Method and system for selecting measurement points based on 3D radar level meter

Abstract

The present invention provides a method and system for selecting measuring points based on a 3D radar level meter, which relates to the field of measurement technology, including obtaining original 3D radar point cloud data of a measurement scene, using an improved octree algorithm to adaptively divide the measurement scene into three-dimensional spaces, obtaining a three-dimensional reconstruction result of the measurement scene, performing normal vector calculation and regional clustering, and combining a regional growing and merging algorithm to determine a plurality of key areas in the three-dimensional reconstruction result; taking maximizing the coverage of the measuring points to the key areas as a first goal, taking minimizing the difference in geometric distribution between the measuring points as a second goal, solving the single-objective optimization problem through an improved differential evolution algorithm, and selecting a solution with the best comprehensive evaluation index as a final 3D radar level meter measurement point optimization arrangement scheme.

Description

Technical Field

[0001] The present invention relates to measurement technology, and in particular to a method and system for selecting measurement points based on a 3D radar level meter. Background Art

[0002] With the rapid development of industrial automation and intelligent technology, real-time monitoring and accurate measurement of materials in silos, storage tanks and other equipment have become increasingly important in many industrial fields (such as chemical, metallurgical, pharmaceutical, food, etc.). In order to achieve effective material management, material level measurement technology is widely used in production process control. By measuring the height or volume of materials inside the silo, the material level measurement device can help companies optimize material storage, transportation and production plans, and avoid the risk of material shortage or excessive accumulation.

[0003] Among the many material level measurement technologies, 3D radar level meters have gradually become the mainstream means of material level measurement in the industrial field due to their high precision, non-contact measurement and ability to adapt to harsh environments. 3D radar level meters generate three-dimensional point cloud data of material distribution inside the silo by transmitting and receiving radar waves, thereby achieving high-precision measurement of the material level. Compared with traditional ultrasonic or capacitive level meters, 3D radar level meters have stronger anti-interference ability and measurement accuracy, and are particularly suitable for complex environments such as high temperature, high pressure, and dust.

[0004] However, the measurement accuracy and efficiency of 3D radar level meters depend to a large extent on the reasonable selection of measurement points. In practical applications, the material distribution inside the silo is often irregular, and there may be accumulation, tilting, adhesion and other phenomena, resulting in large differences in material height in different areas. Therefore, how to reasonably distribute the measurement points within a limited measurement range to ensure effective coverage of key areas while avoiding redundancy and uneven distribution between measurement points is a key issue in improving measurement accuracy.

[0005] Most existing methods for selecting measurement points for 3D radar level meters rely on experience or simple rule-based layouts, such as equally spaced points or fixed path scanning, which may not effectively cover key areas in the silo in some scenarios, especially when the materials in the silo are complexly distributed and have varying shapes. Fixed measurement point layouts may result in insufficient measurement in some key areas, while other areas may have too many or redundant measurement points, resulting in a waste of measurement resources and affecting overall measurement accuracy. Summary of the invention

[0006] The embodiments of the present invention provide a method and system for selecting measurement points based on a 3D radar level meter, which can solve the problems in the prior art.

[0007] According to a first aspect of the embodiments of the present invention,

[0008] Provides a method for selecting measurement points based on 3D radar level meters, including:

[0009] Acquire original 3D radar point cloud data of a measurement scene, wherein the original 3D radar point cloud data includes three-dimensional coordinate information of each point in the measurement scene; based on the original 3D radar point cloud data, use an improved octree algorithm to adaptively divide the measurement scene into multiple sub-areas, and recursively divide the measurement scene into multiple sub-areas until the number of point clouds in each sub-area is lower than a preset threshold or reaches a preset maximum division depth, thereby obtaining a three-dimensional reconstruction result of the measurement scene; perform normal vector calculation and regional clustering on the three-dimensional reconstruction result, and determine multiple key areas in the three-dimensional reconstruction result in combination with a regional growing and merging algorithm;

[0010] Taking maximizing the coverage of the key areas by the measuring points as the first goal and minimizing the geometric distribution difference between the measuring points as the second goal, combined with the constraint conditions that the distance between adjacent measuring points is not less than the preset lower limit and each key area covers at least a preset number of measuring points, the first goal, the second goal and the constraint conditions are comprehensively considered to construct a multi-objective optimization model for the selection of measuring points of the 3D radar level meter;

[0011] Corresponding weight coefficients are set for the first objective and the second objective in the multi-objective optimization model respectively, and the multi-objective optimization problem is converted into a single-objective optimization problem. The single-objective optimization problem is solved by an improved differential evolution algorithm, and the solution with the best comprehensive evaluation index is selected as the final 3D radar level meter measurement point optimization layout scheme, wherein the improved differential evolution algorithm introduces an adaptive mutation strategy to improve the global convergence of the algorithm, and adopts the ε constraint method to deal with multiple constraints of the single-objective optimization problem.

[0012] In an optional embodiment,

[0013] Based on the original 3D radar point cloud data, an improved octree algorithm is used to adaptively divide the measurement scene into multiple sub-areas, and the measurement scene is recursively divided into multiple sub-areas until the number of point clouds in each sub-area is lower than a preset threshold or reaches a preset maximum division depth, and the three-dimensional reconstruction result of the measurement scene is obtained, including:

[0014] Loading the original 3D radar point cloud data into the root node of the octree, the original 3D radar point cloud data including the three-dimensional coordinates of each point in the measurement scene;

[0015] Calculate the point cloud density and normal vector variance of the current octree node as indicators to measure the local point cloud distribution characteristics, where the point cloud density is equal to the number of point clouds in the node divided by the volume of the cube corresponding to the node; the normal vector variance is equal to the sum of the squares of the modulus of the difference between the normal vector of each point in the node and the average value of the normal vectors of all points divided by the number of point clouds in the node;

[0016] According to the point cloud density and the normal vector variance, the segmentation threshold and the segmentation depth are adaptively adjusted, and the specific adjustment method is:

[0017] When the point cloud density is greater than a preset point cloud density threshold and the normal vector variance is greater than a preset normal vector variance threshold, multiplying the division threshold by a first preset coefficient greater than 1, and adding 1 to the division depth;

[0018] When the point cloud density is greater than a preset point cloud density threshold and the normal vector variance is less than or equal to a preset normal vector variance threshold, multiplying the division threshold by a second preset coefficient less than 1, and adding 1 to the division depth;

[0019] When the point cloud density is less than or equal to a preset point cloud density threshold and the normal vector variance is greater than a preset normal vector variance threshold, dividing the division threshold by a third preset coefficient greater than 1, and adding 1 to the division depth;

[0020] When the point cloud density is less than or equal to a preset point cloud density threshold and the normal vector variance is less than or equal to a preset normal vector variance threshold, the node is marked as a leaf node;

[0021] For the current node, if the number of its point clouds is greater than the division threshold and the division depth is less than or equal to the preset maximum division depth, the node is divided into 8 sub-nodes, and the point cloud data of the current node is divided into the corresponding sub-nodes;

[0022] For each newly generated non-empty child node, the partition threshold and partition depth are adaptively adjusted recursively until all nodes are leaf nodes or the partition termination condition is reached;

[0023] The octree partitioning result is output, including the spatial range of each leaf node and the point cloud data contained therein, to obtain a three-dimensional reconstruction result of the measurement scene.

[0024] In an optional embodiment,

[0025] The three-dimensional reconstruction result is subjected to normal vector calculation and regional clustering, and combined with a regional growing and merging algorithm, to determine that a plurality of key regions in the three-dimensional reconstruction result include:

[0026] For each sub-region in the three-dimensional reconstruction result, the normal vectors of all points in the sub-region are calculated by principal component analysis; the point cloud data set in the sub-region is centralized to obtain a decentralized point set, and the coordinates of each point in the decentralized point set are equal to the coordinates of the original 3D radar point cloud data minus the centroid coordinates of the point cloud data set; the covariance matrix of the decentralized point set is calculated, and the covariance matrix is ​​subjected to eigenvalue decomposition to obtain eigenvalues ​​and corresponding eigenvectors arranged in order of size; the eigenvector corresponding to the minimum eigenvalue is taken as the normal vector of the sub-region, indicating the main direction of the sub-region;

[0027] Normalize the normal vectors of all sub-regions in the three-dimensional reconstruction result to obtain a unit normal vector set; cluster the unit normal vector set using a spectral clustering algorithm to obtain multiple normal vector clusters, each normal vector cluster corresponds to a group of sub-regions with similar directions; the specific steps of the spectral clustering algorithm are:

[0028] Calculating a similarity matrix between normal vectors in the unit normal vector set, wherein the elements of the similarity matrix are Gaussian kernel function values ​​corresponding to two normal vectors; normalizing the similarity matrix to obtain a normalized Laplace matrix; performing eigenvalue decomposition on the normalized Laplace matrix to obtain eigenvectors corresponding to the first k largest eigenvalues; stacking the eigenvectors in rows to form a matrix, and normalizing each row of the matrix; taking each row of the normalized matrix as a clustering sample, and applying a K-means algorithm to perform clustering to obtain the normal vector cluster;

[0029] For each normal vector cluster, randomly select a sub-region as the initial seed region, and calculate the average normal vector of the initial seed region; traverse other sub-regions in the current normal vector cluster, and for each sub-region, calculate the distance from its centroid to the centroid of the initial seed region and the angle between its normal vector and the average normal vector of the initial seed region; if the distance is less than a preset distance threshold and the angle is less than a preset angle threshold, merge the sub-region into the initial seed region, update the average normal vector of the initial seed region, and remove the sub-region from the current normal vector cluster; repeat the iteration until no new sub-region can be merged into the current normal vector cluster;

[0030] The final initial seed area is marked as a key area and removed from the current normal vector cluster; the iteration is repeated until the current normal vector cluster is empty, and multiple key areas in the three-dimensional reconstruction result are obtained.

[0031] In an optional embodiment,

[0032] Taking maximizing the coverage of the key areas by the measuring points as the first goal and minimizing the geometric distribution difference between the measuring points as the second goal, combined with the constraint that the distance between adjacent measuring points is not less than the preset lower limit and each key area covers at least a preset number of measuring points, the multi-objective optimization model for selecting the measuring points of the 3D radar level meter is constructed by combining the first goal, the second goal and the constraint, including:

[0033] For each candidate measurement point, calculating the distance from the candidate measurement point to each key area;

[0034] Using a Gaussian kernel function to calculate the coverage of each key area by the candidate measurement points;

[0035] The importance weight of the key area is introduced, and the importance weight is calculated according to the area, volume and shape complexity of the key area. The area weight of the key area is calculated by the area ratio, the volume weight of the key area is calculated by the volume ratio, and the shape complexity weight of the key area is calculated by the ratio of the surface area to the volume of the key area.

[0036] Define the comprehensive coverage of the candidate measurement points to all key areas as the weighted sum of the coverage of each key area, and the weight of the weighted sum is the importance weight;

[0037] A coverage balance factor is introduced to calculate the coverage ratio of each candidate measurement point to each key area and the average coverage ratio of the candidate measurement point to each key area. The comprehensive coverage is corrected based on the difference between the coverage ratio and the average coverage ratio and the coverage balance factor to obtain a coverage evaluation index as the first goal.

[0038] In an optional embodiment,

[0039] Taking maximizing the coverage of the key areas by the measuring points as the first goal and minimizing the geometric distribution difference between the measuring points as the second goal, combined with the constraint that the distance between adjacent measuring points is not less than the preset lower limit and each key area covers at least a preset number of measuring points, the multi-objective optimization model for selecting the measuring points of the 3D radar level meter is constructed by combining the first goal, the second goal and the constraint, and further includes:

[0040] For any two candidate measurement points, calculate the Euclidean distance between the two candidate measurement points;

[0041] Introducing a distance decay function to transform the Euclidean distance to obtain a corrected distance, wherein the distance decay function includes any one of an exponential decay function, a power law decay function, and a threshold decay function;

[0042] Based on the corrected distance, the sum of the corrected distances between all candidate measurement points is calculated to obtain the geometric distribution difference of the candidate measurement point set;

[0043] Calculate the direction vector between any two candidate measurement points, and calculate the direction difference factor of each candidate measurement point based on the sine square value of the angle between any two direction vectors;

[0044] Calculate the height of each candidate measurement point, and calculate the height difference factor of each candidate measurement point based on the sum of the squares of the height differences between the candidate measurement points;

[0045] The weight coefficients of distance difference, direction difference and height difference are introduced, and the weighted sum of the geometric distribution difference, the direction difference factor and the height difference factor is combined to obtain the final evaluation index of the geometric distribution difference between the measurement points as the second goal.

[0046] In an optional embodiment,

[0047] The corresponding weight coefficients are set for the first objective and the second objective in the multi-objective optimization model respectively, and the multi-objective optimization problem is converted into a single-objective optimization problem. The single-objective optimization problem is solved by an improved differential evolution algorithm, and the solution with the best comprehensive evaluation index is selected as the final 3D radar level meter measurement point optimization layout scheme, including:

[0048] Initialize the differential evolution algorithm parameters, including population size NP, mutation factor F, crossover probability CR and maximum number of iterations G max , and simulated annealing algorithm parameters, including the initial temperature T 0 and cooling coefficient α; randomly generate NP candidate solutions as the initial population of the differential evolution algorithm, and let the current temperature T = T 0 ;

[0049] Perform adaptive mutation operation on each candidate solution of differential evolution algorithm to generate mutation vector, and use ε constraint method to deal with multiple constraints of single objective optimization problem;

[0050] Perform crossover operation on each candidate solution of the differential evolution algorithm and the corresponding mutation vector to generate a test vector;

[0051] Calculate the comprehensive evaluation index of each test vector and compare it with the comprehensive evaluation index of the corresponding candidate solution; if the comprehensive evaluation index of the test vector is better than the candidate solution, accept the test vector as a new candidate solution; otherwise, according to the importance sampling criterion, accept the test vector as a new candidate solution with the solution probability corresponding to the importance sampling criterion; at every certain number of iterations, combine the cooling coefficient to decay the current temperature of the simulated annealing algorithm;

[0052] Determine whether the differential evolution algorithm meets the maximum number of iterations or other termination conditions. If so, output the current optimal solution as the optimal layout plan for the measurement points of the 3D radar level meter; otherwise, continue to perform the mutation, crossover and selection operations of the differential evolution algorithm until the termination conditions are met.

[0053] According to a second aspect of the embodiments of the present invention,

[0054] Provide a measurement point selection system based on 3D radar level meter, including:

[0055] The first unit is used to obtain original 3D radar point cloud data of a measurement scene, wherein the original 3D radar point cloud data includes three-dimensional coordinate information of each point in the measurement scene; based on the original 3D radar point cloud data, an improved octree algorithm is used to adaptively divide the measurement scene into multiple sub-areas, and the measurement scene is recursively divided into multiple sub-areas until the number of point clouds in each sub-area is lower than a preset threshold or reaches a preset maximum division depth, thereby obtaining a three-dimensional reconstruction result of the measurement scene; normal vector calculation and regional clustering are performed on the three-dimensional reconstruction result, and a plurality of key areas in the three-dimensional reconstruction result are determined in combination with a regional growing and merging algorithm;

[0056] The second unit is used to maximize the coverage of the key area by the measuring points as the first goal, minimize the geometric distribution difference between the measuring points as the second goal, and take the distance between adjacent measuring points as not less than a preset lower limit and each key area covers at least a preset number of measuring points as constraints, and comprehensively consider the first goal, the second goal and the constraints to construct a multi-objective optimization model for the selection of measuring points of the 3D radar level meter;

[0057] The third unit is used to set corresponding weight coefficients for the first objective and the second objective in the multi-objective optimization model respectively, convert the multi-objective optimization problem into a single-objective optimization problem, solve the single-objective optimization problem through an improved differential evolution algorithm, and select the solution with the best comprehensive evaluation index as the final 3D radar level meter measurement point optimization layout scheme, wherein the improved differential evolution algorithm introduces an adaptive mutation strategy to improve the global convergence of the algorithm, and adopts the ε constraint method to deal with multiple constraints of the single-objective optimization problem.

[0058] According to a third aspect of the embodiments of the present invention,

[0059] An electronic device is provided, comprising:

[0060] processor;

[0061] a memory for storing processor-executable instructions;

[0062] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0063] A fourth aspect of the embodiments of the present invention is:

[0064] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0065] This paper proposes an adaptive space partitioning method based on an improved octree algorithm, which can dynamically adjust the partitioning strategy according to the local distribution characteristics of point cloud data (point cloud density and normal vector variance), and finally achieve efficient 3D reconstruction. Through this method, the details of complex areas can be effectively preserved in unevenly distributed point cloud data, while reducing redundant calculations in smooth areas. This scheme has broad application prospects in the fields of autonomous driving, drone navigation, 3D city modeling, etc., and can significantly improve the efficiency and accuracy of large-scale point cloud data processing.

[0066] This paper proposes a measurement point selection method based on a multi-objective optimization model, which combines the two objectives of maximizing the coverage of key areas and minimizing the geometric distribution difference of measurement points to ensure the reasonable distribution of measurement points. By introducing coverage balance factors and distance constraints, this method can effectively optimize the measurement point selection process of 3D radar level meters and improve measurement accuracy and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a flow chart of a method for selecting measurement points based on a 3D radar level meter according to an embodiment of the present invention;

[0068] Figure 2 It is a structural schematic diagram of a measurement point selection system based on a 3D radar level meter according to an embodiment of the present invention. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0070] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0071] Figure 1FIG. 1 is a flow chart of a method for selecting measurement points based on a 3D radar level meter according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0072] S101. Acquire original 3D radar point cloud data of a measurement scene, wherein the original 3D radar point cloud data includes three-dimensional coordinate information of each point in the measurement scene; based on the original 3D radar point cloud data, use an improved octree algorithm to adaptively divide the measurement scene into multiple sub-regions, and recursively divide the measurement scene into multiple sub-regions until the number of point clouds in each sub-region is lower than a preset threshold or reaches a preset maximum division depth, thereby obtaining a three-dimensional reconstruction result of the measurement scene; perform normal vector calculation and region clustering on the three-dimensional reconstruction result, and determine multiple key regions in the three-dimensional reconstruction result in combination with a region growing and merging algorithm;

[0073] S102. Taking maximizing the coverage of the key area by the measuring points as the first goal and minimizing the difference in geometric distribution between the measuring points as the second goal, taking the distance between adjacent measuring points not less than the preset lower limit and each key area covering at least a preset number of measuring points as the constraint conditions, combining the first goal, the second goal and the constraint conditions, constructing a multi-objective optimization model for selecting measuring points of the 3D radar level meter;

[0074] S103. Set corresponding weight coefficients for the first objective and the second objective in the multi-objective optimization model respectively, convert the multi-objective optimization problem into a single-objective optimization problem, solve the single-objective optimization problem by using an improved differential evolution algorithm, and select the solution with the best comprehensive evaluation index as the final 3D radar level meter measurement point optimization layout scheme, wherein the improved differential evolution algorithm introduces an adaptive mutation strategy to improve the global convergence of the algorithm, and adopts the ε constraint method to handle multiple constraints of the single-objective optimization problem.

[0075] In an optional embodiment,

[0076] Based on the original 3D radar point cloud data, an improved octree algorithm is used to adaptively divide the measurement scene into multiple sub-areas, and the measurement scene is recursively divided into multiple sub-areas until the number of point clouds in each sub-area is lower than a preset threshold or reaches a preset maximum division depth, and the three-dimensional reconstruction result of the measurement scene is obtained, including:

[0077] Loading the original 3D radar point cloud data into the root node of the octree, the original 3D radar point cloud data including the three-dimensional coordinates of each point in the measurement scene;

[0078] Calculate the point cloud density and normal vector variance of the current octree node as indicators to measure the local point cloud distribution characteristics, where the point cloud density is equal to the number of point clouds in the node divided by the volume of the cube corresponding to the node; the normal vector variance is equal to the sum of the squares of the modulus of the difference between the normal vector of each point in the node and the average value of the normal vectors of all points divided by the number of point clouds in the node;

[0079] According to the point cloud density and the normal vector variance, the segmentation threshold and the segmentation depth are adaptively adjusted, and the specific adjustment method is:

[0080] When the point cloud density is greater than a preset point cloud density threshold and the normal vector variance is greater than a preset normal vector variance threshold, multiplying the division threshold by a first preset coefficient greater than 1, and adding 1 to the division depth;

[0081] When the point cloud density is greater than a preset point cloud density threshold and the normal vector variance is less than or equal to a preset normal vector variance threshold, multiplying the division threshold by a second preset coefficient less than 1, and adding 1 to the division depth;

[0082] When the point cloud density is less than or equal to a preset point cloud density threshold and the normal vector variance is greater than a preset normal vector variance threshold, dividing the division threshold by a third preset coefficient greater than 1, and adding 1 to the division depth;

[0083] When the point cloud density is less than or equal to a preset point cloud density threshold and the normal vector variance is less than or equal to a preset normal vector variance threshold, the node is marked as a leaf node;

[0084] For the current node, if the number of its point clouds is greater than the division threshold and the division depth is less than or equal to the preset maximum division depth, the node is divided into 8 sub-nodes, and the point cloud data of the current node is divided into the corresponding sub-nodes;

[0085] For each newly generated non-empty child node, the partition threshold and partition depth are adaptively adjusted recursively until all nodes are leaf nodes or the partition termination condition is reached;

[0086] The octree partitioning result is output, including the spatial range of each leaf node and the point cloud data contained therein, to obtain a three-dimensional reconstruction result of the measurement scene.

[0087] For example, 3D radar point cloud data, as a high-precision scene measurement technology, is widely used in fields such as autonomous driving, robot navigation, geographic information systems, and three-dimensional reconstruction. Point cloud data has the characteristics of large data volume and high data dimension. How to effectively process and characterize these data has become a research hotspot. Octree, as a common space division structure, can effectively organize and manage three-dimensional point cloud data. However, the traditional octree algorithm has the problems of low processing efficiency and insufficient division accuracy when processing large-scale point cloud data. Therefore, this paper proposes an adaptive space division method based on the improved octree algorithm, which can dynamically adjust the division strategy according to the local distribution characteristics of point cloud data, thereby achieving fine space division and efficient three-dimensional reconstruction.

[0088] The present invention provides an adaptive space partitioning method based on an improved octree algorithm. The method introduces point cloud density and normal vector variance as the measurement indicators of local point cloud distribution characteristics, dynamically adjusts the partitioning threshold and partitioning depth, and recursively divides the measurement scene into multiple sub-areas, and finally outputs the three-dimensional reconstruction result of the measurement scene. The specific technical solution includes the following steps:

[0089] Load the original 3D radar point cloud data into the root node of the octree. Each point cloud data contains the 3D coordinate information of each point in the measurement scene. The root node represents the spatial range covered by the entire measurement scene. The point cloud data is first stored in the root node and waits for subsequent processing.

[0090] In order to divide the space more effectively, it is necessary to calculate the point cloud density and normal vector variance within each octree node. These two indicators are used to measure the spatial distribution and surface complexity of the local point cloud.

[0091] Point cloud density: Point cloud density is used to measure the density of the point cloud distribution within a node. It can be calculated by the number of points in a node relative to the volume of the space represented by the node. For example, if an octree node represents a large volume of space, but the number of point clouds in it is relatively small, then the point cloud density of the node is low; on the contrary, if a smaller space contains more point clouds, then the point cloud density of the node is high.

[0092] Normal vector variance: The normal vector refers to the surface normal direction estimated from the point cloud data. The normal vector variance is used to measure the surface complexity of the point cloud within the node. If the point cloud normal vector in an area varies greatly, it means that the surface of the area is more complex or irregular; on the contrary, if the normal vector varies less, the surface is relatively smooth. For example, in a certain area, if the normal vector direction of the point cloud varies greatly, it means that there may be complex surface details in the area, and the variance will also be large.

[0093] According to the point cloud density and normal vector variance in the node, the segmentation threshold and segmentation depth are dynamically adjusted. The specific adjustment methods are as follows:

[0094] High density and complex surface (case 1): When the point cloud density within a node is high and the normal vector variance is large, it means that the point cloud distribution in this area is dense and the surface is complex. In this case, a finer division is required to ensure the preservation of details. Therefore, the division threshold should be increased and the division depth should be increased. For example, in an urban building area with a high point cloud density, the building surface may have a more complex detail structure, so a finer division is required.

[0095] High density but smooth surface (Case 2): When the point cloud density within a node is high but the normal vector variance is small, it indicates that the point cloud in this area is densely distributed but the surface is relatively smooth. For example, in the wall area of ​​a building, the point cloud may be very dense but the surface is relatively smooth. In this case, the fineness of the division can be appropriately reduced to reduce the computational overhead. Therefore, the division threshold is reduced and the division depth is increased.

[0096] Low density but complex surface (case 3): When the point cloud density within a node is low but the normal vector variance is large, it means that the point cloud distribution in this area is sparse, but the surface is relatively complex. For example, in a natural landscape, the point cloud may be sparsely distributed, but the surface curves are complex and varied. In this case, the segmentation threshold should be appropriately reduced, but the segmentation depth still needs to be increased to preserve surface details.

[0097] Low density and smooth surface (Case 4): When the point cloud density within a node is low and the normal vector variance is small, it means that the point cloud distribution in this area is sparse and the surface is relatively smooth. In this case, no further division is required and the node is marked as a leaf node. For example, in a flat ground area, the point cloud is sparse and the surface is smooth, and further division is unnecessary.

[0098] For each octree node, if the number of point clouds in the node is greater than the division threshold and the current division depth has not reached the preset maximum depth, the node is divided into 8 child nodes. Each child node represents 1 / 8 of the spatial range of the current node, and the point cloud data of the current node is divided into the corresponding child node.

[0099] For each newly generated non-empty child node, the steps of adaptively adjusting the partition threshold and partition depth are repeated. This process is recursive until one of the following conditions is met:

[0100] The number of point clouds in the node is lower than the segmentation threshold; the segmentation depth reaches the preset maximum value.

[0101] For example, in a 3D building measurement scenario, a root node contains 10,000 point cloud data. According to the adaptive partitioning strategy, the root node is divided into 8 child nodes. The number of point clouds in some child nodes may be large, so recursive partitioning continues. However, for some areas, such as the flat roof of a building, the partitioning termination condition may be met quickly and no further partitioning is performed.

[0102] When all nodes meet the termination conditions, the final octree structure is output. Each leaf node contains the following information: the spatial range covered by the node; the point cloud data contained in the node.

[0103] The point cloud data represented by these leaf nodes can be used to reconstruct the 3D structure of the entire measurement scene. For example, for urban road reconstruction in autonomous driving scenarios, the output octree structure can accurately represent the 3D spatial distribution of roads, buildings and other obstacles, providing accurate environmental perception data for the autonomous driving system.

[0104] This paper proposes an adaptive space partitioning method based on an improved octree algorithm, which can dynamically adjust the partitioning strategy according to the local distribution characteristics of point cloud data (point cloud density and normal vector variance), and finally achieve efficient 3D reconstruction. Through this method, the details of complex areas can be effectively preserved in unevenly distributed point cloud data, while reducing redundant calculations in smooth areas. This scheme has broad application prospects in the fields of autonomous driving, drone navigation, 3D city modeling, etc., and can significantly improve the efficiency and accuracy of large-scale point cloud data processing.

[0105] In an optional embodiment,

[0106] The three-dimensional reconstruction result is subjected to normal vector calculation and regional clustering, and combined with a regional growing and merging algorithm, to determine that a plurality of key regions in the three-dimensional reconstruction result include:

[0107] For each sub-region in the three-dimensional reconstruction result, the normal vectors of all points in the sub-region are calculated by principal component analysis; the point cloud data set in the sub-region is centralized to obtain a decentralized point set, and the coordinates of each point in the decentralized point set are equal to the coordinates of the original 3D radar point cloud data minus the centroid coordinates of the point cloud data set; the covariance matrix of the decentralized point set is calculated, and the covariance matrix is ​​subjected to eigenvalue decomposition to obtain eigenvalues ​​and corresponding eigenvectors arranged in order of size; the eigenvector corresponding to the minimum eigenvalue is taken as the normal vector of the sub-region, indicating the main direction of the sub-region;

[0108] Normalize the normal vectors of all sub-regions in the three-dimensional reconstruction result to obtain a unit normal vector set; cluster the unit normal vector set using a spectral clustering algorithm to obtain multiple normal vector clusters, each normal vector cluster corresponds to a group of sub-regions with similar directions; the specific steps of the spectral clustering algorithm are:

[0109] Calculating a similarity matrix between normal vectors in the unit normal vector set, wherein the elements of the similarity matrix are Gaussian kernel function values ​​corresponding to two normal vectors; normalizing the similarity matrix to obtain a normalized Laplace matrix; performing eigenvalue decomposition on the normalized Laplace matrix to obtain eigenvectors corresponding to the first k largest eigenvalues; stacking the eigenvectors in rows to form a matrix, and normalizing each row of the matrix; taking each row of the normalized matrix as a clustering sample, and applying a K-means algorithm to perform clustering to obtain the normal vector cluster;

[0110] For each normal vector cluster, randomly select a sub-region as the initial seed region, and calculate the average normal vector of the initial seed region; traverse other sub-regions in the current normal vector cluster, and for each sub-region, calculate the distance from its centroid to the centroid of the initial seed region and the angle between its normal vector and the average normal vector of the initial seed region; if the distance is less than a preset distance threshold and the angle is less than a preset angle threshold, merge the sub-region into the initial seed region, update the average normal vector of the initial seed region, and remove the sub-region from the current normal vector cluster; repeat the iteration until no new sub-region can be merged into the current normal vector cluster;

[0111] The final initial seed area is marked as a key area and removed from the current normal vector cluster; the iteration is repeated until the current normal vector cluster is empty, and multiple key areas in the three-dimensional reconstruction result are obtained.

[0112] For example, in the fields of autonomous driving, robot navigation and geographic information systems, the processing and analysis of 3D radar point cloud data is one of the key technologies. Through further analysis of the 3D reconstruction results, multiple key areas in the scene can be identified, which is of great significance for scene understanding, path planning, etc. This paper proposes a key area identification method based on normal vector calculation, region clustering, and region growing and merging algorithm. This method calculates the normal vector and performs cluster analysis on the sub-areas in the 3D reconstruction results, and finally determines multiple key areas. The specific technical implementation steps are as follows.

[0113] The present invention realizes the recognition of key areas in the three-dimensional reconstruction results by the following steps: first, the normal vector of each sub-region is calculated, then the normal vector is clustered using the spectral clustering algorithm, and then multiple key areas are identified by the region growing and merging algorithm. This method can accurately and quickly identify important areas in the scene, and improve the efficiency and accuracy of three-dimensional scene understanding.

[0114] The 3D reconstruction result is usually composed of multiple sub-regions, each of which contains a certain amount of point cloud data. In order to identify the main directions of these sub-regions and perform subsequent cluster analysis, it is first necessary to calculate the normal vector of each sub-region. Perform principal component analysis (PCA) on the point cloud data in each sub-region to determine the main direction of the sub-region. The specific steps are as follows:

[0115] For each sub-area, first calculate the centroid of all points in the sub-area. The centroid refers to the central position of all points in the sub-area. Subtract the centroid coordinates from the coordinates of each point in the sub-area to obtain a decentralized point set. For example, suppose a sub-area contains 100 points, and the coordinates of each point are obtained by 3D radar. First calculate the average coordinates (centroid coordinates) of these 100 points, and then decentralize each point to obtain the decentralized point set of the sub-area.

[0116] Perform covariance analysis on the decentralized point set. The covariance matrix reflects the changes in the point set in various directions. Perform eigenvalue decomposition on the covariance matrix to obtain multiple eigenvalues ​​and corresponding eigenvectors. The size of the eigenvalue reflects the distribution of the point set in different directions. Take the eigenvector corresponding to the minimum eigenvalue as the normal vector of the sub-region, indicating the main direction of the sub-region. Through the above steps, a normal vector can be calculated for each sub-region, indicating its main direction and surface inclination in three-dimensional space.

[0117] After calculating the normal vectors of each sub-region, these normal vectors need to be clustered to identify sub-regions with similar directions. The steps for normalization and clustering of normal vectors are as follows:

[0118] Since the normal vectors of different sub-regions may have different lengths, all normal vectors need to be normalized to unit vectors. The purpose of normalization is to eliminate the influence of the normal vector length on the clustering results and only consider the direction information of the normal vector. For example, the normal vector of a sub-region is (3, 4, 5), which becomes (0.42, 0.56, 0.70) after normalization. Through normalization, the length of all normal vectors is unified to 1, which is convenient for subsequent similarity calculation.

[0119] The spectral clustering algorithm is used to cluster the normalized normal vectors. Spectral clustering is a graph theory-based method that can effectively handle the clustering problem of high-dimensional data. The specific steps are as follows:

[0120] Calculate the similarity matrix between normal vectors. Each element in the similarity matrix represents the similarity between two normal vectors. The greater the similarity, the closer the directions of the two normal vectors are. Normalize the similarity matrix to obtain a normalized Laplace matrix. The Laplace matrix is ​​used to represent the connection relationship between normal vectors and the internal structure of clustering.

[0121] Perform eigenvalue decomposition on the normalized Laplacian matrix and extract the first few largest eigenvalues ​​and their corresponding eigenvectors. The eigenvectors reflect the clustering structure of the normal vectors. The extracted eigenvectors are normalized, and each eigenvector is used as a sample. The K-means algorithm is applied to cluster these samples. Finally, multiple normal vector clusters are obtained, and each normal vector cluster corresponds to a group of sub-regions with similar directions.

[0122] After the normal vectors are clustered, each normal vector cluster contains multiple sub-regions with similar directions. In order to further identify the key areas in the scene, the region growing and merging algorithm is used to process each normal vector cluster. The specific steps are as follows:

[0123] For each normal vector cluster, randomly select a sub-region as the initial seed region and calculate the average normal vector of the seed region. The initial seed region is the starting point of the entire growth and merging process. For example, suppose a normal vector cluster contains 20 sub-regions, randomly select the first sub-region as the seed region, calculate its normal vector, and use it as the initial average normal vector of the cluster.

[0124] Traverse other sub-regions in the current normal vector cluster, and for each sub-region, calculate the distance from its centroid to the centroid of the initial seed region, as well as the angle between its normal vector and the average normal vector of the initial seed region. If the centroid distance of the sub-region is less than the preset distance threshold, and the normal vector angle is less than the preset angle threshold, then merge the sub-region into the initial seed region.

[0125] For example, if the distance between a sub-region and the centroid of the initial seed region is 5 meters and the preset distance threshold is 10 meters, then the sub-region meets the distance condition. Then calculate the normal vector angle. If the angle is less than the preset 30-degree threshold, the sub-region can be merged into the seed region. After merging, update the average normal vector of the initial seed region and remove the merged sub-region from the normal vector cluster.

[0126] Repeat the above steps and continue to traverse other sub-regions in the normal vector cluster until no new sub-regions can be merged. At this point, the final initial seed region is marked as a key region and removed from the current normal vector cluster.

[0127] For each normal vector cluster, the initial seed selection and merging process is repeated until all sub-regions in the normal vector cluster are processed. Finally, multiple key regions are obtained, which have important directional characteristics and spatial distribution in the 3D reconstructed scene.

[0128] The final output is multiple key areas, which have clear directional characteristics and spatial positions in the 3D reconstruction results. The identification of key areas helps scene understanding, path planning and navigation decisions. For example, in an autonomous driving scenario, identifying key areas such as roads, buildings and obstacles can provide the vehicle with more accurate environmental perception data and improve driving safety and reliability.

[0129] This paper proposes a method for identifying key regions in 3D reconstruction based on normal vector calculation and region clustering. By performing normal vector calculation, spectral clustering analysis, and region growing and merging on subregions in the 3D reconstruction results, multiple key regions are finally identified. This method has high accuracy and computational efficiency and can be effectively applied to fields such as autonomous driving, drone navigation, and 3D city modeling.

[0130] In an optional embodiment,

[0131] Taking maximizing the coverage of the key areas by the measuring points as the first goal and minimizing the geometric distribution difference between the measuring points as the second goal, combined with the constraint that the distance between adjacent measuring points is not less than the preset lower limit and each key area covers at least a preset number of measuring points, the multi-objective optimization model for selecting the measuring points of the 3D radar level meter is constructed by combining the first goal, the second goal and the constraint, including:

[0132] For each candidate measurement point, calculating the distance from the candidate measurement point to each key area;

[0133] Using a Gaussian kernel function to calculate the coverage of each key area by the candidate measurement points;

[0134] The importance weight of the key area is introduced, and the importance weight is calculated according to the area, volume and shape complexity of the key area. The area weight of the key area is calculated by the area ratio, the volume weight of the key area is calculated by the volume ratio, and the shape complexity weight of the key area is calculated by the ratio of the surface area to the volume of the key area.

[0135] Define the comprehensive coverage of the candidate measurement points to all key areas as the weighted sum of the coverage of each key area, and the weight of the weighted sum is the importance weight;

[0136] A coverage balance factor is introduced to calculate the coverage ratio of each candidate measurement point to each key area and the average coverage ratio of the candidate measurement point to each key area. The comprehensive coverage is corrected based on the difference between the coverage ratio and the average coverage ratio and the coverage balance factor to obtain a coverage evaluation index as the first goal.

[0137] For example, in industrial process control and material monitoring, level meters are very important tools for measuring the height, volume and other information of materials in silos or containers. 3D radar level meters generate three-dimensional point cloud data by transmitting and receiving radar signals to reconstruct a three-dimensional model of the material distribution inside the silo. In order to improve measurement accuracy and efficiency, optimizing the distribution of measurement points is one of the key issues. This paper proposes a measurement point selection method based on a multi-objective optimization model, with the main goal of maximizing the coverage of measurement points on key areas, while minimizing the difference in geometric distribution between measurement points. This method combines the constraints of the distance between adjacent measurement points and the coverage requirements of each key area to ensure that the distribution of measurement points is optimized while meeting the constraints.

[0138] The technical solution aims to optimize the selection of measurement points for 3D radar level meters by building a multi-objective optimization model that comprehensively considers the coverage of measurement points to key areas and the geometric distribution differences between measurement points. Specifically, the optimization model includes two main objectives:

[0139] The first goal is to maximize the coverage of the key areas by the measurement points. The second goal is to minimize the difference in geometric distribution between the measurement points.

[0140] At the same time, the model also introduces the following constraints: the distance between adjacent measurement points is not less than the preset lower limit; each key area covers at least a preset number of measurement points.

[0141] In order to select the measurement points reasonably, we first need to calculate the distance from each candidate measurement point to the key area. The key areas are obtained through the previous regional clustering and identification, and their location in the silo is of great significance. The distance between the measurement point and the key area is the basis for the subsequent calculation of coverage.

[0142] Definition of candidate measurement points: Candidate measurement points refer to the measurement locations that can be selected in the silo. Usually, these points are distributed along multiple measurement paths above the silo to form a discrete point set.

[0143] Distance calculation: For each candidate measurement point, calculate its geometric distance to each key area. This distance can be the straight-line distance from the measurement point to the center of mass of the key area, or according to the application scenario, the shortest distance from the measurement point to the boundary of the key area can also be used. For example, assuming that there are multiple key areas in a silo, such as bottom accumulation, side wall attachments, etc., the distances from a candidate measurement point to these key areas can be 5 meters, 7 meters, and 10 meters respectively.

[0144] Coverage refers to the effective coverage of a certain measurement point on the key area. In order to calculate the coverage, the Gaussian kernel function is used to characterize the coverage effect of the candidate measurement point on the key area.

[0145] Gaussian kernel function: Gaussian kernel function is a distance-based weighting function that can calculate the coverage of the key area based on the distance between the candidate measurement point and the key area. The closer the distance, the higher the coverage; the farther the distance, the lower the coverage. For example, the coverage of a measurement point 5 meters away from the key area will be higher than that of a measurement point 10 meters away.

[0146] Coverage calculation: The coverage of each key area is calculated for each candidate measurement point through the Gaussian kernel function. Assuming that the distance between a measurement point and a key area is 5 meters, according to the Gaussian kernel function, the coverage of this point may be 0.8, while the coverage of another measurement point with a distance of 10 meters may be 0.4.

[0147] In practical applications, different key areas have different importance for measurement. Therefore, it is necessary to assign an importance weight to each key area to ensure that the optimization model can give priority to the coverage of important areas. The importance weight can be calculated based on the area, volume and shape complexity of the key area.

[0148] Area weight: The larger the area of ​​a key area, the more measurement points are often needed to cover it. Therefore, the area weight can be calculated by the area ratio of the key areas. The larger the area, the higher the weight. For example, if the area of ​​one key area is 50 square meters and the area of ​​another area is 100 square meters, the area weight of the latter should be relatively higher.

[0149] Volume weight: Volume reflects the size of a critical area in three-dimensional space. A critical area with a larger volume usually carries more material and is more important to measure. Therefore, the volume weight can be calculated by volume ratio. For example, if the volume of one critical area is 500 cubic meters and the volume of another area is 1,000 cubic meters, the latter area will have a higher volume weight.

[0150] Shape complexity weight: The more complex the shape of a critical area is, the more difficult and important it is to measure. Shape complexity can be measured by the ratio of the surface area to the volume of the critical area. Areas with larger surface areas but smaller volumes usually have higher shape complexity. For example, if a critical area has a surface area of ​​200 square meters and a volume of 500 cubic meters, while another area has a surface area of ​​300 square meters and a volume of 1,000 cubic meters, the former has a relatively higher shape complexity.

[0151] The comprehensive coverage is a global evaluation of each candidate measurement point, measuring its total coverage of all key areas. The comprehensive coverage is the weighted sum of the coverage of each key area, and the weight is determined by the importance of each key area.

[0152] Calculation of comprehensive coverage: For each candidate measurement point, multiply its coverage of each key area by the corresponding weight to obtain the weighted coverage. Then, add the weighted coverage of all key areas to obtain the comprehensive coverage of the candidate measurement point. For example, assuming that the coverage of a candidate measurement point for three key areas is 0.8, 0.4, and 0.6, respectively, and the weights of these areas are 0.5, 0.3, and 0.2, respectively, then the comprehensive coverage of the measurement point is 0.8×0.5 + 0.4×0.3 + 0.6×0.2 = 0.56.

[0153] In order to avoid excessive concentration of measurement points in certain key areas, resulting in insufficient coverage of other areas, a coverage balance factor is introduced into the optimization model. The coverage balance factor is used to measure the uniformity of coverage of each key area by the measurement points.

[0154] Calculation of coverage ratio: First, calculate the coverage ratio of each candidate measurement point to each key area. Coverage ratio refers to the ratio of a measurement point's coverage of a key area to its coverage of all areas. For example, if a measurement point covers three areas of 0.8, 0.4, and 0.6, respectively, then its coverage ratio of the first area is 0.8 / (0.8+0.4+0.6) ≈ 0.44.

[0155] Average coverage ratio: Calculate the average coverage ratio of each candidate measurement point to each key area as a comparison benchmark. The average coverage ratio is the theoretical value when the measurement point covers all areas evenly. For example, for three areas, the average coverage ratio of each area should be 1 / 3 ≈ 0.33.

[0156] Balance correction: Adjust the comprehensive coverage of each candidate measurement point based on the difference between the coverage ratio and the average coverage ratio. The larger the difference, the more uneven the coverage, and the comprehensive coverage of the candidate measurement point should be reduced. The coverage balance factor can effectively avoid the concentration of measurement points in certain areas and ensure the uniformity of overall coverage.

[0157] In addition to maximizing coverage, another optimization goal is to minimize the geometric distribution difference between measurement points. The geometric distribution difference is used to measure the uniformity of the distribution of measurement points in space. The smaller the difference, the more uniform the distribution of measurement points.

[0158] Geometric distribution difference calculation: By calculating the distance difference between the measurement points, the geometric distribution is evaluated. For example, if the distance between two measurement points is close to the preset ideal spacing, the geometric distribution difference is small; on the contrary, if the distance between the two points is too large or too small, the difference increases.

[0159] Distance constraint: To avoid the distance between measurement points being too small, a lower limit constraint on the distance between adjacent measurement points is introduced into the model. That is, the distance between any two measurement points must not be less than the preset lower limit, ensuring that the measurement points are evenly distributed and cover the entire critical area.

[0160] By comprehensively maximizing coverage and minimizing geometric distribution differences, combined with the constraints that the distance between adjacent measurement points is not less than the preset lower limit and each key area covers at least a preset number of measurement points, a multi-objective optimization model is finally constructed to select the optimal set of measurement points.

[0161] This paper proposes a measurement point selection method based on a multi-objective optimization model, which combines the two objectives of maximizing the coverage of key areas and minimizing the geometric distribution difference of measurement points to ensure the reasonable distribution of measurement points. By introducing coverage balance factors and distance constraints, this method can effectively optimize the measurement point selection process of 3D radar level meters and improve measurement accuracy and efficiency.

[0162] In an optional embodiment,

[0163] Taking maximizing the coverage of the key areas by the measuring points as the first goal and minimizing the geometric distribution difference between the measuring points as the second goal, combined with the constraint that the distance between adjacent measuring points is not less than the preset lower limit and each key area covers at least a preset number of measuring points, the multi-objective optimization model for selecting the measuring points of the 3D radar level meter is constructed by combining the first goal, the second goal and the constraint, and further includes:

[0164] For any two candidate measurement points, calculate the Euclidean distance between the two candidate measurement points;

[0165] Introducing a distance decay function to transform the Euclidean distance to obtain a corrected distance, wherein the distance decay function includes any one of an exponential decay function, a power law decay function, and a threshold decay function;

[0166] Based on the corrected distance, the sum of the corrected distances between all candidate measurement points is calculated to obtain the geometric distribution difference of the candidate measurement point set;

[0167] Calculate the direction vector between any two candidate measurement points, and calculate the direction difference factor of each candidate measurement point based on the sine square value of the angle between any two direction vectors;

[0168] Calculate the height of each candidate measurement point, and calculate the height difference factor of each candidate measurement point based on the sum of the squares of the height differences between the candidate measurement points;

[0169] The weight coefficients of distance difference, direction difference and height difference are introduced, and the weighted sum of the geometric distribution difference, the direction difference factor and the height difference factor is combined to obtain the final evaluation index of the geometric distribution difference between the measurement points as the second goal.

[0170] For example, in order to evaluate the geometric distribution of the measurement points in the three-dimensional space, it is first necessary to calculate the Euclidean distance between any two candidate measurement points.

[0171] Euclidean distance definition: Euclidean distance is the straight-line distance between two points in three-dimensional space, which is used to measure the distance between measurement points. For example, suppose there are two candidate measurement points A and B, the coordinates of point A are (2, 3, 5), and the coordinates of point B are (5, 7, 1), then the Euclidean distance between the two points reflects the actual distance between point A and point B in three-dimensional space.

[0172] The Euclidean distance between each pair of candidate measurement points is calculated, which is the basis for the subsequent geometric distribution difference calculation.

[0173] In order to more reasonably reflect the geometric distribution between measurement points, the Euclidean distance can be corrected by the distance decay function. Different distance decay functions are used to weight different distance ranges, so that the influence of distant measurement points on the optimization results is weakened.

[0174] Distance decay function type:

[0175] Exponential decay function: As the distance between two measurement points increases, their influence decays rapidly in an exponential form. This is suitable for situations where the influence of close measurement points needs to be highlighted.

[0176] Power-law decay function: As the distance increases, the influence decreases in a power-law manner. Compared with exponential decay, power-law decay is slower and is suitable for scenes with relatively uniform distance distribution.

[0177] Threshold attenuation function: Set a threshold. When the distance between measurement points exceeds this threshold, the attenuation function considers their influence to be extremely low. This is suitable for situations where the influence of distant measurement points needs to be strictly limited.

[0178] For example, assuming that the original Euclidean distance between two measurement points is 10 meters, if an exponential decay function is used, the corrected distance will be much less than 10 meters, perhaps around 2 meters; while if a power-law decay function is used, the corrected distance may be 5 meters.

[0179] Based on the corrected distances, the sum of the corrected distances between all candidate measurement points is calculated to obtain the geometric distribution difference of the candidate measurement point set.

[0180] Meaning of geometric distribution difference: Geometric distribution difference is used to measure whether the distribution between candidate measurement points is uniform. When the corrected distance between measurement points is relatively uniform, the geometric distribution difference is small, and vice versa. By calculating the corrected distance between all candidate measurement points, the uniformity of the overall geometric distribution can be evaluated. For example, assuming there are three candidate measurement points, and the distances between them after correction are 3 meters, 4 meters, and 5 meters, respectively, the geometric distribution difference is the sum of these three values, that is, 12 meters. A larger geometric distribution difference means that the distribution of measurement points is uneven.

[0181] In addition to distance, the direction difference between the measurement points is also an important factor in geometric distribution. The direction difference factor is used to measure whether the directions between any two measurement points are similar.

[0182] Direction vector calculation: For any two candidate measurement points, calculate the direction vector between them. The direction vector reflects the relative direction of the two measurement points. For example, if the coordinates of point A are (2, 3, 5) and the coordinates of point B are (5, 7, 1), the direction vector from A to B can be calculated by the difference between the coordinates of the two points.

[0183] Direction difference factor: Calculates the direction difference factor based on the angle between any two direction vectors. The smaller the angle, the closer the two direction vectors are, and the smaller the direction difference factor; the larger the angle, the larger the direction difference factor. By calculating the sine square value, the direction difference factor reflects the degree of direction difference between the two measurement points. For example, if the angle between the two direction vectors is 30 degrees, the direction difference factor is small; if the angle is 90 degrees, the direction difference factor is large.

[0184] The distribution of measurement points in the vertical direction is also an important component of the geometric distribution difference, so it is necessary to calculate the height difference factor of the candidate measurement points.

[0185] Height calculation: The height of each candidate measurement point is its coordinate value in the vertical direction. Assuming that the height of measurement point A is 5 meters and the height of measurement point B is 8 meters, the height difference between the two points is 3 meters.

[0186] Height difference factor: Based on the height difference between any two measurement points, the height difference factor is calculated. The height difference factor reflects the distribution of the measurement points in the vertical direction. The larger the height difference, the larger the height difference factor, indicating that the distribution of the measurement points in the vertical direction is more uneven. For example, if the height difference between the two measurement points is 1 meter, the height difference factor is small; if the height difference is 10 meters, the height difference factor is large.

[0187] In order to comprehensively measure the geometric distribution differences between measurement points, it is necessary to integrate the influence of distance difference, direction difference and height difference. By introducing weight coefficients, the influence of each difference factor on the final geometric distribution difference evaluation index can be adjusted according to actual needs.

[0188] Introduction of weight coefficients: Different weights are assigned to distance difference, direction difference, and height difference according to specific application scenarios. For example, in some scenarios, the direction difference of the measurement points may be more important, so a larger weight can be assigned to it; in other scenarios, the distance difference may be more important. For example, assume that the weight coefficients are: distance difference weight is 0.5, direction difference weight is 0.3, and height difference weight is 0.2. If the distance difference of a set of measurement points is 5, the direction difference is 3, and the height difference is 2, then the geometric distribution difference evaluation index after weighted summation is: 5×0.5 +3×0.3+ 2×0.2 = 4.1.

[0189] Geometric distribution difference evaluation index: Through weighted summation, the final evaluation index of the geometric distribution difference between measurement points is obtained. This evaluation index is the core of the second objective in the optimization model and is used to measure whether the distribution between measurement points is uniform.

[0190] By taking maximizing coverage as the first goal and minimizing geometric distribution difference as the second goal, combined with the constraints that the distance between adjacent measurement points is not less than the preset lower limit and each key area covers at least a preset number of measurement points, a complete multi-objective optimization model is constructed.

[0191] Objective trade-off: In practical applications, it may be necessary to make a trade-off between coverage and geometric distribution diversity. For example, in some scenarios, priority is given to coverage of key areas, while in other scenarios, more attention is paid to the uniform distribution of measurement points. By adjusting the objective weights in the optimization model, different application requirements can be flexibly responded to.

[0192] The multi-objective optimization model proposed in this paper can effectively optimize the measurement point selection process of the 3D radar level meter by comprehensively considering the coverage of the measurement points to the key areas and the geometric distribution differences between the measurement points. By introducing the distance attenuation function, the direction difference factor and the height difference factor, and combining the weighted summation method, an evaluation index that can measure the rationality of the geometric layout of the measurement points is finally obtained. This model can be widely used in material monitoring and industrial automation scenarios in different industries to help improve measurement accuracy and efficiency.

[0193] In an optional embodiment,

[0194] The corresponding weight coefficients are set for the first objective and the second objective in the multi-objective optimization model respectively, and the multi-objective optimization problem is converted into a single-objective optimization problem. The single-objective optimization problem is solved by an improved differential evolution algorithm, and the solution with the best comprehensive evaluation index is selected as the final 3D radar level meter measurement point optimization layout scheme, including:

[0195] Initialize the differential evolution algorithm parameters, including population size NP, mutation factor F, crossover probability CR and maximum number of iterations G max , and simulated annealing algorithm parameters, including the initial temperature T 0 and cooling coefficient α; randomly generate NP candidate solutions as the initial population of the differential evolution algorithm, and let the current temperature T = T 0 ;

[0196] Perform adaptive mutation operation on each candidate solution of differential evolution algorithm to generate mutation vector, and use ε constraint method to deal with multiple constraints of single objective optimization problem;

[0197] Perform crossover operation on each candidate solution of the differential evolution algorithm and the corresponding mutation vector to generate a test vector;

[0198] Calculate the comprehensive evaluation index of each test vector and compare it with the comprehensive evaluation index of the corresponding candidate solution; if the comprehensive evaluation index of the test vector is better than the candidate solution, accept the test vector as a new candidate solution; otherwise, according to the importance sampling criterion, accept the test vector as a new candidate solution with the solution probability corresponding to the importance sampling criterion; at every certain number of iterations, combine the cooling coefficient to decay the current temperature of the simulated annealing algorithm;

[0199] Determine whether the differential evolution algorithm meets the maximum number of iterations or other termination conditions. If so, output the current optimal solution as the optimal layout plan for the measurement points of the 3D radar level meter; otherwise, continue to perform the mutation, crossover and selection operations of the differential evolution algorithm until the termination conditions are met.

[0200] For example, in the optimization layout of measurement points of 3D radar level meters, the spatial distribution of measurement points needs to meet multiple objectives at the same time: maximizing the coverage of key areas by measurement points and minimizing the geometric distribution differences between measurement points. In order to effectively solve this multi-objective optimization problem, this paper proposes a comprehensive optimization method based on differential evolution (DE) and simulated annealing (SA). By converting the multi-objective optimization problem into a single-objective optimization problem and solving it using the improved differential evolution algorithm, the optimal measurement point layout solution is finally selected.

[0201] The technical solution of this paper includes the following main steps: setting corresponding weight coefficients for the first and second objectives in the multi-objective optimization problem, and converting the multi-objective optimization problem into a single-objective optimization problem. Using an improved differential evolution algorithm to solve the single-objective optimization problem. Combining the temperature attenuation mechanism of the simulated annealing algorithm, the global search capability of the solution is further optimized. Output the solution with the best comprehensive evaluation index as the final optimization layout plan for the measurement points of the 3D radar level meter.

[0202] In the multi-objective optimization problem, the first goal is to maximize the coverage of the key area by the measurement points, and the second goal is to minimize the geometric distribution difference between the measurement points. In order to transform the multi-objective optimization problem into a single-objective optimization problem, it is first necessary to set weight coefficients for these two objectives.

[0203] Setting of weight coefficients: Different weight coefficients are assigned to the first and second objectives according to actual needs. The weight coefficients reflect the importance of different objectives in the optimization process. For some application scenarios, higher coverage may be required, in which case the weight coefficient of the first objective should be larger; in other scenarios, more attention may be paid to the uniformity of the geometric distribution of the measurement points, in which case the weight coefficient of the second objective should be larger. For example, in a silo monitoring system, if the system pays more attention to the comprehensiveness of coverage of key areas, the weight of the first objective may be set to 0.7 and the weight of the second objective to 0.3.

[0204] The differential evolution algorithm is a population-based optimization algorithm that is suitable for complex multi-objective optimization problems. In this algorithm, the algorithm parameters need to be initialized first. Initialization of differential evolution algorithm parameters:

[0205] Population size NP: represents the number of candidate solutions. A larger population size can improve the diversity of solutions, but it will also increase the computational complexity. The population size can be set to a value between 50 and 100 according to the scale of the problem. Mutation factor F: a control parameter used to generate mutation vectors, usually set between 0.5 and 1.0. Crossover probability CR: used to control the crossover operation between candidate solutions and mutation vectors, usually set to around 0.9. Maximum number of iterations Gmax: represents the maximum number of iterations of the algorithm, usually set between 100 and 500 according to the complexity of the problem.

[0206] Initialization of simulated annealing algorithm parameters: Initial temperature T0: represents the initial temperature of simulated annealing, which is used to control the probability of accepting the solution. The initial temperature is usually set to a higher value, such as 100. Cooling coefficient α: controls the temperature decay rate, usually set to about 0.9, that is, the temperature decreases by 10% after each iteration.

[0207] Generation of initial population: Randomly generate NP candidate solutions as the initial population of the differential evolution algorithm. Each candidate solution represents a layout plan of a measurement point, including the spatial coordinates of the measurement point and its corresponding comprehensive evaluation index. Set the current temperature T as the initial temperature T0.

[0208] In each iteration, the differential evolution algorithm generates new solutions through mutation operations. Adaptive mutation operations dynamically adjust the mutation strategy according to the evolutionary state of the current population to improve the algorithm's search ability. Generation of mutation vectors: For each candidate solution, select several random solutions (usually 3) and generate a mutation vector through differential operations. The mutation vector represents the direction and magnitude of the new solution.

[0209] Constraint processing: The measurement point layout needs to meet certain constraints, such as the distance between adjacent measurement points is not less than the preset lower limit, and each key area covers at least a preset number of measurement points. In order to deal with these constraints, the ε constraint method is used. During the mutation process, if the generated candidate solution does not meet the constraints, the ε constraint method is used to modify the solution to make it meet the constraints.

[0210] The new solution generated by the mutation operation needs to be cross-operated with the current candidate solution to generate a test vector.

[0211] Crossover operation: Crossover operation refers to combining the current candidate solution with the mutation vector to generate a new test vector. The main purpose of the crossover operation is to generate a new solution with potential superiority by combining some features of different solutions. The crossover probability CR determines whether each variable comes from a candidate solution or a mutation vector.

[0212] Generation of trial vectors: The trial vectors generated by the crossover operation represent a new measurement point arrangement. The trial vectors will be compared with the current candidate solution to determine whether to replace the candidate solution.

[0213] For each generated trial vector and candidate solution, their comprehensive evaluation index needs to be calculated and a selection operation needs to be performed.

[0214] Calculation of comprehensive evaluation index: Calculate the comprehensive evaluation index of the test vector and candidate solution according to the weight coefficients set above. The comprehensive evaluation index is a weighted sum of the coverage of the key area by the measurement points and the difference in geometric distribution of the measurement points. Solutions with high coverage and small difference in geometric distribution have higher comprehensive evaluation index. For example, if the coverage of a test vector is 80% and the difference in geometric distribution is 10%, if the weight coefficient of coverage is 0.7 and the weight coefficient of geometric distribution difference is 0.3, then the comprehensive evaluation index of the test vector is higher.

[0215] Selection operation: Compare the comprehensive evaluation index of the test vector with the corresponding candidate solution. If the comprehensive evaluation index of the test vector is better than the candidate solution, the test vector is accepted as a new candidate solution; otherwise, the importance sampling criterion is adopted to accept the test vector as a new candidate solution with a certain probability. The importance sampling criterion allows the acceptance of poor solutions through a certain probability mechanism, thereby avoiding falling into the local optimum.

[0216] In order to enhance the global search capability of the algorithm, the temperature decay mechanism of the simulated annealing algorithm is combined to decay the current temperature every certain number of iterations.

[0217] Temperature decay mechanism: During each iteration, the current temperature T decays according to the cooling coefficient α. The higher the temperature, the greater the probability of accepting a poor solution; as the temperature gradually decreases, the algorithm gradually tends to converge. The temperature decay mechanism can prevent the algorithm from converging to the local optimum too early. For example, if the initial temperature is 100 and the cooling coefficient is 0.9, after 10 iterations, the temperature decays to 100×(0.9^10) ≈ 34.87.

[0218] After each iteration, determine whether the differential evolution algorithm meets the maximum number of iterations or other termination conditions. Termination conditions: Common termination conditions include reaching the maximum number of iterations, the solution of the population no longer changing significantly, etc. When the termination conditions are met, the algorithm stops iterating and outputs the current optimal solution as the optimal layout of the measurement points of the 3D radar level meter. For example, if the maximum number of iterations is set to 100, after the algorithm executes 100 iterations, it outputs the solution with the highest current comprehensive evaluation index; or, if the comprehensive evaluation index of the population does not change much in subsequent generations, the algorithm is terminated early.

[0219] This paper proposes a measurement point optimization layout scheme based on differential evolution algorithm and simulated annealing algorithm. By converting the multi-objective optimization problem into a single-objective optimization problem, combined with adaptive mutation operation, crossover operation and the temperature attenuation mechanism of simulated annealing algorithm, the measurement point layout of the 3D radar level meter can be effectively optimized. This method can not only maximize the coverage of the key areas by the measurement points, but also minimize the geometric distribution differences between the measurement points, ensuring the reasonable distribution of the measurement points.

[0220] Figure 2 FIG. 1 is a schematic diagram of a system for selecting measurement points based on a 3D radar level meter according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0221] The first unit is used to obtain original 3D radar point cloud data of a measurement scene, wherein the original 3D radar point cloud data includes three-dimensional coordinate information of each point in the measurement scene; based on the original 3D radar point cloud data, an improved octree algorithm is used to adaptively divide the measurement scene into multiple sub-areas, and the measurement scene is recursively divided into multiple sub-areas until the number of point clouds in each sub-area is lower than a preset threshold or reaches a preset maximum division depth, thereby obtaining a three-dimensional reconstruction result of the measurement scene; normal vector calculation and regional clustering are performed on the three-dimensional reconstruction result, and a plurality of key areas in the three-dimensional reconstruction result are determined in combination with a regional growing and merging algorithm;

[0222] The second unit is used to maximize the coverage of the key area by the measuring points as the first goal, minimize the geometric distribution difference between the measuring points as the second goal, and take the distance between adjacent measuring points as not less than a preset lower limit and each key area covers at least a preset number of measuring points as constraints, and comprehensively consider the first goal, the second goal and the constraints to construct a multi-objective optimization model for the selection of measuring points of the 3D radar level meter;

[0223] The third unit is used to set corresponding weight coefficients for the first objective and the second objective in the multi-objective optimization model respectively, convert the multi-objective optimization problem into a single-objective optimization problem, solve the single-objective optimization problem through an improved differential evolution algorithm, and select the solution with the best comprehensive evaluation index as the final 3D radar level meter measurement point optimization layout scheme, wherein the improved differential evolution algorithm introduces an adaptive mutation strategy to improve the global convergence of the algorithm, and adopts the ε constraint method to deal with multiple constraints of the single-objective optimization problem.

[0224] According to a third aspect of the embodiments of the present invention,

[0225] An electronic device is provided, comprising:

[0226] processor;

[0227] a memory for storing processor-executable instructions;

[0228] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0229] A fourth aspect of the embodiments of the present invention is:

[0230] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0231] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0232] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. The method for selecting measurement points based on 3D radar level meter is characterized in that: include: Acquire original 3D radar point cloud data of a measurement scene, wherein the original 3D radar point cloud data includes three-dimensional coordinate information of each point in the measurement scene; based on the original 3D radar point cloud data, use an improved octree algorithm to adaptively divide the measurement scene into multiple sub-areas by introducing point cloud density and normal vector variance as measurement indicators of local point cloud distribution characteristics, dynamically adjust the division threshold and division depth, and recursively divide the measurement scene into multiple sub-areas until the number of point clouds in each sub-area is lower than a preset threshold or reaches a preset maximum division depth, thereby obtaining a three-dimensional reconstruction result of the measurement scene; perform normal vector calculation and regional clustering on the three-dimensional reconstruction result, and determine multiple key areas in the three-dimensional reconstruction result in combination with a regional growing and merging algorithm; Taking maximizing the coverage of the key areas by the measuring points as the first goal and minimizing the geometric distribution difference between the measuring points as the second goal, combined with the constraint conditions that the distance between adjacent measuring points is not less than the preset lower limit and each key area covers at least a preset number of measuring points, the first goal, the second goal and the constraint conditions are comprehensively considered to construct a multi-objective optimization model for the selection of measuring points of the 3D radar level meter; Corresponding weight coefficients are set for the first objective and the second objective in the multi-objective optimization model respectively, and the multi-objective optimization problem is converted into a single-objective optimization problem. The single-objective optimization problem is solved by an improved differential evolution algorithm, and the solution with the best comprehensive evaluation index is selected as the final 3D radar level meter measurement point optimization layout scheme, wherein the improved differential evolution algorithm introduces an adaptive mutation strategy to improve the global convergence of the algorithm, and adopts the ε constraint method to deal with multiple constraints of the single-objective optimization problem.

2. The method according to claim 1, characterized in that Based on the original 3D radar point cloud data, an improved octree algorithm is used to adaptively divide the measurement scene into multiple sub-areas, and the measurement scene is recursively divided into multiple sub-areas until the number of point clouds in each sub-area is lower than a preset threshold or reaches a preset maximum division depth, and the three-dimensional reconstruction result of the measurement scene is obtained, including: Loading the original 3D radar point cloud data into the root node of the octree, the original 3D radar point cloud data including the three-dimensional coordinates of each point in the measurement scene; Calculate the point cloud density and normal vector variance of the current octree node as indicators to measure the local point cloud distribution characteristics, where the point cloud density is equal to the number of point clouds in the node divided by the volume of the cube corresponding to the node; the normal vector variance is equal to the sum of the squares of the modulus of the difference between the normal vector of each point in the node and the average value of the normal vectors of all points divided by the number of point clouds in the node; According to the point cloud density and the normal vector variance, the segmentation threshold and the segmentation depth are adaptively adjusted, and the specific adjustment method is: When the point cloud density is greater than a preset point cloud density threshold and the normal vector variance is greater than a preset normal vector variance threshold, multiplying the division threshold by a first preset coefficient greater than 1, and adding 1 to the division depth; When the point cloud density is greater than a preset point cloud density threshold and the normal vector variance is less than or equal to a preset normal vector variance threshold, multiplying the division threshold by a second preset coefficient less than 1, and adding 1 to the division depth; When the point cloud density is less than or equal to a preset point cloud density threshold and the normal vector variance is greater than a preset normal vector variance threshold, dividing the division threshold by a third preset coefficient greater than 1, and adding 1 to the division depth; When the point cloud density is less than or equal to a preset point cloud density threshold and the normal vector variance is less than or equal to a preset normal vector variance threshold, the node is marked as a leaf node; For the current node, if the number of its point clouds is greater than the division threshold and the division depth is less than or equal to the preset maximum division depth, the node is divided into 8 sub-nodes, and the point cloud data of the current node is divided into the corresponding sub-nodes; For each newly generated non-empty child node, the partition threshold and partition depth are adaptively adjusted recursively until all nodes are leaf nodes or the partition termination condition is reached; The octree partitioning result is output, including the spatial range of each leaf node and the point cloud data contained therein, to obtain a three-dimensional reconstruction result of the measurement scene.

3. The method according to claim 1, characterized in that The normal vector calculation and regional clustering are performed on the three-dimensional reconstruction result, and a region growing and merging algorithm is combined to determine that a plurality of key regions in the three-dimensional reconstruction result include: For each sub-region in the three-dimensional reconstruction result, the normal vectors of all points in the sub-region are calculated by principal component analysis; the point cloud data set in the sub-region is centralized to obtain a decentralized point set, and the coordinates of each point in the decentralized point set are equal to the coordinates of the original 3D radar point cloud data minus the centroid coordinates of the point cloud data set; the covariance matrix of the decentralized point set is calculated, and the covariance matrix is ​​subjected to eigenvalue decomposition to obtain eigenvalues ​​and corresponding eigenvectors arranged in order of size; the eigenvector corresponding to the minimum eigenvalue is taken as the normal vector of the sub-region, indicating the main direction of the sub-region; Normalize the normal vectors of all sub-regions in the three-dimensional reconstruction result to obtain a unit normal vector set; cluster the unit normal vector set using a spectral clustering algorithm to obtain multiple normal vector clusters, each normal vector cluster corresponds to a group of sub-regions with similar directions; the specific steps of the spectral clustering algorithm are: Calculating a similarity matrix between normal vectors in the unit normal vector set, wherein the elements of the similarity matrix are Gaussian kernel function values ​​corresponding to two normal vectors; normalizing the similarity matrix to obtain a normalized Laplace matrix; performing eigenvalue decomposition on the normalized Laplace matrix to obtain eigenvectors corresponding to the first k largest eigenvalues; stacking the eigenvectors in rows to form a matrix, and normalizing each row of the matrix; taking each row of the normalized matrix as a clustering sample, and applying a K-means algorithm to perform clustering to obtain the normal vector cluster; For each normal vector cluster, randomly select a sub-region as the initial seed region, and calculate the average normal vector of the initial seed region; traverse other sub-regions in the current normal vector cluster, and for each sub-region, calculate the distance from its centroid to the centroid of the initial seed region and the angle between its normal vector and the average normal vector of the initial seed region; if the distance is less than a preset distance threshold and the angle is less than a preset angle threshold, merge the sub-region into the initial seed region, update the average normal vector of the initial seed region, and remove the sub-region from the current normal vector cluster; repeat the iteration until no new sub-region can be merged into the current normal vector cluster; The final obtained initial seed area is marked as a key area, and is removed from the current normal vector cluster; the iteration is repeated until the current normal vector cluster is empty, and multiple key areas in the three-dimensional reconstruction result are obtained.

4. The method according to claim 1, characterized in that: Taking maximizing the coverage of the key areas by the measuring points as the first goal and minimizing the geometric distribution difference between the measuring points as the second goal, combined with the constraint that the distance between adjacent measuring points is not less than the preset lower limit and each key area covers at least a preset number of measuring points, the multi-objective optimization model for selecting the measuring points of the 3D radar level meter is constructed by combining the first goal, the second goal and the constraint, including: For each candidate measurement point, calculating the distance from the candidate measurement point to each key area; Using a Gaussian kernel function to calculate the coverage of each key area by the candidate measurement points; The importance weight of the key area is introduced, and the importance weight is calculated according to the area, volume and shape complexity of the key area. The area weight of the key area is calculated by the area ratio, the volume weight of the key area is calculated by the volume ratio, and the shape complexity weight of the key area is calculated by the ratio of the surface area to the volume of the key area. Define the comprehensive coverage of the candidate measurement points to all key areas as the weighted sum of the coverage of each key area, and the weight of the weighted sum is the importance weight; A coverage balance factor is introduced to calculate the coverage ratio of each candidate measurement point to each key area and the average coverage ratio of the candidate measurement point to each key area. The comprehensive coverage is corrected based on the difference between the coverage ratio and the average coverage ratio and the coverage balance factor to obtain a coverage evaluation index as the first goal.

5. The method according to claim 4, characterized in that Taking maximizing the coverage of the key areas by the measuring points as the first goal and minimizing the geometric distribution difference between the measuring points as the second goal, combined with the constraint that the distance between adjacent measuring points is not less than the preset lower limit and each key area covers at least a preset number of measuring points, the multi-objective optimization model for selecting the measuring points of the 3D radar level meter is constructed by combining the first goal, the second goal and the constraint, and further includes: For any two candidate measurement points, calculate the Euclidean distance between the two candidate measurement points; Introducing a distance decay function to transform the Euclidean distance to obtain a corrected distance, wherein the distance decay function includes any one of an exponential decay function, a power law decay function, and a threshold decay function; Based on the corrected distance, the sum of the corrected distances between all candidate measurement points is calculated to obtain the geometric distribution difference of the candidate measurement point set; Calculate the direction vector between any two candidate measurement points, and calculate the direction difference factor of each candidate measurement point based on the sine square value of the angle between any two direction vectors; Calculate the height of each candidate measurement point, and calculate the height difference factor of each candidate measurement point based on the sum of the squares of the height differences between the candidate measurement points; The weight coefficients of distance difference, direction difference and height difference are introduced, and the weighted sum of the geometric distribution difference, the direction difference factor and the height difference factor is combined to obtain the final evaluation index of the geometric distribution difference between the measurement points as the second goal.

6. The method according to claim 1, characterized in that The corresponding weight coefficients are set for the first objective and the second objective in the multi-objective optimization model respectively, and the multi-objective optimization problem is converted into a single-objective optimization problem. The single-objective optimization problem is solved by an improved differential evolution algorithm, and the solution with the best comprehensive evaluation index is selected as the final 3D radar level meter measurement point optimization layout scheme, including: Initialize the differential evolution algorithm parameters, including population size NP, mutation factor F, crossover probability CR and maximum number of iterations G max , and simulated annealing algorithm parameters, including initial temperature T0 and cooling coefficient α; randomly generate NP candidate solutions as the initial population of the differential evolution algorithm, and set the current temperature T=T0; Perform adaptive mutation operation on each candidate solution of differential evolution algorithm to generate mutation vector, and use ε constraint method to deal with multiple constraints of single objective optimization problem; Perform crossover operation on each candidate solution of the differential evolution algorithm and the corresponding mutation vector to generate a test vector; Calculate the comprehensive evaluation index of each test vector and compare it with the comprehensive evaluation index of the corresponding candidate solution; if the comprehensive evaluation index of the test vector is better than the candidate solution, accept the test vector as a new candidate solution; otherwise, according to the importance sampling criterion, accept the test vector as a new candidate solution with the solution probability corresponding to the importance sampling criterion; at every certain number of iterations, combine the cooling coefficient to decay the current temperature of the simulated annealing algorithm; Determine whether the differential evolution algorithm meets the maximum number of iterations or other termination conditions. If so, output the current optimal solution as the optimal layout plan for the measurement points of the 3D radar level meter; otherwise, continue to perform the mutation, crossover and selection operations of the differential evolution algorithm until the termination conditions are met.

7. A 3D radar level meter measurement point selection system, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to obtain original 3D radar point cloud data of a measurement scene, wherein the original 3D radar point cloud data includes three-dimensional coordinate information of each point in the measurement scene; based on the original 3D radar point cloud data, an improved octree algorithm is used to adaptively divide the measurement scene into multiple sub-areas, and the measurement scene is recursively divided into multiple sub-areas until the number of point clouds in each sub-area is lower than a preset threshold or reaches a preset maximum division depth, thereby obtaining a three-dimensional reconstruction result of the measurement scene; normal vector calculation and regional clustering are performed on the three-dimensional reconstruction result, and a plurality of key areas in the three-dimensional reconstruction result are determined in combination with a regional growing and merging algorithm; The second unit is used to maximize the coverage of the key area by the measuring points as the first goal, minimize the geometric distribution difference between the measuring points as the second goal, and take the distance between adjacent measuring points as not less than a preset lower limit and each key area covers at least a preset number of measuring points as constraints, and comprehensively consider the first goal, the second goal and the constraints to construct a multi-objective optimization model for the selection of measuring points of the 3D radar level meter; The third unit is used to set corresponding weight coefficients for the first objective and the second objective in the multi-objective optimization model respectively, convert the multi-objective optimization problem into a single-objective optimization problem, solve the single-objective optimization problem through an improved differential evolution algorithm, and select the solution with the best comprehensive evaluation index as the final 3D radar level meter measurement point optimization layout scheme, wherein the improved differential evolution algorithm introduces an adaptive mutation strategy to improve the global convergence of the algorithm, and adopts the ε constraint method to deal with multiple constraints of the single-objective optimization problem.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images