A laser radar point cloud ground segmentation method and system based on adaptive graph signal processing
By employing adaptive Voronoi diagrams and graph signal processing methods, the problems of topological connectivity and feature smoothing in sparse regions of LiDAR point cloud ground segmentation are solved, achieving high-precision ground segmentation and improved perception security.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENYANG AEROSPACE UNIVERSITY
- Filing Date
- 2026-03-24
- Publication Date
- 2026-05-29
AI Technical Summary
Existing LiDAR point cloud ground segmentation methods suffer from topological connectivity collapse and over-smoothing of features due to sparsity and non-uniformity in unstructured outdoor environments, making it difficult to maintain high-precision segmentation in sparse regions.
A Voronoi diagram-based and graph signal processing approach is adopted to recover the ground height model of sparse regions by adaptively constructing the graph topology, performing spectral filtering and convex optimization, and combining sensor and vehicle physical constraints.
Achieving high-precision ground segmentation under sparse lidar conditions, maintaining topological connectivity, suppressing noise, restoring key geometric features, and improving the perception safety and adaptability of autonomous driving systems.
Smart Images
Figure CN122116322A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of autonomous driving and robot environmental perception technology, and in particular to a method and system for ground segmentation of LiDAR point clouds based on adaptive graph signal processing (GSP). Background Technology
[0002] In key technologies for autonomous vehicles, unmanned delivery vehicles, and field mobile robots, robustly segmenting ground points from 3D LiDAR point clouds is the fundamental geometric perception layer for accessibility analysis, path planning, and LiDAR odometry localization. While extracting ground surfaces is relatively straightforward in structured urban environments, significant challenges remain in unstructured outdoor scenarios. These environments typically feature uneven terrain, sensor noise, and, most critically, the inherent sparsity and non-uniformity of LiDAR sampling at long distances, as well as the sparsity of low-beam, low-cost LiDAR laser lines.
[0003] Existing ground segmentation methods are mainly divided into learning-based and non-learning methods. While learning-based methods perform well on specific benchmarks, they often suffer from high computational costs and poor cross-domain generalization ability. Therefore, non-learning geometric methods are still widely used due to their efficiency and interpretability. Among them, state-of-the-art mesh-based methods (such as Patchwork++, GroundGrid, etc.) perform reasonably well in dense urban areas, but face a fundamental dilemma related to fixed resolution in unstructured field environments, namely the "spatial rigidity" problem.
[0004] On the one hand, in dense near-field regions, coarse fixed meshes contain a variety of terrain elements, making it difficult to accurately fit the ground or calculate ground height, thus making it difficult to capture fine-grained geometric features such as curbs and slopes (leading to undersegmentation).
[0005] On the other hand, in sparse far-field regions or when using low-beam radar, the subdivided mesh suffers from "connectivity collapse." Due to the non-uniformity and sparsity of LiDAR sampling, effective ground points often fall into non-adjacent discrete mesh cells, severing the topological connections required for the propagation of ground geometry features. This prevents the algorithm from making effective ground inferences based on local geometric priors; or, within a single effective region, the number of points is too small to accurately reflect the terrain features of that region. This neighborhood partitioning based on metric distance is inherently unsuitable for handling the irregular sparsity of 3D LiDAR at long distances, leading to perception blind spots in the downstream path planning module.
[0006] Therefore, there is an urgent need for a ground segmentation method that can adapt to scene complexity and maintain topological connectivity under sparse conditions. Summary of the Invention
[0007] To address the aforementioned problems, the core objective of this invention is to provide a LiDAR point cloud ground segmentation framework based on Voronoi diagrams and graph signal processing. This framework reformulates the LiDAR point cloud ground segmentation problem as a low-rank manifold extraction task on an adaptive scene topology, effectively overcoming the problems of connection and representation failures and excessive feature smoothing in existing technologies. This achieves high-precision, sparsity-resistant LiDAR point cloud ground segmentation. This invention also provides a LiDAR point cloud ground segmentation method based on adaptive graph signal processing, comprising the following steps:
[0008] The raw environmental point cloud data collected by lidar is acquired and preprocessed to remove outliers and reflection noise, resulting in purified point cloud data.
[0009] Seed points are dynamically generated based on the local geometric complexity of the purified point cloud data, and a scene-adaptive graph topology is established by constructing a Voronoi graph. The purified point cloud data is then assigned to the corresponding Voronoi units. An initial elevation map signal is defined on the nodes of the graph topology. The initial elevation map signal is converted to the spectral domain using graph Fourier transform, and spectral filtering is performed using an adaptive low-pass filter to extract the low-frequency smooth signal representing the global ground manifold.
[0010] By introducing sensor height tolerance and vehicle maximum climbing gradient as physical constraints, convex optimization is performed on the low-frequency smoothed signal to recover the smoothed sharp terrain geometric features and obtain a refined ground height model.
[0011] Based on a refined ground height model, the original point cloud is adaptively classified, and ground point sets and non-ground point sets are generated through topological interpolation and dynamic threshold determination.
[0012] Furthermore, the point cloud preprocessing includes region of interest (ROI) filtering and reflection noise removal (RNR). ROI filtering removes points within the vehicle's range and distant outliers exceeding the maximum detection threshold by setting minimum and maximum radial distance detection thresholds. Reflection noise removal sets a physical lower bound threshold based on the lidar installation height; if the vertical height of a point in the point cloud is less than this threshold, it is treated as a noise point generated by multipath reflection and removed. In outdoor environments, multipath reflection often results in false points located below the ground. A threshold based on the lidar installation height is set. The physical lower bound threshold, if the vertical height of a point in the point cloud The following conditions must be met:
[0013] ;
[0014] Then remove it, among which To allow for tolerance, preventing it from lowering the ground envelope estimate.
[0015] Furthermore, the constructed scene-adaptive graph topology aims to address the sparsity problem. The point cloud space is divided into cylindrical sections, and local regions are calculated... Geometric complexity index:
[0016] ;
[0017] in, The standard deviation of the elevation of the original point cloud in this region. For points, The number of original point clouds in the region; the sampling interval of the seed points is inversely proportional to the geometric complexity index, making the seed points dense in complex regions and sparse in flat regions.
[0018] The sampling interval for seed points is determined based on this index. :
[0019] ;
[0020] in, It is the basic interval. This is the normalized gain. Using this seed point sampling interval formula, dense seed points are generated in complex regions to preserve detail, while sparse seed points are generated in flat or sparse regions to maintain connectivity. Subsequently, a Voronoi diagram is constructed based on the seed points to ensure that topological connectivity remains robust as point cloud density decreases.
[0021] Furthermore, the construction of the graph signal based on the Voronoi graph and the subsequent spectral filtering utilize graph signal processing theory. First, the confidence level of each Voronoi unit is calculated. This confidence level combines point density and structural consistency. (Graph edge weights) Taking into account spatial distance, height difference, and node confidence, the global smooth ground manifold signal is extracted by performing eigenvalue decomposition on the graph Laplacian matrix and using an adaptive low-pass filter to truncate high-frequency components (corresponding to obstacle noise).
[0022] The calculation formula is as follows:
[0023] ;
[0024] in, For the first Number of point clouds within a Voronoi cell For elevation standard deviation, The attenuation coefficient is... The minimum confidence level is preset; then the edge weights of the graph are constructed, taking into account spatial distance, height difference and node confidence; finally, the graph Fourier transform is performed using the eigenvalue decomposition of the normalized Laplacian matrix, and high-frequency noise is filtered out by an adaptive cutoff frequency low-pass filter.
[0025] Furthermore, the convex optimization refinement of the physical constraints aims to address the oversmoothing problem caused by spectral filtering. An objective function is constructed that includes smoothing and data fidelity terms, and a set of constraints is introduced. The following optimization problem is solved using the Alternating Direction Multiplier Method (ADMM):
[0026] ;;
[0027] in, Let the ground height be the solution. The signal after spectral filtering. The confidence matrix is... For Laplace matrix, For smoothing parameters; the constraint set It includes height constraints and maximum slope constraints between adjacent nodes, and its specific definition is: ;
[0028] in, express The real space of dimension 1 The total number of nodes in the graph topology; The set of edges of the graph topology; The ground height vector to be solved The elevation scalar element corresponding to the adjacent node. The lowest elevation point during unit initialization; and These are the preset lower and upper bound tolerances, respectively; The Euclidean distance between adjacent nodes; This is the preset maximum traversable gradient for vehicles.
[0029] The ADMM algorithm is used to solve the problem, so that the final ground model can maintain global smoothness and recover sharp geometric features such as road edges.
[0030] Furthermore, the adaptive classification propagates the geometric constraints on the sparse graph back to the original dense point cloud. Inverse distance weighted (IDW) interpolation is performed using Voronoi topological adjacency relations to obtain the estimated ground height for each point, and the height residual is calculated. A context-aware hysteresis thresholding strategy is employed, dynamically adjusting the classification threshold based on the local maximum slope. This classification threshold decreases as the local maximum slope increases, and the dynamic threshold is calculated. When the height residual is less than the dynamic threshold, it is marked as a ground point.
[0031] The formula for calculating the dynamically adjusted classification threshold is:
[0032] ;
[0033] in, For the midpoint of the point cloud Dynamic classification threshold; These are the upper and lower bounds of the preset basic threshold interval, respectively; For point The local maximum slope at that location; The penalty function decreases as the local slope increases. For point The confidence level of the Voronoi unit.
[0034] The threshold decreases with increasing local slope and increases with increasing confidence, thus tolerating sensor noise in flat areas and strictly distinguishing the ground from obstacles at steep edges.
[0035] According to another aspect of the present invention, a lidar point cloud ground segmentation system based on adaptive graph signal processing is also provided, comprising:
[0036] The point cloud preprocessing module is used to acquire the raw environmental point cloud data collected by the lidar, and to preprocess it to remove outliers and reflection noise, so as to obtain purified point cloud data.
[0037] The scene adaptive graph construction module is used to dynamically generate seed points based on the local geometric complexity of the point cloud, construct a Voronoi graph, and assign the original point cloud to the corresponding Voronoi units;
[0038] The graph signal extraction module is used to construct a graph signal based on the Voronoi graph and perform spectral filtering. It uses graph Fourier transform to convert the elevation signal to the spectral domain and separates the low-frequency signal representing the ground manifold.
[0039] The convex optimization refinement module is used to perform physical constraint convex optimization refinement on the filtered signal, introduce physical constraint conditions to recover geometric features, and obtain a refined ground height model.
[0040] The adaptive classification module is used to adaptively classify the original point cloud based on a refined ground height model. It generates a set of ground points and a set of non-ground points through topological interpolation and dynamic threshold determination.
[0041] The LiDAR point cloud ground segmentation method based on adaptive graph signal processing of this invention effectively solves the failure problem of fixed mesh methods in sparse regions through the above steps. Adaptive Voronoi topology naturally possesses density invariance, maintaining geometric propagation even in extremely sparse wire bundles (such as 16 lines). Simultaneously, by combining graph filtering and physical constraint optimization, it achieves both robust noise suppression and preservation of key roadside features, significantly improving the perception safety of autonomous driving systems and their adaptability to different environments and sensor configurations. Attached Figure Description
[0042] Figure 1 This is a flowchart illustrating the present invention;
[0043] Figure 2 This is a schematic diagram of scene complexity perception and seed point generation;
[0044] Figure 3 This is a schematic diagram of adaptive Voronoi topology construction;
[0045] Figure 4 This is a schematic diagram illustrating the effects of spectral filtering and convex optimization. Detailed Implementation
[0046] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.
[0047] refer to Figure 1 This invention provides a ground segmentation method for lidar point clouds based on adaptive graph signal processing (GSP-Voronoi) to address the connectivity collapse and feature oversmoothing problems of existing grid-based methods in sparse environments. It includes the following steps:
[0048] The raw environmental point cloud data collected by lidar is acquired and preprocessed to remove outliers and reflection noise, resulting in purified point cloud data.
[0049] Construct a scene-adaptive graph topology, dynamically generate seed points based on the local geometric complexity of the point cloud, construct a Voronoi graph, and assign the original point cloud to the corresponding Voronoi unit;
[0050] Based on the Voronoi diagram, a graph signal is constructed and spectral filtering is performed. The elevation signal is converted to the spectral domain using graph Fourier transform, and the low-frequency signal representing the ground manifold is separated out.
[0051] The filtered signal is refined by convex optimization with physical constraints. Physical constraints are introduced to recover geometric features, resulting in a refined ground height model.
[0052] Based on a refined ground height model, the original point cloud is adaptively classified, and ground point sets and non-ground point sets are generated through topological interpolation and dynamic threshold determination.
[0053] Specifically,
[0054] 1. Point cloud preprocessing
[0055] First, in order to establish a reliable foundation for graph construction, the original point cloud... Sensor-specific artifacts need to be removed. Apply radial distance constraints. Remove noise points from the vehicle body and excessively distant noise points.
[0056] To address the multipath reflection problem in outdoor environments (i.e., beams hitting wet ground or the chassis creating virtual points below ground level), a reflection noise reduction (RNR) filter is employed. If the point... elevation If the physical lower bound is violated, it will be suppressed:
[0057] ;
[0058] in For radar installation height, This is the tolerance margin.
[0059] 2. Scene Adaptive Graph Construction
[0060] The construction of a scene-adaptive graph topology aims to address the sparsity problem. Traditional fixed meshes fail in sparse regions; therefore, this embodiment employs a scene-adaptive Voronoi topology.
[0061] First, to reduce the initial complexity of data processing, the original point cloud is downsampled using a columnar voxelization method, retaining the lowest point within each columnar voxel as a candidate point.
[0062] Next, the geometric complexity index of the local region is calculated. This index is used to quantify the degree of terrain undulation and point cloud density:
[0063] ;
[0064] in, This represents the elevation standard deviation (i.e., roughness) of the original point cloud within the region. This represents the original number of point clouds within the region. The logarithmic term in the formula serves to compress the dynamic range, effectively preventing the complexity score of high-density near-field regions from being too high and thus obscuring terrain features.
[0065] The local seed point sampling interval is defined based on this index. :
[0066] ;
[0067] in, This represents the basic sampling interval on flat terrain (in this embodiment, it can be, for example, 5.0 meters). It is a normalized gain parameter calibrated based on the sensor's angular resolution.
[0068] This formula implements an "efficiency-fidelity trade-off" mechanism:
[0069] High geometric complexity ( Large areas (such as roadside areas, vegetation boundaries), sampling interval Automatic reduction generates dense seed points, constructing high-resolution Voronoi cells to accurately preserve terrain details;
[0070] In flat and open ( Small areas (such as the center of the road surface), sampling interval The larger Voronoi cells automatically increase in size, generating sparse seed points. This not only saves computational resources, but more importantly, the larger Voronoi cells can bridge the gaps between sparse point clouds, naturally maintaining the topological connectivity of the graph and preventing connection interruption problems in traditional meshing methods.
[0071] Finally, based on the generated non-uniform seed point set The Lloyd relaxation algorithm is applied to optimize the spatial distribution of seed points, making the distribution more uniform and conforming to geometric features. Subsequently, a gridded breadth-first search (BFS) is used to efficiently construct a discrete Voronoi partition, establishing an undirected weighted graph. Each original point is assigned to a corresponding Voronoi cell, such as... Figure 2 and Figure 3 As shown.
[0072] 3. Spectral signal extraction
[0073] Based on the constructed Voronoi topology, this embodiment remodels the ground segmentation problem as a signal separation problem on a graph.
[0074] First, define the definition diagram signal. Each node signal value Initialize to this Voronoi unit The minimum elevation of all original points within the area, i.e. To capture the lower envelope of the terrain.
[0075] To quantify the differences in node reliability caused by the non-uniformity of lidar scanning, a joint confidence metric is introduced. The calculation formula is as follows:
[0076] ;
[0077] in, For unit The number of point clouds within, The elevation standard deviation (i.e., roughness) of the point cloud within the cell. The structural consistency attenuation coefficient (which can be 1.5 in this embodiment) is used. This is a preset minimum confidence level (e.g., 0.10). This metric combines statistical evidence (more points, more confidence) and structural consistency (flatter, more confidence), and utilizes... To ensure numerical stability in extremely sparse regions.
[0078] Constructing the edge weights of the graph To reduce computational complexity, weights are calculated only for direct neighbors in the Voronoi topology. For connected node pairs... The edge weights are calculated using a multimodal two-sided kernel function. :
[0079] ;
[0080] in, These are the geometric center coordinates of the Voronoi element; and This is the initial elevation signal for the corresponding unit; and These are the hyperparameters controlling spatial distance sensitivity and elevation difference sensitivity, respectively (in this embodiment, they can be taken as...). ).
[0081] This weighting formula not only considers traditional spatial proximity and elevation continuity, but also innovatively introduces joint node confidence, ensuring that information is strongly propagated only between nodes with reliable observations, thereby effectively blocking the spread of erroneous signals between the ground and obstacles in sparse far-field conditions.
[0082] Subsequently, based on the weight matrix Calculate the normalized graph Laplacian matrix (in (This is the degree matrix), and its eigenvalues are decomposed to obtain the eigenvector matrix. and eigenvalue diagonal matrix .
[0083] The elevation signal is transformed to the spectral domain using the graphical Fourier transform (GFT): .
[0084] In the spectral domain, to separate the low-frequency components representing smooth terrain from the high-frequency components representing noise and obstacles, this embodiment employs an adaptive low-pass filter. To avoid the failure of traditional fixed cutoff frequencies under different terrain conditions, this embodiment dynamically determines the cutoff frequency based on the energy preservation criterion. The eigenvalue index corresponding to the cutoff frequency is shown below. The calculation formula is as follows:
[0085] ;
[0086] in, The image shows the elevation signal after transformation to the spectral domain. One frequency component; The total number of nodes in the graph; A preset energy retention threshold (e.g., in this embodiment, it is set to...) That is, retaining 99% of the main energy.
[0087] This mechanism dynamically finds the minimum cutoff index that satisfies the principal energy percentage. This achieves the technical effect of adaptively filtering high-frequency noise based on terrain complexity without manual parameter tuning. Finally, the smoothed ground signal is recovered using inverse graphical Fourier transform (IGFT). This effectively separates the global low-frequency ground manifold.
[0088] 4. Physically Constrained Convex Optimization
[0089] While map filtering effectively removes high-frequency noise, it can also lead to over-smoothing of sharp terrain features such as curbs. This embodiment restores these geometric details by introducing a physically constrained convex optimization step.
[0090] Construct a confidence-weighted variational minimization objective function:
[0091] ;
[0092] in, It is the refined ground elevation vector to be solved; It is the smoothed signal output from the previous stage of spectral filtering; It is a confidence diagonal matrix, which serves as a spatial attention mechanism, ensuring that high-confidence regions strictly adhere to the spectral prior, while low-confidence regions rely more on smoothing terms. Perform regularization ( (For smoothing parameters).
[0093] Constraint Set Sensor characteristics and vehicle kinematic constraints are explicitly encoded and defined as follows:
[0094] ;
[0095] This set of constraints contains two parts of physical meaning:
[0096] Height range constraints: This is the original lowest point elevation. and These are upper and lower bound tolerances set based on sensor measurement errors. This prevents the optimized surface from deviating too far from the original measurement data, ensuring data consistency.
[0097] Maximum slope constraint: The Euclidean distance between the centroids of adjacent Voronoi elements is given. The maximum passable gradient for vehicles (e.g.) This constraint mandates that the elevation difference between adjacent units must match the vehicle's physical climbing ability, thereby ensuring that the generated ground model is practically traversable.
[0098] The problem described above is a typical constrained quadratic programming problem. This embodiment employs the Alternating Direction Multiplier Method (ADMM) for iterative solution. This method decouples the smooth objective function from hard geometric constraints, thereby rapidly converging to obtain a ground surface that maintains global coherence while successfully recovering local sharp features (such as curbs), as shown below. Figure 4 As shown.
[0099] 5. Adaptive classification
[0100] Finally, the refined geometric constraints calculated on the sparse graph are propagated back to the original dense point cloud to achieve point-by-point semantic segmentation.
[0101] For any query point in the original point cloud Using its Voronoi cell and its topological neighbors, topological inverse distance weighted (IDW) interpolation is performed to calculate the estimated ground height of the point. .
[0102] To effectively handle complex terrain, this embodiment abandons fixed thresholds and adopts a "context-aware lag threshold" strategy. For each point, its dynamic determination threshold is calculated. :
[0103] ;
[0104] in, For a preset basic threshold range (e.g.) ); For local confidence levels; This is the slope penalty function.
[0105] Slope penalty function The specific definitions are as follows:
[0106] ;
[0107] in This represents the local maximum slope at that point.
[0108] This dynamic threshold mechanism achieves dual adaptive adjustment:
[0109] Slope penalty mechanism: local slope When added, the function The value decreased, causing the threshold to drop. Downbound Tighten. This means that in steep areas (such as at the base of walls or along curbs), the algorithm will use the strictest judgment criteria to effectively prevent vertical obstacles from being misjudged as the ground.
[0110] Confidence reward mechanism: In high-confidence areas with flat terrain and dense point clouds, the threshold will be adjusted upwards. Relaxing the limits allows for tolerance of the sensor's inherent measurement noise, avoiding oversegmentation.
[0111] The final classification criterion is: the height residual of the calculated point. ,like If the condition is met, the point is marked as a ground point; otherwise, it is marked as a non-ground point.
[0112] The method provided in this invention has been validated on the SemanticKITTI benchmark and a real robot platform. Experimental results show that the method achieves extremely high precision (96.25%) while maintaining high recall, ensuring the safety of autonomous driving. In particular, in the stress test of downsampling 64-line data to 16-line, the performance of traditional mesh methods significantly decreased, while the present invention, with its adaptive Voronoi topology, maintained stable performance (F1-score decreased by only 0.77%), demonstrating its excellent robustness to sparsity and applicability to various low-cost or long-range LiDAR sensing applications in autonomous driving and robotics.
[0113] 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 segmentation of lidar point clouds based on adaptive graph signal processing, characterized in that, Includes the following steps, The raw environmental point cloud data collected by lidar is acquired and preprocessed to remove outliers and reflection noise, resulting in purified point cloud data. Seed points are dynamically generated based on the local geometric complexity of the purified point cloud data, and a scene-adaptive graph topology is established by constructing a Voronoi graph. The purified point cloud data is then assigned to the corresponding Voronoi units. An initial elevation map signal is defined on the nodes of the graph topology. The initial elevation map signal is converted to the spectral domain using graph Fourier transform, and spectral filtering is performed using an adaptive low-pass filter to extract the low-frequency smooth signal representing the global ground manifold. By introducing sensor height tolerance and vehicle maximum climbing gradient as physical constraints, convex optimization is performed on the low-frequency smoothed signal to recover the smoothed sharp terrain geometric features and obtain a refined ground height model. Based on a refined ground height model, the original point cloud is adaptively classified, and ground point sets and non-ground point sets are generated through topological interpolation and dynamic threshold determination.
2. The method for ground segmentation of lidar point clouds based on adaptive graph signal processing as described in claim 1, characterized in that, The preprocessing of the original environmental point cloud data includes region of interest filtering and reflection noise removal. The region of interest filtering removes points located within the vehicle body and distant outliers exceeding the maximum detection threshold by setting a minimum and a maximum detection threshold for radial distance. The reflection noise removal sets a physical lower bound threshold based on the installation height of the lidar. If the vertical height of a point in the point cloud is less than this threshold, it is treated as a noise point generated by multipath reflection and removed.
3. The method for ground segmentation of lidar point clouds based on adaptive graph signal processing as described in claim 1, characterized in that, In the construction of the scene-adaptive graph topology, the method for calculating local geometric complexity is as follows: the point cloud space is divided into cylindrical shapes, and the geometric complexity index is calculated for local regions. The calculation formula is as follows: ； in, The standard deviation of the elevation of the original point cloud in this region. The number of original point clouds in the region; the sampling interval of the seed points is inversely proportional to the geometric complexity index, making the seed points dense in complex regions and sparse in flat regions.
4. The method for ground segmentation of lidar point clouds based on adaptive graph signal processing as described in claim 1, characterized in that, In the step of constructing the graph signal based on the Voronoi graph and performing spectral filtering, the confidence level of each Voronoi unit node is first calculated using the following formula: ； in, For the first Number of point clouds within a Voronoi cell For elevation standard deviation, The attenuation coefficient is... The minimum confidence level is preset; then the edge weights of the graph are constructed, taking into account spatial distance, height difference and node confidence; finally, the graph Fourier transform is performed using the eigenvalue decomposition of the normalized Laplacian matrix, and high-frequency noise is filtered out by an adaptive cutoff frequency low-pass filter.
5. The lidar point cloud ground segmentation method based on adaptive graph signal processing as described in claim 1, characterized in that, In the convex optimization refinement step of the physical constraints, an objective function containing a smoothing term and a data fidelity term is constructed, and a constraint set is introduced. The following optimization problem is solved using the Alternating Direction Multiplier Method (ADMM): ； in, Let the ground height be the solution. The signal after spectral filtering. The confidence matrix is... For Laplace matrix, For smoothing parameters; the constraint set Includes height constraints and maximum slope constraints for adjacent nodes.
6. The lidar point cloud ground segmentation method based on adaptive graph signal processing as described in claim 5, characterized in that, The constraint set The maximum slope constraint between adjacent nodes is specifically defined as follows: ； in, express The real space of dimension 1 The total number of nodes in the graph topology; The set of edges of the graph topology; The ground height vector to be solved The elevation scalar element corresponding to the adjacent node. The lowest elevation point during unit initialization; and These are the preset lower and upper bound tolerances, respectively; The Euclidean distance between adjacent nodes; This is the preset maximum passable gradient for vehicles.
7. The method for ground segmentation of lidar point clouds based on adaptive graph signal processing as described in claim 1, characterized in that, In the step of adaptively classifying the cleaned point cloud data based on a refined ground height model: The original point cloud is subjected to inverse distance weighted interpolation using Voronoi topological adjacency relations to obtain the estimated ground height, and the height residual of the points is calculated. A context-aware lag threshold strategy is adopted to dynamically adjust the classification threshold based on the local maximum slope, wherein the classification threshold decreases as the local maximum slope increases; When the height residual is less than the dynamic classification threshold, the corresponding point is marked as a ground point.
8. The lidar point cloud ground segmentation method based on adaptive graph signal processing as described in claim 7, characterized in that, The formula for calculating the dynamically adjusted classification threshold is: ； in, For the midpoint of the point cloud The dynamic classification threshold; These are the upper and lower bounds of the preset basic threshold interval, respectively; For point The local maximum slope at that location; The penalty function decreases as the local slope increases. For point The confidence level of the Voronoi unit to which it belongs.
9. A lidar point cloud ground segmentation system based on adaptive graph signal processing, characterized in that, include, The point cloud preprocessing module is used to acquire the raw environmental point cloud data collected by the lidar, and to preprocess it to remove outliers and reflection noise, so as to obtain purified point cloud data. The scene adaptive graph construction module is used to dynamically generate seed points based on the local geometric complexity of the point cloud, construct a Voronoi graph, and assign the original point cloud to the corresponding Voronoi units; The graph signal extraction module is used to construct a graph signal based on the Voronoi graph and perform spectral filtering. It uses graph Fourier transform to convert the elevation signal to the spectral domain and separates the low-frequency signal representing the ground manifold. The convex optimization refinement module is used to perform physical constraint convex optimization refinement on the filtered signal, introduce physical constraint conditions to recover geometric features, and obtain a refined ground height model. The adaptive classification module is used to adaptively classify the original point cloud based on a refined ground height model. It generates a set of ground points and a set of non-ground points through topological interpolation and dynamic threshold determination.