A method for automatically calibrating the installation deviation of an airborne laser radar based on voxel sharpness

Abstract

This invention discloses an automatic calibration method for airborne lidar placement deviation based on voxel sharpness. The method includes: selecting a natural calibration site containing building facades, linear markers, and point targets; collecting point cloud and POS data via a UAV flying in multiple headings; modeling the placement deviation to be calibrated using six parameters; optimizing the parameters using a genetic algorithm; using voxel sharpness as the fitness function; voxelizing the point cloud in the region of interest; counting the number of points within each voxel and calculating the sharpness score; a higher sharpness score indicates a more pronounced clustering of point clouds in the feature region; and iteratively optimizing the method through selection, crossover, and mutation operations to output the optimal placement deviation parameters. This invention uses an ordinary basketball court as a natural calibration site, eliminating the need for artificial targets. After calibration, the point cloud sharpness is improved by more than 150%, and the positioning accuracy reaches the centimeter level.

Description

Technical Field

[0001] This invention relates to the field of lidar calibration technology, and more specifically, to an automatic calibration method for the installation deviation of airborne lidar based on voxel sharpness. Background Technology

[0002] In UAV-borne LiDAR systems, the Position Detector (POS) consists of GPS and an IMU (Integrated Device Unit) to provide the carrier's position and attitude information. A spatial installation relationship exists between the laser scanner and the POS; their relative position is called the placement line offset, and their relative angle is called the placement angle error; both are collectively referred to as the POS placement deviation. During actual installation, placement deviations are unavoidable due to factors such as mechanical installation errors and misalignment of electronic centers. These systematic errors directly affect the final point cloud coordinate calculation results, severely impacting the point cloud positioning accuracy.

[0003] Traditional methods for checking POS (Positioning Parameter Correction) deviations mainly rely on setting up a high-precision calibration field and solving for the calibration parameters through least squares adjustment. However, these methods have the following shortcomings: (1) A calibration field with artificial targets or control points needs to be specially set up, which is costly and time-consuming, and the calibration field usually needs to be maintained regularly. (2) The least squares method is highly dependent on the initial value and is prone to getting trapped in local optima, especially in the highly nonlinear conical scanning mode; (3) Traditional methods usually rely on manual selection of feature points, which has a low degree of automation and is cumbersome to operate; (4) The complex geometric relationship of the conical scanning mode increases the difficulty of calibration, and traditional models often cannot fully describe it.

[0004] To address the aforementioned issues, there is an urgent need to develop a joint calibration method for POS placement deviation that is not required for dedicated targets, has a high degree of automation, is robust, and is applicable to conical scanning photon counting lidar.

[0005] Chinese patent document CN117111041A discloses an automatic calibration method for the placement angle of airborne lidar based on point cloud conjugate matching. This method calibrates the heading angle by extracting conjugate point clouds of buildings at two moments before and after the scan trajectory of a single flight path, and then extracts conjugate point clouds of buildings with the same name from two opposing flight paths for roll and pitch angle calibration. It uses the point cloud grid area method to calculate the point cloud coverage area and optimizes the placement angle parameters with the highest point cloud overlap as the objective. This method reduces the requirements for the calibration site and minimizes manual intervention. However, this method still has the following limitations: First, it relies on buildings as calibration markers, limiting its application in scenarios lacking obvious buildings; second, heading angle calibration requires accurate differentiation of point clouds at moments before and after the scan trajectory within the same flight path, while roll and pitch angle calibration requires the presence of buildings with the same name on two opposing flight paths, still placing high demands on the flight path and the distribution of ground features; third, the separate calibration of heading angle and roll / pitch angles may affect overall consistency due to parameter decoupling.

[0006] In their paper "An Automatic Correction Method for the Placement Angle Considering the Similarity Point Matching Model" (pp. 133-135+140), published in *Geospatial Information*, Vol. 23, No. 09, 2025, Wu Yongcai et al. proposed an iterative matching method using similar features of surface points with the same name within overlapping areas of adjacent flight strips. The optimal correction solution for the placement angle was obtained by minimizing the overall mean square error of the sample distance, simplifying the calibration workflow. In their paper "Positioning Angle Error Calibration for Airborne LiDAR Without Field Control" (pp. 1042-1053), published in *Acta Geodaetica et al.*, Vol. 54, No. 6, 2025, Su Dianpeng et al. addressed the problem of sparse point clouds and difficulty in identifying corresponding features in coastal areas without field control points. They proposed a placement angle error calibration algorithm based on planar feature constraints, using the minimum distance between two planes within a single flight strip, the minimum angle between the normal vectors of two planes within two flight strips, and the minimum center distance between building apexes within two flight strips as constraints to calibrate pitch, roll, and heading angles. However, the above methods still rely on the extraction of homonymous feature points, plane segmentation, or building recognition, which limits their application in scenarios with sparse features. Summary of the Invention

[0007] This invention provides an automatic calibration method for the placement deviation of airborne lidar based on voxel sharpness, which can solve the technical problems of existing UAV-borne lidar POS placement deviation calibration methods that rely on dedicated calibration fields, have high degree of manual intervention, and are prone to getting trapped in local optima.

[0008] An automatic calibration method for installation deviation of airborne lidar based on voxel sharpness, comprising: (1) Select a natural verification site that includes at least building facades, linear markers and point targets. The airborne lidar flies over the verification site in a multi-heading flight mode to obtain the original point cloud data and original POS observation data in the body coordinate system with the laser emission position as the origin. (2) Based on the geometric positioning model of airborne lidar, under the condition that the initial value of POS placement deviation is zero, the original body point cloud is solved to the geodetic coordinate system to obtain the point cloud before correction; (3) Perform coarse denoising on the point cloud before correction, and obtain the corresponding denoised body point cloud data and denoised POS observation data according to the timestamp index; (4) The POS placement deviation is parameterized into a six-degree-of-freedom model, including three rotation angle errors and three translation offsets; (5) Based on the genetic algorithm, the six parameters of POS placement bias are used as individual genes, and the initial population is generated by real number encoding. (6) For each individual in the current population, perform the following operations: Substitute the denoised body point cloud data, the denoised POS observation data and the placement deviation parameters corresponding to the current individual into the geometric positioning model, recalculate the point cloud coordinates in the geodetic coordinate system, and obtain a set of corrected point clouds. For the corrected point cloud, three-dimensional voxel division is performed within the preset region of interest. The number of points in each voxel is counted and the sharpness score is calculated. This sharpness score is used as the fitness value of the current individual. (7) Based on the fitness value, perform selection, crossover and mutation operations on the population to generate a new generation of population, and determine whether the preset termination condition is met; if not, return to step (6) to continue iterating; if it is met, output the current optimal placement deviation parameter.

[0009] In step (1), a basketball court is selected as a natural calibration site. When the airborne lidar flies over the calibration site in a multi-heading flight mode, the number of flight paths is no less than 5, to ensure that the fine features of the basketball court (lines, baskets, backboards, etc.) are covered multiple times from different perspectives.

[0010] In step (3), the coarse denoising process uses a statistical outlier filtering algorithm to calculate the relationship between each point cloud point and its corresponding outlier. If the average distance between neighboring points exceeds the threshold range of the global mean plus three standard deviations, it is judged as a noise point and removed. The value is 10.

[0011] In step (4), the three rotation angle errors include the roll angle error. Pitch angle error and heading angle error The three translation offsets include the X-axis offset. Y-axis offset and Z-axis offset .

[0012] In step (5), the real number encoding method directly represents each parameter to be optimized as a real number, and the individual gene form is: The population size is set to 50-200, and the initial population is randomly generated within the preset parameter range.

[0013] In step (6), the point cloud coordinates are recalculated using the following formula: ; in, The coordinates of the laser footprint in the geodetic coordinate system are: The coordinates of the GPS antenna phase center. The rotation matrix formed by the attitude angles provided for the POS. To form a rotation matrix for the angular deviations, This represents the vector representation of the conical scanning laser ranging value in the body coordinate system. This is the offset of the placement line.

[0014] In step (6), three-dimensional voxel division is performed within the preset region of interest, specifically as follows: Set the voxel size and divide the region of interest into a 3D voxel grid; count the number of points in each voxel and select voxels with a number of points greater than or equal to the effective voxel threshold as the effective voxel set.

[0015] In step (6), the number of points within each voxel is counted and the sharpness score is calculated using the following formula: ; in, The sharpness score is the fitness value of an individual. For an effective set of voxels; For the first The number of points within an individual element; This is a saturation threshold to prevent abnormally high values ​​in a particular corner from dominating the score.

[0016] Furthermore, fitness calculations can be performed on the entire scenic cloud or on local calculations for fine-structured regions within the basketball court to improve computational efficiency and sensitivity to fine features; the fine-structured regions can be manually defined.

[0017] The voxel size was determined based on the following considerations: Too small a voxel is susceptible to local noise, and the insufficient number of points within a single voxel makes it impossible to effectively characterize the clustering degree of feature regions; too large a voxel smooths out fine structural features, causing point clouds of key structures such as baskets and lines to be merged into the same voxel, reducing the sensitivity of the evaluation function. Through comparative experiments, a voxel size of 0.1m achieves the best balance between sensitivity and robustness in point cloud clustering evaluation.

[0018] In step (7), the selection operation adopts tournament selection or roulette selection, and the probability of an individual being selected is proportional to the reciprocal of its fitness value; the crossover operation adopts simulated binary crossover or arithmetic crossover, and the crossover probability is set to 0.7-0.9; the mutation operation adopts polynomial mutation or Gaussian mutation, and the mutation probability is set to 0.1-0.2.

[0019] Furthermore, to improve the algorithm's convergence performance, an adaptive crossover and mutation strategy is adopted: a higher crossover and mutation probability is used in the early stages of evolution to enhance the global search capability; as the number of generations increases, the crossover and mutation probabilities are gradually reduced to enhance the local fine-grained search capability.

[0020] In step (7), the preset termination condition is: no improvement in fitness for 10 consecutive generations or reaching the maximum number of iterations of 30 generations.

[0021] Furthermore, the method also includes a verification step for the calibration results: the optimal placement parameters obtained by the genetic algorithm and the original POS observation data are substituted into the geometric positioning model, the optimal corrected point cloud is recalculated, the voxel sharpness of the point cloud before and after calibration is calculated, and the improvement ratio is compared; at the same time, the ghosting elimination of fine structural areas such as the basket is observed and visually verified.

[0022] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention constructs a six-parameter joint calibration model that includes placement angle error and placement line offset for the special scanning mode of conical scanning photon counting lidar, realizing a complete description and compensation of system error.

[0023] 2. This invention utilizes natural calibration sites, such as basketball courts, which include building facades, linear markers, and point targets, as calibration sites. By leveraging their inherent fine structural features as natural calibration markers, it eliminates the need for dedicated artificial targets or control points, significantly reducing calibration costs and improving operational convenience. Basketball courts are widely distributed throughout the country, requiring no special approval and enabling rapid calibration.

[0024] 3. This invention innovatively proposes using point cloud voxel sharpness as a fitness function, with a voxel size of 0.1m. The physical meaning of this fitness function is clear: when the calibration parameters are accurate, the point cloud is highly concentrated in fine structures such as baskets and lines, resulting in a large and concentrated number of points within each voxel and a high voxel sharpness value; when calibration is insufficient, the point cloud is scattered or even exhibits ghosting, resulting in a sparse number of points within each voxel and a low voxel sharpness value. This indicator does not require the identification of specific targets or a true reference, making it applicable to any scene containing fine structures and achieving fully automatic, reference-free calibration.

[0025] 4. This invention uses a genetic algorithm for parameter optimization, which does not require providing accurate initial values ​​and has strong global search capabilities, effectively avoiding the problem of traditional least squares methods getting trapped in local optima; at the same time, it adopts an adaptive crossover and mutation strategy, which takes into account both global search and local fine search.

[0026] 5. This invention, through multi-angle and multi-directional flight path design, ensures that fine structures are fully observed from different perspectives, providing rich geometric constraints for parameter calculation and improving the reliability and accuracy of calibration parameters.

[0027] 6. This invention adopts a voxel-based parallel statistical architecture, combined with CUDA parallel computing, which can efficiently process large-scale point cloud data and reduce the single calibration calculation time from several hours to several minutes, meeting the needs of engineering applications.

[0028] 7. The voxel sharpness values ​​of the point cloud data after calibration are significantly improved by this invention, the ghosting phenomenon of fine structures such as the basket disappears, and the point cloud positioning accuracy is significantly improved, which can provide a high-quality data foundation for subsequent target recognition and 3D reconstruction. Experiments show that the voxel sharpness values ​​of the point cloud are improved by more than 70% after calibration, and the positioning accuracy reaches the centimeter level. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a flowchart of an automatic calibration method for the installation deviation of an airborne lidar based on voxel sharpness, according to an embodiment of the present invention.

[0031] Figure 2 This is a schematic diagram of the basketball court calibration site and flight path design in an embodiment of the present invention.

[0032] Figure 3 This is a fitness convergence curve during the optimization process of the genetic algorithm.

[0033] Figure 4 It is the evolutionary curve of the angle parameter during the optimization process of the genetic algorithm.

[0034] Figure 5 It is the evolutionary curve of the offset parameter during the optimization process of the genetic algorithm.

[0035] Figure 6 It is the box plot of the parameter distribution in the last 10 generations during the optimization process of the genetic algorithm.

[0036] Figure 7 It is the comparison diagram of the point clouds in the basketball court area before and after calibration in the embodiment of the present invention. Specific Embodiments

[0037] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0038] It should be noted that, without conflict, the features in the following embodiments and implementation manners can be combined with each other.

[0039] The UAV used in this embodiment is the DJI RTK 400 rotary-wing UAV, equipped with a self-developed conical scanning photon counting lidar system, with a laser wavelength of 532 nm, a scanning frequency of 20 Hz, and a ranging accuracy of ±0.05 m. A standard outdoor basketball court is selected as the natural calibration site, which is 28 m long and 15 m wide, and includes complete three-point lines (radius 6.75 m), free throw lines (distance from the end line 5.8 m), basketball hoops (diameter 0.45 m, height from the ground 3.05 m), backboards (1.8 m × 1.05 m), and other fine features. As Figure 2 shown, the overall flight route design adopts a combination of a "rice" - shaped cross flight route and a polygon flight route along the sidelines, with a flight height of 100 m and a flight route overlap rate of 80%. During data collection, the UAV flight speed is 2 m / s, the POS system sampling frequency is 100 Hz, and the lidar point cloud density is about 200 points / ㎡. The original point cloud data and the original POS observation data in the body coordinate system with the laser emission position as the origin are obtained.

[0040] As Figure 1 shown, an automatic calibration method for the placement deviation of airborne lidar based on voxel sharpness includes the following steps: S1, based on the geometric positioning model of the airborne lidar, under the condition that the initial value of the POS placement deviation is zero, the original body point cloud is solved to the geodetic coordinate system to obtain the point cloud before calibration.

[0041] First, establish the airborne lidar scanning coordinate system, and then establish the sensor coordinate system based on the elliptical scanning method. Taking the northeast-sky coordinate system as an example, the origin is the center of the reflector, and the X-axis is along the direction of the emitted laser ray. The Y-axis is along the flight direction, and the Z-axis is vertically upward. The unit vector L of the outgoing ray is represented as: L ; in, The azimuth angle of the emitted ray. The zenith angle of the outgoing ray.

[0042] S2 performs coarse denoising on the uncorrected point cloud to prevent noise from interfering with the voxel sharpness judgment, and obtains the corresponding denoised body point cloud data and denoised POS observation data according to the timestamp index, which are used as inputs for subsequent steps.

[0043] In the coarse point cloud denoising stage, a statistical filtering algorithm is used with k=10. The calculated global average distance is 0.32m and the standard deviation is 0.08m. Points with an average distance greater than 0.5m are identified as noise points. KNN distance formula and retained points are also discussed. as follows: ; ; The placement deviation parameterization adopts a six-parameter model, where the initial range of the angle error is set to... The initial range of line offset is set to The expression for the six-parameter model is: In the formula, These are the errors in roll angle, pitch angle, and yaw angle, respectively. These represent the offsets along the X, Y, and Z axes, respectively. The genetic algorithm population size is set to 50, with the initial population randomly generated within the above range.

[0044] S3. Substitute the denoised body point cloud data, the denoised POS observation data, the original POS observation data, and the placement deviation parameters corresponding to the current individual into the geometric positioning model, recalculate the point cloud coordinates, and obtain a set of corrected point clouds.

[0045] Point cloud recalculation: where P is the coordinate of the laser footprint in the geodetic coordinate system. The coordinates of the GPS antenna phase center. The rotation matrix formed by the attitude angles provided for the POS. Let L be the rotation matrix formed by the angular deviation, and let L be the vector representation of the conical scanning laser ranging value in the body coordinate system. This is for the placement line offset. The rotation matrix uses a ZYX order Euler angle transformation, and its calculation formula is: ; In the formula, This is the heading angle rotation matrix. This is the pitch angle rotation matrix. The roll angle rotation matrix is ​​expressed as follows: ; ; in, For heading angle, The pitch angle, This is the roll angle. The placement angle deviation rotation matrix. The calculation form and The problem is consistent; simply replace the attitude angle with the placement angle error. .

[0046] S4, Fitness Calculation: Extract point clouds from the corrected point cloud data within the region of interest (ROI) to obtain a point cloud dataset containing N point clouds. Set the voxel size and divide the ROI into a uniform 3D voxel grid. Traverse all point clouds, calculate the voxel index based on their spatial coordinates, and count the number of points within each voxel to obtain the occupied voxel set and its corresponding count. Preliminary judgment: if N is less than the minimum point count threshold or the occupied voxel count is less than the minimum voxel threshold, the current individual's fitness value is 0. Filter individuals that meet the fitness criteria... For voxels used in the voxel sharpness calculation, voxels with a number of points greater than or equal to the effective voxel threshold are selected as the effective voxel set. For each voxel in the effective voxel set, its number of points is compared with the saturation threshold, and the smaller value is taken as the number of points to be included in the calculation. The voxel sharpness score is calculated by summing the squares of the number of points included in the calculation within the effective voxels and then dividing by the number of effective voxels. The voxel sharpness score is used as the fitness value of the current individual. The larger the fitness function value, the more obvious the point cloud clusters in the feature region, and the more accurate the corresponding calibration parameters.

[0047] ; in, For an effective set of voxels, For the first The number of points within an individual element. The saturation threshold is set to 50 to prevent abnormally high values ​​in a particular corner from dominating the score. A larger fitness function value indicates a more pronounced clustering of point clouds in the feature region, resulting in more accurate calibration parameters.

[0048] Figures 3 to 6 The diagram illustrates the optimization process of the genetic algorithm in an embodiment of the present invention. Figure 3 For the fitness convergence curve, Figure 4 Evolution curve of angle parameter, Figure 5 The evolution curve for the offset parameter.Figure 6 Box plot of parameter distribution for the last 10 generations.

[0049] S5, Selection Operation: Tournament selection (tournament size 3) is adopted, and the selection probability formula is as follows: ; in, A tournament set consisting of three randomly selected individuals. For individuals The fitness value.

[0050] S6: The crossover operation uses simulated binary crossover, and the crossover formula is: ; in, and For the first generation of the parent individual One gene, and For the first generation of individuals One gene, It is generated by the following formula: ; In the formula, for Random numbers This is the distribution index.

[0051] S7: The mutation operation uses polynomial mutation, and the mutation formula is: ; in, For the first generation of the parent individual One gene, For the first generation of individuals One gene, and The first The lower and upper bounds of a gene. It is generated by the following formula: ; In the formula, for Random numbers The mutation distribution index is used. The termination condition is set as the fitness value change being less than 1e-5 for 20 consecutive generations or reaching 50 generations. In this embodiment, convergence occurs at generation 23, and the optimal fitness value increases from the initial 334.8 to 865.

[0052] The optimal calibration parameter is: roll angle error. = -0.011°, pitch angle error =0.492°, heading angle error =1.043°, eccentricity in the X direction = 0.016 m, eccentricity in the Y direction = 0.030 m, Z-direction eccentricity =-0.014 m.

[0053] After substituting this parameter into the recalculation, the voxel sharpness score increased from 327.12 before calibration to 850.51, an improvement of more than 150%, and the ghosting of point clouds in the basket area was completely eliminated, with all kinds of details clearly visible.

[0054] Figure 7 As shown, the point cloud comparison of the basketball court area before and after calibration is presented in an embodiment of the present invention. (a) shows obvious ghosting before calibration, and (b) shows the elimination of ghosting after calibration.

[0055] This embodiment fully verifies the effectiveness of the method of the present invention: low-cost, target-free calibration is achieved using an ordinary basketball court; by combining the voxel sharpness fitness function with the genetic algorithm, the six-dimensional placement deviation parameters are accurately solved, significantly improving the point cloud quality and positioning accuracy, and providing a reliable calibration solution for the engineering application of UAV-borne conical scanning photon counting lidar.

[0056] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automatic calibration method for installation deviation of airborne lidar based on voxel sharpness, characterized in that, include: (1) Select a natural verification site that includes at least building facades, linear markers and point targets. The airborne lidar flies over the verification site in a multi-heading flight mode to obtain the original point cloud data and original POS observation data in the body coordinate system with the laser emission position as the origin. (2) Based on the geometric positioning model of airborne lidar, under the condition that the initial value of POS placement deviation is zero, the original body point cloud is solved to the geodetic coordinate system to obtain the point cloud before correction; (3) Perform coarse denoising on the point cloud before correction, and obtain the corresponding denoised body point cloud data and denoised POS observation data according to the timestamp index; (4) The POS placement deviation is parameterized into a six-degree-of-freedom model, including three rotation angle errors and three translation offsets; (5) Based on the genetic algorithm, the six parameters of POS placement bias are used as individual genes, and the initial population is generated by real number encoding. (6) For each individual in the current population, perform the following operations: Substitute the denoised body point cloud data, the denoised POS observation data and the placement deviation parameters corresponding to the current individual into the geometric positioning model, recalculate the point cloud coordinates in the geodetic coordinate system, and obtain a set of corrected point clouds. For the corrected point cloud, three-dimensional voxel division is performed within the preset region of interest. The number of points in each voxel is counted and the sharpness score is calculated. This sharpness score is used as the fitness value of the current individual. (7) Based on the fitness value, perform selection, crossover and mutation operations on the population to generate a new generation of population, and determine whether the preset termination condition is met; if not, return to step (6) to continue iterating; if it is met, output the current optimal placement deviation parameter.

2. The automatic calibration method for airborne lidar placement deviation based on voxel sharpness according to claim 1, characterized in that, In step (1), a basketball court is selected as a natural calibration site; when the airborne lidar flies over the calibration site in a multi-heading flight mode, the number of flight paths is no less than 5.

3. The automatic calibration method for airborne lidar placement deviation based on voxel sharpness according to claim 1, characterized in that, In step (4), the three rotation angle errors include the roll angle error. Pitch angle error and heading angle error The three translation offsets include the X-axis offset. Y-axis offset and Z-axis offset .

4. The automatic calibration method for airborne lidar placement deviation based on voxel sharpness according to claim 1, characterized in that, In step (6), the point cloud coordinates are recalculated using the following formula: ； in, These are coordinates in the geodetic coordinate system. The coordinates of the GPS antenna phase center. The rotation matrix formed by the attitude angles provided for the POS. To form a rotation matrix for the angular deviations, This represents the vector representation of the conical scanning laser ranging value in the body coordinate system. This is the offset of the placement line.

5. The automatic calibration method for airborne lidar placement deviation based on voxel sharpness according to claim 1, characterized in that, In step (6), three-dimensional voxel division is performed within the preset region of interest, specifically as follows: Set the voxel size and divide the region of interest into a 3D voxel grid; count the number of points in each voxel and select voxels with a number of points greater than or equal to the effective voxel threshold as the effective voxel set.

6. The automatic calibration method for airborne lidar placement deviation based on voxel sharpness according to claim 5, characterized in that, In step (6), the number of points within each voxel is counted and the sharpness score is calculated using the following formula: ； in, The sharpness score is the fitness value of an individual. For an effective set of voxels; For the first The number of points within an individual element; This is a saturation threshold to prevent abnormally high values ​​in a particular corner from dominating the score.

7. The automatic calibration method for airborne lidar placement deviation based on voxel sharpness according to claim 1, characterized in that, In step (7), the selection operation adopts the tournament selection method or the roulette selection method, the crossover operation adopts the simulated binary crossover or the arithmetic crossover, and the mutation operation adopts the polynomial mutation or the Gaussian mutation.

8. The automatic calibration method for airborne lidar placement deviation based on voxel sharpness according to claim 1, characterized in that, In step (7), the preset termination condition is: no improvement in fitness for 10 consecutive generations or reaching the maximum number of iterations of 30 generations.

Images