Unmanned vehicle fast point cloud segmentation algorithm in complex environment

Abstract

The application discloses a kind of unmanned vehicle fast point cloud segmentation algorithm under complex environment, by original laser radar point cloud is encoded as initial equilateral triangle grid chart structure, introduce regularized space representation, to enhance the expression ability of point cloud local geometry;For each point cloud in equilateral triangle grid, the local geometric feature is extracted and the grid flatness is calculated using principal component analysis method, so as to realize the effective distinction of ground area and non-ground area;For the grid area with higher flatness, directly determine the internal point cloud as ground point cloud;For the grid area with lower flatness and complex terrain, then improve the local modeling accuracy by adaptive subdivision operation;In the non-ground point cloud segmentation stage, an adaptive euclidean clustering algorithm based on laser radar angle resolution is used to accelerate the segmentation, and the clustering scale is set adaptively for point clouds in different distance ranges, so as to significantly reduce the computational complexity while ensuring the accuracy of segmentation.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, specifically to a fast point cloud segmentation algorithm for unmanned vehicles in complex environments. Background Technology

[0002] The realization of autonomous driving relies on the vehicle's perception of external information. Complete and accurate environmental information is the foundation for path planning, behavioral decision-making, and motion control, and obstacle detection is an important research area in environmental perception. Currently, the sensors used for obstacle detection include LiDAR and cameras. Cameras are greatly affected by lighting conditions, and it is difficult to obtain depth information. LiDAR, on the other hand, has strong anti-interference capabilities and can obtain accurate distance and reflection intensity information, and is therefore widely used in autonomous vehicles.

[0003] The realization of autonomous vehicles depends on effective perception of the surrounding environment. Point cloud segmentation is one of the key technologies in LiDAR environmental perception, and its results directly affect the reliability of functions such as obstacle detection, path planning and motion control.

[0004] Currently, the environmental perception sensors commonly used in autonomous vehicles mainly include cameras and LiDAR. Cameras are easily affected by lighting and weather conditions, and it is difficult to obtain accurate three-dimensional spatial information; in contrast, LiDAR has advantages such as high ranging accuracy and strong anti-interference ability, and can directly acquire three-dimensional point cloud data, so it is widely used in autonomous vehicles.

[0005] Existing LiDAR-based point cloud segmentation methods mostly rely on strategies such as fixed height thresholds, regular grids, or fixed-scale clustering, which can achieve point cloud segmentation in certain scenarios. However, in practical applications, point cloud density varies significantly with distance, and there are differences between ground morphology and target scale. Segmentation methods with fixed parameters struggle to simultaneously achieve both segmentation accuracy and processing efficiency, easily leading to inaccurate segmentation or insufficient real-time performance. Therefore, it is necessary to propose a LiDAR point cloud segmentation method that can improve processing speed while ensuring point cloud segmentation accuracy, to meet the real-time and reliability requirements of autonomous vehicles. Summary of the Invention

[0006] The purpose of this invention is to provide a fast point cloud segmentation algorithm for unmanned vehicles in complex environments.

[0007] To achieve the above-mentioned technical effects, the present invention is implemented through the following technical solution:

[0008] A fast point cloud segmentation algorithm for unmanned vehicles in complex environments, characterized by the following steps:

[0009] S1: Encode the original radar point cloud into a larger initial equilateral triangular grid structure—ETG (equilateral triangular grid), specifically including:

[0010] An equilateral triangular regular grid field is pre-constructed with a fixed spatial resolution on the XY plane of the Cartesian coordinate system. Each equilateral triangular grid... Corresponding local point cloud ETG consists of a set of nodes and a set of edges:

[0011]

[0012] In the formula: This represents the set of nodes in an ETG, where each node corresponds to a triangular unit. Indicates the first The node corresponding to the graph node. One equilateral triangle unit; Represents the first equilateral triangular mesh in the initial triangular mesh. One triangular unit; Represents the initial set of equilateral triangles; Represents the set of edges in a graph, used to connect adjacent nodes; Connecting nodes and A side of a triangle indicates that the two triangles are adjacent; Indicates the index number of the triangle;

[0013] S2: Perform principal component analysis on the point cloud in each initial equilateral triangular mesh and calculate the mesh flatness, specifically including:

[0014] For the initial equilateral triangular mesh Point clouds inside Perform principal component analysis and set the point cloud mean. ,set up Here is the point cloud covariance matrix:

[0015]

[0016] In the formula: T is the matrix transpose symbol;

[0017] For covariance matrix Performing eigenvalue decomposition, we have:

[0018]

[0019] In the formula, for eigenvalues; for eigenvectors, Indicates the number of eigenvalues; eigenvalues Arranged in descending order, we have:

[0020] S3: Calculate the triangular mesh based on the desired eigenvalues. Structural characteristic coefficients:

[0021]

[0022] In the formula: , , These represent triangular meshes. Linearity, flatness, and cohesion; To prevent extremely small positive numbers with a denominator of zero;

[0023] S4: The flatness of an equilateral triangular mesh is defined as follows:

[0024]

[0025] In the formula: Indicates the flatness of the equilateral triangular mesh; These represent the log-normalized results of linearity, flatness, and aggregation, respectively. These represent the weighting coefficients of linearity, flatness, and cohesion in flatness, respectively.

[0026] The log normalization expression is as follows:

[0027]

[0028] Essentially, it involves first compressing the dynamic range using a logarithmic function, and then mapping it to the [0,1] interval;

[0029] S5: Determine the flatness of the initial equilateral triangular mesh. If the terrain of a region mesh is flat, the point cloud within that region is considered a ground point cloud. If the region mesh is not flat, perform a subdivision operation on the mesh, specifically including:

[0030] Set flatness threshold ,when At that time, determine the initial equilateral triangular mesh. The covered area has complex terrain, requiring adaptive subdivision of the grid; when When the terrain of the grid area is determined to be flat, the point cloud in that area is considered as a ground point cloud;

[0031] S6: When For the initial equilateral triangular mesh Perform a subdivision operation (Level-0 → Level-1) to set the grid. The triangular element in the XY plane of the Cartesian coordinate system is:

[0032]

[0033] In the formula: Let be the coordinates of the three vertices of the equilateral triangle;

[0034] Taking the midpoints of the three sides of the triangular element, we get:

[0035]

[0036] In the formula: Representing the midpoints of the three sides, the triangular unit is formed by connecting these three midpoints. Divided into 4 sub-triangular units, namely:

[0037]

[0038]

[0039]

[0040]

[0041] In the formula, the superscript (1) indicates the grid level obtained from the first subdivision;

[0042] Each sub-triangle element is treated as a subdivided equilateral triangular mesh node. And based on its projection area in the XY plane, the original point cloud set is... The points are redistributed to the sub-grids to obtain the corresponding local point clouds:

[0043]

[0044] In the formula: Indicates the first in the parent grid One grid cell; This indicates the first subdivision obtained after subdividing the parent mesh cell. Sub-grid cells; This indicates that the point set belongs to the first-level equilateral triangular mesh; Represents a point cloud index; Represents the projection operator;

[0045] S7: After completing one subdivision, for each sub-mesh node Repeat steps S2, 3, 4, 5, and 6 to perform principal component analysis and calculate its flatness. Set the same flatness threshold as the initial layer. And decide whether to continue subdividing based on the following rules:

[0046] when At this point, further subdivision is performed on the submesh to obtain Level-2 subdivided submesh nodes. ;when When the subgrid is subdivided, the point cloud contained in that subgrid is considered as a candidate set of ground point clouds; the above calculation and discrimination process is repeated for the Level-2 subgrid nodes. Further subdivision yields Level-3 sub-mesh nodes. ,when Stop subdividing at that time;

[0047] S8: Set the following termination conditions during the subdivision process. If any condition is met, the subdivision of the mesh will stop:

[0048] 1. Maximum subdivision level constraint: Subdivision stops when the level reaches Level-3;

[0049] 2. Point Count Constraint: When the number of points H in the sub-grid is less than the minimum point count threshold. When this happens, mark it as an unknown mesh and stop subdividing;

[0050] 3. Minimum side length constraint: When the sub-mesh side length is less than the minimum resolution Stop subdividing at that time;

[0051] S9: Segmentation of non-ground point clouds, specifically including:

[0052] Based on the adaptive distance threshold setting of LiDAR angular resolution, let a frame of non-ground point cloud set be... The vertical angular resolution of the lidar is For the radial distance of the distance sensor, The spatial position of two adjacent laser scanning channels, and the theoretical distance in the vertical direction, can be expressed as:

[0053]

[0054] In the formula: This indicates the adaptive Euclidean distance threshold used for clustering; ;

[0055] S10: Divide the horizontal space into several concentric annular regions according to the radial distance of the distance sensor, and perform Euclidean clustering in each annular region using a fixed distance threshold;

[0056] Define a set of steps with a fixed step size Increasing distance threshold sequence:

[0057]

[0058] In the formula: Indicates the first The clustering distance threshold corresponding to each annular region; according to the formula Calculate each distance threshold The corresponding maximum clustering detection distance The outer radius of the annular region is defined as:

[0059]

[0060] In the formula: It is a rounding function;

[0061] The width between two adjacent annular regions is defined as:

[0062]

[0063] S11: After completing the ring division of the non-ground point cloud, Euclidean clustering is performed within each ring region using a corresponding fixed distance threshold, and combined with a neighborhood search structure to achieve rapid growth of point cloud clusters:

[0064] For the Let there be a ring-shaped region, and let its corresponding non-ground point cloud set be . The fixed Euclidean clustering distance threshold used within this annular region is For any two points If the following Euclidean distance constraint is satisfied:

[0065]

[0066] Then it is assumed that the two points have spatial connectivity within the annular region, where Representing the 3D Euclidean distance; based on the above distance determination relationship, constructing the adjacency relationship of the point cloud within the annular region, and performing Euclidean clustering using the neighborhood expansion method; arbitrarily selecting an unvisited point. As a seed point, iterative search is performed to find its satisfaction. The adjacent points are continuously expanded outwards to form a point cloud cluster, which can be formally represented as:

[0067]

[0068] In the formula: Indicates the first The i-th point cloud cluster is generated within each annular region; Use this as the seed point, and start performing neighborhood expansion and growth clustering from this point; This represents a connected path consisting of points. The endpoint of the path; any two adjacent points on the path satisfy the example threshold constraint.

[0069] Furthermore, in S10, the point cloud set within the annular region... Construct a kd-tree spatial index structure and perform a radius neighborhood search operation based on the index structure.

[0070] The fast point cloud segmentation algorithm for unmanned vehicles in complex environments provided by this invention has the following beneficial effects:

[0071] This invention achieves rapid differentiation between ground and non-ground point clouds by encoding the original LiDAR point cloud into an equilateral triangular mesh structure and introducing principal component analysis and flatness discrimination at the mesh scale. By performing mesh subdivision only on areas with complex terrain, the number of meshes is effectively reduced while maintaining the accuracy of ground modeling, thereby lowering the overall computational complexity. Furthermore, in the non-ground point cloud segmentation stage, this invention employs an adaptive Euclidean clustering method based on LiDAR angular resolution, combined with a ring-splitting acceleration strategy. This allows the clustering scale to adaptively adjust with changes in point cloud distance, while transforming global variable-threshold clustering into local fixed-threshold clustering, significantly improving point cloud clustering efficiency and segmentation stability. In summary, this invention effectively improves processing speed while maintaining point cloud segmentation accuracy, enabling rapid and accurate segmentation of LiDAR point clouds, and is suitable for unmanned vehicle perception applications requiring both real-time performance and accuracy. Attached Figure Description

[0072] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0073] Figure 1 This is a flowchart of the fast point cloud segmentation algorithm for unmanned vehicles in complex environments according to the present invention;

[0074] Figure 2 This is a schematic diagram of the equilateral triangular mesh and its subdivisions in this invention;

[0075] Figure 3 This is a schematic diagram of the non-ground point cloud ring segmentation strategy of the present invention;

[0076] Figure 4 This is the experimental platform described in the embodiments of the present invention;

[0077] Figure 5These are experimental results of the present invention under different environments, where white point clouds represent the ground and other point clouds of different colors represent different obstacles; Figure (a) is an experimental result of the experimental platform described in this embodiment at a complex crossroads; Figure (b) is an experimental result of the experimental platform described in this embodiment in an unstructured jungle environment. Detailed Implementation

[0078] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0079] Example 1

[0080] See Figures 1 to 5 As shown, a fast point cloud segmentation algorithm for unmanned vehicles in complex environments is characterized by the following steps:

[0081] S1: Encode the original radar point cloud into a larger initial equilateral triangular grid structure—ETG (equilateral triangular grid), specifically including:

[0082] ETG is a pre-constructed equilateral triangular regular grid field with a fixed spatial resolution on the XY plane of the Cartesian coordinate system. Each equilateral triangular grid... This corresponds to a portion of the local point cloud. ETG consists of a set of nodes and a set of edges:

[0083]

[0084] In the formula: This represents the set of nodes in an ETG, where each node corresponds to a triangular unit. Indicates the first The node corresponding to the graph node. One equilateral triangle unit; Represents the first equilateral triangular mesh in the initial triangular mesh. One triangular unit; Represents the initial set of equilateral triangles; Represents the set of edges in a graph, used to connect adjacent nodes; Connecting nodes and A side of a triangle indicates that the two triangles are adjacent; Indicates the index number of the triangle;

[0085] S2: Perform principal component analysis on the point cloud in each initial equilateral triangular mesh and calculate the mesh flatness, specifically including:

[0086] For the initial equilateral triangular mesh Point clouds inside Perform principal component analysis and set the point cloud mean. ,set up Here is the point cloud covariance matrix:

[0087]

[0088] In the formula: T is the matrix transpose symbol;

[0089] For covariance matrix Performing eigenvalue decomposition, we have:

[0090]

[0091] In the formula, for eigenvalues; for eigenvectors, Indicates the number of eigenvalues; eigenvalues Arranged in descending order, we have:

[0092] S3: Calculate the triangular mesh based on the desired eigenvalues. Structural characteristic coefficients:

[0093]

[0094] In the formula: , , These represent triangular meshes. Linearity, flatness, and cohesion; To prevent extremely small positive numbers with a denominator of zero;

[0095] S4: The flatness of an equilateral triangular mesh is defined as follows:

[0096]

[0097] In the formula: Indicates the flatness of the equilateral triangular mesh; These represent the log-normalized results of linearity, flatness, and aggregation, respectively. These represent the weighting coefficients of linearity, flatness, and cohesion in flatness, respectively, in this embodiment. Take values ​​of 0.5, 2, and 1 respectively;

[0098] Since the structural indices constructed based on eigenvalues ​​have a large dynamic range and are easily affected by local anomalous point clouds, logarithmic normalization is applied to each structural index before index fusion. Through logarithmic mapping, the numerical range of the indices can be effectively compressed, the influence of extreme values ​​can be suppressed, and the relative magnitude relationship between the indices can be maintained, thereby improving the stability and robustness of ground flatness assessment.

[0099] The log normalization expression is as follows:

[0100]

[0101] Essentially, it involves first compressing the dynamic range using a logarithmic function, and then mapping it to the [0,1] interval;

[0102] S5: Determine the flatness of the initial equilateral triangular mesh. If the terrain of a region mesh is flat, the point cloud within that region is considered a ground point cloud. If the region mesh is not flat, perform a subdivision operation on the mesh, specifically including:

[0103] Set flatness threshold In this embodiment Take 0.567, when At that time, determine the initial equilateral triangular mesh. The covered area has complex terrain, requiring adaptive subdivision of the grid; when When the terrain of the grid area is determined to be flat, the point cloud in that area is considered as a ground point cloud;

[0104] S6: When For the initial equilateral triangular mesh Perform a subdivision operation (Level-0 → Level-1) to set the grid. The triangular element in the XY plane of the Cartesian coordinate system is:

[0105]

[0106] In the formula: The coordinates of the three vertices of the equilateral triangle are given. In this embodiment, the side length of the initial equilateral triangle mesh in Level-0 is set to 6.

[0107] Taking the midpoints of the three sides of the triangular element, we get:

[0108]

[0109] In the formula: Representing the midpoints of the three sides, the triangular unit is formed by connecting these three midpoints. Divided into 4 sub-triangular units, namely:

[0110]

[0111]

[0112]

[0113]

[0114] In the formula, the superscript (1) indicates the grid level obtained from the first subdivision;

[0115] Each sub-triangle element is treated as a subdivided equilateral triangular mesh node. And based on its projection area in the XY plane, the original point cloud set is... The points are redistributed to the sub-grids to obtain the corresponding local point clouds:

[0116]

[0117] In the formula: Indicates the first in the parent grid One grid cell; This indicates the first subdivision obtained after subdividing the parent mesh cell. Sub-grid cells; This indicates that the point set belongs to the first-level equilateral triangular mesh; Represents a point cloud index; Represents the projection operator;

[0118] S7: After completing one subdivision, for each sub-mesh node Repeat steps S2, 3, 4, 5, and 6 to perform principal component analysis and calculate its flatness. In this embodiment, the same flatness threshold as the initial layer is set. And decide whether to continue subdividing based on the following rules:

[0119] when At this point, further subdivision is performed on the submesh to obtain Level-2 subdivided submesh nodes. ;when When the subgrid is subdivided, the point cloud contained in that subgrid is considered as a candidate set of ground point clouds; the above calculation and discrimination process is repeated for the Level-2 subgrid nodes. Further subdivision yields Level-3 sub-mesh nodes. ,when Stop subdividing at that time;

[0120] In a specific embodiment of the present invention, in order to balance computational efficiency and ground modeling accuracy, the maximum number of iterations of the subdivision operation is set to 3, that is, the maximum subdivision level is Level-3; when the grid level reaches Level-3, subdivision is no longer performed, and subsequent ground point cloud determination and output are directly performed on the Level-3 grid cells;

[0121] S8: Set the following termination conditions during the subdivision process. If any condition is met, the subdivision of the mesh will stop:

[0122] 1. Maximum subdivision level constraint: Subdivision stops when the level reaches Level-3;

[0123] 2. Point Count Constraint: When the number of points H in the sub-grid is less than the minimum point count threshold. When this happens, mark it as an unknown mesh and stop subdividing;

[0124] 3. Minimum side length constraint: When the sub-mesh side length is less than the minimum resolution Stop subdividing at that time;

[0125] After completing the above subdivision, the point clouds in all grid cells that have stopped subdivision are aggregated to form a non-ground point cloud, which is used for subsequent ground object segmentation.

[0126] S9: Segmentation of non-ground point clouds, specifically including:

[0127] Based on the adaptive distance threshold setting of LiDAR angular resolution, let a frame of non-ground point cloud set be... The vertical angular resolution of the lidar is In this embodiment for For the radial distance of the distance sensor, The spatial position of two adjacent laser scanning channels, and the theoretical distance in the vertical direction, can be expressed as:

[0128]

[0129] In the formula: This indicates the adaptive Euclidean distance threshold used for clustering; ;

[0130] S10: Divide the horizontal space into several concentric annular regions according to the radial distance of the distance sensor, and perform Euclidean clustering in each annular region using a fixed distance threshold;

[0131] Define a set of steps with a fixed step size An increasing distance threshold sequence, in this embodiment Take 4:

[0132]

[0133] In the formula: Indicates the first The clustering distance threshold corresponding to each annular region; according to the formula Calculate each distance threshold The corresponding maximum clustering detection distance The outer radius of the annular region is defined as:

[0134]

[0135] In the formula: It is a rounding function;

[0136] The width between two adjacent annular regions is defined as:

[0137]

[0138] In this way, the entire horizontal space is divided into multiple concentric annular regions centered on the sensor, with each annular region using a corresponding fixed distance threshold. Euclidean clustering is performed to discretize the originally continuously changing adaptive threshold into a finite number of region thresholds;

[0139] S11: After completing the ring division of the non-ground point cloud, Euclidean clustering is performed within each ring region using a corresponding fixed distance threshold, and combined with a neighborhood search structure to achieve rapid growth of point cloud clusters:

[0140] For the Let there be a ring-shaped region, and let its corresponding non-ground point cloud set be . The fixed Euclidean clustering distance threshold used within this annular region is For any two points If the following Euclidean distance constraint is satisfied:

[0141]

[0142] Then it is assumed that the two points have spatial connectivity within the annular region, where Representing the 3D Euclidean distance; based on the above distance determination relationship, constructing the adjacency relationship of the point cloud within the annular region, and performing Euclidean clustering using the neighborhood expansion method; arbitrarily selecting an unvisited point. As a seed point, iterative search is performed to find its satisfaction. The adjacent points are continuously expanded outwards to form a point cloud cluster, which can be formally represented as:

[0143]

[0144] In the formula: Indicates the first The i-th point cloud cluster is generated within each annular region; Use this as the seed point, and start performing neighborhood expansion and growth clustering from this point; This represents a connected path consisting of points. The endpoint of the path; any two adjacent points on the path satisfy the example threshold constraint.

[0145] In step S10, the point cloud set within the annular region... Construct a kd-tree spatial index structure and perform a radius neighborhood search operation based on the index structure.

[0146] use Figure 4 The experimental platform shown verifies the effectiveness of the fast point cloud segmentation algorithm for unmanned vehicles in complex environments proposed in this invention. The experimental platform is a self-designed all-terrain unmanned vehicle system, equipped with a monocular camera, a 16-line LiDAR, an inertial measurement unit (IMU), and an industrial control computer. All sensors are connected to and synchronized with the industrial control computer, acquiring and processing multi-source sensor data through a topic communication mechanism within the ROS framework.

[0147] To comprehensively evaluate the applicability and robustness of the method of this invention in different application scenarios, experimental scenarios were selected, including a complex intersection in a structured road environment and a jungle scene in an unstructured environment. In these scenarios, the lidar point cloud data collected by the unmanned vehicle during actual driving was processed to verify the effectiveness of the method of this invention in obstacle detection and point cloud segmentation.

[0148] In Figure (a), even at a complex intersection, the algorithm described in this invention can still accurately identify the ground and walls. Even when three vehicles are close together, the algorithm can still distinguish between them. In the complex jungle environment shown in Figure (b), the algorithm described in this invention can clearly and accurately identify pedestrians, sloping ground, and trees.

[0149] Experimental results show that the point cloud segmentation and obstacle detection method described in this invention exhibits excellent segmentation performance in typical scenarios such as complex intersections and jungles. Even when multiple vehicles are distributed close together, it can still accurately distinguish different vehicle targets, while achieving stable recognition of various environmental elements such as the ground, walls, pedestrians, sloping ground, and trees. Furthermore, by introducing adaptive modeling and loop-splitting acceleration strategies, the computational complexity is effectively reduced while ensuring segmentation accuracy, enabling the algorithm to run stably at a frequency of approximately 50Hz in this embodiment, meeting the real-time and accuracy requirements of autonomous vehicles.

Claims

1. An unmanned vehicle fast point cloud segmentation algorithm in a complex environment, characterized in that, Includes the following steps: S1: Encode the original radar point cloud into a larger initial equilateral triangular grid structure—ETG (equilateral triangular grid), specifically including: A regular grid field of equilateral triangles pre-constructed on the XY plane of a Cartesian coordinate system with a fixed spatial resolution, each equilateral triangle grid corresponding local point cloud An ETG consists of a set of nodes and a set of edges: In the formula: This represents the set of nodes in an ETG, where each node corresponds to a triangular unit. Indicates the first The node corresponding to the graph node. One equilateral triangle unit; Represents the first equilateral triangular mesh in the initial triangular mesh. One triangular unit; Represents the initial set of equilateral triangles; Represents the set of edges in a graph, used to connect adjacent nodes; Connecting nodes and A side of a triangle indicates that the two triangles are adjacent; Indicates the index number of the triangle; S2: Perform principal component analysis on the point cloud in each initial equilateral triangular mesh and calculate the mesh flatness, specifically including: on an initial equilateral triangular mesh point cloud within perform principal component analysis, let the point cloud mean be the point cloud covariance matrix: In the formula: T is the matrix transpose symbol; For covariance matrix Performing eigenvalue decomposition, we have: In the formula, for eigenvalues; for eigenvectors, Indicates the number of eigenvalues; eigenvalues Arranged in descending order, we have: S3: Calculate the triangular mesh based on the desired eigenvalues. Structural characteristic coefficients: In the formula: , , These represent triangular meshes. Linearity, flatness, and cohesion; To prevent extremely small positive numbers with a denominator of zero; S4: The flatness of an equilateral triangular mesh is defined as follows: In the formula: Indicates the flatness of the equilateral triangular mesh; These represent the log-normalized results of linearity, flatness, and aggregation, respectively. These represent the weighting coefficients of linearity, flatness, and cohesion in flatness, respectively. The log normalization expression is as follows: Essentially, it involves first compressing the dynamic range using a logarithmic function, and then mapping it to the [0,1] interval; S5: Determine the flatness of the initial equilateral triangular mesh. If the terrain of a region mesh is flat, the point cloud within that region is considered a ground point cloud. If the region mesh is not flat, perform a subdivision operation on the mesh, specifically including: Set flatness threshold ,when At that time, determine the initial equilateral triangular mesh. The covered area has complex terrain, requiring adaptive subdivision of the grid; when When the terrain of the grid area is determined to be flat, the point cloud in that area is considered as a ground point cloud; S6: When For the initial equilateral triangular mesh Perform a subdivision operation (Level-0 → Level-1) to set the grid. The triangular element in the XY plane of the Cartesian coordinate system is: In the formula: Let be the coordinates of the three vertices of the equilateral triangle; Taking the midpoints of the three sides of the triangular element, we get: In the formula: Representing the midpoints of the three sides, the triangular unit is formed by connecting these three midpoints. Divided into 4 sub-triangular units, namely: In the formula, the superscript (1) indicates the grid level obtained from the first subdivision; Each sub-triangle element is treated as a subdivided equilateral triangular mesh node. And based on its projection area in the XY plane, the original point cloud set is... The points are redistributed to the sub-grids to obtain the corresponding local point clouds: In the formula: Indicates the first in the parent grid One grid cell; This indicates the first subdivision obtained after subdividing the parent mesh cell. Sub-grid cells; This indicates that the point set belongs to the first-level equilateral triangular mesh; Represents a point cloud index; Represents the projection operator; S7: After completing one subdivision, for each sub-mesh node Repeat steps S2, 3, 4, 5, and 6 to perform principal component analysis and calculate its flatness. Set the same flatness threshold as the initial layer. And decide whether to continue subdividing based on the following rules: when At this point, further subdivision is performed on the submesh to obtain Level-2 subdivided submesh nodes. ;when When the subgrid is subdivided, the point cloud contained in that subgrid is considered as a candidate set of ground point clouds; the above calculation and discrimination process is repeated for the Level-2 subgrid nodes. Further subdivision yields Level-3 sub-mesh nodes. ,when Stop subdividing at that time; S8: Set the following termination conditions during the subdivision process. If any condition is met, the subdivision of the mesh will stop:

1. Maximum subdivision level constraint: Subdivision stops when the level reaches Level-3; 2. Point Count Constraint: When the number of points H in the sub-grid is less than the minimum point count threshold. When this happens, mark it as an unknown mesh and stop subdividing; 3. Minimum side length constraint: When the sub-mesh side length is less than the minimum resolution Stop subdividing at that time; S9: Segmentation of non-ground point clouds, specifically including: Based on the adaptive distance threshold setting of LiDAR angular resolution, let a frame of non-ground point cloud set be... The vertical angular resolution of the lidar is For the radial distance of the distance sensor, The spatial position of two adjacent laser scanning channels, and the theoretical distance in the vertical direction, can be expressed as: In the formula: This indicates the adaptive Euclidean distance threshold used for clustering; ; S10: Divide the horizontal space into several concentric annular regions according to the radial distance of the distance sensor, and perform Euclidean clustering in each annular region using a fixed distance threshold; Define a set of steps with a fixed step size Increasing distance threshold sequence: In the formula: Indicates the first The clustering distance threshold corresponding to each annular region; according to the formula Calculate each distance threshold The corresponding maximum clustering detection distance The outer radius of the annular region is defined as: In the formula: It is a rounding function; The width between two adjacent annular regions is defined as: S11: After completing the ring division of the non-ground point cloud, Euclidean clustering is performed within each ring region using a corresponding fixed distance threshold, and combined with a neighborhood search structure to achieve rapid growth of point cloud clusters: For the Let there be a ring-shaped region, and let its corresponding non-ground point cloud set be . The fixed Euclidean clustering distance threshold used within this annular region is For any two points If the following Euclidean distance constraint is satisfied: Then it is assumed that the two points have spatial connectivity within the annular region, where Representing the 3D Euclidean distance; based on the above distance determination relationship, constructing the adjacency relationship of the point cloud within the annular region, and performing Euclidean clustering using a neighborhood expansion method; arbitrarily selecting an unvisited point. As a seed point, iterative search is performed to find its satisfaction. The adjacent points are continuously expanded outwards to form a point cloud cluster, which can be formally represented as: In the formula: Indicates the first The i-th point cloud cluster is generated within each annular region; Use this as the seed point, and start performing neighborhood expansion and growth clustering from this point; This represents a connected path consisting of points. The endpoint of the path; any two adjacent points on the path satisfy the example threshold constraint.

2. The fast point cloud segmentation algorithm for unmanned vehicles in complex environments according to claim 1, characterized in that, In step S10, the point cloud set within the annular region... Construct a kd-tree spatial index structure and perform a radius neighborhood search operation based on the index structure.

Images