A local anomaly perception-based laser radar point cloud ground extraction method

Abstract

This invention proposes a ground point cloud extraction method based on local anomaly perception using LiDAR, belonging to the field of point cloud data processing. The method includes: acquiring raw LiDAR point cloud data and dividing it into fan-shaped and ring-shaped grids to form gridded point cloud data; performing local anomaly analysis on the gridded point cloud data and using dual threshold filtering based on initial confidence levels to obtain a core seed point set and a candidate seed point set; constructing a two-layer criterion verification method; establishing a first priority queue based on the initial confidence level of the core seed point set for region growing to obtain a preliminary ground point set; establishing a second priority queue for candidate seed points that have passed environmental verification for region growing to obtain an expanded ground point set; filtering the expanded ground point set to obtain a refined ground point set; and adding ground point labels to each point in the refined ground point set to obtain labeled ground point cloud data. This invention achieves highly robust and high-precision ground point extraction in complex scenarios.

Description

Technical Field

[0001] This invention relates to the field of point cloud data processing technology, and in particular to a method for ground extraction of LiDAR point clouds based on local anomaly perception. Background Technology

[0002] LiDAR (Light Detection and Ranging) acquires high-precision 3D point cloud data of the surrounding environment by emitting and receiving laser beams. In fields such as map building, autonomous driving, and robot navigation, accurately distinguishing between ground points and non-ground points is a crucial preprocessing step for subsequent tasks such as obstacle detection and path planning. Existing ground point extraction methods mainly include those based on mesh generation, plane fitting, and region growing.

[0003] However, these methods generally suffer from a lack of adaptability due to fixed thresholds. Region-growing methods typically use fixed slope and height difference thresholds, which cannot adapt to both flat roads and undulating terrain simultaneously. In areas with significant terrain undulations, fixed thresholds can easily lead to missed extraction of ground points; while in flat areas, overly lenient thresholds can introduce incorrect extraction of non-ground points. Furthermore, traditional methods also suffer from insufficient reliability in the selection of initial seed points.

[0004] Chinese patent CN114170149A discloses a method for extracting road geometry information based on laser point clouds. This method preprocesses the point cloud acquired by a vehicle-mounted laser scanning system, extracts its elevation features and local normal vector features to distinguish between ground points and non-ground points, and then uses a region growing method to connect road surface points on the ground points. However, it uses a random selection method to determine seed points, which may mistakenly select noise points or the bottom points of obstacles as seed points, leading to incorrect initial judgments and propagating errors, making it difficult to adapt to complex and changing terrain environments. Summary of the Invention

[0005] In view of this, the present invention proposes a ground extraction method for lidar point clouds based on local anomaly perception, which solves the problem that the existing technology cannot adapt to complex and variable terrain environments, and achieves highly robust and high-precision ground point extraction in complex scenes.

[0006] The technical solution of the present invention is implemented as follows: The present invention provides a method for ground extraction of lidar point clouds based on local anomaly perception, including the following steps;

[0007] S1. Acquire raw lidar point cloud data, preprocess the raw lidar point cloud data and divide it into fan-shaped and ring-shaped grids to form gridded point cloud data;

[0008] S2. Perform local anomaly analysis on the gridded point cloud data, calculate the feature vector of each point in the grid cell, calculate the anomaly score distribution of the grid cell through the local anomaly detection model, convert the anomaly score into the initial confidence score of each point, and perform double threshold screening based on the initial confidence score to obtain the core seed point set and the candidate seed point set.

[0009] S3. Construct a two-layer criterion test method, which includes a constraint criterion based on local slope and a coherence constraint criterion based on the global terrain model.

[0010] S4. Based on the initial confidence of the core seed point set, establish a first priority queue for region growth, obtain the neighboring points of the growth center point and verify them using a two-layer criterion test method to obtain a preliminary ground point set; based on the preliminary ground point set, perform environmental verification on the candidate seed point set, and establish a second priority queue for the candidate seed points that pass the environmental verification to perform region growth, thereby obtaining an extended ground point set;

[0011] S5. Filter the extended ground point set to obtain a refined ground point set. Add ground point labels to each point in the refined ground point set and recombine them with non-ground points to obtain labeled ground point cloud data.

[0012] Based on the above technical solutions, preferably, the local anomaly detection model uses an isolated forest unsupervised learning model to train the feature vectors of all points within a grid cell. Specifically, the isolated forest unsupervised learning model optimizes the feature weights and adaptively optimizes the density of the feature vectors of each point within the grid cell, including:

[0013] The feature vector of each point in the grid cell is calculated to distinguish the Gini importance of ground / non-ground points, and then normalized to obtain the feature weights.

[0014] Construct an isolation tree, where each internal node of the isolation tree is randomly selected based on a weighted average of its feature weights;

[0015] For any point in the preprocessed point cloud data, obtain its local point cloud density;

[0016] Set the density-adaptive path length, and extend the path length of points in sparse regions using the density-adaptive path length.

[0017] Based on the above technical solutions, preferably, step S2 specifically includes:

[0018] For point P in each grid cell of the gridded point cloud data, search for the set of k nearest neighbors of point P within a set radius centered on point P.

[0019] Calculate the feature vector of each point based on the nearest neighbor set. , where the feature vector Including height features Flatness characteristics linear features Curvature characteristics and verticality features ;

[0020] Iterate through all points within the grid cell and extract the feature vector of each point. Constructing a local feature space , ,in This represents the total number of points within a grid cell.

[0021] Local anomaly detection model is used to analyze the local feature space. All feature vectors The calculation is performed to obtain the anomaly score for all points in the grid cell. ;

[0022] The negative exponential function is used to convert outlier scores into initial confidence levels. Based on the initial confidence level Sort all points in the grid in descending order, and select the k points with the highest initial confidence to form the core seed point set; where the negative exponential function is:

[0023] ;

[0024] in, Indicates the sensitivity coefficient;

[0025] Set a confidence threshold and form a candidate seed point set from the points that are higher than the confidence threshold.

[0026] Based on the above technical solutions, preferably, the constraint criterion based on local slope specifically includes:

[0027] For the point Q to be judged, determine the local reference region where point Q is located, and calculate the standard deviation of the height values ​​of all points in the local reference region and the point cloud density of the local reference region.

[0028] Constructing a dynamic slope threshold function :

[0029] ;

[0030] in, Indicates the basic slope threshold. This represents the terrain undulation weighting coefficient. Indicates the local height standard deviation. This represents the point cloud density weighting coefficient. Indicates local point cloud density. Represents horizontal distance in polar coordinates;

[0031] During the region growth process, the slope between the point Q to be judged and the growth center point is calculated, and it is determined whether the slope satisfies the constraint condition of the dynamic slope threshold function. If it does, it indicates that the point Q to be judged and the growth center point have local compatibility.

[0032] Based on the above technical solutions, preferably, the coherence constraint criterion based on the global terrain model specifically includes:

[0033] An irregular triangular mesh model is initialized. During the region growing process, whenever a new point is identified as a ground point, it is immediately inserted into the irregular triangular mesh model and the triangular mesh structure is dynamically updated. The irregular triangular mesh model represents the continuous surface of the currently extracted terrain.

[0034] For the point to be judged Find points in the current real-time irregular triangular mesh model. The triangular face containing the projection point, the calculation point perpendicular distance to the triangular face ;

[0035] Set adaptive coherence constraint threshold Among them, the adaptive coherence constraint threshold With point cloud density and the average confidence level of the ground confirmed at the current growth stage Negative correlation;

[0036] Judgment point Does the perpendicular distance to the triangular facet satisfy the adaptive coherence constraint threshold? If it does, then the point... It has global consistency with the existing terrain model.

[0037] Based on the above technical solutions, preferably, the step of establishing a first priority queue based on the initial confidence of the core seed point set for region growth, obtaining the neighboring points of the growth center point, and verifying them using a two-layer criterion verification method specifically includes:

[0038] Step a: Initialize the first priority queue , core seed point set All points are inserted into the first priority queue in descending order of initial confidence level. An initial decay factor is applied to each seed point. ;

[0039] Step b, from the first priority queue The point with the highest confidence level is selected as the current seed point. Search for the current seed point All unmarked points within the spatial neighborhood;

[0040] Step c: Verify each unmarked point using a two-layer criterion test method. Calculate the propagation attenuation factor for neighborhood points that simultaneously satisfy the constraint criteria and the coherence constraint criteria, and calculate the final confidence level based on the initial confidence level and the initial attenuation factor of the neighborhood points.

[0041] Step d: Mark the verified neighborhood points as ground points and add them to the initial ground point set. The neighboring points are then inserted into the real-time irregular triangular mesh model to update the terrain surface, and the final confidence scores of the neighboring points are added to the first priority queue. As a new growth center;

[0042] Repeat steps b through d until the first priority queue is reached. If the value is empty, a preliminary ground point set is obtained.

[0043] Based on the above technical solutions, preferably, the step of performing environmental verification on the candidate seed point set based on the preliminary ground point set, and establishing a second priority queue for the candidate seed points that pass the environmental verification for region growing, specifically includes:

[0044] B1, when the first priority queue After it is empty, initialize the second priority queue. ;

[0045] B2. On the candidate seed point set Unmarked points Statistical points The number of points within the surrounding spherical neighborhood that have been marked as ground points and whose final confidence level is higher than the threshold is used as the environmental support.

[0046] B3. Dynamically calculate the startup confidence level based on environmental support. ;

[0047] B4. Confidence level will be activated. Insert it into the second priority queue as a priority. And reset the initial decay factor ;

[0048] B5, from the second priority queue Remove the point with the highest confidence level and use it as the current growth center point P2. Then search for all unlabeled points in the spatial neighborhood of the current growth center point P2.

[0049] B6. Verify each unmarked point using a two-layer criterion test method. Calculate the propagation attenuation factor for neighboring points that simultaneously satisfy the constraint criteria based on local slope and the coherence constraint criteria based on the global terrain model. Calculate the final confidence of neighboring points based on the initial confidence of neighboring points and the propagation attenuation factor.

[0050] B7. Mark the verified neighboring points as ground points and add them to the final ground point set. Insert the neighboring points into the real-time irregular triangular mesh model to update the terrain surface. At the same time, add the neighboring points to the second priority queue according to their final confidence scores. As a new growth center;

[0051] B8. Repeat steps B5 through B7 until the second priority queue is reached. If empty, the extended ground point set is obtained.

[0052] Based on the above technical solutions, preferably, the formula for calculating the initiation confidence level is as follows:

[0053] ；

[0054] in, The table shows the initial confidence levels of the candidate seed points. Indicates the strength enhancement coefficient. Indicates environmental support. This represents the minimum number of points threshold.

[0055] Based on the above technical solutions, preferably, step S5 specifically includes:

[0056] Iterate through each point in the expanded ground point set and count the number of ground points within its specified radius neighborhood. Points with fewer than a preset threshold are identified as isolated points and removed from the expanded ground point set to obtain an isolated point set.

[0057] Set threshold Iterate through all points in the outlier set and remove those with a final confidence level below the threshold. From these points, a refined set of ground points is obtained;

[0058] Add ground point labels to each point in the refined ground point set, and recombine the refined ground point set with the non-ground points in the original point cloud data to form a complete labeled point cloud data.

[0059] More preferably, step S1 specifically includes:

[0060] The raw lidar point cloud data is acquired and invalid points with distances exceeding the preset effective range are filtered out to obtain effective point cloud data;

[0061] Transform the effective point cloud data from Cartesian coordinates to polar coordinates, where the horizontal distance... Horizontal azimuth Vertical pitch angle ;

[0062] Based on horizontal azimuth The effective point cloud data is evenly divided into N sector regions. The angular span of each sector is ;

[0063] Within each sector, along the horizontal distance The direction of increase, at fixed intervals Divided into M concentric annular sub-zones, forming Two-dimensional mesh structure Each grid cell contains points falling within that region, where the grid... Indicates the first The sector, the first Subregions of the ring.

[0064] The ground extraction method for lidar point clouds based on local anomaly perception of the present invention has the following advantages over the prior art:

[0065] (1) By dividing the point cloud data into fan-shaped and ring-shaped grids, and using a local anomaly detection model to perform local feature analysis in each grid cell, it is possible to effectively identify the abnormal points in the local feature distribution, convert the anomaly score into an initial confidence score, and use double threshold screening based on the confidence score to obtain the core seed point set and candidate seed point set, which significantly improves the accuracy of ground point extraction and achieves high robustness and high precision ground point extraction in complex scenarios;

[0066] (2) Based on the constraint criteria of local slope and the coherence constraint criteria based on the global terrain model, the constraints are tightened in flat terrain or dense point cloud areas to ensure extraction accuracy. By constructing and updating the irregular triangular mesh model in real time, the deviation between the point to be judged and the constructed terrain surface is constrained, ensuring the global coherence of ground point extraction and effectively avoiding the accumulation of errors caused by local judgment errors.

[0067] (3) By using the two-stage region growth mechanism of the first priority queue and the second priority queue, the reliability of the initial growth is guaranteed, and the omission of ground areas due to the local absence of core seed points is avoided, thereby improving the coverage integrity of the algorithm in complex scenarios. Attached Figure Description

[0068] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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.

[0069] Figure 1 This is a flowchart of a lidar point cloud ground extraction method based on local anomaly perception according to the present invention;

[0070] Figure 2 This is a block diagram of a lidar point cloud ground extraction method based on local anomaly perception according to the present invention. Detailed Implementation

[0071] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0072] like Figure 1 and Figure 2 As shown, this invention provides a method for ground extraction of lidar point clouds based on local anomaly perception, comprising the following steps:

[0073] S1. Acquire raw lidar point cloud data, preprocess the raw lidar point cloud data and divide it into fan-shaped and ring-shaped grids to form gridded point cloud data;

[0074] Specifically, step S1 includes:

[0075] The raw lidar point cloud data is acquired and invalid points beyond the preset effective range are filtered out to obtain valid point cloud data; invalid points include those beyond the preset effective range of the lidar. The points, and the noise caused by the characteristics of the radar itself;

[0076] Transform the effective point cloud data from Cartesian coordinates to polar coordinates, where the horizontal distance... Horizontal azimuth Vertical pitch angle ;

[0077] Based on horizontal azimuth The effective point cloud data is evenly divided into N sector regions. The angular span of each sector is ;

[0078] Within each sector, along the horizontal distance The direction of increase, at fixed intervals Divided into M concentric annular sub-zones, forming Two-dimensional mesh structure Each grid cell contains points falling within that region, where the grid... Indicates the first The sector, the first Subregions of the ring.

[0079] Understandably, a data acquisition vehicle containing sensors such as lidar, inertial navigation system and global positioning system is used to acquire raw lidar point cloud data. The lidar can be a 16-line mechanical lidar with a horizontal field of view of 360 degrees, a vertical field of view of -15 degrees to +15 degrees, and a maximum range of 100 meters.

[0080] In one embodiment of the present invention, N can be set to 180, then the angular span of each sector is 2 degrees, based on each point. The angle assigns it to the corresponding sector. Medium. Starting distance meters, fixed interval Meters (M) can be set to 106 to form a 180×106 two-dimensional grid structure.

[0081] S2. Perform local anomaly analysis on the gridded point cloud data, calculate the feature vector of each point in the grid cell, calculate the anomaly score distribution of the grid cell through the local anomaly detection model, convert the anomaly score into an initial confidence score for each point, and perform double threshold screening based on the initial confidence score to obtain the core seed point set and candidate seed point set; wherein, the local anomaly detection model uses an isolated forest unsupervised learning model to train the feature vectors of all points in the grid cell, wherein the isolated forest unsupervised learning model performs feature weight optimization and density adaptive optimization on the feature vector of each point in the grid cell, specifically including:

[0082] The Gini importance of the feature vector of each point in the grid cell for distinguishing between ground and non-ground points is calculated and normalized to obtain the feature weights. ,in, The weights of the height feature z are represented. Indicates flatness characteristics The weight, The weights represent the linear feature L. The weights of the curvature feature C are represented. The weights of the verticality feature V are represented.

[0083] Construct an isolation tree, where each internal node of the isolation tree is randomly selected based on a weighted average of its feature weights;

[0084] For any point in the preprocessed point cloud data, obtain its local point cloud density. :

[0085] ;

[0086] in, This represents the number of points in the neighborhood of radius R. The threshold representing the number of empirical points within a neighborhood of radius R;

[0087] Set density adaptive path length The path length of points in sparse regions is extended by using density-adaptive path length:

[0088] ;

[0089] in, This represents the path length required for point P to be isolated in the isolation tree. This indicates that the preset density compensation coefficient is used to control the intensity of the compensation.

[0090] Understandably, intentionally extending the path length of points in sparse regions prevents normal points in these regions from being easily identified as anomalies simply because of their short path lengths, significantly improving the robustness of the local anomaly detection model under different point cloud densities. After training, the local anomaly detection model outputs an anomaly score for each point P within a grid cell. The score ranges from 0 to 1. A higher score indicates that point P is more abnormal in the local feature distribution of its grid.

[0091] In one embodiment of the present invention, the number of isolation trees is 50. After inference by the local anomaly detection model, the anomaly score of the flat road surface points is... , and the pebble of .

[0092] Specifically, step S2 includes:

[0093] For point P in each grid cell of the gridded point cloud data, search for the set of k nearest neighbors of point P within a set radius centered on point P.

[0094] Calculate the feature vector of each point based on the nearest neighbor set. , where the feature vector Including height features Flatness characteristics linear features Curvature characteristics and verticality features ;

[0095] Iterate through all points within the grid cell and extract the feature vector of each point. Constructing a local feature space , ,in This represents the total number of points within a grid cell.

[0096] Local anomaly detection model is used to analyze the local feature space. All feature vectors The calculation is performed to obtain the anomaly score for all points in the grid cell. ;

[0097] The negative exponential function is used to convert outlier scores into initial confidence levels. Based on the initial confidence level Sort all points in the grid in descending order, and select the k points with the highest initial confidence to form the core seed point set; where the negative exponential function is:

[0098] ;

[0099] in, This represents the sensitivity coefficient, used to control the rate at which the confidence level decays with respect to outlier scores;

[0100] Set a confidence threshold and form a candidate seed point set from the points that are higher than the confidence threshold.

[0101] Understandably, the height feature z represents the absolute height value of point P, while the flatness feature... The minimum eigenvalue representation obtained by principal component analysis of the neighboring point set. A smaller value indicates a flatter region. The formula for calculating the linear feature L is: This is used to distinguish the planar structure of the ground from linear objects such as utility poles. This represents the largest eigenvalue obtained by performing principal component analysis on the nearest neighbor set. This represents the intermediate eigenvalues ​​obtained from principal component analysis of the neighboring point set. The formula for calculating the curvature feature C is: This reflects the degree of curvature of a local surface. The perpendicularity feature V is calculated by taking the normal vector of point P. And calculate Received, among which This is a vertical vector; the smaller the value, the more horizontal the local surface. , , The three eigenvalues ​​obtained from principal component analysis and .

[0102] In one embodiment of the present invention, the radius of the center at point P is 0.3m, and k=10.

[0103] S3. Construct a two-layer criterion test method, which includes a constraint criterion based on local slope and a coherence constraint criterion based on the global terrain model.

[0104] Specifically, the constraint criteria based on local slope include:

[0105] For the point Q to be judged, determine the local reference region where point Q is located, and calculate the standard deviation of the height values ​​of all points in the local reference region and the point cloud density of the local reference region.

[0106] Constructing a dynamic slope threshold function :

[0107] ;

[0108] in, Indicates the basic slope threshold. This represents the terrain undulation weighting coefficient. Indicates the local height standard deviation. This represents the point cloud density weighting coefficient. Indicates local point cloud density. Represents horizontal distance in polar coordinates;

[0109] During the region growth process, the slope between the point Q to be judged and the growth center point is calculated, and it is determined whether the slope satisfies the constraint condition of the dynamic slope threshold function. If it does, it indicates that the point Q to be judged and the growth center point have local compatibility.

[0110] In one embodiment of the present invention, the local reference region where the point Q to be determined is located is typically its annular sub-zone and its adjacent sub-zones. The point cloud density is calculated by dividing the number of points in the region by the area of ​​the region, and then normalized.

[0111] Basic slope threshold Based on experience with flat roads, the value can be set to 0.15. Terrain undulation weighting coefficient. It can be set to 0.8. A larger value indicates a more uneven terrain and a higher permissible slope threshold. Point cloud density weighting coefficient. It can be set to 0.2. The smaller the value, the sparser the point cloud, and the larger the allowable slope threshold, in order to compensate for the height jump caused by the sparseness.

[0112] Furthermore, the coherence constraint criterion based on the global terrain model specifically includes:

[0113] An irregular triangular mesh model is initialized. During the region growing process, whenever a new point is identified as a ground point, it is immediately inserted into the irregular triangular mesh model and the triangular mesh structure is dynamically updated. The irregular triangular mesh model represents the continuous surface of the currently extracted terrain.

[0114] For the point to be judged Find points in the current real-time irregular triangular mesh model. The triangular face containing the projection point, the calculation point perpendicular distance to the triangular face It represents the point Degree of deviation from the existing terrain model:

[0115] ;

[0116] in, Point The actual height Indicates theoretical depth;

[0117] Set adaptive coherence constraint threshold Among them, the adaptive coherence constraint threshold With point cloud density and the average confidence level of the ground confirmed at the current growth stage Negative correlation:

[0118] ;

[0119] Judgment point Does the perpendicular distance to the triangular facet satisfy the adaptive coherence constraint threshold? If it does, then the point... It has global consistency with the existing terrain model.

[0120] Understandably, the adaptive coherence constraint threshold is negatively correlated with the point cloud density and the average confidence of the confirmed ground in the current growth stage. This means that in areas with sparse point clouds or low confidence of confirmed ground, the coherence constraint is relaxed to enhance the algorithm's exploration ability, while the constraint is tightened to ensure the smoothness of the terrain.

[0121] This invention uses a constraint criterion based on local slope and a coherence constraint criterion based on a global terrain model. It tightens constraints in flat terrain or dense point cloud areas to ensure extraction accuracy. By constructing and updating an irregular triangular mesh model in real time, it constrains the deviation between the point to be judged and the constructed terrain surface, ensuring the global coherence of ground point extraction and effectively avoiding the accumulation of errors caused by local judgment mistakes.

[0122] S4. Based on the initial confidence of the core seed point set, establish a first priority queue for region growth, obtain the neighboring points of the growth center point and verify them using a two-layer criterion test method to obtain a preliminary ground point set; based on the preliminary ground point set, perform environmental verification on the candidate seed point set, and establish a second priority queue for the candidate seed points that pass the environmental verification to perform region growth, thereby obtaining an extended ground point set;

[0123] Specifically, the process of establishing a first priority queue based on the initial confidence level of the core seed point set for region growth, obtaining neighboring points of the growth center point, and verifying them using a two-layer criterion verification method includes:

[0124] Step a: Initialize the first priority queue , core seed point set All points are inserted into the first priority queue in descending order of initial confidence level. An initial decay factor is applied to each seed point. ;

[0125] Step b, from the first priority queue The point with the highest confidence level is selected as the current seed point. Search for the current seed point All unmarked points within the spatial neighborhood;

[0126] Step c: Verify each unmarked point using a two-layer criterion test method, and calculate the propagation attenuation factor for neighborhood points that simultaneously satisfy the constraint criteria and the coherence constraint criteria. The final confidence level is calculated based on the initial confidence level and initial decay factor of the neighboring points. :

[0127] ;

[0128] ;

[0129] in, This represents the attenuation coefficient, used to control the rate of attenuation. express;

[0130] Step d: Mark the verified neighborhood points as ground points and add them to the initial ground point set. The neighboring points are then inserted into the real-time irregular triangular mesh model to update the terrain surface, and the final confidence scores of the neighboring points are added to the first priority queue. As a new growth center;

[0131] Repeat steps b through d until the first priority queue is reached. If the value is empty, a preliminary ground point set is obtained.

[0132] In one embodiment of the present invention, the initial attenuation factor If the value is 1, the element in the queue is a triple. The spatial neighborhood can be set to within 1 meter.

[0133] Two-layer criterion testing methods include:

[0134] Calculate the current seed point With neighboring points Slope between:

[0135] ;

[0136] in, Indicates the current seed point With neighboring points The slope value between, Indicates the current seed point The height coordinates, Representing neighborhood points The height coordinates, Indicates the current seed point x-axis coordinates Representing neighborhood points x-axis coordinates Indicates the current seed point y-axis coordinate, Representing neighborhood points y-axis coordinate;

[0137] Determine if it satisfies Simultaneously calculate neighboring points Vertical distance to the real-time irregular triangular mesh model Determine whether the condition is met. .

[0138] Calculate the propagation attenuation factor for neighborhood points that simultaneously satisfy the two-layer criteria. .

[0139] Furthermore, the step of performing environmental verification on the candidate seed point set based on the preliminary ground point set, and establishing a second priority queue for the candidate seed points that pass the environmental verification for region growing, specifically includes:

[0140] B1, when the first priority queue After it is empty, initialize the second priority queue. ;

[0141] B2. On the candidate seed point set Unmarked points Statistical points The number of points within the surrounding spherical neighborhood that have been marked as ground points and whose final confidence level is higher than the threshold is used as the environmental support.

[0142] B3. Dynamically calculate the startup confidence level based on environmental support. The formula for calculating the initiation confidence level is as follows:

[0143] ；

[0144] in, The table shows the initial confidence levels of the candidate seed points. Indicates the strength enhancement coefficient. Indicates environmental support. This represents the minimum number of points threshold;

[0145] B4. Confidence level will be activated. Insert it into the second priority queue as a priority. And reset the initial decay factor ;

[0146] B5, from the second priority queue Remove the point with the highest confidence level and use it as the current growth center point P2. Then search for all unlabeled points in the spatial neighborhood of the current growth center point P2.

[0147] B6. Verify each unmarked point using a two-layer criterion test method. Calculate the propagation attenuation factor for neighboring points that simultaneously satisfy the constraint criteria based on local slope and the coherence constraint criteria based on the global terrain model. Calculate the final confidence of neighboring points based on the initial confidence of neighboring points and the propagation attenuation factor.

[0148] B7. Mark the verified neighboring points as ground points and add them to the final ground point set. Insert the neighboring points into the real-time irregular triangular mesh model to update the terrain surface. At the same time, add the neighboring points to the second priority queue according to their final confidence scores. As a new growth center;

[0149] B8. Repeat steps B5 through B7 until the second priority queue is reached. If empty, the extended ground point set is obtained.

[0150] This invention employs a two-stage region growth mechanism with a first priority queue and a second priority queue. While ensuring the priority growth of high-confidence seed points, it verifies the environment of candidate seed points and dynamically calculates the start confidence based on the environment support. This allows medium-confidence points located near reliable ground areas to obtain a second growth opportunity. This not only ensures the reliability of the initial growth but also avoids the omission of ground areas due to the local absence of core seed points, thereby improving the coverage integrity of the algorithm in complex scenarios.

[0151] S5. Filter the extended ground point set to obtain a refined ground point set. Add ground point labels to each point in the refined ground point set and recombine them with non-ground points to obtain labeled ground point cloud data.

[0152] Specifically, step S5 includes:

[0153] Iterate through each point in the expanded ground point set and count the number of ground points within its specified radius neighborhood. Points with fewer than a preset threshold are identified as isolated points and removed from the expanded ground point set to obtain an isolated point set.

[0154] Set threshold Iterate through all points in the outlier set and remove those with a final confidence level below the threshold. From these points, a refined set of ground points is obtained;

[0155] Add ground point labels to each point in the refined ground point set, and recombine the refined ground point set with the non-ground points in the original point cloud data to form a complete labeled point cloud data.

[0156] This invention divides point cloud data into fan-shaped and ring-shaped grids and uses a local anomaly detection model to perform local feature analysis within each grid cell. This effectively identifies anomalies in the local feature distribution, converts anomaly scores into initial confidence scores, and uses dual threshold screening based on the confidence scores to obtain core seed point sets and candidate seed point sets. This provides a reliable starting point for subsequent region growing, significantly improves the accuracy of ground point extraction, and achieves highly robust and high-precision ground point extraction in complex scenarios.

[0157] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for ground extraction from a LiDAR point cloud based on local anomaly perception, characterized in that: The method comprises the following steps: S1, obtaining original laser radar point cloud data, preprocessing the original laser radar point cloud data, and performing fan-ring grid division to form grid point cloud data; S2, performing local anomaly analysis on the grid point cloud data, calculating the feature vector of each point in the grid unit, calculating the anomaly score distribution of the grid unit through a local anomaly detection model, converting the anomaly score into an initial confidence of each point, and performing double-threshold screening according to the initial confidence to obtain a core seed point set and a candidate seed point set; S3, constructing a double-layer criterion verification method, wherein the double-layer criterion verification method comprises a constraint criterion based on local slope and a continuity constraint criterion based on a global terrain model; the constraint criterion based on local slope specifically comprises: For a point to be judged Q, a local reference region in which the point Q is located is determined, the standard deviation of the height values of all points in the local reference region and the point cloud density of the local reference region are calculated; Constructing dynamic slope threshold function : ； wherein, denotes a base slope threshold, denotes a terrain roughness weight coefficient, denotes a local height standard deviation, denotes a point cloud density weight coefficient, denotes a local point cloud density, denotes a horizontal distance in polar coordinate system; In the region growing process, the slope between the point to be judged Q and the growing center point is calculated, and it is judged whether the slope satisfies the constraint condition of the dynamic slope threshold function, if yes, it is indicated that the point to be judged Q has local compatibility with the growing center point; The continuity constraint criterion based on the global terrain model specifically comprises: An irregular triangle mesh model is initialized, and every time a new point is confirmed as a ground point in the region growing process, the new point is inserted into the irregular triangle mesh model and the triangle mesh structure is dynamically updated, and the irregular triangle mesh model represents the continuous surface of the current extracted terrain; For the point to be judged In the current real-time irregular triangle net model, find the triangle face where the point is located Calculate the vertical distance from the point to the triangle face ;​ Setting an adaptive coherence constraint threshold wherein the adaptive coherence constraint threshold with the density of the point cloud and the average confidence that the ground has been confirmed for the current growth stage is negatively correlated; Judgment point whether the perpendicular distance to the triangular patch satisfies the adaptive coherence constraint threshold, if yes, it means that the point has global coherence with the constructed terrain model; S4, a first priority queue is established based on the initial confidence of the core seed point set to perform region growing, the neighboring points of the growing center point are obtained, and the double-layer criterion verification method is used for verification to obtain a preliminary ground point set; the environment of the candidate seed point set is verified based on the preliminary ground point set, and a second priority queue is established for the candidate seed points that pass the environment verification to perform region growing to obtain an expanded ground point set; S5, filtering the expanded ground point set to obtain a refined ground point set, adding a ground point label to each point in the refined ground point set, and recombining the ground point and the non-ground point to obtain labeled ground point cloud data.

2. The local outlier aware ground extraction method for LiDAR point clouds of claim 1, wherein: The local anomaly detection model uses an isolation forest unsupervised learning model to train the feature vectors of all points in the grid unit, wherein the isolation forest unsupervised learning model optimizes the feature weight of the feature vector of each point in the grid unit and performs density adaptive optimization, and specifically comprises: The Gini importance of the feature vector of each point in the grid unit for distinguishing ground / non-ground points is calculated, and normalization processing is performed to obtain a feature weight; A isolation tree is constructed, wherein each internal node of the isolation tree is randomly selected according to the feature weight; For any point in the preprocessed point cloud data, the local point cloud density of the point is obtained; A density adaptive path length is set, and the path length of the points in the sparse area is extended through the density adaptive path length.

3. The method of claim 2, wherein: Step S2 specifically comprises: For a point P in each grid unit of the grid point cloud data, the k-nearest neighbor point set within a set radius is searched with the point P as the center; calculating a feature vector for each point based on a set of neighboring points where the feature vector includes a height feature , a flatness feature , a line feature , a curvature feature , and a perpendicularity feature ; traversing all points within the grid cell, a feature vector for each point is computed pooling the local feature space , wherein is the total number of points within the grid cell; By the local anomaly detection model on the local feature space All feature vectors in the grid cell Compute the anomaly score of all points in the grid cell ; The abnormal score is converted into initial confidence by using a negative exponential function , according to the initial confidence All points in the grid are sorted in descending order, and the top k points with the highest initial confidence are selected to form a core seed point set; wherein the negative exponential function is: ； wherein represents a sensitivity coefficient; A confidence threshold is set, and points higher than the confidence threshold are formed into a candidate seed point set.

4. The local outlier aware LiDAR point cloud ground extraction method of claim 1, wherein: The initial confidence based on the core seed point set establishes a first priority queue for region growing, acquires neighboring points of the growing center point, and adopts a double-layer criterion verification method for verification, specifically including: Step a, initialize the first priority queue , insert all points in the core seed point set into the first priority queue in descending order of initial confidence , and attach an initial decay factor to each seed point ; Step b. taking the point with the highest confidence from the first priority queue as the current seed point Step c. searching all unmarked points within the spatial neighborhood of the current seed point Step d. marking the found points as labeled Step e. repeating steps b. to d. until all points are labeled Step c, verifying each unmarked point using a double-layer criterion verification method, calculating a propagation attenuation factor for neighboring points that simultaneously satisfy the constraint criterion and the continuity constraint criterion, and calculating a final confidence based on the initial confidence of the neighboring points and the initial attenuation factor; Step d, mark the validated neighbor points as ground points and add them to the preliminary ground point set and insert the neighbor points into the real-time irregular triangulation model, update the terrain surface, and add the final confidence of the neighbor points to the first priority queue as a new growing center point; Repeat steps b-d until the first priority queue is empty For empty, get the preliminary ground point set.

5. The method of claim 4, wherein: The candidate seed point set is verified based on the preliminary ground point set, and a second priority queue is established for region growing for candidate seed points that pass the environmental verification, specifically including: B1. When the first priority queue is empty, initialize the second priority queue ; B2, to the candidate seed point set points in the surrounding spherical neighborhood that have been labeled as ground points and have a final confidence above a threshold , statistics point the number of points in the surrounding spherical neighborhood that have been labeled as ground points and have a final confidence above a threshold as environment support B3, dynamically calculating the confidence level based on the environmental support ; B4, the start confidence as a priority to insert into the second priority queue and reset the initial decay factor ; B5. from the second priority queue A point with the highest confidence is removed as the current growth center point P2, and all unmarked points in the spatial neighborhood of the current growth center point P2 are searched. B6, verifying each unmarked point using a double-layer criterion verification method, calculating a propagation attenuation factor for neighboring points that simultaneously satisfy the constraint criterion based on local slope and the continuity constraint criterion based on the global terrain model, and calculating a final confidence of the neighboring points based on the initial confidence of the neighboring points and the propagation attenuation factor; B7. Mark the validated neighboring points as ground points to the final ground points set and insert the neighboring points into the real-time irregular triangulation model to update the terrain surface, while insert the neighboring points into the second priority queue according to their final confidence as a new growing center point; B8. Repeat B5 - Step B7 until the second priority queue is empty For empty, get the extended ground point set.

6. The method of claim 5, wherein: The calculation formula of the starting confidence is: ； wherein, an initial confidence score of a table candidate seed point, denotes a boosting strength coefficient, denotes an environmental support degree, denotes a minimum point threshold.

7. The local anomaly perception based ground extraction method for LiDAR point clouds of claim 1, wherein: Step S5 specifically includes: Counting the number of ground points in the specified radius neighborhood of each point in the extended ground point set Determining the points with the number of ground points lower than the preset threshold as isolated points and removing them from the extended ground point set to obtain an isolated point set Setting a threshold , traversing all points in the outlier point set and removing points with final confidence lower than the threshold , obtaining a refined ground point set; Adding a ground point label to each point in the refined ground point set, and recombining the refined ground point set with the non-ground points in the original point cloud data to form complete labeled point cloud data.

8. The local outlier aware LiDAR point cloud ground extraction method of claim 1, wherein: Step S1 specifically includes: Obtaining original laser radar point cloud data and filtering out invalid points with a distance exceeding a preset effective range to obtain effective point cloud data; Converting the effective point cloud data from a Cartesian coordinate system to a polar coordinate system, wherein a horizontal distance , an azimuth angle , a vertical pitch angle ; Azimuth angle based Uniformly dividing effective point cloud data into N sector regions Each sector has an angular span of ; Within each sector, along the horizontal distance The direction of increase, at fixed intervals Divided into M concentric annular sub-zones, forming Two-dimensional mesh structure Each grid cell contains points falling within that region, where the grid... Indicates the first The sector, the first Subregion of the ring.

Images