Method and system for centerline determination based on neural network and point cloud spatial features

Abstract

The application discloses a road centerline determination method and system based on a neural network and point cloud space features, and relates to the technical field of neural networks; the method comprises the following steps: generating a passable road domain mask and a layer identification map based on original three-dimensional point cloud data of an N-layer overlapping road, and extracting a passable area in the layer; high-confidence skeletons are obtained through non-maximum suppression and three-dimensional refinement; a candidate road centerline set is generated through connectivity clustering and length constraint; comprehensive anomaly detection is carried out based on geometric, topological and semantic features, the anomaly type of the road centerline is identified, and an anomaly priority queue is established; the candidate road centerline is divided into a batch repair set and a single-point repair set; geometric shape adjustment, topological structure reconstruction and semantic attribute updating are respectively performed on the two sets to obtain a corrected road centerline set; the application realizes high-precision optimization and anomaly classification and repair of the geometry, topology and semantics of the road centerline, and solves the conflict between local optimization and global optimization of the road centerline.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of neural networks, and more particularly, to a road centerline determination method and system based on neural networks and point cloud spatial features. BACKGROUND

[0002] In the field of high-precision map making and intelligent transportation, the accurate acquisition of road centerlines is one of the basic works. Traditional road centerline extraction methods usually rely on manual annotation or geometric fitting based on two-dimensional images. In dealing with complex road scenes such as multi-lane intersections, irregular curvature roads, and road occlusions, there are often problems such as insufficient positioning accuracy, poor shape continuity, and unstable topological connectivity. In recent years, with the development of laser radar and three-dimensional point cloud technology, road centerline extraction based on point cloud data has gradually become the mainstream. However, massive point cloud data inevitably produces noise points, missing segments, and geometric distortions during acquisition, and there are differences in semantic attributes, coordinate systems, and data resolution between different acquisition batches, making the extraction and repair of road centerlines face higher requirements for accuracy and consistency.

[0003] Existing technologies generally rely on single geometric or topological repair methods to deal with the above problems, lacking a comprehensive optimization mechanism for road centerline geometry, topology, and semantic attributes. For example, existing technologies only smooth the curve for geometric offset of the road centerline, but cannot simultaneously repair the broken topological structure of the road centerline. For another example, although existing technologies can achieve node connectivity optimization, they do not consider the unification of semantic labels and the resolution of attribute conflicts, resulting in inconsistencies in the repaired road centerline in terms of vision or logic. In addition, existing batch repair and single-point fine repair often cannot be executed collaboratively in the same framework, resulting in conflicts between local optimization and global optimization and affecting the overall quality of the final results.

[0004] In view of this, the present application proposes a road centerline determination method and system based on neural networks and point cloud spatial features to solve the above problems. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art and achieve the above-mentioned purposes, the present application provides the following technical solution: a road centerline determination method based on neural networks and point cloud spatial features, comprising:

[0006] Processing the original three-dimensional point cloud data of N layers of overlapping roads to obtain a passable road domain mask and a layer identification map;

[0007] Extracting an intra-layer passable area based on the passable road domain mask and the layer identification map, obtaining a high-confidence skeleton through non-maximum suppression and three-dimensional thinning, and generating a candidate road centerline set through connectivity clustering and length constraint;

[0008] comprehensive anomaly detection processing is performed on the geometric, topological and semantic features corresponding to the candidate road centerline set to obtain a road centerline anomaly type corresponding to each candidate road centerline, the road centerline anomaly type including a geometric anomaly type, a topological anomaly type and a semantic anomaly type;

[0009] An abnormal priority queue is established based on the road centerline anomaly types corresponding to the candidate road centerline set; the road centerline anomaly types of the candidate road centerlines are divided into a batch repair set and a single-point repair set in sequence according to the abnormal priority queue; geometric shape adjustment, topological structure reconstruction and semantic attribute updating are performed on the batch repair set and the single-point repair set respectively to obtain a corrected road centerline set.

[0010] Further, the method for establishing the abnormal priority queue comprises:

[0011] An abnormal influence weight table is pre-constructed;

[0012] Each abnormal type in the road centerline anomaly type is assigned a weight according to the abnormal influence weight table to obtain an abnormal type weight corresponding to each abnormal type;

[0013] For each candidate road centerline in the candidate road centerline set, the abnormal type weights contained in the candidate road centerline are accumulated to obtain an abnormal comprehensive influence score of the candidate road centerline;

[0014] All abnormal comprehensive influence scores corresponding to the candidate road centerlines are constructed into an abnormal comprehensive influence score set; the abnormal comprehensive influence score set is sorted in descending order to obtain the abnormal priority queue.

[0015] Further, the method for dividing the road centerline anomaly types of the candidate road centerlines into the batch repair set and the single-point repair set in sequence according to the abnormal priority queue comprises:

[0016] For each candidate road centerline in the abnormal priority queue, the road centerline anomaly types, the abnormal type weights and the corresponding abnormal comprehensive influence scores contained in the candidate road centerline are input into an abnormal shunting model to obtain a batch repair set and a single-point repair set corresponding to each candidate road centerline; the abnormal shunting model is a deep neural network with an input layer, a hidden layer and an output layer as the core.

[0017] Further, the method for obtaining the corrected road centerline set comprises:

[0018] For each candidate road centerline in the abnormal priority queue, if the candidate road centerline contains a batch repair set, the spatial positions and the abnormal type weights corresponding to each abnormal type in the batch repair set are extracted and constructed into corresponding batch abnormal nodes; the batch abnormal nodes are assembled into a batch repair subgraph, and the batch repair subgraph is input into a first parameter setting model to obtain a first parameter set;

[0019] If the candidate road centerline contains a single-point repair set, the spatial position and anomaly type weight corresponding to each anomaly type in the single-point repair set are extracted and constructed into corresponding single-point anomaly data; each single-point anomaly data is input into a second parameter setting model to obtain a second parameter set;

[0020] According to the first parameter set and the second parameter set, the candidate road centerline is geometrically adjusted, topologically reconstructed, and semantically updated to obtain a corrected road centerline;

[0021] All corrected road centerlines are collected into a corrected road centerline set.

[0022] Further, the method for judging the geometric anomaly type comprises:

[0023] The candidate road centerline is scattered into B sequence positioning points according to a preset first interval distance; for each sequence positioning point, the shortest Euclidean distance from the sequence positioning point to one side boundary of the road is calculated and recorded as a first Euclidean distance, and the shortest Euclidean distance from the sequence positioning point to the other side boundary of the road is calculated and recorded as a second Euclidean distance; the first Euclidean distance and the second Euclidean distance are subtracted and the absolute value is taken to obtain a bidirectional distance residual corresponding to the sequence positioning point; the first Euclidean distance and the second Euclidean distance are summed to obtain a local road width corresponding to the sequence positioning point;

[0024] The B bidirectional distance residuals are constructed into a bidirectional distance residual set; and the B local road widths are constructed into a local road width set;

[0025] A symmetry index is calculated based on the bidirectional distance residual set and the local road width set; and a road width consistency index is calculated based on the local road width set;

[0026] If the symmetry index is less than a preset symmetry index threshold, a geometric offset anomaly is added to the geometric anomaly type of the candidate road centerline; and if the road width consistency index is less than a preset road width consistency index threshold, a width anomaly is added to the geometric anomaly type of the candidate road centerline.

[0027] Further, the method for judging the topological anomaly type comprises:

[0028] The candidate road centerline is scattered into a line element sequence according to a preset second interval distance, and the spacing of adjacent line elements is calculated based on spatial adjacency relationship; when the spacing of any adjacent line elements is greater than a preset breaking threshold, it is determined that the candidate road centerline has a breaking anomaly, and a breaking anomaly is added to the topological anomaly type of the candidate road centerline;

[0029] A global intersection search is performed on the candidate road centerline in three-dimensional space. When a candidate road centerline is detected to have a geometric intersection at a non-adjacent line element and the distance is less than a preset self-intersection judgment threshold, it is determined that the candidate road centerline has a self-intersection anomaly, and a self-intersection anomaly is added to the topological anomaly type of the candidate road centerline.

[0030] If adjacent line elements of the same candidate road centerline are located in different road layers and the distance between adjacent line elements is less than the preset cross-layer determination threshold, it is determined that the candidate road centerline has a cross-layer short circuit anomaly, and a short circuit anomaly is added to the topology anomaly type of the candidate road centerline.

[0031] Furthermore, methods for determining semantic anomaly types include:

[0032] A pre-constructed road semantic feature set is provided; the road semantic feature set includes road level information and road lane number information; the road level information includes expressways, urban expressways, arterial roads, secondary arterial roads, and local roads.

[0033] The spatial position of the candidate road centerline is spatially superimposed with the real-time collected road vector data or road surface data to extract the road object with the highest spatial overlap with the candidate road centerline, and the road grade and number of lanes corresponding to the road object are obtained.

[0034] If the road level of the road object is different from the road level corresponding to the road semantic feature set, it is determined that the candidate road centerline has a road level mismatch; road level mismatch is added to the semantic anomaly type of the candidate road centerline.

[0035] If the number of lanes of a road object is different from the number of lanes corresponding to the road semantic feature set, it is determined that the candidate road centerline has a lane number mismatch; lane number mismatch is added to the semantic anomaly type of the candidate road centerline.

[0036] Furthermore, the method for extracting the candidate road centerline set includes:

[0037] Based on the traffic area mask and layer identifier map, the passable area within each candidate road layer number is extracted; for the spatial boundary points of the passable area within the layer, a three-dimensional contour tracking algorithm is used to construct a set of boundary patches.

[0038] Based on the passable region within the layer and the corresponding set of boundary patches, the three-dimensional Euclidean distance field and the corresponding distance field gradient are calculated; based on the distance field gradient, non-maximum suppression operation is performed on the three-dimensional Euclidean distance field along the gradient direction to obtain a skeleton candidate set, and three-dimensional thinning processing is performed on the passable region within the layer to obtain the morphological skeleton.

[0039] Point cloud data points that exist simultaneously in the candidate skeleton set and the morphological skeleton are constructed into a high-confidence skeleton point set. The interconnected skeleton points in the high-confidence skeleton point set are clustered into continuous skeleton segments using a three-dimensional connectivity search algorithm. Length constraint processing is performed on each skeleton segment, and skeleton segments with a length lower than the preset minimum skeleton length threshold are removed.

[0040] All retained skeleton line segments are aggregated to construct a set of candidate road centerlines.

[0041] Furthermore, the method for obtaining the traffic area mask and layer identifier map includes:

[0042] Standard point cloud data is obtained by preprocessing the original 3D point cloud data;

[0043] For each point cloud data point in the standard point cloud data, point-level spatial features are extracted; the point-level spatial features include normal vector, curvature indicator, mean echo intensity and variance of echo intensity, and a neighborhood height histogram of the point cloud data point is statistically plotted along the gravity direction; the point-level spatial features of all point cloud data points in the standard point cloud data are constructed into a point-level spatial feature set.

[0044] The point-level spatial feature set is input into the road domain probability assessment model to obtain the road domain probability score corresponding to each point cloud data point. Based on the preset road domain probability score threshold, the point cloud data points are labeled with road domain labels or non-road domain labels.

[0045] By combining all road domain labels and non-road domain labels with the 3D spatial coordinates of the corresponding point cloud data points, a connected component is constructed to obtain a passable road domain mask. After combining all road domain labels and non-road domain labels with the 3D spatial coordinates of the corresponding point cloud data points, a connected component is constructed, and after connectivity analysis, region culling and hole filling, a spatially topologically continuous passable road domain mask is obtained.

[0046] Layer identification maps are obtained by performing hierarchical recognition processing based on the traffic area mask.

[0047] Furthermore, the method for obtaining a layer identification map by performing hierarchical recognition processing based on the traffic area mask includes:

[0048] For each point cloud data point within the traffic area mask, the position, neighborhood height histogram, and absolute height value in the gravity direction of the point cloud data point are read, and height distribution characteristics are statistically formed.

[0049] Based on the preset height segmentation threshold, the traffic area mask is divided into M height bands. Within each height band, the local plane characteristics are analyzed by combining the normal vector and curvature indicator. Point cloud data points with normal vector similarity higher than the corresponding threshold and height variance lower than the corresponding threshold are clustered into the same connected piece, which is used as a candidate road layer and assigned a connected piece number.

[0050] For candidate road layers that overlap significantly or partially overlap, they are separated by using the horizontal position distribution and the minimum height difference between layers, and a candidate road layer number is assigned to each point cloud data point.

[0051] Map all candidate road layer numbers to their corresponding 3D coordinates to generate a layer identifier map aligned with the traffic area mask space.

[0052] A path centerline determination system based on neural networks and point cloud spatial features, used to implement the path centerline determination method based on neural networks and point cloud spatial features, includes:

[0053] The layered road segmentation module is used to process the original 3D point cloud data of N layers of overlapping roads to obtain the traffic road mask and layer identifier map;

[0054] The candidate centerline extraction module extracts the passable area within the layer based on the passable road domain mask and layer identifier map, obtains a high-confidence skeleton through non-maximum suppression and three-dimensional refinement, and generates a set of candidate road centerlines through connectivity clustering and length constraints.

[0055] The centerline verification module performs comprehensive anomaly detection processing based on the geometric, topological, and semantic features corresponding to the candidate road centerline set to obtain the corresponding road centerline anomaly types, which include geometric anomaly types, topological anomaly types, and semantic anomaly types.

[0056] The centerline hierarchical correction module establishes an anomaly priority queue based on the anomaly types of the candidate road centerlines. According to the anomaly priority queue, the candidate road centerlines are divided into batch repair sets and single-point repair sets according to their anomaly types. Geometric morphology adjustment, topology reconstruction, and semantic attribute update are performed on the batch repair sets and single-point repair sets respectively to obtain the corrected road centerline sets.

[0057] Compared with existing technologies, the path centerline determination method and system based on neural networks and point cloud spatial features proposed in this invention have the following technical effects and advantages:

[0058] This invention enables high-precision, automated centerline extraction and anomaly repair in large-scale point cloud data. By introducing a dual-set repair strategy—batch repair sets and single-point repair sets—the technical solution establishes a parameterized adjustment mechanism across geometric, topological, and semantic dimensions. This fundamentally improves the integrity and consistency of centerline data and resolves the conflict between local and global centerline optimization. Compared to existing technologies, this invention not only boasts higher automation and adaptability but also reduces manual intervention, significantly improving road network data processing efficiency.

[0059] Furthermore, this invention utilizes a first set of parameters to achieve efficient global repair of batch anomalies, encompassing geometric adjustment parameters, topology optimization parameters, and semantic consistency parameters, thereby ensuring comprehensive optimization of the path centerline correction results in terms of morphological continuity, structural connectivity, and semantic consistency; simultaneously, it combines a second set of parameters to perform fine-grained localization and optimization of single-point anomalies, avoiding interference with the global results and achieving precise repair of small but critical anomalies.

[0060] In summary, the technical solution of this invention significantly improves the geometric continuity, topological rationality, and semantic accuracy of the final output road centerline data, which is beneficial for high-precision map construction, road element recognition, and intelligent driving applications. Attached Figure Description

[0061] Figure 1 This is a schematic diagram of the path centerline determination system based on neural networks and point cloud spatial features according to Embodiment 1 of the present invention;

[0062] Figure 2 This is a flowchart of the path centerline determination method based on neural networks and point cloud spatial features according to Embodiment 2 of the present invention;

[0063] Figure 3 This is a flowchart of the method for establishing an exception priority queue according to Embodiment 1 of the present invention;

[0064] Figure 4 This is a flowchart of the method for obtaining the modified path centerline set according to Embodiment 1 of the present invention;

[0065] Figure 5 This is a flowchart of the method for determining the geometric anomaly type in Embodiment 1 of the present invention;

[0066] Figure 6 This is a flowchart of the method for determining the topological anomaly type in Embodiment 1 of the present invention;

[0067] Figure 7 This is a flowchart of the semantic anomaly type determination method in Embodiment 1 of the present invention. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be described in detail, clearly, and completely below with reference to the accompanying drawings. It should be particularly noted that the specific embodiments described below are only for better illustrating and explaining the technical solutions of the present invention, and are intended to enable those skilled in the art to better understand and implement the present invention, and should not be construed as limiting the scope of protection of the present invention. Without departing from the spirit and substance of the present invention, those skilled in the art can modify, adjust, or make equivalent substitutions based on the content disclosed in the present invention, and these should all be considered within the scope of protection of the present invention.

[0069] Example 1:

[0070] Please see Figure 1 As shown, this embodiment discloses a road centerline determination system based on neural networks and point cloud spatial features, including a hierarchical road domain segmentation module, a candidate centerline extraction module, a centerline verification module, and a centerline hierarchical correction module. Each module is connected by wired and / or wireless means to achieve data transmission.

[0071] The layered road segmentation module is used to process the original 3D point cloud data of N layers of overlapping roads to obtain the traffic road mask and layer identifier map.

[0072] The method for obtaining the traffic area mask and layer identifier map includes:

[0073] Standard point cloud data is obtained by preprocessing the original 3D point cloud data;

[0074] For each point in the standard point cloud data, principal component analysis is performed on the neighborhood of the point to obtain the normal vector and curvature indicator. The mean and variance of the echo intensity in the neighborhood of the point are calculated. A neighborhood height histogram of the point is plotted along the gravity direction. The position, normal vector, curvature indicator, mean echo intensity, variance of echo intensity, and neighborhood height histogram of the point are used to construct the corresponding point-level spatial features. The point-level spatial features of all point data points in the standard point cloud data are used to construct a point-level spatial feature set.

[0075] The point-level spatial feature set is input into the road domain probability assessment model to obtain the road domain probability score corresponding to each point cloud data point. If the road domain probability score is greater than or equal to the preset road domain probability score threshold, the corresponding point cloud data point is labeled as a road domain label. If the road domain probability score is less than the preset road domain probability score threshold, the corresponding point cloud data point is labeled as a non-road domain label.

[0076] By combining all road domain labels and non-road domain labels with the 3D spatial coordinates of the corresponding point cloud data points to construct a connected component, a passable road domain mask is obtained. After combining all road domain labels and non-road domain labels with the 3D spatial coordinates of the corresponding point cloud data points to construct a connected component, a passable road domain mask with continuous spatial topology is obtained through connectivity analysis, small region removal and hole filling.

[0077] Layer identification maps are obtained by performing hierarchical recognition processing based on the traffic area mask.

[0078] Specifically, the method for acquiring the standard point cloud data involves first aligning the original 3D point cloud data with the IMU / GNSS pose interpolation to the same reference time according to the timestamp; then, mapping the original 3D point cloud data from the sensor coordinate system to the world coordinate system using the gravity direction as a reference, resulting in a gravity-aligned standard point cloud where the height axis is consistent with the vertical direction. Based on the gravity-aligned standard point cloud, statistical outlier cleaning and dynamic object screening are performed to remove isolated noise points and the influence of short-term moving targets, yielding a static scene point cloud. The IMU (Inertial Measurement Unit) is used to acquire inertial measurement data such as the angular velocity and acceleration of the measurement vehicle; GNSS (Global Navigation Satellite System) is used to acquire the 3D position and velocity information of the measurement vehicle.

[0079] It should be noted that the method for obtaining the normal vector specifically involves selecting a local neighborhood centered on each point cloud data point using the K-nearest neighbor algorithm, performing local plane fitting on the local neighborhood, and obtaining the normal direction of the fitted plane, which is then used as the normal vector of that point cloud data point. To improve stability, least squares fitting or plane estimation based on principal component analysis can be used, and consistency constraints can be applied to the normal direction. Consistency constraints can include unifying the normal direction to the opposite direction of gravity or adjusting it along the dominant direction of the neighborhood. In noisy scenarios, the initial normal vector can be processed by neighborhood weighted smoothing or bilateral filtering to reduce the impact of random disturbances on subsequent decisions.

[0080] Specifically, the method for obtaining the curvature indicator involves first estimating the local plane or local principal direction based on the local neighborhood of each point cloud data point. Then, the degree of deviation from the local plane or principal direction is used as a measure of the surface change of that point cloud data point, denoted as the curvature indicator. The curvature indicator can be obtained from the local fitting residual, the discreteness of the neighborhood geometric distribution, or the degree of non-uniformity of the energy distribution along the principal direction.

[0081] Specifically, the method for obtaining neighborhood echo intensity involves using the echo intensity values ​​recorded in the original 3D point cloud data to collect intensity samples of each point cloud data point within its local neighborhood, and calculating the mean and variance of the echo intensity. To reduce systematic biases caused by factors such as sensor distance attenuation and slippery reflections, distance compensation or device-level normalization can be performed on the intensity before statistical analysis.

[0082] Those skilled in the art are familiar with the acquisition process and engineering implementation of normal vector, curvature indicator and neighborhood echo intensity, which provides mature function interfaces and implementation references for existing computer vision and 3D reconstruction toolchains, and is sufficient to support the implementation of the technical solution of this application.

[0083] The training method for the road domain probability assessment model includes:

[0084] A road domain probability assessment dataset is pre-constructed, which includes road domain probability assessment data of the LY group and the corresponding road domain probability scores of the LY group road domain probability assessment data, where LY is a positive integer; the road domain probability assessment data includes a set of point-level spatial features; the road domain probability assessment dataset is divided into a training set and a validation set, the training set is used for learning the parameters of the road domain probability assessment model, and the validation set is used to monitor the generalization performance and overfitting degree of the road domain probability assessment model in real time;

[0085] A node-output graph neural network architecture is adopted as the road domain probability assessment model. The road domain probability assessment data is standardized and vectorized before being input into the node-output graph neural network architecture, which consists of an input layer, hidden layers, and an output layer. Each hidden layer uses a nonlinear activation function to extract features, and the output layer uses a softmax activation function to obtain the probability distribution corresponding to each road domain probability score. Finally, the road domain probability score corresponding to the highest probability is taken as the prediction result of the road domain probability assessment model. During training, the cross-entropy loss function is used as the optimization objective, and a gradient descent-type optimization algorithm is used to update the network weights. An early stopping strategy is set: when the prediction accuracy on the validation set reaches or exceeds a preset threshold, the road domain probability assessment model is determined to have converged and training is terminated.

[0086] It should be noted that the road domain probability score is used to quantify the probability that each point cloud data point belongs to a road domain, and the value range of the road domain probability score is [missing value]. The closer the road domain probability score is to 1, the higher the probability that the point cloud data point belongs to the road domain; the closer the road domain probability score is to 0, the higher the probability that the point cloud data point belongs to the non-road domain.

[0087] For example, in a preferred embodiment of this application, for urban road point cloud data, the road domain probability scoring threshold can be set to 0.5. That is, when the road domain probability score is greater than or equal to 0.5, the corresponding point cloud data point is labeled as a road domain, and when the road domain probability score is less than 0.5, the corresponding point cloud data point is labeled as a non-road domain. For example, for high-precision map acquisition scenarios, the road domain probability scoring threshold can be set to 0.55 to improve the ability to distinguish complex boundary areas. For example, for unpaved roads or scenarios with low point cloud density, the road domain probability scoring threshold can be set to 0.45 to avoid road domain omissions caused by noise. Furthermore, those skilled in the art can flexibly configure the road domain probability scoring threshold according to the performance of the acquisition equipment, road type, and task accuracy requirements, so as to obtain stable road domain and non-road domain determination results in different application environments, supporting the implementation of the technical solution of this application.

[0088] Methods for obtaining layer identification maps based on traffic area masks include:

[0089] For each point cloud data point within the traffic area mask, read the position and neighborhood height histogram in the point-level spatial features corresponding to the point cloud data point, and combine the absolute height value of the point cloud data point in the gravity direction to statistically obtain the height distribution characteristics of all point cloud data points in the traffic area mask.

[0090] Based on the height distribution characteristics, the traffic area mask is divided into M height bands using a preset height segmentation threshold. Different height bands correspond to different road levels. Within each height band, local plane characteristics are further analyzed based on the normal vector and curvature indicator. Point cloud data points with normal vector similarity greater than a preset normal vector similarity threshold and height variance lower than a preset height variance threshold are clustered into the same connected piece, which serves as the candidate road layer within the height band. The connected piece is then assigned a corresponding connected piece number.

[0091] For candidate road layers with overlapping heights or partial overlaps, layer separation is performed by combining the horizontal position distribution of point cloud data points and the minimum height difference between layers to ensure that different road layers are not confused in space. Each point cloud data point is assigned a unique candidate road layer number according to its height band and connected piece number.

[0092] The candidate road layer numbers of all point cloud data points within the traffic area mask are combined with their corresponding 3D spatial locations to generate a layer identifier map aligned with the traffic area mask, providing basic data support for subsequent layer binding and road centerline extraction.

[0093] For candidate road layers with overlapping or partial overlap, the method of separating layers by combining the horizontal position distribution of point cloud data points and the minimum height difference between layers is specifically as follows: For each obtained candidate road layer, statistical analysis is performed based on the gravity direction height values ​​of all point cloud data points in the candidate road layer. If the height ranges of different candidate road layers overlap, or if the height difference in the height direction is less than a preset minimum height difference threshold (e.g., 1.5 meters), then it is determined that there is overlapping or partial overlap. This situation is more common in parallel sections of interchange ramps, auxiliary roads under bridges, or multi-layer structures with insufficient road elevation differences. Relying solely on height band segmentation can easily lead to point cloud data points from different layers being classified into the same layer.

[0094] After determining the existence of overlapping heights or local overlaps, inter-layer separation is performed based on the horizontal planar distribution of point cloud data points. Specifically, the point cloud data of the candidate road layer is projected onto a horizontal plane, and the horizontal distribution range of each point cloud data point relative to the candidate layer's centerline or geometric center is calculated. A planar clustering algorithm is used to divide points with significant differences in horizontal distribution into different sub-clusters; the planar clustering algorithm could be, for example, region growing based on Euclidean distance or DBSCAN. For point cloud data points with overlapping horizontal distributions but height differences close to the minimum height difference threshold between layers, further discrimination is performed based on their normal vectors and curvature indicators: points with large differences in normal vectors or significant differences in curvature characteristics are assigned to different candidate road layers.

[0095] Through the above-mentioned joint analysis of horizontal position distribution and height difference, the candidate road layers that originally overlapped or partially overlapped can be effectively separated into independent layers, and a unique candidate road layer number is assigned to each separated candidate road layer. This ensures that the layer identification map still has clear inter-layer boundaries and spatial consistency in complex road scenarios, thereby supporting the stable application of the technical solution of this application in multi-layer road environments.

[0096] It should be noted that the height segmentation threshold is used to divide the traffic area mask into zones according to the height value in the direction of gravity, with different height zones corresponding to different road levels. For example, in this application, 1.5 meters can be selected as the height segmentation threshold, that is, when the center height difference between adjacent height zones is greater than or equal to 1.5 meters, they are determined to be different road levels; the height segmentation threshold is applicable to common urban interchanges and overpass / underpass scenarios, and can effectively distinguish between upper and lower road surfaces.

[0097] The normal vector similarity threshold is used to determine whether the normal vectors of different point cloud data points are close enough to be classified into the same local plane. For example, in this application, the cosine similarity of the normal vectors can be used for measurement, and the threshold is set to 0.95. That is, when the cosine similarity of the normal vectors of two points is greater than or equal to 0.95, their local plane orientations are considered to be consistent. The normal vector similarity threshold can ensure the stability of planar clustering within the same layer and prevent directional interference between different layers.

[0098] The height variance threshold is used to evaluate the height consistency of point cloud data points within the same cluster. In a specific example, the height variance threshold is set to 0.10 m², meaning that when the height variance within a cluster is less than or equal to 0.10 m², the cluster is considered to have good height consistency and can be used as a candidate road layer. The height variance threshold can filter out non-planar areas with significant height fluctuations caused by road surface damage, vehicle obstruction, or acquisition noise.

[0099] The candidate centerline extraction module extracts the passable area within the layer based on the passable road domain mask and layer identifier map. It obtains a high-confidence skeleton through non-maximum suppression and three-dimensional refinement, and generates a set of candidate road centerlines through connectivity clustering and length constraints.

[0100] The method for extracting the candidate road centerline set includes:

[0101] Based on the traffic area mask and layer identifier map, the passable area within each candidate road layer number is extracted; for the spatial boundary points of the passable area within the layer, a three-dimensional contour tracking algorithm is used to construct a set of boundary patches.

[0102] Based on the passable region within the layer and the corresponding set of boundary patches, the three-dimensional Euclidean distance field and the corresponding distance field gradient are calculated; based on the distance field gradient, non-maximum suppression operation is performed on the three-dimensional Euclidean distance field along the gradient direction to obtain a skeleton candidate set, and three-dimensional thinning processing is performed on the passable region within the layer to obtain the morphological skeleton.

[0103] Point cloud data points that exist simultaneously in the candidate skeleton set and the morphological skeleton are constructed into a high-confidence skeleton point set. The interconnected skeleton points in the high-confidence skeleton point set are clustered into continuous skeleton line segments using a three-dimensional connectivity search algorithm. Length constraint processing is performed on each skeleton line segment, and skeleton line segments with lengths lower than the preset minimum skeleton length threshold are removed to eliminate false road centerlines caused by breaks, offsets or isolated small areas due to local noise.

[0104] All retained skeleton line segments are aggregated to construct a candidate road centerline set, thereby providing a highly reliable input for the accurate generation of subsequent road centerlines.

[0105] It should be noted that the method for obtaining the 3D Euclidean distance field includes: given the known traversable regions within a layer and their corresponding set of boundary patches, the traversable regions within the layer are voxels and constructed into a unified 3D raster coordinate system; for each voxel, the Euclidean distance from the center of the voxel to the nearest boundary patch is calculated, forming a distance value distribution in voxel units; the Euclidean distances of all voxels together constitute the 3D Euclidean distance field in 3D space. The 3D Euclidean distance field can reflect the nearest spatial distance from any voxel to the boundary of the traversable region, providing a metric basis for subsequent skeleton extraction.

[0106] The method for obtaining the distance field gradient includes: after obtaining the three-dimensional Euclidean distance field, using a three-dimensional difference operator to calculate the spatial first-order partial derivatives of the distance field in the horizontal, vertical, and longitudinal directions, respectively, to obtain the gradient vector corresponding to each voxel unit; the three-dimensional difference operator is, for example, the three-dimensional Sobel operator or the three-dimensional central difference. The direction of the gradient vector points to the direction of the fastest change in distance value, that is, the direction of the nearest boundary, and the magnitude of the gradient vector reflects the rate of distance change. Using the distance field gradient, non-maximum suppression can be performed along the gradient direction in subsequent steps, thereby extracting candidate skeleton points with geometrical center significance.

[0107] The method for obtaining the skeleton candidate set includes: after obtaining the three-dimensional Euclidean distance field and its corresponding distance field gradient, searching along both the positive and negative directions of the gradient vector at each voxel position, comparing the distance value of the current position with the distance values ​​of adjacent positions in that direction; if the distance value of the current position reaches a local maximum in the gradient direction, i.e., is greater than the distance values ​​of adjacent positions, then the position is determined to be a potential skeleton point and added to the skeleton candidate set. By performing the above-mentioned non-maximum suppression operation along the gradient direction on the entire distance field, redundant points at non-central positions can be removed, and skeleton candidate points located at local maxima and having geometrical central significance can be retained, thereby forming the skeleton candidate set.

[0108] The method for obtaining the morphological skeleton includes: for voxel units in the passable regions within a layer, a morphology-based 3D thinning algorithm is used to remove boundary voxels that do not affect connectivity layer by layer until the remaining voxel set cannot be further shrunk. The result is the morphological skeleton. The 3D thinning algorithm can be, for example, a 3D iterative erosion-dilation algorithm or a 3D topology-preserving thinning algorithm. The morphological skeleton can preserve the overall topological structure of the passable regions within a layer while eliminating boundary redundancy.

[0109] Furthermore, in the technical solution of this application, the purpose of first obtaining the candidate road centerline set is to provide a high-confidence geometric basis for the subsequent accurate road centerline generation and verification. By combining multiple information sources such as boundary patches, three-dimensional Euclidean distance fields, and morphological skeletons under the constraints of the traffic area mask and layer identifier map, the candidate road centerlines extracted can effectively preserve the spatial center trend of the main road structure and eliminate most false road centerlines caused by noise, local missing points, or outliers. Using the candidate road centerline set as an intermediate result reduces the search space and computational complexity of subsequent processing steps, and improves the stability and accuracy of the final road centerline fitting and optimization, thereby ensuring the continuity and consistency of road geometric features globally.

[0110] The centerline verification module performs comprehensive anomaly detection processing based on the geometric, topological, and semantic features corresponding to the candidate road centerline set to obtain the road centerline anomaly type corresponding to each candidate road centerline. The road centerline anomaly type includes geometric anomaly type, topological anomaly type, and semantic anomaly type.

[0111] like Figure 5 As shown, the methods for determining geometric anomaly types include:

[0112] The spatial geometric information of candidate road centerlines and corresponding road boundaries is obtained, and the road boundaries are extracted based on the set of boundary patches of passable areas within the layer.

[0113] The candidate road centerline is discretized into B sequential positioning points according to a preset first spacing. For each sequential positioning point, the shortest Euclidean distance from the sequential positioning point to one side boundary of the road is calculated and denoted as the first Euclidean distance. The shortest Euclidean distance from the sequential positioning point to the other side boundary of the road is calculated and denoted as the second Euclidean distance. The difference between the first Euclidean distance and the second Euclidean distance is taken as the absolute value to obtain the bidirectional distance residual corresponding to the sequential positioning point. The sum of the first Euclidean distance and the second Euclidean distance is used to obtain the local road width corresponding to the sequential positioning point. For example, in this application, the first spacing can be set to 20cm.

[0114] Construct a set of two-way distance residuals from B bidirectional distance residuals; construct a set of local road widths from B local road widths;

[0115] The symmetry index is calculated based on the bidirectional distance residual set and the local road width set; the road width consistency index is calculated based on the local road width set.

[0116] If the symmetry index is less than a preset symmetry index threshold, a geometric offset anomaly is added to the geometric anomaly types of the candidate road centerline; if the road width consistency index is less than a preset road width consistency index threshold, a width anomaly is added to the geometric anomaly types of the candidate road centerline. For example, in this application, the symmetry index threshold and the road width consistency index threshold can be set to 0.75.

[0117] The method for calculating the symmetry index includes:

[0118] ;

[0119] in, The symmetry index has a range of values ​​of [value missing]. ; For the index variable of the summation formula, Indicates the first Two-way distance residuals, This represents the width of the b-th local road. Indicates will The results obtained are truncated at The closer the symmetry index is to 1, the more balanced the distance between the two sides of the candidate road centerline and the more centered the geometry; the closer the symmetry index is to 0, the more obvious the overall eccentricity of the candidate road centerline.

[0120] The calculation method for the road width consistency index includes:

[0121] ;

[0122] in, As an indicator of road width consistency, Represents the set of local road widths The average width of local roads, This represents the standard deviation of the local road widths of the set of local road widths. Indicates will The results obtained are truncated at .

[0123] like Figure 6 As shown, the methods for determining topological anomaly types include:

[0124] The candidate road centerline is discretized into a sequence of line elements according to a preset second spacing. The spacing between each adjacent line element is calculated based on spatial adjacency. When the spacing between any adjacent line element is greater than a preset breakage threshold, the candidate road centerline is determined to have a breakage anomaly, and a breakage anomaly is added to the topological anomaly type of the candidate road centerline. For example, in this application, the second spacing can be set to 20cm.

[0125] A global intersection search is performed on the candidate road centerlines in three-dimensional space. When a candidate road centerline is detected to have a geometric intersection at a non-adjacent line element and the distance is less than a preset self-intersection judgment threshold, it is determined that the candidate road centerline has a self-intersection anomaly, and a self-intersection anomaly is added to the topological anomaly type of the candidate road centerline.

[0126] If adjacent line elements of the same candidate road centerline are located in different road layers and the distance between adjacent line elements is less than the preset cross-layer determination threshold, it is determined that the candidate road centerline has a cross-layer short circuit anomaly, and a short circuit anomaly is added to the topology anomaly type of the candidate road centerline.

[0127] It should be noted that, for example, in a preferred embodiment of this application, the breakage threshold can be set to 0.5m. This breakage threshold is used to limit the maximum permissible spatial interval between adjacent skeleton line elements in the candidate road centerline. When the three-dimensional spatial distance between adjacent line elements is greater than the breakage threshold, a breakage anomaly is determined to exist at that location. The breakage threshold can be adapted and adjusted according to different map precision and road width characteristics.

[0128] The self-intersection threshold can be set to 0.3m. This threshold limits the minimum distance between different segments of the same candidate road centerline on the spatial projection plane. When the spatial distance between non-adjacent segments is less than the self-intersection threshold and there is an intersection topology relationship, a self-intersection anomaly is determined to exist at that location. The selection of the self-intersection threshold can be optimized by combining the road centerline extraction accuracy and the minimum lane width.

[0129] The cross-layer determination threshold can be set to 1.0m. This threshold is used to limit the minimum reasonable distance between different road layers in three-dimensional space. When adjacent line elements in the candidate road centerline belong to different road layers and their three-dimensional spatial distance is lower than the cross-layer determination threshold, a cross-layer short-circuit anomaly is determined to exist at that location. The cross-layer determination threshold can be set in conjunction with the inter-layer safety distance of elevated bridges, tunnels, and other three-dimensional transportation structures.

[0130] In this application, identifying self-intersection anomalies, break anomalies, and cross-layer short-circuit anomalies serves to perform geometric and topological quality control on the candidate road centerline set, ensuring the continuity, uniqueness, and road layer consistency of the generated road centerline results. Specifically, detecting self-intersection anomalies prevents unreasonable spatial intersections of road centerlines, thus preventing incorrect routes from being generated in navigation paths; detecting break anomalies promptly identifies gaps in road centerlines, facilitating repair through interpolation or re-extraction of skeleton line segments, improving road connectivity; and detecting cross-layer short-circuit anomalies effectively prevents unreasonable direct connections between different road layers, ensuring the correctness of the road topology in three-dimensional traffic scenarios. By identifying these three types of anomalies—self-intersection anomalies, break anomalies, and cross-layer short-circuit anomalies—the geometric accuracy and topological consistency of the road centerline data generated in this application can be significantly improved.

[0131] like Figure 7 As shown, the methods for determining semantic anomaly types include:

[0132] A pre-constructed road semantic feature set is provided; the road semantic feature set includes road level information and road lane number information; the road level information includes expressways, urban expressways, arterial roads, secondary arterial roads and local roads; the road semantic feature set is derived from map databases or historical high-precision map data.

[0133] The spatial position of the candidate road centerline is spatially superimposed with the real-time collected road vector data or road surface data to extract the road object with the highest spatial overlap with the candidate road centerline, and the road grade and number of lanes corresponding to the road object are obtained.

[0134] If the road level of the road object is different from the road level corresponding to the road semantic feature set, it is determined that the candidate road centerline has a road level mismatch; road level mismatch is added to the semantic anomaly type of the candidate road centerline.

[0135] If the number of lanes of a road object is different from the number of lanes corresponding to the road semantic feature set, it is determined that the candidate road centerline has a lane number mismatch; lane number mismatch is added to the semantic anomaly type of the candidate road centerline.

[0136] It should be noted that in this application, identifying discrepancies in road class and lane number plays a crucial role in ensuring the accuracy of road centerline results and the consistency of road semantics. Specifically, identifying discrepancies in road class can effectively detect differences in class attributes between candidate road centerlines and actual roads. For example, if the target road level of the candidate road centerline obtained based on skeleton extraction is the highway level, but the spatially matched road object is an urban branch road, it can be determined as a road class discrepancy. Identifying road class discrepancies can avoid erroneous class assignments caused by incorrect road layer identification, skeleton extraction offset, or data source defects, preventing mismatches in traffic rule matching, navigation planning, and traffic control strategy levels in subsequent practical applications.

[0137] Lane count discrepancies can be identified by detecting differences between the candidate road centerline and the actual road in terms of lane capacity description. For example, if a candidate road centerline is expected to be associated with two lanes, but the matched road object attributes show four lanes, this indicates an anomaly in lane count discrepancy. This type of identification can effectively avoid distortion of capacity assessment caused by incorrect lane counts, preventing unreasonable planning decisions in traffic flow simulation, right-of-way allocation, or road reconstruction design.

[0138] In summary, identifying discrepancies in road grade and lane number can serve as a core step in the semantic quality control of candidate road centerlines. On the one hand, it provides a direct basis for the correction, replacement, or removal of abnormal centerlines; on the other hand, it enhances the ability of the final generated road centerline data to ensure consistency in road semantics, accuracy in traffic simulation, and reliability in engineering applications. This ensures that the road centerline results generated in this application not only have high geometric accuracy but also completeness and correctness in road attribute description.

[0139] The centerline hierarchical correction module establishes an anomaly priority queue based on the anomaly types of the candidate road centerlines. According to the anomaly priority queue, the candidate road centerlines are divided into batch repair sets and single-point repair sets according to their anomaly types. Geometric morphology adjustment, topology reconstruction, and semantic attribute update are performed on the batch repair sets and single-point repair sets respectively to obtain the corrected road centerline sets.

[0140] like Figure 3 As shown, the methods for establishing an exception priority queue include:

[0141] A pre-constructed anomaly impact weight table is used to define the degree of impact of different anomaly types on road centerline geometric accuracy, topological connectivity, and semantic consistency. The anomaly impact weight table is set according to the degree of impact of anomaly type on road geometric accuracy, connectivity, and semantic accuracy, with anomaly types that affect driving safety and road recognition accuracy being assigned higher weights.

[0142] Based on the aforementioned abnormal impact weight table, weights are assigned to each abnormal type in the road centerline abnormality type to obtain the abnormal type weight corresponding to each abnormal type;

[0143] For each candidate road centerline in the candidate road centerline set, the weights of the anomaly types contained in the candidate road centerline are accumulated to obtain the comprehensive anomaly impact score of the candidate road centerline.

[0144] An anomaly comprehensive impact score set is constructed from the anomaly comprehensive impact scores corresponding to all candidate road centerlines; the anomaly comprehensive impact score set is sorted in descending order to obtain the anomaly priority queue.

[0145] For example, the weighting table of anomalies is shown in Table 1.

[0146] Table 1 Weighting Table of Abnormal Impacts

[0147]

[0148] The method for classifying candidate road centerline anomaly types into batch repair sets and single-point repair sets based on an anomaly priority queue includes:

[0149] For each candidate road centerline in the anomaly priority queue, the anomaly type, anomaly type weight, and corresponding comprehensive anomaly impact score of the candidate road centerline are input into the anomaly triage model to obtain the batch repair set and single-point repair set corresponding to each candidate road centerline.

[0150] The training method for the abnormal traffic splitting model includes:

[0151] An anomaly diversion dataset is pre-constructed, comprising FL group anomaly diversion data and corresponding batch repair sets and single-point repair sets, where FL is a positive integer; the anomaly diversion data includes the road centerline anomaly type, anomaly type weight, and anomaly comprehensive impact score; the anomaly diversion dataset is divided into a training set and a validation set, the training set is used for learning the parameters of the anomaly diversion model, and the validation set is used for real-time monitoring of the generalization performance and overfitting degree of the anomaly diversion model;

[0152] A deep neural network with a multilayer perceptron as its core is used as the anomaly offshoot model. Anomaly offshoot data is standardized and vectorized before being input into the deep neural network, which consists of an input layer, hidden layers, and an output layer. Each hidden layer extracts features using a non-linear activation function, and the output layer uses a softmax activation function to obtain the probability distributions corresponding to each batch repair set and single-point repair set. Finally, the batch repair set and single-point repair set corresponding to the highest probability are taken as the prediction results of the anomaly offshoot model. During training, the cross-entropy loss function is used as the optimization objective, and a gradient descent-type optimization algorithm is used to update the network weights. An early stopping strategy is implemented: when the prediction accuracy on the validation set reaches or exceeds a prediction accuracy threshold, the anomaly offshoot model is considered to have converged and training is terminated. For example, the prediction accuracy threshold can be set to 95% in this application.

[0153] It should be noted that the anomaly type of the road centerline reflects the nature of the anomaly and the complexity of its repair; the anomaly type weight quantifies the degree of impact of the anomaly on the overall geometric, topological, or semantic consistency; and the comprehensive anomaly impact score integrates the anomaly distribution and impact range of the candidate road centerline, reflecting the urgency and global importance of repair. When the road centerline anomaly type, anomaly type weight, and comprehensive anomaly impact score are used as inputs to the anomaly diversion model, the model can automatically determine whether the anomaly of the candidate road centerline is more suitable for centralized batch repair (i.e., consistency adjustment of adjacent road segments) or independent single-point repair (i.e., local mutation correction) based on feature patterns, thereby achieving dynamic partitioning without the need for manual rule configuration.

[0154] By dividing the repair into batch and single-point repair sets, static classification methods based on a single threshold or fixed rules can be avoided, improving the adaptability and accuracy of the repair strategy. The anomaly triage model can flexibly select the optimal repair path by combining the actual road data distribution and anomaly characteristic changes. This improves the efficiency of batch repair when anomalies are concentrated, and reduces unnecessary global adjustment overhead when anomalies are scattered. Furthermore, it maintains stable repair quality and processing efficiency in different scenarios.

[0155] like Figure 4 As shown, the methods for obtaining the corrected set of road centerlines include:

[0156] For each candidate road centerline in the anomaly priority queue, if the candidate road centerline contains a batch repair set, the spatial location and anomaly type weight corresponding to each anomaly type in the batch repair set are extracted and constructed into corresponding batch anomaly nodes; the batch anomaly nodes are assembled into a batch repair subgraph, and the batch repair subgraph is input into the first parameter setting model to obtain the first parameter set;

[0157] If the candidate road centerline contains a single-point repair set, then extract the spatial location and anomaly type weight corresponding to each anomaly type in the single-point repair set, and construct the corresponding single-point anomaly data; input each single-point anomaly data into the second parameter setting model to obtain the second parameter set;

[0158] Based on the first parameter set and the second parameter set, the candidate road centerline is geometrically adjusted, topologically reconstructed, and semantically updated to obtain the corrected road centerline.

[0159] All corrected road centerlines are combined into a corrected road centerline set.

[0160] The training methods for the first parameter setting model include:

[0161] A first parameter setting dataset is pre-constructed, which includes first parameter setting data of group DY and the first parameter set corresponding to the first parameter setting data of group DY, where DY is a positive integer; the first parameter setting data includes batch repair sub-images; the first parameter setting dataset is divided into a training set and a validation set, the training set is used for learning the parameters of the first parameter setting model, and the validation set is used to monitor the generalization performance and overfitting degree of the first parameter setting model in real time.

[0162] A node-output graph neural network architecture is adopted as the first parameter setting model. The first parameter setting data is standardized and vectorized before being input into the node-output graph neural network architecture, which consists of an input layer, hidden layers, and an output layer. Each hidden layer uses a non-linear activation function to extract features, and the output layer uses a softmax activation function to obtain the probability distribution corresponding to each first parameter set. Finally, the first parameter set corresponding to the highest probability is taken as the prediction result of the first parameter setting model. During training, the cross-entropy loss function is used as the optimization objective, and a gradient descent-type optimization algorithm is used to update the network weights. An early stopping strategy is set: when the prediction accuracy on the validation set reaches or exceeds a preset threshold, the first parameter setting model is determined to have converged and training is terminated.

[0163] The training methods for the second parameter setting model include:

[0164] A second parameter setting dataset is pre-constructed, which includes the second parameter setting data of group DE and the second parameter set corresponding to the second parameter setting data of group DE, where DE is a positive integer; the second parameter setting data includes single-point outlier data; the second parameter setting dataset is divided into a training set and a validation set, the training set is used for learning the parameters of the second parameter setting model, and the validation set is used to monitor the generalization performance and overfitting of the second parameter setting model in real time.

[0165] A deep neural network with a multilayer perceptron as its core is used as the second parameter setting model. The second parameter setting data is standardized and vectorized before being input into the deep neural network, which consists of an input layer, hidden layers, and an output layer. Each hidden layer uses a nonlinear activation function to extract features, and the output layer uses a softmax activation function to obtain the probability distribution corresponding to each second parameter set. Finally, the second parameter set corresponding to the highest probability is taken as the prediction result of the second parameter setting model. During training, the cross-entropy loss function is used as the optimization objective, and a gradient descent-type optimization algorithm is used to update the network weights. An early stopping strategy is set: when the prediction accuracy on the validation set reaches or exceeds the prediction accuracy threshold, the second parameter setting model is determined to have converged and training is terminated.

[0166] The first parameter set includes geometric adjustment parameters, topology optimization parameters, and semantic attribute consistency parameters. The geometric adjustment parameters include the lateral displacement of the road centerline, the longitudinal displacement of the road centerline, and the curvature smoothing coefficient. Specifically, the lateral and longitudinal displacements of the road centerline are used to control the overall offset of a batch of abnormal nodes in the planar coordinate system along the lateral and longitudinal directions, respectively, to correct spatial misalignment caused by acquisition errors or accumulated offsets. The curvature smoothing coefficient is used to constrain the curvature change range of the path of the batch of nodes. By adjusting the weight of the curve control points, the smoothing of a large range of curves is achieved, eliminating geometric abrupt changes and ensuring a natural transition.

[0167] The topology optimization parameters include node connection reconstruction weights and break merging thresholds. Specifically, the node connection reconstruction weights are used to measure the necessity and priority of rebuilding connection edges between adjacent batch of abnormal nodes. When the connection reconstruction weight is higher than the preset connection reconstruction weight threshold, for example, 0.30, new connection edges are automatically established or existing connection edges are replaced to improve the overall path coherence. The break merging threshold is used to determine whether to perform a merging operation on topology break segments. When the spatial distance, directional deviation, and semantic consistency between break points all meet the break merging threshold, connection and merging are automatically performed to eliminate topology breakage and reduce redundant branches.

[0168] The semantic attribute consistency parameters include functional label unification rules, attribute conflict priority weights, and node semantic inheritance ratios. Specifically, functional label unification rules are used to standardize core semantic labels such as road function type and lane number during batch repair to avoid semantic inconsistencies between adjacent road segments. Attribute conflict priority weights are used to determine the order of selection of conflicting attributes during multi-source data fusion to ensure the consistency of the final semantic results. Node semantic inheritance ratios are used to control the degree to which batch of abnormal nodes inherit semantic information from the original nodes after adjustment, to ensure the continuity and reliability of semantic information while optimizing the geometric and topological structures.

[0169] The first parameter set includes single-point geometric adjustment offset, single-point smoothing radius, node connection weight threshold, direction correction angle, and semantic attribute update priority. Specifically, the single-point geometric adjustment offset is used to determine the fine-tuning magnitude of the single-point anomaly's position in the planar coordinate system. By applying offset correction based on the original position, the node can be accurately aligned to the expected position. The single-point smoothing radius is used to constrain the curve smoothness within the range around the node, eliminating abrupt changes or sharp corners during correction. The node connection weight threshold is used to filter the retention or reconstruction conditions of the node in the topological relationship. When the connection weight with adjacent nodes is lower than the node connection weight threshold, disconnection or replacement connection operations are performed to improve the rationality of the local structure. The direction correction angle is used to limit the direction adjustment range of the path segment to which the node belongs. When the angle with adjacent paths exceeds a set angle value, such as 15°, the direction is corrected by rotating or fine-tuning the local geometric vector. The semantic attribute update priority is used to determine the order in which nodes update semantic information such as road grade, number of lanes, and function type, ensuring that key semantics are matched first.

[0170] Example 2:

[0171] Please see Figure 2 As shown, this embodiment provides a method for determining the centerline of a path based on neural networks and point cloud spatial features, including:

[0172] The original 3D point cloud data of N overlapping roads is processed to obtain the traffic area mask and layer label map;

[0173] Based on the traffic area mask and layer identifier map, the passable area within the layer is extracted. A high-confidence skeleton is obtained through non-maximum suppression and three-dimensional refinement. A set of candidate road centerlines is generated through connectivity clustering and length constraints.

[0174] Based on the geometric, topological, and semantic features corresponding to the candidate road centerline set, a comprehensive anomaly detection process is performed to obtain the road centerline anomaly type corresponding to each candidate road centerline. The road centerline anomaly type includes geometric anomaly type, topological anomaly type, and semantic anomaly type.

[0175] An anomaly priority queue is established based on the anomaly types of the candidate road centerlines. The anomaly types of the candidate road centerlines are then divided into a batch repair set and a single-point repair set according to the anomaly priority queue. Geometric morphology adjustment, topology reconstruction, and semantic attribute update are performed on the batch repair set and the single-point repair set, respectively, to obtain the corrected road centerline set.

[0176] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0177] In conclusion, 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 determining the centerline of a path based on neural networks and point cloud spatial features, characterized in that, include: The original 3D point cloud data of N overlapping roads is processed to obtain the traffic area mask and layer label map; Based on the traffic area mask and layer identifier map, the passable area within the layer is extracted. A high-confidence skeleton is obtained through non-maximum suppression and three-dimensional refinement. A set of candidate road centerlines is generated through connectivity clustering and length constraints. Based on the geometric, topological, and semantic features corresponding to the candidate road centerline set, a comprehensive anomaly detection process is performed to obtain the road centerline anomaly type corresponding to each candidate road centerline. The road centerline anomaly type includes geometric anomaly type, topological anomaly type, and semantic anomaly type. An anomaly priority queue is established based on the anomaly types of the candidate road centerlines. The anomaly types of the candidate road centerlines are then divided into a batch repair set and a single-point repair set according to the anomaly priority queue. Geometric morphology adjustment, topology reconstruction, and semantic attribute update are performed on both batch repair sets and single-point repair sets to obtain the corrected path centerline set; The methods for obtaining the modified path centerline set include: For each candidate road centerline in the anomaly priority queue, if the candidate road centerline contains a batch repair set, the spatial location and anomaly type weight corresponding to each anomaly type in the batch repair set are extracted and constructed into corresponding batch anomaly nodes; the batch anomaly nodes are assembled into a batch repair subgraph, and the batch repair subgraph is input into the first parameter setting model to obtain the first parameter set; If the candidate road centerline contains a single-point repair set, then extract the spatial location and anomaly type weight corresponding to each anomaly type in the single-point repair set, and construct the corresponding single-point anomaly data; input each single-point anomaly data into the second parameter setting model to obtain the second parameter set; Based on the first parameter set and the second parameter set, the candidate road centerline is geometrically adjusted, topologically reconstructed, and semantically updated to obtain the corrected road centerline. All corrected road centerlines are combined into a corrected road centerline set.

2. The method for determining the centerline based on neural networks and point cloud spatial features according to claim 1, characterized in that, The methods for establishing an exception priority queue include: Pre-construct an anomaly impact weight table; Based on the aforementioned abnormal impact weight table, weights are assigned to each abnormal type in the road centerline abnormality type to obtain the abnormal type weight corresponding to each abnormal type; For each candidate road centerline in the candidate road centerline set, the weights of the anomaly types contained in the candidate road centerline are accumulated to obtain the comprehensive anomaly impact score of the candidate road centerline. An anomaly comprehensive impact score set is constructed from the anomaly comprehensive impact scores corresponding to all candidate road centerlines; the anomaly comprehensive impact score set is sorted in descending order to obtain the anomaly priority queue.

3. The method for determining the centerline based on neural networks and point cloud spatial features according to claim 1, characterized in that, The method for classifying candidate road centerline anomaly types into batch repair sets and single-point repair sets based on an anomaly priority queue includes: For each candidate road centerline in the anomaly priority queue, the anomaly type, anomaly type weight, and corresponding comprehensive anomaly impact score of the candidate road centerline are input into the anomaly diversion model to obtain the batch repair set and single-point repair set corresponding to each candidate road centerline; the anomaly diversion model is a deep neural network with the input layer, hidden layer and output layer as the core.

4. The method for determining the centerline based on neural networks and point cloud spatial features according to claim 1, characterized in that, Methods for determining geometric anomaly types include: The candidate road centerline is discretized into B sequential positioning points according to a preset first spacing. For each sequential positioning point, the shortest Euclidean distance from the sequential positioning point to one side boundary of the road is calculated and denoted as the first Euclidean distance. The shortest Euclidean distance from the sequential positioning point to the other side boundary of the road is calculated and denoted as the second Euclidean distance. The difference between the first Euclidean distance and the second Euclidean distance is taken as the absolute value to obtain the bidirectional distance residual corresponding to the sequential positioning point. The sum of the first Euclidean distance and the second Euclidean distance is used to obtain the local road width corresponding to the sequential positioning point. Construct a set of two-way distance residuals from B bidirectional distance residuals; construct a set of local road widths from B local road widths; The symmetry index is calculated based on the bidirectional distance residual set and the local road width set; the road width consistency index is calculated based on the local road width set. If the symmetry index is less than the preset symmetry index threshold, then add a geometric offset anomaly to the geometric anomaly type of the candidate road centerline; if the road width consistency index is less than the preset road width consistency index threshold, then add a width anomaly to the geometric anomaly type of the candidate road centerline.

5. The method for determining the centerline based on neural networks and point cloud spatial features according to claim 1, characterized in that, Methods for determining topological anomaly types include: The candidate road centerline is discretized into a sequence of line elements according to a preset second spacing, and the spacing between each adjacent line element is calculated based on the spatial adjacency relationship. When the spacing between any adjacent line element is greater than a preset breakage threshold, it is determined that the candidate road centerline has a breakage anomaly, and a breakage anomaly is added to the topological anomaly type of the candidate road centerline. A global intersection search is performed on the candidate road centerline in three-dimensional space. When a candidate road centerline is detected to have a geometric intersection at a non-adjacent line element and the distance is less than a preset self-intersection judgment threshold, it is determined that the candidate road centerline has a self-intersection anomaly, and a self-intersection anomaly is added to the topological anomaly type of the candidate road centerline. If adjacent line elements of the same candidate road centerline are located in different road layers and the distance between adjacent line elements is less than the preset cross-layer determination threshold, it is determined that the candidate road centerline has a cross-layer short circuit anomaly, and a short circuit anomaly is added to the topology anomaly type of the candidate road centerline.

6. The method for determining the centerline based on neural networks and point cloud spatial features according to claim 1, characterized in that, Methods for determining semantic anomaly types include: A pre-constructed road semantic feature set is provided; the road semantic feature set includes road level information and road lane number information; the road level information includes expressways, urban expressways, arterial roads, secondary arterial roads, and local roads. The spatial position of the candidate road centerline is spatially superimposed with the real-time collected road vector data or road surface data to extract the road object with the highest spatial overlap with the candidate road centerline, and the road grade and number of lanes corresponding to the road object are obtained. If the road level of the road object is different from the road level corresponding to the road semantic feature set, it is determined that the candidate road centerline has a road level mismatch; road level mismatch is added to the semantic anomaly type of the candidate road centerline. If the number of lanes of a road object is different from the number of lanes corresponding to the road semantic feature set, it is determined that the candidate road centerline has a lane number mismatch; lane number mismatch is added to the semantic anomaly type of the candidate road centerline.

7. The method for determining the centerline based on neural networks and point cloud spatial features according to claim 1, characterized in that, The method for extracting the candidate road centerline set includes: Based on the traffic area mask and layer identifier map, the passable area within each candidate road layer number is extracted; for the spatial boundary points of the passable area within the layer, a three-dimensional contour tracking algorithm is used to construct a set of boundary patches. Based on the passable region within the layer and the corresponding set of boundary patches, the three-dimensional Euclidean distance field and the corresponding distance field gradient are calculated; based on the distance field gradient, non-maximum suppression operation is performed on the three-dimensional Euclidean distance field along the gradient direction to obtain a skeleton candidate set, and three-dimensional thinning processing is performed on the passable region within the layer to obtain the morphological skeleton. Point cloud data points that exist simultaneously in the candidate skeleton set and the morphological skeleton are constructed into a high-confidence skeleton point set. The interconnected skeleton points in the high-confidence skeleton point set are clustered into continuous skeleton segments using a three-dimensional connectivity search algorithm. Length constraint processing is performed on each skeleton segment, and skeleton segments with a length lower than the preset minimum skeleton length threshold are removed. All retained skeleton line segments are aggregated to construct a set of candidate road centerlines.

8. The method for determining the centerline based on neural networks and point cloud spatial features according to claim 1, characterized in that, The method for obtaining the traffic area mask and layer identifier map includes: Standard point cloud data is obtained by preprocessing the original 3D point cloud data; For each point cloud data point in the standard point cloud data, point-level spatial features are extracted; the point-level spatial features include normal vector, curvature indicator, mean echo intensity and variance of echo intensity, and a neighborhood height histogram of the point cloud data point is statistically plotted along the gravity direction; the point-level spatial features of all point cloud data points in the standard point cloud data are constructed into a point-level spatial feature set. The point-level spatial feature set is input into the road domain probability assessment model to obtain the road domain probability score corresponding to each point cloud data point. Based on the preset road domain probability score threshold, the point cloud data points are labeled with road domain labels or non-road domain labels. By combining all road domain labels and non-road domain labels with the 3D spatial coordinates of the corresponding point cloud data points, a connected component is constructed to obtain a passable road domain mask. After combining all road domain labels and non-road domain labels with the 3D spatial coordinates of the corresponding point cloud data points, a connected component is constructed, and after connectivity analysis, region culling and hole filling, a spatially topologically continuous passable road domain mask is obtained. Layer identification maps are obtained by performing hierarchical recognition processing based on the traffic area mask.

9. The method for determining the centerline based on neural networks and point cloud spatial features according to claim 8, characterized in that, Methods for obtaining layer identification maps based on traffic area masks include: For each point cloud data point within the traffic area mask, the position, neighborhood height histogram, and absolute height value in the gravity direction of the point cloud data point are read, and height distribution characteristics are statistically formed. Based on the preset height segmentation threshold, the road area mask is divided into M height bands. Within each height band, the local plane characteristics are analyzed by combining the normal vector and curvature indicator. Point cloud data points with normal vector similarity higher than the corresponding threshold and height variance lower than the corresponding threshold are clustered into the same connected piece, which is used as a candidate road layer and assigned a connected piece number. For candidate road layers that overlap significantly or partially overlap, they are separated by using the horizontal position distribution and the minimum height difference between layers, and a candidate road layer number is assigned to each point cloud data point. Map all candidate road layer numbers to their corresponding 3D coordinates to generate a layer identifier map aligned with the traffic area mask space.

10. A path centerline determination system based on neural networks and point cloud spatial features, used to implement the path centerline determination method based on neural networks and point cloud spatial features as described in any one of claims 1-9, characterized in that, include: The layered road segmentation module is used to process the original 3D point cloud data of N layers of overlapping roads to obtain the traffic road mask and layer identifier map; The candidate centerline extraction module extracts the passable area within the layer based on the passable road domain mask and layer identifier map, obtains a high-confidence skeleton through non-maximum suppression and three-dimensional refinement, and generates a set of candidate road centerlines through connectivity clustering and length constraints. The centerline verification module performs comprehensive anomaly detection processing based on the geometric, topological, and semantic features corresponding to the candidate road centerline set to obtain the corresponding road centerline anomaly types, which include geometric anomaly types, topological anomaly types, and semantic anomaly types. The centerline hierarchical correction module establishes an anomaly priority queue based on the anomaly types of the candidate road centerlines. According to the anomaly priority queue, the anomaly types of the candidate road centerlines are divided into batch repair sets and single-point repair sets. Geometric morphology adjustment, topology reconstruction, and semantic attribute update are performed on both batch repair sets and single-point repair sets to obtain the corrected path centerline set.

Images