Real-time ground segmentation method and system

Abstract

The invention relates to the technical field of robot environment perception, in particular to a real-time ground segmentation method and system. The method comprises the following steps: receiving and preprocessing laser radar point cloud data, and dividing the data into a plurality of concentric annular areas; carrying out adaptive ground fitting on each area and processing the multilayer structure through vertical plane fitting to obtain an initial ground point set; correcting an under-segmentation region of the current frame by using a time sequence backspacing mechanism and updating parameters; constructing an elevation map based on the initial point set, and obtaining a fine ground point set through adaptive resolution CSF simulation; and carrying out confidence evaluation on the two types of point sets, and generating a final result through multi-criterion fusion. Accurate segmentation of the complex terrain is realized, and robustness and real-time performance of a dynamic environment are improved.

Description

Technical Field

[0001] This invention relates to the field of robot environmental perception technology, and more specifically, to a real-time ground segmentation method and system. Background Technology

[0002] With its high-precision distance measurement capabilities, LiDAR has become a core sensor for environmental perception in mobile robots. Ground extraction is a key technology in environmental perception, which directly affects the safety and reliability of robot navigation.

[0003] Existing ground extraction methods mainly include: height threshold-based methods, but these are sensitive to terrain undulations and are prone to misclassifying concave areas as non-ground or convex areas as ground in rugged terrain; RANSAC plane fitting methods assume that the ground is an ideal plane, but cannot adapt to complex terrains such as mountains and steps; and morphological filtering methods rely on manual parameter adjustment, but have poor robustness in different environments. Therefore, a real-time ground segmentation method and system are designed.

[0004] The existing technology has the following shortcomings, specifically:

[0005] 1. Traditional ground segmentation methods based on single-plane fitting are prone to misclassifying some ground points as non-ground points in complex terrain scenarios, resulting in missed ground detections. When relying solely on the CSF algorithm for ground segmentation, although the accuracy is guaranteed if the entire scene is processed at high resolution, the computational load is large, and it will cause a waste of computing power in some simple shape areas.

[0006] 2. Some ground segmentation algorithms do not differentiate point clouds at different distances and adopt a uniform high-resolution processing strategy, which increases the computation time when processing large scene point clouds and makes it difficult to meet the needs of real-time applications. They also lack a mechanism to use historical frame information to correct the segmentation results of the current frame. When the environment changes temporarily, the segmentation results are prone to fluctuation and lack robustness.

[0007] 3. The key parameters of existing ground segmentation algorithms are mostly fixed values ​​set by human experience. In different scenarios, fixed parameters are difficult to adapt to scene changes, have poor generalization ability, and require frequent manual adjustment of parameters, which is inconvenient to use. Summary of the Invention

[0008] The purpose of this invention is to provide a real-time ground segmentation method and system to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention aims to provide a real-time ground segmentation method, comprising: S1. Point cloud preprocessing and partitioning: receiving point cloud data from a lidar radar, and after coordinate transformation, intensity filtering and reflection noise removal, dividing the point cloud into multiple concentric ring regions based on the distance from the point cloud to the sensor.

[0010] S2. Fast initial segmentation: For each concentric ring region, an adaptive ground likelihood estimation is used to perform independent ground fitting. The initial classification is completed by combining the height threshold and normal vector constraints. The multi-layer ground structure is processed by fitting the vertical plane of the region to obtain the initial ground point set.

[0011] S3. Temporal optimization processing: Using the ground segmentation results of the previous and next frames, the instantaneous undersegmented regions of the current frame are identified and corrected through a temporal ground backoff mechanism, and the adaptive parameters are updated.

[0012] S4.CSF Local Fine Fitting: Based on the initial ground point set, extract the initial ground points and their neighborhoods, construct a local elevation map, perform adaptive resolution cloth simulation, and obtain a fine ground model and the corresponding fine ground point set.

[0013] S5. Multi-level fusion decision: Confidence assessment is performed on the initial ground point set and the refined ground point set, and the final classification result is generated and output based on a multi-criteria fusion strategy.

[0014] As a further improvement to this technical solution, the concentric annular region is specifically implemented as follows:

[0015] Using the lidar sensor as the center, at least two distance thresholds are preset as the radius boundaries of the annular area, and the point cloud space is divided into at least three concentric annular areas. At the same time, taking the direction directly in front of the sensor as the starting reference, each concentric annular area is evenly divided into multiple sectors along the 360° circumference, forming a two-dimensional grid partition of "annular + sector". Each partition is assigned an independent identifier, and a mapping relationship between the partition and the point cloud data is established. The index information of the point cloud in each partition and the height statistical characteristics of the ground points are stored.

[0016] As a further improvement to this technical solution, the adaptive ground likelihood estimation performs independent ground fitting. The specific implementation method is as follows: For the point cloud in each partition, calculate the height value of all points and statistically analyze the height mean and variance. Based on the height variance, determine the terrain complexity of the partition. If the height variance is less than a preset complexity threshold, use a plane fitting model to calculate the ground plane coefficient. If the height variance is greater than or equal to the preset complexity threshold, use a piecewise fitting model to divide the point cloud in the partition into subsets according to the height interval. Fit a local ground plane for each subset. For each point in the partition, calculate its distance to the corresponding fitted ground plane. Combine the angle between the normal vector of the point and the normal vector of the ground plane to construct a ground likelihood function and output the ground likelihood value of each point. Determine the classification threshold based on the distribution of the ground likelihood values. Preliminarily determine points with likelihood values ​​higher than the preset likelihood threshold as ground candidate points and points with likelihood values ​​lower than the preset likelihood threshold as non-ground candidate points.

[0017] As a further improvement to this technical solution, the vertical plane fitting is specifically implemented as follows:

[0018] For non-ground candidate points obtained after rapid initial segmentation, spatially continuous point cloud clusters are extracted. The minimum bounding box of each point cloud cluster is calculated, and the ratio of the height to the horizontal dimension of the bounding box is determined. If the ratio is greater than a preset verticality threshold, it is determined to be a suspected vertical plane structure. For point cloud clusters with suspected vertical plane structures, a plane fitting algorithm is used to calculate their plane coefficients. The angle between the normal vector of the fitted plane and the vertical direction is extracted. If the angle between the normal vector of the fitted plane and the vertical direction is less than a preset angle threshold, it is determined to be a vertical plane. Then, the positional relationship between the vertical plane and the surrounding ground candidate points is analyzed. If there is a continuous height connection between the bottom of the vertical plane and the ground candidate points, the point cloud above the vertical plane is re-included in the range of ground candidate points.

[0019] As a further improvement to this technical solution, the temporal ground backoff mechanism is specifically implemented as follows: Store the ground point set of the previous frame and the ground fitting parameters of each partition; compare the initial segmentation result of the current frame with the stored ground point set of the previous frame; calculate the intersection-union ratio (IU) of the ground point set of the current frame and the ground point set of the previous frame; if the IU is greater than or equal to a preset IU threshold, the current frame segmentation result is determined to be stable, and the historical frame data is directly retained and updated; if the IU is less than the preset IU threshold, locate the target partition with an IU lower than the threshold, extract the ground fitting parameters of the target partition from the previous frame, combine them with the point cloud data of the target partition in the current frame, recalculate the ground likelihood value, and based on the recalculated ground likelihood value, correct the initial segmentation result of the target partition in the current frame; use the corrected ground point set as the final initial segmentation result of the current frame, and update the ground parameter library of the historical frames.

[0020] As a further improvement to this technical solution, the adaptive resolution cloth simulation is specifically implemented as follows: Based on the spatial distribution of the initial ground points, a local elevation map is constructed. An initial cloth simulation plane is preset above the local elevation map. The cloth plane is divided into a grid of mass points, with the grid spacing matching the elevation map resolution. For each cloth mass point, virtual gravity is applied to make it move towards the elevation map direction. At the same time, the internal elastic force between the mass points is calculated to maintain the cloth shape. When a mass point moves to a distance less than a preset collision threshold from the ground point in the elevation map, it is determined to be a collision. The position of the mass point is fixed, and the gravity movement and collision detection process is repeated iteratively until the movement speed of all mass points is less than a preset convergence threshold. The surface formed by the fixed cloth mass points is used as a fine ground model, and the point cloud below the model is extracted as a fine ground point set.

[0021] As a further improvement to this technical solution, the confidence assessment is specifically implemented as follows: For the ground candidate points obtained from coarse segmentation, the height score of each ground candidate point is calculated based on the average distance from each ground candidate point to the ground model corresponding to the two segmentation results; the normal vector score of each ground candidate point is calculated based on the average angle between the normal vector of each ground candidate point and the normal vector of the two ground models; the neighborhood score of each ground candidate point is calculated based on the proportion of the preset number of neighborhood points of each ground candidate point that are simultaneously identified as ground points by the two segmentation results; the comprehensive confidence of each ground candidate point is calculated by weighted calculation; and the comprehensive confidence of each fine ground point is calculated similarly for the fine ground points obtained by CSF fine fitting.

[0022] As a further improvement to this technical solution, the specific implementation method for generating and outputting the final classification result based on the multi-criteria fusion strategy is as follows: extract the comprehensive confidence scores of each candidate ground point and the comprehensive confidence scores of each fine-grained ground point; based on the comprehensive confidence scores of each candidate ground point and the comprehensive confidence scores of each fine-grained ground point, set a final classification threshold; compare the comprehensive confidence scores of each candidate ground point and the comprehensive confidence scores of each fine-grained ground point with the final classification threshold to determine whether each candidate ground point is a ground point and whether each fine-grained ground point is a ground point. For a ground point, if a point is the intersection of the initial ground point set and the refined ground point set, and that point is a ground point, then that point is determined to be the final ground point. If a point is the intersection of the initial ground point set and the refined ground point set, and that point is not a ground point, then that point is determined to be a non-ground point. If a point is the difference between the initial ground point set and the refined ground point set, and that point is a ground point, then the CSF surface distance is checked for a second determination. If a point is the difference between the refined ground point set and the initial ground point set, and that point is a ground point, then the neighborhood consistency is checked for a second determination.

[0023] A second aspect of the present invention provides a system for a real-time ground segmentation method, comprising: a preprocessing module for receiving lidar point cloud data and performing reflection noise removal, discrete point filtering, and dividing the point cloud into concentric ring regions.

[0024] The initial segmentation module, connected to the preprocessing module, is used to perform independent ground fitting for each partition using adaptive ground likelihood estimation, and to process the multi-layer ground structure by combining regional vertical plane fitting, and output the initial ground point set.

[0025] The temporal optimization module, connected to the initial segmentation module, is used to compare the segmentation results of the preceding and following frames using a temporal ground back-off mechanism to correct the instantaneous undersegmented regions of the current frame.

[0026] The fine fitting module, connected to the time-series optimization module, is used to construct a local elevation map based on the optimized initial ground point set and perform an adaptive resolution cloth physics simulation to obtain a fine ground model and the corresponding fine ground point set.

[0027] The fusion decision module, connected to the fine fitting module, is used to evaluate the confidence of the initial ground point set and the fine ground point set, and generate the final classification result based on the multi-criteria fusion strategy.

[0028] The timing information update module is connected to the initial segmentation module, the timing optimization module, and the fusion decision module, respectively. It is used to adaptively update the algorithm parameters according to the current frame processing result and store historical information for use in the next frame processing.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0030] 1. This invention employs a hierarchical processing logic of "rapid initial segmentation + CSF local fine fitting + multi-level fusion decision," utilizing rapid initial segmentation for efficient processing of simple terrains, and CSF local fine fitting for high-precision fitting of complex terrains. Finally, multi-criteria fusion decision integrates the advantages of both types of results, enabling accurate ground segmentation in various scenarios. The regional vertical plane fitting function effectively identifies multi-layered ground structures, re-including points above the vertical plane that conform to ground characteristics into ground candidate points, avoiding the misclassification of multi-layered ground as non-ground in existing technologies.

[0031] 2. Based on the parallel processing mechanism of concentric ring partitioning, different processing strategies are adopted for regions at different distances. While ensuring the accuracy of near-distance ground segmentation, the computational load of far-distance point cloud processing is reduced, meeting the real-time requirements. The temporal optimization processing uses the temporal ground back-off mechanism to correct the instantaneously under-segmented regions of the current frame using historical frame information. It can also dynamically update adaptive parameters, making the algorithm more robust to dynamic environments.

[0032] 3. Adaptive ground likelihood estimation can automatically select a planar fitting or piecewise fitting model based on the complexity of the terrain in a region, without the need for manual adjustment of parameters for different scenarios. This avoids the drawbacks of existing technologies that rely on human experience to set fixed parameters and have poor generalization in different scenarios. Attached Figure Description

[0033] 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.

[0034] Figure 1 This is a schematic diagram of the implementation steps of the method of the present invention;

[0035] Figure 2 This is a schematic diagram of the system connection of the present invention;

[0036] Figure 3 This is a schematic diagram of the concentric ring partition of the present invention;

[0037] Figure 4 This is a schematic diagram of the fabric simulation process of the present invention;

[0038] Figure 5 This is a flowchart illustrating the fusion decision-making process of the present invention. Detailed Implementation

[0039] 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.

[0040] Example: Please refer to Figure 1 As shown, a real-time ground segmentation method is provided, including: S1. Point cloud preprocessing and partitioning: receiving LiDAR point cloud data, after coordinate transformation, intensity filtering and reflection noise removal, the point cloud is divided into multiple concentric ring regions based on the distance from the point cloud to the sensor.

[0041] In one specific embodiment, the concentric annular region is implemented as follows: (See attached document) Figure 3As shown, with the lidar sensor as the center, at least two distance thresholds are preset as the radius boundaries of the annular area, and the point cloud space is divided into at least three concentric annular areas. At the same time, with the direction directly in front of the sensor as the starting reference, each concentric annular area is evenly divided into multiple sectors along the 360° circumference to form a two-dimensional grid partition of "annular + sector". Each partition is assigned an independent identifier, and a mapping relationship between the partition and the point cloud data is established. The index information of the point cloud in each partition and the height statistical characteristics of the ground points are stored.

[0042] Partitioning strategy explanation:

[0043] Distance-adaptive partitioning:

[0044] Close-range area (0-10m): High resolution, fine-grained processing

[0045] Medium range (10-35m): Medium resolution, balancing accuracy and efficiency.

[0046] Long distance areas (35-50m): Low resolution, fast processing

[0047] Sector division:

[0048] The 360-degree area is divided into N sectors (typically 36 sectors, each 10 degrees).

[0049] Each sector is processed independently, supporting parallel computing.

[0050] S2. Fast initial segmentation: For each concentric ring region, an adaptive ground likelihood estimation is used to perform independent ground fitting. The initial classification is completed by combining the height threshold and normal vector constraints. The multi-layer ground structure is processed by fitting the vertical plane of the region to obtain the initial ground point set.

[0051] In one specific embodiment, the adaptive ground likelihood estimation performs independent ground fitting. The specific implementation method is as follows: For the point cloud in each partition, calculate the height value of all points and statistically analyze the height mean and variance. Based on the height variance, determine the terrain complexity of the partition. If the height variance is less than a preset complexity threshold, use a plane fitting model to calculate the ground plane coefficient. If the height variance is greater than or equal to the preset complexity threshold, use a piecewise fitting model to divide the point cloud in the partition into subsets according to the height interval. Fit a local ground plane for each subset. For each point in the partition, calculate its distance to the corresponding fitted ground plane. Combine the angle between the normal vector of the point and the normal vector of the ground plane to construct a ground likelihood function and output the ground likelihood value of each point. Determine the classification threshold based on the distribution of the ground likelihood values. Preliminarily determine points with likelihood values ​​higher than the preset likelihood threshold as ground candidate points and points with likelihood values ​​lower than the preset likelihood threshold as non-ground candidate points.

[0052] It should be noted that the preset complexity threshold is set through "historical data statistics + scene adaptation optimization": first, based on a large number of point cloud samples with different terrains (flat, rugged, multi-layered structures, etc.), the critical distribution value of ground height variance is statistically analyzed to determine the initial threshold; then, combined with real-time segmentation, the classification accuracy is dynamically fine-tuned to ensure accurate differentiation between simple terrains that can be fitted to a plane and complex terrains that require piecewise fitting; the plane fitting model and the piecewise fitting model are existing technologies and will not be elaborated on here; the ground likelihood function is constructed with the distance from the point to the fitted ground plane and the angle between the point's normal vector and the ground plane's normal vector as core inputs: the smaller the distance (indicating the point is closer) The smaller the angle between the ground model and the ground (indicating that the normal vector of the point is more consistent with the ground normal vector), the higher the likelihood value of the point belonging to the ground. By normalizing the distance and the angle and assigning preset weights, the ground likelihood value of each point is obtained by weighted summation, realizing a probabilistic and quantitative assessment of the event that "the point belongs to the real ground". The preset likelihood value threshold is set through "sample statistics + dynamic calibration": first, based on a large number of labeled point cloud samples, the likelihood value distribution of ground points and non-ground points is statistically analyzed, and the value at the intersection of the two distributions is selected as the initial threshold; then, combined with the feedback of classification accuracy in real-time segmentation, the threshold is dynamically fine-tuned to ensure that ground candidate points and non-ground candidate points can be accurately distinguished.

[0053] The physical meaning of height variance is "the degree of deviation between the height value of all points in the partition and the height mean": the definition of "preset complexity threshold" is not an independent dimension parameter independent of height variance, but a "complexity classification boundary" based on height variance. Its essence is: through a large number of experiments, statistically determine the "critical value of height variance between flat terrain and complex terrain".

[0054] In one specific embodiment, the vertical plane fitting is implemented as follows: for non-ground candidate points obtained after fast initial segmentation, extract point cloud clusters with continuous spatial positions, calculate the minimum bounding box of each point cloud cluster, determine the ratio of the height of the bounding box to the horizontal dimension, if the ratio is greater than a preset verticality threshold, it is determined to be a suspected vertical plane structure, for point cloud clusters with suspected vertical plane structures, use a plane fitting algorithm to calculate its plane coefficient, extract the angle between the normal vector of the fitting plane and the vertical direction, if the angle between the normal vector of the fitting plane and the vertical direction is less than a preset angle threshold, it is determined to be a vertical plane, and then analyze the positional relationship between the vertical plane and the surrounding ground candidate points, if there is a continuous height connection between the bottom of the vertical plane and the ground candidate points, then the point cloud above the vertical plane is re-included in the range of ground candidate points.

[0055] The minimum bounding box is the smallest cuboid that can completely enclose all points in the cluster. The preset verticality threshold is set through "sample calibration + scene adaptation": First, based on a large number of point cloud cluster samples of vertical planes (such as steps and walls) and non-vertical structures (such as ground protrusions and low obstacles), the critical distribution of the ratio of their height to horizontal dimensions is statistically analyzed to determine the initial threshold; then, combined with the recognition accuracy of vertical planes in real-time segmentation, the threshold is dynamically fine-tuned to ensure accurate differentiation between vertical and non-vertical structures. The plane fitting algorithm is existing technology and will not be elaborated on here. The preset angle threshold is set through "sample statistics + dynamic calibration": First, based on a large number of labeled vertical and non-vertical plane samples, the distribution range of the angle between the fitting plane normal vector and the vertical direction is statistically analyzed, and the value that maximizes the differentiation between the two types of planes is selected as the initial threshold; then, combined with the feedback of the recognition accuracy of vertical planes in real-time segmentation, the threshold is fine-tuned to adapt to the plane determination requirements in different scenarios.

[0056] S3. Temporal optimization processing: Using the ground segmentation results of the previous and next frames, the instantaneous undersegmented regions of the current frame are identified and corrected through a temporal ground backoff mechanism, and the adaptive parameters are updated.

[0057] In one specific embodiment, the temporal ground backoff mechanism is implemented as follows: storing the ground point set of the previous frame and the ground fitting parameters of each partition; comparing the initial segmentation result of the current frame with the stored ground point set of the previous frame; calculating the intersection-union ratio (IU) of the ground point set of the current frame and the ground point set of the previous frame; if the IU is greater than or equal to a preset IU threshold, the current frame segmentation result is determined to be stable, and the historical frame data is directly retained and updated; if the IU is less than the preset IU threshold, the target partition with an IU lower than the threshold is located, the ground fitting parameters of the target partition in the previous frame are extracted, and the ground likelihood value is recalculated in combination with the point cloud data of the target partition in the current frame; based on the recalculated ground likelihood value, the initial segmentation result of the target partition in the current frame is corrected, and the corrected ground point set is used as the final initial segmentation result of the current frame, and the ground parameter library of the historical frames is updated.

[0058] Intersection over Union (IoU) is used to quantify the degree of overlap between the ground point set of the current frame and the previous frame. The core is to calculate the overlap ratio by "intersection points / union points", which reflects the temporal stability of the segmentation results. The preset IoU threshold is set by "historical data statistics + real-time feedback calibration": first, based on the IoU distribution of stable segmentation scenarios in multiple frames, a critical value that can distinguish between "segmentation stability" and "segmentation fluctuation" is selected as the initial threshold; then, combined with the dynamic fine-tuning of the accuracy after correction in real-time segmentation, it ensures accurate identification of unstable partitions that need temporal backtracking correction.

[0059] S4.CSF Local Fine Fitting: Based on the initial ground point set, extract the initial ground points and their neighborhoods, construct a local elevation map, perform adaptive resolution cloth simulation, and obtain a fine ground model and the corresponding fine ground point set.

[0060] In one specific embodiment, the adaptive resolution cloth simulation is implemented as follows: (See [reference]) Figure 4 As shown, a local elevation map is constructed based on the spatial distribution of the initial ground points. The resolution of the elevation map is related to the concentric ring region where the point cloud is located. The closer the region is to the sensor, the higher the resolution of the elevation map. A cloth is preset above the local elevation map to simulate the initial plane. The cloth plane is divided into a grid of mass points with the grid spacing matching the elevation map resolution. For each cloth mass point, virtual gravity is applied to make it move in the direction of the elevation map. At the same time, the internal elastic force between the mass points is calculated to maintain the shape of the cloth. When the mass point moves to a distance less than the ground point in the elevation map, it is determined to be a collision. The position of the mass point is fixed, and the gravity motion and collision detection process is repeated iteratively until the movement speed of all mass points is less than the preset convergence threshold. The surface formed by the fixed cloth mass points is used as a fine ground model, and the point cloud below the model is extracted as a fine ground point set.

[0061] Constructing a local elevation map based on the spatial distribution of initial ground points involves dividing the point cloud space into grid cells with an adaptive resolution, recording the highest elevation value (or average elevation value) of ground points within each grid, and forming a two-dimensional matrix containing elevation information in grid units. This visually reflects the undulating terrain of a local area. The calculation principle for the internal elastic force between mass points is as follows: each fabric mass point establishes an elastic connection only with its neighboring mass points (e.g., 4-neighbor or 8-neighbor mass points). The internal elastic force includes elastic restoring force (resisting deformation and returning the mass points to their initial spacing) and damping force (suppressing oscillations and simulating energy loss). The resultant force of these two forces is the internal elastic force between the mass points, ensuring that the fabric conforms to the terrain without excessive deformation under gravity. The specific process is as follows: Define parameters: Let the mass points... With adjacent particles The initial spacing (the distance when the fabric is not deformed) is The current spacing is ( (Calculation of Euclidean distance between two point masses), elastic modulus. (Controlling tensile / compressive stiffness), damping coefficient (Restriction energy loss rate) is a preset parameter.

[0062] Calculate elastic restoring force: when It is generated over time, and its direction is along the line connecting the two particles. When stretched, it is a tensile force, and when compressed, it is a thrust force.

[0063] Calculate the damping force: it is opposite to the direction of the relative velocity between the two particles, and suppresses the morphological oscillation caused by excessive relative motion between the particles;

[0064] Composite internal elastic force: point mass Subjected to the point mass The internal elastic force is the vector sum of the elastic restoring force and the damping force; similarly, the force of a particle can be calculated. The reaction force received;

[0065] Elastic recovery force: ;

[0066] Damping force: ;

[0067] , Let be the velocity vector of the two particles. It is the projection of the relative velocity onto the line connecting the two particles, and the overall system ensures that the damping force is opposite to the direction of the relative motion.

[0068] Resultant force of internal elasticity: ;

[0069] The preset collision threshold is set based on "accuracy requirements + grid resolution": The initial value is determined by using the local elevation map grid resolution as a base, combined with the ground fitting accuracy target (e.g., millimeter level). For complex terrain, the threshold is appropriately reduced to improve fit, while for flat terrain, it is increased to reduce computational consumption. The preset convergence threshold is set based on "stability requirements + iteration efficiency": The initial value is first set with reference to the minimum value of the particle motion velocity (e.g., approaching 0), and then dynamically fine-tuned according to the convergence speed of the cloth simulation. Iteration stops when the shape is stable, balancing fitting effect and real-time performance.

[0070] Fabric simulation parameters:

[0071] Fabric properties:

[0072] Grid resolution: 0.3-1.0m (adaptive)

[0073] Stiffness coefficient: 1.0-3.0 (related to terrain complexity)

[0074] Particle mass: 1.0 (normalized)

[0075] Physics simulation:

[0076] Gravitational acceleration: 9.8 m / s² (inverted space)

[0077] Time step: 0.65

[0078] Convergence threshold: Maximum velocity < 0.001

[0079] Collision detection:

[0080] Comparison of fabric mass points and topographic elevation map

[0081] After the collision, the particle remains stationary, and its velocity returns to zero.

[0082] S5. Multi-level fusion decision: Confidence assessment is performed on the initial ground point set and the refined ground point set, and the final classification result is generated and output based on a multi-criteria fusion strategy.

[0083] In one specific embodiment, the confidence assessment is implemented as follows: for the ground candidate points obtained from coarse segmentation, the height score of each ground candidate point is calculated based on the average distance from each ground candidate point to the ground model corresponding to the two segmentation results; the normal vector score of each ground candidate point is calculated based on the average angle between the normal vector of each ground candidate point and the normal vector of the two ground models; the neighborhood score of each ground candidate point is calculated based on the proportion of each ground candidate point's neighborhood points that are simultaneously identified as ground points by the two segmentation results; the comprehensive confidence of each ground candidate point is calculated by weighted calculation; and the comprehensive confidence of each fine ground point is calculated similarly for the fine ground points obtained from CSF fine fitting.

[0084] Model 1, Plane Fitting Model: Ground Equation ;

[0085] normal vector ;

[0086] Model 2, CSF fabric simulation model: Ground equations ;

[0087] normal vector ;

[0088] Ground candidate points are Its preset number of neighborhood points is:

[0089] , which is the preset number of neighboring points, such as 8 or 16.

[0090] Height score: Reflects the average distance from the point to the two ground models; the smaller the distance, the higher the score (closer to the ground). Distance from the point to model 1: ;

[0091] Distance from point to model 2:

[0092] Average distance: ;

[0093] Height score (normalized to [0,1], the smaller the distance, the higher the score): ,in, This is the distance attenuation coefficient, which controls the sensitivity of the score to distance.

[0094] Normal vector score: Reflects the average angle between the point's normal vector and the normal vectors of the two models. The smaller the angle, the higher the score (the more consistent the normal vectors). The point's normal vector is... (Preset calculated value), Angle with the normal vector of Model 1: ;

[0095] Angle with the normal vector of Model 2: ;

[0096] Average included angle: ;

[0097] Normal vector score (normalized to [0,1], the smaller the angle, the higher the score): ;in The angle attenuation coefficient controls the sensitivity of the score to the included angle.

[0098] Neighborhood score: Reflects the consistency between the two segmentation results among neighborhood points; a higher consistency ratio results in a higher score (stronger neighborhood support). The number of neighborhood points simultaneously identified as ground points by both models: Neighborhood score (the proportion is directly used as the score): , This represents a preset constant (total number of neighborhood points).

[0099] The overall confidence level of ground candidate points is calculated using the following formula:

[0100] ;

[0101] in , , The system includes preset weighting factors for height score, normal vector score, and neighborhood score. The height score weighting factor is set with "ground elevation fit priority" as its core principle: first, based on a large number of labeled samples, the correlation between height score and the true labels of ground points is statistically analyzed. If height distance error has the greatest impact on classification accuracy (e.g., elevation deviation is the main cause of misclassification in complex terrain), then a higher initial weight is set (e.g., 0.4). Then, through feedback from the confusion matrix of real-time segmentation, the weight is dynamically fine-tuned to ensure accurate differentiation between ground and non-ground points in height-sensitive terrain (e.g., gently undulating ground). The normal vector score weighting factor focuses on "planar consistency recognition requirements": initially referencing empirical values ​​from similar ground segmentation algorithms (e.g., 0.3). For scenarios involving multi-layered ground and slopes, if a misjudgment of the angle between the normal vectors leads to missed detection of points above the vertical plane, the weight should be appropriately increased (e.g., adjusted to 0.35). If the influence of the normal vector is small in flat terrain, the weight should be reduced to balance the role of the normal vector constraint with other scores. The neighborhood score weight factor is based on "scene stability adaptation": in dynamic environments (such as sensor jitter, temporary obstacles), a higher initial weight (e.g., 0.3) is set to strengthen the correction effect of neighborhood consistency on instantaneous misjudgments. In static flat scenarios, the weight can be reduced (e.g., adjusted to 0.25) to reduce the interference of redundant neighborhood information on accurate segmentation, and finally ensure that the total weight is 1, so as to calculate the comprehensive confidence of each fine ground point.

[0102] In one specific embodiment, the method for generating and outputting the final classification result based on the multi-criteria fusion strategy is as follows: (See [reference]). Figure 5 As shown, the comprehensive confidence scores of each candidate ground point and each fine-grained ground point are extracted. Based on these scores, a final classification threshold is set. The comprehensive confidence scores of each candidate ground point and each fine-grained ground point are compared with the final classification threshold to determine whether each candidate ground point is a ground point. Similarly, for each fine-grained ground point, if a point is the intersection of the initial and fine-grained ground point sets and is a ground point, it is determined to be a final ground point. If a point is the intersection of the initial and fine-grained ground point sets and is not a ground point, it is determined to be a non-ground point. If a point is the difference between the initial and fine-grained ground point sets and is a ground point, the CSF surface distance is checked for a second determination. If a point is the difference between the fine-grained and initial ground point sets and is a ground point, neighborhood consistency is checked for a second determination.

[0103] A secondary judgment is made by checking the distance to the CSF surface. If the overall confidence level is higher than the final classification threshold and the vertical distance from the point to the ground model generated by the fine fitting of CSF is less than the preset distance threshold, it is determined to be the final ground point; otherwise, it is determined to be a non-ground point. A secondary judgment is made by checking the neighborhood consistency. If the overall confidence level is higher than the final classification threshold and the proportion of the preset number of neighborhood points that were coarsely segmented and determined to be ground candidate points is greater than the preset neighborhood consistency threshold, it is determined to be the final ground point; otherwise, it is determined to be a non-ground point.

[0104] It should be noted that there is another case where the point is not in the union of the initial ground point set and the fine ground point set, in which case it is directly determined to be a non-ground point.

[0105] See Figure 2 As shown, a system for providing a real-time ground segmentation method includes: a preprocessing module for receiving lidar point cloud data and performing reflection noise removal, discrete point filtering, and dividing the point cloud into concentric ring regions.

[0106] The initial segmentation module, connected to the preprocessing module, is used to perform independent ground fitting for each partition using adaptive ground likelihood estimation, and to process the multi-layer ground structure by combining regional vertical plane fitting, and output the initial ground point set.

[0107] The temporal optimization module, connected to the initial segmentation module, is used to compare the segmentation results of the preceding and following frames using a temporal ground back-off mechanism to correct the instantaneous undersegmented regions of the current frame.

[0108] The fine fitting module, connected to the time-series optimization module, is used to construct a local elevation map based on the optimized initial ground point set and perform an adaptive resolution cloth physics simulation to obtain a fine ground model and the corresponding fine ground point set.

[0109] The fusion decision module, connected to the fine fitting module, is used to evaluate the confidence of the initial ground point set and the fine ground point set, and generate the final classification result based on the multi-criteria fusion strategy.

[0110] The timing information update module is connected to the initial segmentation module, the timing optimization module, and the fusion decision module, respectively. It is used to adaptively update the algorithm parameters according to the current frame processing result and store historical information for use in the next frame processing.

[0111] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A real-time ground segmentation method, characterized in that, include: S1. Point cloud preprocessing and partitioning: After receiving the point cloud data from the lidar, the point cloud is divided into multiple concentric ring regions based on the distance from the point cloud to the sensor after coordinate transformation, intensity filtering and reflection noise removal. S2. Fast initial segmentation: For each concentric ring region, an adaptive ground likelihood estimation is used to perform independent ground fitting. The initial classification is completed by combining the height threshold and normal vector constraints. The multi-layer ground structure is processed by fitting the vertical plane of the region to obtain the initial ground point set. S3. Temporal optimization processing: Using the ground segmentation results of the previous and next frames, the instantaneous undersegmented regions of the current frame are identified and corrected through the temporal ground back-off mechanism, and the adaptive parameters are updated. S4.CSF Local Fine Fitting: Based on the initial ground point set, the initial ground points and their neighborhoods are extracted, a local elevation map is constructed, and an adaptive resolution cloth simulation is performed to obtain a fine ground model and the corresponding fine ground point set; S5. Multi-level fusion decision: Confidence assessment is performed on the initial ground point set and the refined ground point set, and the final classification result is generated and output based on a multi-criteria fusion strategy.

2. The real-time ground segmentation method according to claim 1, characterized in that, The specific implementation method of the concentric annular region is as follows: Using the lidar sensor as the center, at least two distance thresholds are preset as the radius boundaries of the annular area to divide the point cloud space into at least three concentric annular areas. At the same time, taking the direction directly in front of the sensor as the starting reference, each concentric annular area is uniformly divided into multiple sectors along the 360° circumference to form a two-dimensional grid partition of "annular + sector". Each partition is assigned an independent identifier, and a mapping relationship between the partition and the point cloud data is established. The index information of the point cloud in each partition and the height statistics of the ground points are stored.

3. The real-time ground segmentation method according to claim 2, characterized in that, The adaptive ground likelihood estimation performs independent ground fitting, and the specific implementation method is as follows: For the point cloud within each partition, calculate the height values ​​of all points and statistically analyze the mean and variance of the heights. Based on the height variance, determine the terrain complexity of the partition. If the height variance is less than a preset complexity threshold, use a plane fitting model to calculate the ground plane coefficient. If the height variance is greater than or equal to the preset complexity threshold, use a piecewise fitting model to divide the point cloud within the partition into subsets according to height intervals. Fit a local ground plane to each subset. For each point within the partition, calculate its distance to the corresponding fitted ground plane. Combine the angle between the normal vector of the point and the normal vector of the ground plane to construct a ground likelihood function and output the ground likelihood value of each point. Determine the classification threshold based on the distribution of the ground likelihood values. Points with likelihood values ​​higher than the preset likelihood threshold are initially identified as ground candidate points, and points with likelihood values ​​lower than the preset likelihood threshold are initially identified as non-ground candidate points.

4. The real-time ground segmentation method according to claim 3, characterized in that, The vertical plane fitting is specifically implemented as follows: For non-ground candidate points obtained after rapid initial segmentation, spatially continuous point cloud clusters are extracted. The minimum bounding box of each point cloud cluster is calculated, and the ratio of the height to the horizontal dimension of the bounding box is determined. If the ratio is greater than a preset verticality threshold, it is determined to be a suspected vertical plane structure. For point cloud clusters with suspected vertical plane structures, a plane fitting algorithm is used to calculate their plane coefficients. The angle between the normal vector of the fitted plane and the vertical direction is extracted. If the angle between the normal vector of the fitted plane and the vertical direction is less than a preset angle threshold, it is determined to be a vertical plane. Then, the positional relationship between the vertical plane and the surrounding ground candidate points is analyzed. If there is a continuous height connection between the bottom of the vertical plane and the ground candidate points, the point cloud above the vertical plane is re-included in the range of ground candidate points.

5. The real-time ground segmentation method according to claim 4, characterized in that, The specific implementation method of the time-series ground rollback mechanism is as follows: The system stores the ground point set and ground fitting parameters of each partition from the previous frame. It compares the initial segmentation result of the current frame with the stored ground point set from the previous frame and calculates the intersection-union ratio (IU) between the ground point set of the current frame and the ground point set of the previous frame. If the IU is greater than or equal to a preset IU threshold, the current frame segmentation result is considered stable, and the historical frame data is directly retained and updated. If the IU is less than the preset IU threshold, the system locates the target partition with an IU lower than the threshold, extracts the ground fitting parameters of the target partition from the previous frame, and recalculates the ground likelihood value based on the point cloud data of the target partition in the current frame. Based on the recalculated ground likelihood value, the initial segmentation result of the target partition in the current frame is corrected, and the corrected ground point set is used as the final initial segmentation result of the current frame. The system also updates the ground parameter library of the historical frames.

6. The real-time ground segmentation method according to claim 5, characterized in that, The adaptive resolution cloth simulation is specifically implemented as follows: Based on the spatial distribution of the initial ground points, a local elevation map is constructed. A cloth simulation initial plane is preset above the local elevation map, and the cloth plane is divided into a grid of mass points with the grid spacing matching the elevation map resolution. For each cloth mass point, virtual gravity is applied to make it move towards the elevation map direction. At the same time, the internal elastic force between the mass points is calculated to maintain the cloth shape. When a mass point moves to a distance less than the ground point in the elevation map, it is determined to be a collision. The position of the mass point is fixed, and the gravity motion and collision detection process is repeated iteratively until the movement speed of all mass points is less than the preset convergence threshold. The surface formed by the fixed cloth mass points is used as a fine ground model, and the point cloud below the model is extracted as a fine ground point set.

7. The real-time ground segmentation method according to claim 6, characterized in that, The confidence assessment is specifically implemented as follows: For the ground candidate points obtained from coarse segmentation, the height score of each ground candidate point is calculated based on the average distance from each ground candidate point to the ground model corresponding to the two segmentation results. The normal vector score of each ground candidate point is calculated based on the average angle between the normal vector of each ground candidate point and the normal vector of the two ground models. The neighborhood score of each ground candidate point is calculated based on the proportion of the preset number of neighborhood points of each ground candidate point that are simultaneously identified as ground points by the two segmentation results. The comprehensive confidence of each ground candidate point is calculated by weighted calculation. Similarly, the comprehensive confidence of each fine ground point is calculated for the fine ground points obtained by CSF fine fitting.

8. The real-time ground segmentation method according to claim 7, characterized in that, The specific implementation method for generating and outputting the final classification result based on the multi-criteria fusion strategy is as follows: The comprehensive confidence scores of each candidate ground point and each fine-grained ground point are extracted. Based on these scores, a final classification threshold is set. The comprehensive confidence scores of each candidate ground point and each fine-grained ground point are compared with the final classification threshold to determine whether each candidate ground point is a ground point. Similarly, each fine-grained ground point is determined to be a ground point. If a point is the intersection of the initial and fine-grained ground point sets and is a ground point, it is classified as a final ground point. If a point is the intersection of the initial and fine-grained ground point sets and is not a ground point, it is classified as a non-ground point. If a point is the difference between the initial and fine-grained ground point sets and is a ground point, the CSF surface distance is checked for a second determination. If a point is the difference between the fine-grained and initial ground point sets and is a ground point, neighborhood consistency is checked for a second determination.

9. A system for performing the real-time ground segmentation method according to any one of claims 1-8, characterized in that, include: The preprocessing module is used to receive lidar point cloud data and perform reflection noise removal, discrete point filtering, and divide the point cloud into concentric ring regions. The initial segmentation module, connected to the preprocessing module, is used to perform independent ground fitting for each partition using adaptive ground likelihood estimation, and to process multi-layer ground structures by combining regional vertical plane fitting, and output the initial ground point set. The temporal optimization module, connected to the initial segmentation module, is used to compare the segmentation results of the previous and next frames using a temporal ground back-off mechanism to correct the instantaneous undersegmented regions of the current frame. The fine fitting module, connected to the time-series optimization module, is used to construct a local elevation map based on the optimized initial ground point set and perform an adaptive resolution physical simulation of the ground to obtain a fine ground model and the corresponding fine ground point set. The fusion decision module, connected to the fine fitting module, is used to evaluate the confidence of the initial ground point set and the fine ground point set, and generate the final classification result based on the multi-criteria fusion strategy. The timing information update module is connected to the initial segmentation module, the timing optimization module, and the fusion decision module, respectively. It is used to adaptively update the algorithm parameters according to the current frame processing result and store historical information for use in the next frame processing.

Images