicle point cloud; S3, matching the point pairs by using an iterative closest point ICP algorithm on the basis of the initial transformation matrix to obtain a final transformation matrix; S4, performing space coordinate system conversion on the ground-based point cloud based on the unmanned aerial vehicle point cloud according to the final transformation matrix.

2. The method of claim 1, wherein, Step S1 specifically comprises: S11, acquiring forest area point cloud data by using an unmanned aerial vehicle laser scanning system and a ground-based laser scanning system; S12, dividing the point clouds of the two data sources into ground point clouds and non-ground point clouds respectively by using a cloth simulation filtering algorithm CSF, and obtaining normalized point clouds based on the ground point clouds respectively; S13, obtaining slice point clouds by intercepting the normalized point clouds according to preset height positions of single trees; S14, extracting single tree coordinate information from the slice point clouds of the two data sources respectively by using a DBSCAN clustering algorithm.

3. The method of claim 2, wherein, In step S14, specifically: S141, processing the point cloud slices of each data source by using a DBSCAN clustering method to obtain point cloud clusters of each single tree; S142, taking the average value of the point cloud plane coordinates in the point cloud cluster as the coordinate position of each single tree.

4. The method of claim 1, wherein, Step S3 specifically comprises: S31, taking a feature set of the unmanned aerial vehicle point cloud as an initial point cloud feature set, and constructing a corresponding k-dimensional space Kd tree; S32, searching the nearest points in the target point cloud feature set in the initial point cloud feature set based on the Kd tree; the target point cloud feature set is a feature set of a coarse matching ground point cloud determined based on the ground point cloud and the initial transformation matrix; S33, calculating the similarity of the nearest point pairs, selecting a threshold t, and recording the matching point pairs if the similarity is less than the threshold to obtain a matching point pair set; S34, calculating the rotation matrix and the translation vector by using the singular value decomposition method according to the matching point pair set, and then obtaining the coordinates of the initial point cloud feature set after one-time transformation; setting the iteration number n and the convergence threshold, and obtaining the point set of the initial point cloud after final transformation until the calculation number is n or the iteration converges, and the convergence condition is the distance variance of the matching point sets of two times of transformation.

5. The method of claim 4, wherein, In step S32, the judgment criterion of the nearest points is the Euclidean distance.

6. The method of claim 4, wherein, In step S33, the judgment criteria of the similarity are the minimum eigenvalue of the covariance matrix and the length of the triangle side of the nearest point neighborhood.

7. A targetless automatic registration device for forestry UAV point clouds and ground-based point clouds, characterized in that, The device comprises: The extraction module is configured to extract point cloud slices of a preset height position of each data source for point clouds in a forest area from two data sources of a ground-based laser scanning system and a UAV laser scanning system, process the point cloud slices by using a density-based spatial clustering algorithm DBSCAN, and extract single tree coordinate information of each data source. The initial transformation matrix determination module is configured to: randomly select coordinate information of a single tree from single-tree coordinate information of two data sources respectively, sequentially connect each single tree in homologous data to form a starburst diagram, and repeat the above operation until starburst diagrams of all single trees in each data source are constructed; determine a corresponding relationship of the starburst diagrams according to shapes of the starburst diagrams in the two data sources; and solve an initial transformation matrix by using a least square fitting method based on the corresponding relationship. The final transformation matrix acquisition module is configured to: acquire a final transformation matrix by matching the point pairs by using an iterative closest point (ICP) algorithm on a basis of the initial transformation matrix. The spatial coordinate system conversion module is configured to: convert the spatial coordinate system of the unmanned aerial vehicle point cloud to the spatial coordinate system of the ground-based point cloud according to the final transformation matrix.

Images