Multi-information airborne laser radar point cloud registration method based on GICP

Abstract

The invention provides a GICP-based multi-information airborne laser radar point cloud registration method. The method comprises the following steps of 1, calculating a normal vector of each local point cloud through a principal component analysis method; step 2, performing coarse registration on the local point cloud to be registered through a sampling consistency initial registration method; step 3, calculating an error function of the point cloud after coarse registration; 4, solving the error function to obtain an optimal value of a rigid body transformation matrix T *; and step 5, according to the optimal value of the rigid body transformation matrix T *, performing pose transformation on the local point cloud to be registered to complete registration. According to the invention, for the autonomous landing situation of the unmanned aerial vehicle, the registration precision and convergence speed of airborne laser point cloud are improved, and the method has the advantages of high precision and fast convergence, and is suitable for a real-time landing scene.

Description

Technical Field

[0001] The present invention relates to a method for registering lidar point clouds, and particularly to a multi-information airborne lidar point cloud registration method based on GICP. Background Art

[0002] The information provided in this section is only background information related to the present disclosure, and it is not necessarily prior art.

[0003] Due to the complexity and unknown nature of the ground conditions, the autonomous landing of an unmanned aerial vehicle (UAV) is a task that is very prone to accidents. The core problem is to select a safe landing point. UAVs usually have high requirements for the terrain of the landing point. If the terrain is uneven, accidents will occur. The mainstream method is to use airborne sensors for depth estimation, such as monocular cameras, binocular cameras, or ultrasonic radars, etc., to obtain the relative distance between the ground and the UAV, as well as the slope and roughness of the terrain, and find a flat ground as the landing point. However, due to the lack of reliable distance sensors and algorithms, these methods have poor generalization ability in unfamiliar environments.

[0004] Lidar is a non-contact active technology for quickly obtaining a three-dimensional dense point cloud of the object surface. Its principle is the time-of-flight measurement method, that is, a laser signal is emitted towards the target, and the relative distance between the sensor and the target is calculated by using the round-trip time. Airborne lidar is one of the geospatial information detection technologies, and it is a system that integrates three technologies: laser, global positioning system, and inertial navigation system. Specifically, the lidar is carried on an aircraft, and the precise three-dimensional spatial coordinates of the ground object surface are obtained by combining them. Compared with other detection means, due to the advantages of high brightness, monochromaticity, and good directivity of the laser itself, and because it uses a laser with a shorter wavelength, it can quickly and efficiently obtain relevant information of small targets with high resolution. As one of the most effective and reliable means for collecting terrain data, airborne lidar is widely used in remote sensing detection of various scenarios.

[0005] Point cloud registration is to splice the point cloud data of the target part scanned from multiple angles to obtain the complete three-dimensional point cloud of the target in a unified coordinate system.

[0006] In the prior art, there are problems with the accuracy and convergence speed of the point cloud registration algorithm for processing airborne lidar point cloud registration.

[0007] It should be noted that the information disclosed in the above background art section is only used to strengthen the understanding of the background of the present disclosure, and thus may include information that does not constitute prior art known to those of ordinary skill in the art. Summary of the Invention

[0008] Objective of the Invention: The technical problem to be solved by the present invention is to provide a multi-information airborne lidar point cloud registration method based on GICP in view of the deficiencies of the prior art.

[0009] To solve the above technical problem, the present invention discloses a multi-information airborne lidar point cloud registration method based on GICP, including the following steps:

[0010] Step 1, calculating the normal vectors of each local point cloud through the principal component analysis method;

[0011] Step 2, performing rough registration on the local point cloud to be registered through the sampling consistency initial registration method;

[0012] Step 3, calculating the error function of the point cloud after rough registration;

[0013] Step 4, solving the error function to obtain the optimal value of the rigid body transformation matrix T * ;

[0014] Step 5, performing pose transformation on the local point cloud to be registered according to the optimal value of the rigid body transformation matrix T * to complete the registration.

[0015] Further, the calculation of the normal vectors of each local point cloud in Step 1 includes:

[0016] Step 1-1, setting any local point cloud as P n , where n = 1, 2,..., N, is the local point cloud number, and N is the number of local point clouds; setting a point in the local point cloud P n as p a , where a is the point number; obtaining the k-nearest neighbor points of the point p a as n i , where i = 1, 2,..., K, k is the nearest neighbor point number, and K is the preset number of nearest neighbor points;

[0017] Step 1-2, using the point set N i composed of all k-nearest neighbor points to fit the tangent plane at the point p a , which is expressed as follows:

[0018] Ax + By + Cz = D (D > 0)

[0019] where (x, y, z) represents the points in the tangent plane, A, B, and C are the components of the normal vector of the tangent plane in the x, y, and z coordinate axes directions, and D is the distance between the tangent plane and the origin;

[0020] Step 1-3, defining the objective function J as the distance between each point in the point set N and the fitted tangent plane, minimizing it, and transforming it into an extreme value problem, which is expressed as follows:

[0021]

[0022] Among them, min represents minimization, and (x i , y i , z i ) represents the coordinates of the i-th point, and λ is the Lagrange operator;

[0023] Steps 1-4: Take the partial derivatives of the above objective function to obtain:

[0024]

[0025] Among them, the intermediate variables and are calculated as follows:

[0026]

[0027] Calculate the covariance matrix S of the fitted tangent plane as follows:

[0028]

[0029] Among them, p i is the i-th point in the point cloud P n , and is the geometric center point of the point cloud P n .

[0030] Step 1-5: Perform eigenvalue decomposition on the covariance matrix S to calculate the eigenvalues and eigenvectors, specifically as follows:

[0031] Minimize the objective function to obtain the eigenvector n corresponding to the minimum eigenvalue of the covariance matrix S i as the normal vector at the point p i .

[0032] Furthermore, the rough registration described in Step 2 includes:

[0033] Step 2-1: Arbitrarily select two point clouds from all the local point clouds to be roughly registered, and set them as the source point cloud P0 and the target point cloud Q0 respectively;

[0034] Step 2-2: Select s sample points from the source point cloud P0 and the target point cloud Q0 respectively, and ensure that the initial distance between any two sample points in the same point cloud does not exceed the preset minimum distance d min ;

[0035] Step 2-3: For each sample point in the source point cloud P0, randomly select the corresponding point in the target point cloud Q0;

[0036] Step 2-4: Calculate the rigid body transformation matrix based on the sample points and corresponding points, and calculate the distance error of the point cloud and the quality for evaluating the current registration transformation.

[0037] Step 2-5: Repeat Step 2-1 to Step 2-3 until the value of the distance error sum of the point cloud is minimized. At this time, the rigid body transformation matrix is the final rigid body transformation matrix, completing the rough registration of the source point cloud P0 and the target point cloud Q0, and obtaining the roughly registered point clouds P and Q.

[0038] Step 2-6: Repeat Step 2-2 to 2-5 until the rough registration of all local point clouds is completed.

[0039] Further, the minimum distance d min in Step 2-2 is set to 0.05.

[0040] Further, calculating the error function of the point cloud after rough registration in Step 3 includes:

[0041] Define the error function as a probability distribution model. Let the roughly registered point clouds P and Q be the source point cloud and the target point cloud to be registered, and they follow a Gaussian distribution, expressed as follows:

[0042]

[0043] where, p i and q i represent the i-th points in the source point cloud P and the target point cloud Q, and represent the sets of points in the two point clouds, and are the covariance matrices of the two point clouds. Let the rigid body transformation matrix T * be:

[0044]

[0045] Then the corresponding relationship between the source point cloud P and the target point cloud Q is:

[0046]

[0047] According to the GICP point cloud registration method, the error equation is obtained as:

[0048] e i = q i - T * p i

[0049] where, e i represents the error; according to the probability model, the probability distribution of the error e i is obtained and the rigid body transformation matrix T is solved by the maximum likelihood estimation method.* , and simplify the result to obtain the error function as follows:

[0050]

[0051] where, represents the T value that maximizes * .

[0052] Furthermore, the error function for calculating the coarsely registered point cloud described in step 3 further includes:

[0053] introduce the La * a * b * color information of the local point cloud and the echo intensity information during airborne lidar detection, calculate the feature information covariance Σ d and mean μ p , and use them to recalculate the covariance matrix of the point cloud and update the error function with it.

[0054] Furthermore, the recalculation of the covariance matrix of the point cloud described in step 3 includes:

[0055] Represent a single point p in any point cloud P j as follows:

[0056]

[0057] where j is the point number, is the position coordinate information, x j , y j and z j are the coordinate values respectively, is the point cloud feature information, α I is the intensity channel weight, I i is the echo intensity information of the point cloud, α c is the color channel weight, L i , a i and b i are the La * a * b * color information of the point cloud;

[0058] For any point p j , after searching for neighboring points through the kd-tree method and projecting them onto a plane perpendicular to the normal, the projection is obtained, which is represented as follows:

[0059]

[0060] where u j is p jProjection on a plane perpendicular to the normal is the position coordinate information is the point cloud feature information, U p is the projection of the point cloud P on a plane perpendicular to the normal; after the projection transformation, U p The overall covariance of is set to Σ w ;

[0061] For the point cloud feature information Introduce the Gaussian kernel function and the feature information measurement covariance Λ, and calculate the kernel weight w for the projection u j ∈U, specifically as follows: j Specifically as follows:

[0062]

[0063] Calculate the feature information covariance Σ j and the mean μ d based on the kernel weight w, specifically as follows: p Specifically as follows:

[0064]

[0065] Normalize according to the feature information covariance and calculate the correlation coefficient matrix Ω as follows:

[0066]

[0067] where; represents the set of real numbers; the covariance matrix obtained is:

[0068]

[0069] where ∈ is a constant representing the covariance along the normal.

[0070] Furthermore, the intensity channel weight α I is set to 0.25, and the color channel weight α c is set to 0.75.

[0071] Furthermore, for the error function solved in step 4, that is, the BFGS quasi-Newton method is used to solve the error function obtained in step 3 to obtain the transformation matrix T * .

[0072] Furthermore, the local point cloud described in step 1 is that the airborne lidar of the unmanned aerial vehicle scans the landing ground from different angles to obtain multiple independent and partially overlapping point clouds.

[0073] Beneficial effects:

[0074] 1. In view of the scenario of the autonomous landing of an unmanned aerial vehicle (UAV) and the characteristics of the terrain point cloud obtained by the airborne lidar, the color information and echo intensity information of the point cloud are incorporated into the registration algorithm, effectively improving the problems of poor registration accuracy and slow convergence speed of the traditional point cloud registration algorithm for processing airborne lidar point clouds.

[0075] 2. The designed point cloud registration algorithm of the present invention has the advantages of high accuracy and fast convergence in the real-time landing scenario. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] The following further specifically describes the present invention in conjunction with the drawings and specific embodiments, and the above and / or other advantages of the present invention will become clearer.

[0077] Figure 1 is the overall flow schematic diagram of the present invention.

[0078] Figure 2 is the schematic diagram of calculating the normal vectors of each local point cloud by the principal component analysis method.

[0079] Figure 3 is the schematic diagram of the specific registration effect provided by an embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0080] The overall idea of the present invention is as follows: During the landing process of the UAV, the airborne lidar scans the landing ground from different angles to obtain multiple independent and partially overlapping partial point clouds (local point clouds). Each part of the three-dimensional point cloud can be regarded as a rigid body, so the point cloud registration problem is transformed into the coordinate transformation problem of a three-dimensional rigid body, that is, calculating the optimal matching rule and performing coordinate transformation on the point cloud for registration. Therefore, point cloud registration can be regarded as a search problem for a certain point set in continuous space, and its solution can be converted into solving the corresponding transformation relationship. By calculating the coordinate transformation relationship, the multi-view three-dimensional point clouds are aligned to obtain the overall information of the target.

[0081] The present invention improves the traditional GICP algorithm (Generalized-ICP, a point cloud registration method) to obtain the complete three-dimensional point cloud of the target in a unified coordinate system in real time. In view of the problems of poor registration accuracy and slow convergence speed of the traditional point cloud registration algorithm for processing airborne lidar point clouds, the color information and echo intensity information of the point cloud are incorporated into the registration algorithm. The designed point cloud registration algorithm of the present invention has high accuracy and fast convergence and is applicable to the real-time landing scenario.

[0082] The specific technical solution is as follows: A multi-information airborne lidar point cloud registration method based on GICP is proposed, including the following steps:

[0083] Step 1: Calculate the normal vectors of each local point cloud through Principal Components Analysis (PCA).

[0084] Step 2: Coarsely register the point cloud to be registered through the Sample Consensus Initial Alignment (SAC-IA) algorithm.

[0085] Step 3: Introduce * a * b * color information and echo intensity information to calculate the error function.

[0086] Step 4: Solve the error function through the BFGS quasi-Newton method to obtain the optimal value of the rigid body transformation matrix T * .

[0087] Step 5: Perform pose transformation on the point cloud to be registered to complete the registration.

[0088] In one implementation, Step 1 includes:

[0089] For a point p i (i = 1, 2,..., n) in the three-dimensional point cloud P i , find its k nearest neighbors as N i (i = 1, 2,..., k). The algorithm uses the point set N i to fit the tangent plane Ax + By + Cz = D (D > 0) at p i , and defines the objective function as the distance between each point in N i and the fitted plane, minimizes it, and transforms it into an extreme value problem:

[0090]

[0091] Taking the partial derivative gives:

[0092]

[0093] where

[0094] The covariance matrix of the plane fitted by the k nearest neighbors is:

[0095]

[0096] Finally, perform eigenvalue decomposition on S to calculate its eigenvalues and eigenvectors. The PCA algorithm minimizes the objective distance function and takes the eigenvector n i corresponding to the smallest eigenvalue of the covariance matrix S as the normal vector at point p i .

[0097] In one implementation, step 2 includes:

[0098] 1) Select s sample points in the source point cloud P. To ensure that the Feature Point Feature Histogram (FPFH) features of the sample points are different as much as possible, the initial distance between any two of them should not exceed the preset minimum distance d min ;

[0099] 2) For each sampled point, points that meet the similar FPFH feature conditions can be found in the target point cloud Q, and the corresponding points of the sampled points in the point cloud Q are randomly selected from these points;

[0100] 3) Calculate the rigid body transformation matrix through the sampled points and the corresponding points, and evaluate the quality of the current registration transformation using the distance error sum of the point clouds.

[0101] Repeat the above steps until a set of optimal transformations that minimize the value of the distance error sum is finally found, and the corresponding transformation is the final rigid body transformation matrix.

[0102] In one implementation, step 3 includes:

[0103] The Generalized-ICP (GICP) algorithm defines the error function as a probability distribution model, assuming that the source point cloud P and the target point cloud Q to be registered follow a Gaussian distribution:

[0104]

[0105] In the formula and are the point cloud covariance matrices. For the corresponding point set and Since there is no scale transformation between the point clouds, let the rigid body transformation matrix have the corresponding relationship From this, the error equation in GICP is e i = q i - T * p i = f(T * ). According to the probability model, the probability distribution of e i can be obtained:

[0106]

[0107] Using the Maximum Likelihood Estimation (MLE) method, T * can be solved:

[0108]

[0109] Simplifying the above equation gives:

[0110]

[0111] A single point in the point cloud is the position coordinate information, is the point cloud feature information, α I is the intensity channel weight, I is the point cloud intensity information, α c is the color channel weight. Then, for the points in point clouds P and Q, the covariance C P and C Q are calculated respectively. Specifically, for a point p in point clouds P and Q j , after searching for the neighborhood points through the kd-tree algorithm and projecting them onto the plane perpendicular to the normal, the projection is the overall covariance of the transformed U p is ∑ w .

[0112] For the point cloud feature information p d , the Gaussian kernel function is introduced the covariance of the feature information measurement is Λ, and for u j ∈U, the kernel weights are calculated as follows:

[0113]

[0114] Using the kernel weights, the covariance Σ d and the mean μ p of the feature information can be calculated as:

[0115]

[0116] Normalizing through the overall covariance, let the correlation coefficient matrix reflects the correlation between the weighted data of the feature information and the overall data. Therefore, the covariance in the multi-information registration algorithm is:

[0117]

[0118] In the formula, ∈ is a constant representing the covariance along the normal.

[0119] In one implementation, step 4 uses the BFGS quasi-Newton method to solve the error function obtained in step 3 to obtain the transformation matrix T * .

[0120] In one implementation, step 5 transforms the point cloud to be registered according to the transformation matrix T * obtained in step 4, unifying it to the same coordinate system to complete the point cloud registration.

[0121] Example:

[0122] This application uses a specific example to further illustrate a multi - information airborne lidar point cloud registration method based on GICP proposed in this application. As Figure 1 shown, it includes the following steps:

[0123] Step 1: Calculate the normal vector of each local point cloud through the Principal Components Analysis (PCA), as Figure 2 shown;

[0124] For a point p i (i = 1, 2,..., n) in the three - dimensional point cloud P i , find its k - nearest neighbors N i (i = 1, 2,..., k). The algorithm uses the point set N i to fit the tangent plane Ax + By + Cz = D (D > 0) at p i , and defines the objective function as the distance between each point in N i and the fitted plane, minimizes it, and transforms it into an extreme - value problem:

[0125]

[0126] Taking the partial derivatives gives:

[0127]

[0128] In the formula

[0129] The covariance matrix of the plane fitted by the k - nearest neighbors is:

[0130]

[0131] Finally, perform eigenvalue decomposition on S to calculate its eigenvalues and eigenvectors. The PCA algorithm minimizes the objective distance function and takes the eigenvector n i corresponding to the minimum eigenvalue of the covariance matrix S as the normal vector of the point p i at that point.

[0132] Step 2: Coarsely register the point cloud to be registered through the Sample Consensus Initial Alignment (SAC - IA) algorithm. Taking the registration of two point clouds as shown in Figure 3 as an example, the specific method is as follows:

[0133] 1) Select 200 sample points from the source point cloud P. To ensure that the Feature Histograms of Point Features (FPFH) of the sample points are different as much as possible, the initial distance between any two of them should not exceed the preset minimum distance of 0.05 m.

[0134] 2) For each sampled point, points that meet the similar FPFH feature conditions can be found from the target point cloud Q, and the corresponding points of the sampled points in the point cloud Q are randomly selected from these points.

[0135] 3) Calculate the rigid body transformation matrix through the sampled points and the corresponding points, and evaluate the quality of the current registration transformation using the distance error sum of the point cloud.

[0136] Repeat the above steps until a set of optimal transformations that minimize the value of the distance error sum is finally found, and the corresponding transformation is the final rigid body transformation matrix.

[0137] Step 3: Introduce L * a * b * Calculate the error function using the color information and the echo intensity information.

[0138] Define the error function as a probability distribution model, assuming that the source point cloud P and the target point cloud Q to be registered follow a Gaussian distribution:

[0139]

[0140]

[0141] In the formula and are the point cloud covariance matrices. For the corresponding point sets and Since there is no scale transformation between the point clouds, let the rigid body transformation matrix There is a corresponding relationship Thus, the error equation in GICP is e i = q i - T * p i = f(T * ). According to the probability model, the probability distribution of e i can be obtained:

[0142]

[0143] The maximum likelihood estimation method (MLE) can be used to solve for T * :

[0144]

[0145] Simplifying the above equation gives:

[0146]

[0147] A single point in the point cloud is the position coordinate information, is the point cloud feature information, α I is the intensity channel weight, I is the point cloud intensity information, α c is the color channel weight. Then, for the points in point clouds P and Q, the covariance C P and C Q are calculated respectively. Specifically, for a point p in point clouds P and Q j , after searching for the neighborhood points through the kd-tree algorithm and projecting them onto the plane perpendicular to the normal, the projection is The overall covariance of the transformed U p is ∑ w .

[0148] For the point cloud feature information p d , the Gaussian kernel function is introduced The covariance of the feature information measurement is Λ, and for u j ∈U, the kernel weights are calculated as follows:

[0149]

[0150] Using the kernel weights, the covariance ∑ d and the mean μ p of the feature information can be calculated as:

[0151]

[0152] Normalize through the overall covariance, and let the correlation coefficient matrix reflects the correlation between the weighted feature information data and the overall data. Therefore, the covariance in the multi-information registration algorithm is:

[0153]

[0154] In the formula, ∈ is a constant representing the covariance along the normal.

[0155] Step 4: For the error function Solve this function through the BFGS quasi-Newton method to obtain the optimal value of the rigid body transformation matrix T * :

[0156]

[0157] Step 5: According to the transformation matrix T obtained in Step 4 *, the point cloud to be registered is transformed and unified into the same coordinate system to complete point cloud registration.

[0158] Corresponding to the point cloud to be registered and Transformation process:

[0159]

[0160] Use the sensor to collect multiple frames of data. The three-dimensional point cloud data of two frames and the transformation results are as Figure 3 shown. The two frames are represented by red and green respectively, with a total of 11,282 data points.

[0161] In specific implementation, the present application provides a computer storage medium and a corresponding data processing unit. Among them, the computer storage medium can store a computer program. When the computer program is executed by the data processing unit, it can run the invention content of a multi-information airborne lidar point cloud registration method based on GICP provided by the present invention and some or all of the steps in each embodiment. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0162] Those skilled in the art can clearly understand that the technical solutions in the embodiments of the present invention can be implemented by means of a computer program and its corresponding general hardware platform. Based on such an understanding, the technical solutions in the embodiments of the present invention, in essence, or the part that contributes to the prior art can be embodied in the form of a computer program, that is, a software product. The computer program software product can be stored in the storage medium, including several instructions for causing a device (which can be a personal computer, a server, a single-chip microcomputer, an MCU, or a network device, etc.) including a data processing unit to execute the methods described in each embodiment or some parts of the embodiments of the present invention.

[0163] The present invention provides an idea and method for a multi-information airborne lidar point cloud registration method based on GICP. There are many methods and ways to specifically implement this technical solution. The above is only the preferred implementation manner of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and retouches can be made, and these improvements and retouches should also be regarded as the protection scope of the present invention. Each component not clearly defined in this embodiment can be implemented by the prior art.

Claims

1. A multi-information airborne laser radar point cloud registration method based on GICP, characterized in that: The following steps are involved: Step 1, calculate the normal vector of each local point cloud by principal component analysis; Step 2, coarsely align the local point cloud to be aligned by using the sampling consistency initial alignment method; Step 3, calculating the error function of the point cloud after rough registration; Step 4: Solve the error function to obtain the rigid body transformation matrix T * The optimal value of Step 5: According to the rigid body transformation matrix T * The optimal value of is used to transform the position and posture of the local point cloud to be registered and complete the registration.

2. The multi-information airborne laser radar point cloud registration method based on GICP according to claim 1 is characterized in that: The calculation of the normal vector of each local point cloud described in step 1 includes: Step 1-1, let any local point cloud be P n , where n = 1, 2, ..., N, is the local point cloud number, and N is the number of local point clouds; let the local point cloud P n A point in the a , where a is the number of the point; find the point p a The k nearest neighbors of i , where i = 1, 2, ..., K, k is the number of the neighboring point, and K is the preset number of neighboring points; Step 1-2, use all k nearest neighbor points N i The point set N is composed of, and the fitting point p a The tangent plane at is expressed as follows: Ax+By+Cz=D(D>0) Where (x, y, z) represents a point in the tangent plane, A, B, and C are the components of the normal vector of the tangent plane in the directions of the three coordinate axes x, y, and z, and D is the distance between the tangent plane and the origin; Step 1-3, define the objective function J as the distance between each point in the point set N and the fitted tangent plane, minimize it, and transform it into an extreme value problem, which is expressed as follows: Among them, min means minimization, (x i ,y i ,z i ) represents the coordinates of the i-th point, λ is the Lagrangian operator; Steps 1-4, take partial derivatives of the above objective function and get: Among them, the intermediate variable and The calculation method is as follows: Calculate the covariance matrix S of the fitted tangent plane as follows: Among them, p i is the point cloud P n The i-th point in is the point cloud P n The geometric center point of Steps 1-5, perform eigendecomposition on the covariance matrix S and calculate the eigenvalues ​​and eigenvectors, as follows: Minimize the objective function and obtain the eigenvector n corresponding to the minimum eigenvalue of the covariance matrix S i As point p i The normal vector at .

3. The multi-information airborne laser radar point cloud registration method based on GICP according to claim 1 is characterized in that: The coarse registration described in step 2 includes: Step 2-1, among all the local point clouds to be roughly registered, select two point clouds at random and set them as the source point cloud P0 and the target point cloud Q0 respectively; Step 2-2, select s sample points in the source point cloud P0 and the target point cloud Q0 respectively, and satisfy that the initial distance between any two sample points in the same point cloud does not exceed the preset minimum distance d min ; Step 2-3, for each sample point in the source point cloud P0, randomly select a corresponding point in the target point cloud Q0; Step 2-4, calculate the rigid body transformation matrix based on the sample points and corresponding points, and calculate the distance error of the point cloud and the quality of the current registration transformation; Step 2-5, repeating steps 2-1 to 2-3 until the distance error and value of the point cloud are minimized, the rigid body transformation matrix at this time is the final rigid body transformation matrix, completing the coarse registration of the source point cloud P0 and the target point cloud Q0, and obtaining the coarsely registered point clouds P and Q; Step 2-6: Repeat steps 2-2 to 2-5 until all local point clouds are roughly aligned.

4. The multi-information airborne laser radar point cloud registration method based on GICP according to claim 3 is characterized in that: The minimum distance d described in step 2-2 min Set to 0.

05.

5. The multi-information airborne laser radar point cloud registration method based on GICP according to claim 1 is characterized in that: The error function of the point cloud after the rough registration is calculated as described in step 3, including: The error function is defined as a probability distribution model, and the point clouds P and Q after rough registration are the source point cloud and target point cloud to be registered, and they obey Gaussian distribution, which is expressed as follows: Among them, p i and q i represents the i-th point in the source point cloud P and the target point cloud Q, and represents the set of points in two point clouds, and is the covariance matrix of the two point clouds, let the rigid body transformation matrix T * for: Then the corresponding relationship between the source point cloud P and the target point cloud Q is: According to the GICP point cloud registration method, the error equation is: e i =q i -T * p i Among them, e i Represents the error; according to the probability model, the error e is obtained i The probability distribution of the rigid body transformation matrix T is solved by using the maximum likelihood estimation method. * , and simplify the result to get the error function as follows: in, Indicates that The largest T * value.

6. The multi-information airborne laser radar point cloud registration method based on GICP according to claim 5 is characterized in that: The error function of the point cloud after the rough registration is calculated as described in step 3, and also includes: Introducing the local point cloud L * a * b * The color information and the echo intensity information during airborne laser radar detection are used to calculate the feature information covariance ∑ d and mean μ p , used to recalculate the covariance matrix of the point cloud and use it to update the error function.

7. The multi-information airborne laser radar point cloud registration method based on GICP according to claim 6 is characterized in that: Recalculate the covariance matrix of the point cloud as described in step 3, including: Take any single point p in the point cloud P j It is expressed as follows: Among them, j is the number of the point, p i p is the position coordinate information, x j ,y j and z j are the coordinate values, is the point cloud feature information, α I is the intensity channel weight, I i is the echo intensity information of the point cloud, α c is the color channel weight, L i 、a i and b i L is the point cloud * a * b * Color information; For any point p j , search the neighborhood points through the kd-tree method and project them onto the plane perpendicular to the normal line to get the projection, which is expressed as follows: Among them, u j For p j Projection onto the plane perpendicular to the normal, is the location coordinate information, is the point cloud feature information, U p is the projection of the point cloud P on the plane perpendicular to the normal; after the projection transformation, U p The population covariance of w ; Point cloud feature information Introducing Gaussian kernel function and feature information measurement covariance Λ, the projection u j ∈U calculates the kernel weight w j , as follows: According to the kernel weight w j Calculate the feature information covariance ∑ d With mean μ p , as follows: Normalization is performed according to the feature information covariance, and the correlation coefficient matrix Ω is calculated as follows: in; represents a set of real numbers; the covariance matrix is ​​obtained as: where ∈ is a constant representing the covariance along the normal.

8. The multi-information airborne laser radar point cloud registration method based on GICP according to claim 7 is characterized in that: Intensity channel weight α I Set to 0.25, color channel weight α c Set to 0.

75.

9. The multi-information airborne laser radar point cloud registration method based on GICP according to claim 1 is characterized in that: Step 4 solves the error function, that is, uses the BFGS quasi-Newton method to solve the error function obtained in step 3 to obtain the transformation matrix T * .

10. The multi-information airborne laser radar point cloud registration method based on GICP according to claim 1, characterized in that: The local point cloud described in step 1 is obtained by scanning the landing ground from different angles by the drone's airborne laser radar, thereby obtaining multiple independent point clouds with overlapping areas.

Images