Neural network and point cloud spatial feature-based road center line determination method and system

Abstract

The invention discloses a neural network and point cloud spatial feature-based road center line determination method and system, and relates to the technical field of neural networks. The method comprises the steps of generating a passing road domain mask and a layer identification graph based on original three-dimensional point cloud data of N layers of overlapped roads, extracting passable areas in the layers, obtaining a high-confidence skeleton through non-maximum suppression and three-dimensional refinement, and generating a candidate road center line set through connectivity clustering and length constraint; performing comprehensive anomaly detection based on geometric, topological and semantic features, identifying the anomaly type of the road and center lines, establishing an anomaly priority queue, and dividing candidate road and center lines into a batch repair set and a single-point repair set; respectively executing geometric shape adjustment, topological structure reconstruction and semantic attribute updating on the two types of sets to obtain a corrected road center line set; according to the method, high-precision optimization and anomaly classification repair of geometry, topology and semantics of the road center line are realized, and the conflict between local optimization and global optimization of the road center line is solved.

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] The global intersection point search is performed on the candidate road centerline in three-dimensional space, when it is detected that the candidate road centerline has geometric intersection at non-adjacent line elements, and the interval is less than a preset self-intersection judgment threshold, it is judged that the candidate road centerline has self-intersection anomaly, and the self-intersection anomaly is added in the topological anomaly type of the candidate road centerline;

[0030] If the adjacent line elements of the same candidate road centerline are located in different road layers respectively, and the interval of the adjacent line elements is less than a preset cross-layer judgment threshold, it is judged that the candidate road centerline has cross-layer short-circuit anomaly, and the short-circuit anomaly is added in the topological anomaly type of the candidate road centerline.

[0031] Further, the judgment method of the semantic anomaly type comprises:

[0032] A pre-constructed road semantic feature set is constructed; the road semantic feature set comprises road grade information and road lane number information; the road grade information comprises expressway, urban expressway, main road, secondary road and branch road;

[0033] The spatial position of the candidate road centerline is subjected to spatial overlay operation with the real-time collected road vector data or road surface data, the road object with the highest spatial coincidence degree with the candidate road centerline is extracted, and the road grade and lane number corresponding to the road object are obtained;

[0034] If the road grade of the road object is different from the corresponding road grade in the road semantic feature set, it is judged that the candidate road centerline has road grade inconsistency; the road grade inconsistency is added in the semantic anomaly type of the candidate road centerline;

[0035] If the lane number of the road object is different from the corresponding lane number in the road semantic feature set, it is judged that the candidate road centerline has lane number inconsistency; the lane number inconsistency is added in the semantic anomaly type of the candidate road centerline.

[0036] Further, the extraction method of the candidate road centerline set comprises:

[0037] Based on the passable road domain mask and the layer identification map, the passable area in each layer corresponding to each candidate road layer number is extracted; for the spatial boundary points of the passable area in the layer, a three-dimensional contour tracking algorithm is used to construct a boundary surface patch set;

[0038] Based on the passable area in the layer and the corresponding boundary surface patch set, a three-dimensional Euclidean distance field and a corresponding distance field gradient are calculated; the non-maximum suppression operation along the gradient direction is performed on the three-dimensional Euclidean distance field based on the distance field gradient, to obtain a skeleton candidate set, and a three-dimensional thinning process is performed on the passable area in the layer to obtain a morphological skeleton;

[0039] The point cloud data points existing in the skeleton candidate set and the morphological skeleton are constructed into a high-confidence skeleton point set, and a three-dimensional connectivity search algorithm is used to cluster the skeleton points connected to each other in the high-confidence skeleton point set into continuous skeleton line segments; length constraint processing is performed on each skeleton line segment, and skeleton line segments with a length lower than a preset minimum skeleton length threshold are removed;

[0040] All retained skeleton line segments are collected and constructed into a candidate road centerline set.

[0041] Further, the method for obtaining the passable road domain mask and the layer identification map comprises:

[0042] The original three-dimensional point cloud data is preprocessed to obtain standard point cloud data;

[0043] For each point cloud data point in the standard point cloud data, a point-level spatial feature is extracted; the point-level spatial feature includes a normal vector, a curvature indicator, an echo intensity mean value, and an echo intensity variance, and a neighborhood height histogram of the point cloud data point along the gravity direction is counted; 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 a road domain probability evaluation model to obtain a road domain probability score corresponding to each point cloud data point, and a road domain label or a non-road domain label is labeled for the point cloud data point according to a preset road domain probability score threshold;

[0045] All road domain labels and non-road domain labels are combined with the three-dimensional spatial coordinates of the corresponding point cloud data points to construct a connected domain, and a passable road domain mask is obtained; all road domain labels and non-road domain labels are combined with the three-dimensional spatial coordinates of the corresponding point cloud data points to construct a connected domain, and a passable road domain mask with a spatial topological continuity is obtained through connectivity analysis, region removal, and hole filling processing;

[0046] Layer identification processing is performed based on the passable road domain mask to obtain a layer identification map.

[0047] Further, the method for obtaining a layer identification map based on the passable road domain mask comprises:

[0048] For each point cloud data point in the passable road domain mask, the position, neighborhood height histogram, and absolute height value in the gravity direction of the point cloud data point are read, and a height distribution feature is counted;

[0049] According to a preset height segmentation threshold, the passable road domain mask is divided into M height bands, and the local plane characteristics are analyzed in each height band in combination with the normal vector and the curvature indicator; the point cloud data points with a high normal vector similarity higher than a corresponding threshold and a low height variance lower than a corresponding threshold are clustered into the same connected piece as a candidate road layer and are assigned a connected piece number;

[0050] For the highly overlapped or partially coincident candidate road layers, separate by horizontal position distribution and minimum height difference between layers, and determine the candidate road layer number for each point cloud data point;

[0051] Map all candidate road layer numbers and corresponding three-dimensional coordinates to generate a layer identification map spatially aligned with the passable road mask.

[0052] The road centerline determination system based on neural network and point cloud spatial features is used to implement the road centerline determination method based on neural network and point cloud spatial features, comprising:

[0053] The hierarchical road area segmentation module is used to process the original three-dimensional point cloud data of N layers of overlapping roads to obtain a passable road mask and a layer identification map;

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

[0055] The centerline checking 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 type, which includes geometric anomaly type, topological anomaly type and semantic anomaly type;

[0056] The centerline hierarchical correction module establishes an abnormal priority queue based on the road centerline anomaly type corresponding to the candidate road centerline set; according to the abnormal priority queue, the road centerline anomaly type of 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 update are performed on the batch repair set and the single-point repair set respectively to obtain a corrected road centerline set.

[0057] Compared with the prior art, the road centerline determination method and system based on neural network and point cloud spatial features has the following technical effects and advantages:

[0058] The present application can realize high-precision and automatic road centerline extraction and anomaly repair in large-scale point cloud data. By introducing the double-set repair strategy of batch repair set and single-point repair set, the technical scheme of the present application establishes a parameterized adjustment mechanism in the geometric, topological and semantic dimensions, improves the integrity and consistency of the road centerline data from the source, and solves the conflict between local optimization and global optimization of the road centerline. Compared with the prior art, the present application not only has higher automation and adaptability, but also can reduce manual intervention and significantly improve the road network data processing efficiency.

[0059] Further, the present application uses the first parameter set to realize efficient global repair of batch abnormalities, covering geometric adjustment parameters, topology optimization parameters and semantic consistency parameters, thereby ensuring the comprehensive optimization of the correction results of the road centerline in terms of morphological continuity, structural connectivity and semantic consistency; at the same time, the second parameter set is combined to finely position and locally optimize single-point abnormalities, avoid interference with the global results, and realize accurate repair of small but key abnormal points.

[0060] In summary, the technical scheme of the present application significantly improves the road centerline data in terms of geometric continuity, topological rationality and semantic accuracy, which is beneficial to high-precision map construction, road feature recognition and intelligent driving applications. BRIEF DESCRIPTION OF DRAWINGS

[0061] Figure 1 FIG. 1 is a schematic diagram of a road centerline determination system based on a neural network and point cloud spatial features according to an embodiment of the present application;

[0062] Figure 2 FIG. 2 is a flowchart of a road centerline determination method based on a neural network and point cloud spatial features according to an embodiment of the present application;

[0063] Figure 3 FIG. 3 is a flowchart of a method for establishing an abnormal priority queue according to an embodiment of the present application;

[0064] Figure 4 FIG. 4 is a flowchart of a method for obtaining a corrected road centerline set according to an embodiment of the present application;

[0065] Figure 5 FIG. 5 is a flowchart of a method for judging geometric abnormal types according to an embodiment of the present application;

[0066] Figure 6 FIG. 6 is a flowchart of a method for judging topological abnormal types according to an embodiment of the present application;

[0067] Figure 7 FIG. 7 is a flowchart of a method for judging semantic abnormal types according to an embodiment of the present application. DETAILED DESCRIPTION

[0068] The technical solutions in the embodiments of the present application will be described in detail below with reference to the accompanying drawings. It should be particularly noted that the specific embodiments described below are only used to better illustrate and describe the technical solutions of the present application, and are intended to enable those skilled in the art to better understand and implement the present application, and should not be understood as limiting the scope of protection of the present application. Those skilled in the art can modify, adjust or equivalently replace the present application according to the content disclosed in the present application without departing from the spirit and essence of the present application, which should be considered as falling within the scope of protection of the present application.

[0069] Embodiment 1:

[0070] Referring to Figure 1 As shown in the drawings, the embodiment discloses a centerline determination system based on neural network and point cloud spatial features, comprising a hierarchical road domain segmentation module, a candidate centerline extraction module, a centerline checking module and a centerline hierarchical correction module, each module is connected through wired and / or wireless connection to realize data transmission.

[0071] The hierarchical road domain segmentation module is used for processing the original three-dimensional point cloud data of N-layer overlapping roads to obtain a passing road domain mask and a layer identification map.

[0072] The method for obtaining the passing road domain mask and the layer identification map comprises:

[0073] The original three-dimensional point cloud data is preprocessed to obtain standard point cloud data;

[0074] For each point cloud data point in the standard point cloud data, the neighborhood of the point cloud data point is subjected to principal component analysis to obtain a normal vector and a curvature indicator, the neighborhood echo intensity of the point cloud data point is calculated to obtain an echo intensity mean value and an echo intensity variance, the neighborhood height histogram of the point cloud data point is counted along the gravity direction, and the position, normal vector, curvature indicator, echo intensity mean value, echo intensity variance and neighborhood height histogram of the point cloud data point are constructed into corresponding point-level spatial features; 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;

[0075] The point-level spatial feature set is input into a road domain probability evaluation model to obtain a road domain probability score corresponding to each point cloud data point, if the road domain probability score is greater than or equal to a 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] All road domain labels and non-road domain labels are combined with the three-dimensional spatial coordinates of the corresponding point cloud data points to construct a connected domain to obtain a passing road domain mask; all road domain labels and non-road domain labels are combined with the three-dimensional spatial coordinates of the corresponding point cloud data points to construct a connected domain, which is subjected to connectivity analysis, small area elimination and hole filling processing to obtain a spatial topologically continuous passing road domain mask;

[0077] The layer identification map is obtained based on the hierarchical identification processing of the passing road domain mask.

[0078] The method for obtaining the standard point cloud data specifically comprises the following steps: aligning original three-dimensional point cloud data and IMU / GNSS pose interpolation to the same reference time according to timestamps; and mapping the original three-dimensional point cloud data from a sensor coordinate system to a world coordinate system based on a gravity direction to obtain gravity-aligned standard point cloud, so that the height axis is consistent with the vertical direction. The IMU is an inertial measurement unit, which is used to obtain inertial measurement data such as angular velocity and acceleration of a measurement carrier; and the GNSS is a global navigation satellite system, which is used to obtain three-dimensional position and speed information of the measurement carrier.

[0079] It should be noted that the method for obtaining the normal vector specifically comprises the following steps: selecting a local neighborhood according to the K-nearest neighbor algorithm with each point cloud data point as the center; performing local plane fitting on the local neighborhood to obtain the normal direction of the fitted plane as the normal vector of the point cloud data point. To improve stability, least squares fitting or plane estimation based on principal component analysis can be used, and consistency constraints such as adjusting the uniform normal direction to the opposite direction of gravity or along the main direction of the neighborhood are applied. In a scene with relatively large noise, the initial normal can be smoothed or bilateral filtered with neighborhood weighting to reduce the influence of random disturbance on subsequent determination.

[0080] The method for obtaining the curvature indicator specifically comprises the following steps: based on the local neighborhood of each point cloud data point, first, estimating the local plane or the local main direction, and then taking the degree of deviation from the local plane or the main direction as the surface change measure of the point cloud data point, which is denoted as the curvature indicator. The curvature indicator can be obtained from the local fitting residual, the dispersion of the neighborhood geometric distribution, or the uneven degree of the main direction energy distribution.

[0081] The method for obtaining the neighborhood echo intensity specifically comprises the following steps: using the echo intensity value recorded in the original three-dimensional point cloud data, collecting the intensity samples of each point cloud data point in the local neighborhood of the point cloud data point, and calculating the echo intensity mean and the echo intensity variance. To reduce systematic bias caused by factors such as sensor distance attenuation and wet slippery reflection, distance compensation or device-level normalization can be performed on the intensity before statistics.

[0082] Those skilled in the art know the acquisition process and engineering implementation of the normal vector, the curvature indicator and the neighborhood echo intensity, which provides mature function interfaces and implementation references for existing computer vision and three-dimensional reconstruction tool chains, and is sufficient to support the landing of the technical solutions of the present application.

[0083] The method for training the road area probability evaluation model comprises the following steps:

[0084] Pre-construct a road domain probability evaluation dataset, which includes LY group road domain probability evaluation data and road domain probability scores corresponding to the LY group road domain probability evaluation data, LY being a positive integer; the road domain probability evaluation data includes a point-level spatial feature set; divide the road domain probability evaluation dataset into a training set and a validation set, the training set being used for road domain probability evaluation model parameter learning, and the validation set being used for real-time monitoring of the generalization performance and overfitting degree of the road domain probability evaluation model;

[0085] The node output type graph neural network architecture is adopted as the road domain probability evaluation model, and the road domain probability evaluation data is input into the node output type graph neural network architecture after being standardized and vectorized. The node output type graph neural network architecture is composed 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 a probability distribution corresponding to each road domain probability score. Finally, the road domain probability score corresponding to the maximum probability is taken as the prediction result of the road domain probability evaluation model. In the training process, the cross-entropy loss function is used as the optimization objective, the gradient descent type optimization algorithm is used to update the network weights, and an early stopping strategy is set: when the prediction accuracy on the validation set reaches or exceeds a preset threshold, it is determined that the road domain probability evaluation model has converged and the training is terminated.

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

[0087] For example, in the preferred embodiments of the present application, for example, for urban road point cloud data, the road domain probability score 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 label, 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 label. For example, for high-precision map acquisition scenarios, the road domain probability score threshold can be set to 0.55 to improve the ability to distinguish complex boundary regions. For example, for non-paved roads or scenarios with low point cloud density, the road domain probability score threshold can be set to 0.45 to avoid road domain detection caused by noise. Further, those skilled in the art can flexibly configure the road domain probability score threshold according to the performance of the acquisition device, the type of road and the accuracy requirements of the task, so that stable road domain and non-road domain determination results can be obtained in different application environments, supporting the implementation of the technical solutions of the present application.

[0088] The method for obtaining a layer identification map based on the passing road domain mask includes:

[0089] For each point cloud data point in the passing road mask, the position and neighborhood height histogram in the point-level spatial feature corresponding to the point cloud data point are read, combined with the absolute height value of the point cloud data point in the gravity direction, and the height distribution characteristics of all point cloud data points in the passing road mask are obtained by statistical analysis;

[0090] According to the height distribution characteristics, the passing road mask is divided into M height bands by using a preset height segmentation threshold, and different height bands correspond to different road levels. In each height band, the local plane characteristics are further analyzed based on the normal vector and the curvature indicator. The point cloud data points with a normal vector similarity greater than a preset normal vector similarity threshold and a height variance lower than a preset height variance threshold are clustered into the same connected piece as a candidate road layer in the height band, and the connected piece is assigned a corresponding connected piece number.

[0091] For the candidate road layers with height overlap or local coincidence, the horizontal position distribution of the point cloud data points and the minimum height difference between layers are combined to separate the layers, so as 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 the height band and the connected piece number to which it belongs.

[0092] The candidate road layer numbers of all point cloud data points in the passing road mask are mapped to generate a layer identification map aligned with the passing road mask, providing basic data support for subsequent layer binding road centerline extraction.

[0093] For the candidate road layers with height overlap or local coincidence, the horizontal position distribution of the point cloud data points and the minimum height difference between layers are combined to separate the layers. Specifically, for each obtained candidate road layer, the gravity direction height values of all point cloud data points in the candidate road layer are statistically analyzed. If the height ranges of different candidate road layers overlap or are separated by less than a preset minimum height difference threshold, for example, 1.5 meters, in the height direction, it is determined that there is height overlap or local coincidence. Such situations are common in parallel sections of interchange ramps, under-bridge auxiliary roads, or multi-layer structures with insufficient road height differences. Relying only on height band segmentation can easily lead to point cloud data points of different layers being classified into the same level.

[0094] After determining that there is high overlap or partial overlap, the interlayer separation is performed in combination with the horizontal position distribution of the point cloud data points. Specifically, the point cloud data of the candidate road layer is projected to a horizontal plane, and the horizontal distribution range of each point cloud data point relative to the center line or geometric center of the candidate layer is calculated; the points with obvious horizontal distribution difference are divided into different sub-clusters by using a plane clustering algorithm, such as region growing or DBSCAN based on Euclidean distance. For the point cloud data points with horizontal distribution overlap but height difference close to the minimum height difference threshold between layers, the normal vector and curvature indicator are combined to determine: the points with large normal vector difference or significant curvature feature difference are assigned to different candidate road layers.

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

[0096] It should be noted that the height segmentation threshold is used to divide the passing road domain mask into different height bands according to the height value in the gravity direction, and different height bands correspond to different road levels. For example, in the present application, 1.5 meters can be selected as the height segmentation threshold, i.e., when the center height difference between adjacent height bands is greater than or equal to 1.5 meters, it is determined as different road levels; the height segmentation threshold is suitable for common urban overpasses and underpasses, and can effectively distinguish the 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 assigned to the same local plane. For example, in the present application, the cosine similarity of the normal vector can be used for measurement, and the threshold is set to 0.95, i.e., when the cosine similarity of the normal vectors of two points is greater than or equal to 0.95, it is considered that the local plane directions are consistent; the normal vector similarity threshold can ensure the stability of the plane clustering in the same layer, and prevent direction interference between different layers.

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

[0099] The candidate centerline extraction module extracts the traversable area within the layer based on the traversable road area mask and layer identification 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 constraint.

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

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

[0102] Based on the passable area in the layer and the corresponding boundary patch set, a three-dimensional Euclidean distance field and a corresponding distance field gradient are calculated; based on the distance field gradient, a non-maximum suppression operation is performed on the three-dimensional Euclidean distance field along the gradient direction to obtain a skeleton candidate set, and a three-dimensional thinning process is performed on the passable area in the layer to obtain a morphological skeleton;

[0103] Point cloud data points that exist in both the skeleton candidate set and the morphological skeleton are constructed into a high-confidence skeleton point set, and 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 constraints are applied to each skeleton line segment, and skeleton line segments with lengths below a preset minimum skeleton length threshold are removed to eliminate false centerlines caused by breaks, offsets, or isolated small areas due to local noise.

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

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

[0106] The method for obtaining the distance field gradient comprises: after obtaining a three-dimensional Euclidean distance field, performing spatial first-order partial derivation calculation on the distance field in three directions of horizontal, vertical and perpendicular directions respectively by using a three-dimensional difference operator to obtain a gradient vector corresponding to each voxel unit; the three-dimensional difference operator is, for example, a three-dimensional Sobel operator or a three-dimensional central difference. The direction of the gradient vector points to the direction in which the distance value changes fastest, that is, the direction of the nearest boundary, and the modulus value of the gradient vector reflects the fastness or slowness of the distance change. By the distance field gradient, non-maximum suppression can be performed in the subsequent step along the gradient direction, so as to extract the skeleton candidate points with geometric center significance.

[0107] The method for obtaining the skeleton candidate set comprises: after obtaining the three-dimensional Euclidean distance field and the corresponding distance field gradient, searching in two directions of positive and negative gradient vectors respectively at each voxel position, comparing the distance value of the current position with the distance value of the adjacent position in the direction; if the distance value of the current position reaches a local maximum in the gradient direction, that is, is greater than the distance value of the adjacent position, it is determined that the position is a potential skeleton point, and is added to the skeleton candidate set. By performing the above non-maximum suppression operation along the gradient direction on the entire distance field, the redundant points at non-central positions can be removed, the skeleton candidate points at the local maximum positions and with geometric center significance are retained, and then the skeleton candidate set is formed.

[0108] The method for obtaining the morphological skeleton comprises: for the voxel units of the in-layer passable area, removing the boundary voxels that do not affect connectivity layer by layer by using a three-dimensional morphological thinning algorithm based on morphology until the remaining voxel set cannot be further contracted, and the obtained result is the morphological skeleton. The three-dimensional morphological thinning algorithm is, for example, a three-dimensional iterative corrosion-expansion algorithm or a three-dimensional topological preservation thinning algorithm. The morphological skeleton can retain the overall topological structure of the in-layer passable area and eliminate the boundary redundancy at the same time.

[0109] Further, in the technical solution of the present application, the candidate road centerline set is obtained in advance, which provides a high-confidence geometric basis for the subsequent accurate road centerline generation and verification. The candidate road centerline extracted by combining the boundary surface, the three-dimensional Euclidean distance field and the morphological skeleton and the like under the constraint of the passable road domain mask and the layer identification map can effectively retain the spatial center trend of the road main body and eliminate most of the pseudo road centerlines caused by noise, local defects or abnormal points. Taking the candidate road centerline set as an intermediate result can reduce the search space and the calculation complexity of the subsequent processing link, and improve the stability and accuracy of the final road centerline fitting and optimization, so as to ensure the continuity and consistency of the road geometric features in the global range.

[0110] The middle line checking module performs comprehensive abnormality detection processing based on geometric, topological and semantic features corresponding to the candidate road middle line set to obtain a road middle line abnormality type corresponding to each candidate road middle line. The road middle line abnormality type includes a geometric abnormality type, a topological abnormality type and a semantic abnormality type.

[0111] As shown in Figure 5 , the judgment method of the geometric abnormality type includes:

[0112] The spatial geometric information of the candidate road middle line and the corresponding road boundary is obtained. The road boundary is extracted based on a boundary facet set of the passable area in the layer.

[0113] The candidate road middle line 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 difference between the first Euclidean distance and the second Euclidean distance is taken to obtain the bidirectional distance residual corresponding to the sequence positioning point. The sum of the first Euclidean distance and the second Euclidean distance is taken to obtain the local road width corresponding to the sequence positioning point. In this application, the first interval distance can be set to 20 cm, for example.

[0114] 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.

[0115] The symmetry index is calculated based on the bidirectional distance residual set and the local road width set, and 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 abnormality is added to the geometric abnormality type of the candidate road middle line. If the road width consistency index is less than a preset road width consistency index threshold, a width abnormality is added to the geometric abnormality type of the candidate road middle line. In this application, the symmetry index threshold can be set to 0.75, and the road width consistency index threshold can be set to 0.75, for example.

[0117] The calculation method of the symmetry index includes:

[0118] ;

[0119] Wherein, The symmetry index is an index, and the value range of the symmetry index is ; is an index variable of the summation formula, represents the first bidirectional distance residual, denotes the b-th local road width, denotes the truncation of the result obtained in The closer the symmetry index is to 1, the more balanced the distance of the candidate road centerline along the entire centerline is, and the more centered the geometry is. The closer the symmetry index is to 0, the more obvious the overall eccentricity of the candidate road centerline is.

[0120] The method for calculating the road width consistency index comprises:

[0121]

[0122] wherein, is the road width consistency index, denotes the mean of the local road width of the local road width set denotes the standard deviation of the local road width of the local road width set. denotes the truncation of the result obtained in

[0123] As shown in Figure 6 , the method for judging the topological anomaly type comprises:

[0124] The candidate road centerline is scattered into a line element sequence according to a preset second interval distance, and the distance between adjacent line elements is calculated based on the spatial adjacency relationship. When the distance between any adjacent line elements is greater than a preset breaking threshold, it is determined that the candidate road centerline has a breaking anomaly, and the breaking anomaly is added to the topological anomaly type of the candidate road centerline. Exemplarily, in the present application, the second interval distance can be set to 20 cm.

[0125] The candidate road centerline is searched for global intersection points in a three-dimensional space. When it is detected that the candidate road centerline has a geometric intersection point at non-adjacent line elements, 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 the 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 the adjacent line elements is less than a preset cross-layer judgment threshold, it is determined that the candidate road centerline has a cross-layer short-circuit anomaly, and the short-circuit anomaly is added to the topological anomaly type of the candidate road centerline.

[0127] It should be noted that exemplarily, in the preferred embodiment of the present application, the breaking threshold can be set to 0.5 m, and the breaking threshold is used to limit the maximum allowed 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 breaking threshold, it is determined that there is a breaking anomaly at this position. The breaking threshold can be adjusted according to different map accuracy and road width characteristics. ​​​​​

[0128] The self-intersection determination threshold can be set to 0.3 m, and the self-intersection determination threshold is used to limit the minimum distance between different line segments of the same candidate road centerline on the spatial projection plane. When the spatial distance between non-adjacent line segments is less than the self-intersection determination threshold and there is a crossing topological relationship, it is determined that there is a self-intersection anomaly at this position. The selection of the self-intersection determination threshold can be optimized in combination with the road centerline extraction accuracy and the minimum lane width.

[0129] The cross-layer determination threshold can be set to 1.0 m, and the cross-layer determination threshold is used to limit the minimum reasonable spacing in three-dimensional space between different road layers. 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, it is determined that there is a cross-layer short circuit anomaly at this position. The cross-layer determination threshold can be set in combination with the interlayer safety spacing of three-dimensional traffic structures such as viaducts and tunnels.

[0130] In the present application, the self-intersection anomaly, the fracture anomaly and the cross-layer short circuit anomaly are identified to perform quality control on the geometric and topological structure of the candidate road centerline set, so as to ensure that the generated road centerline result has continuity, uniqueness and road layer consistency. Among them, the detection of the self-intersection anomaly can avoid unreasonable intersection of the road centerline in space, thereby preventing errors in the navigation path generation; the detection of the fracture anomaly can timely find the gap in the road centerline, so as to facilitate repair by interpolation or re-extraction of the skeleton line segment, and improve the road connectivity; the detection of the cross-layer short circuit anomaly can effectively prevent unreasonable direct connection between different road layers, and ensure the correctness of the road topological structure in the three-dimensional traffic scene. Through the identification of the above three types of anomalies, the geometric accuracy and topological consistency of the road centerline data generated by the present application can be significantly improved.

[0131] As shown in Figure 7 The judgment method of the semantic anomaly type includes:

[0132] A pre-constructed road semantic feature set is constructed; the road semantic feature set includes road grade information and road lane number information; the road grade information includes expressway, urban expressway, main road, secondary road and branch road; the road semantic feature set is derived from a map database or historical high-precision map data.

[0133] The spatial position of the candidate road centerline is subjected to spatial overlay operation with the real-time collected road vector data or road surface data, the road object with the highest spatial coincidence degree with the candidate road centerline is extracted, and the road grade and the number of lanes corresponding to the road object are obtained;

[0134] If the road grade of the road object is different from the corresponding road grade in the road semantic feature set, it is determined that the candidate road centerline has inconsistent road grades; the inconsistent road grades are added to the semantic anomaly type of the candidate road centerline.

[0135] If the number of lanes of the road object is not the same as the number of lanes corresponding to the road semantic feature set, it is determined that the candidate road centerline has inconsistent lane numbers, and lane number inconsistency is added to the semantic abnormal type of the candidate road centerline.

[0136] It should be noted that in the present application, the identification of road level inconsistency and lane number inconsistency plays an important role in ensuring the accuracy of the road centerline result and the consistency of the road semantics. Specifically, the identification of road level inconsistency can effectively find the difference between the candidate road centerline and the actual road in the level attribute. For example, when the target road layer to which the candidate road centerline based on the skeleton extraction belongs is a highway layer, but the spatially matched road object is a city branch, it can be determined that the road level is inconsistent. The identification of road level inconsistency can avoid incorrect level assignment caused by road layer identification error, skeleton extraction deviation or data source defect, and prevent level mismatch in traffic rule matching, navigation planning and traffic control strategy in subsequent practical application process.

[0137] The identification of lane number inconsistency can find the difference between the candidate road centerline and the actual road in the lane capacity description. For example, the candidate road centerline is expected to be associated with two lanes, but the matched road object attribute shows four lanes, indicating that there is lane number inconsistency. Such identification can effectively avoid the distortion of traffic capacity evaluation caused by lane number error, and prevent unreasonable planning decisions in traffic flow simulation, road right allocation or road reconstruction design.

[0138] In summary, the identification of road level inconsistency and lane number inconsistency can be used as a core link of semantic quality control of the candidate road centerline, which can provide direct basis for the correction, replacement or rejection of abnormal centerline, and improve the guarantee ability of the finally generated road centerline data in terms of road semantic consistency, traffic simulation accuracy and engineering application reliability, so as to ensure that the road centerline result generated by the present application not only has high geometric precision, but also has integrity and correctness in road attribute description.

[0139] The centerline hierarchical correction module establishes an abnormal priority queue based on the road centerline abnormal type corresponding to the candidate road centerline set; divides the road centerline abnormal type of the candidate road centerline into a batch repair set and a single-point repair set according to the abnormal priority queue; respectively executes geometric shape adjustment, topological structure reconstruction and semantic attribute update on the batch repair set and the single-point repair set, and obtains a corrected road centerline set.

[0140] As shown in Figure 3 The establishment method of the abnormal priority queue includes:

[0141] A pre-constructed abnormality influence weight table is provided, which is used to define the influence degree of different abnormality types on the geometric accuracy, topological connectivity and semantic consistency of the road centerline; the abnormality influence weight table sets the influence degree of abnormality types on the geometric accuracy, connectivity and semantic accuracy of the road, and the abnormality types that affect driving safety and road recognition accuracy are given a higher weight.

[0142] According to the abnormality influence weight table, each abnormality type in the road centerline abnormality type is given a weight value, and the abnormality type weight corresponding to each abnormality type is obtained.

[0143] For each candidate road centerline in the candidate road centerline set, the abnormality type weight contained in the candidate road centerline is accumulated to obtain the abnormality comprehensive influence score of the candidate road centerline.

[0144] All candidate road centerline corresponding abnormality comprehensive influence score is constructed into an abnormality comprehensive influence score set; the abnormality comprehensive influence score set is sorted in descending order to obtain an abnormality priority queue.

[0145] For example, the abnormality influence weight table is shown in Table 1.

[0146] Table 1 Abnormality influence weight table

[0147]

[0148] According to the abnormality priority queue, the road centerline abnormality types of the candidate road centerline are divided into a batch repair set and a single-point repair set.

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

[0150] The training method of the abnormality shunting model includes:

[0151] An abnormality shunting dataset is pre-constructed, which includes FL group abnormality shunting data and the batch repair set and single-point repair set corresponding to the FL group abnormality shunting data, FL being a positive integer; the abnormality shunting data includes road centerline abnormality type, abnormality type weight and abnormality comprehensive influence score; the abnormality shunting dataset is divided into a training set and a validation set, the training set is used for abnormality shunting model parameter learning, and the validation set is used for real-time monitoring of the generalization performance and overfitting degree of the abnormality shunting model;

[0152] The deep neural network taking the multilayer perceptron as the core is used as the anomaly shunting model. The anomaly shunting data is input into the deep neural network after being processed by standardization and vectorization. The deep neural network is composed of an input layer, a hidden layer 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 batch repair set and single-point repair set. Finally, the batch repair set and single-point repair set corresponding to the maximum probability are taken as the prediction result of the anomaly shunting model. In the training process, the cross-entropy loss function is used as the optimization objective, the gradient descent optimization algorithm is used to update the network weights, and the early stopping strategy is set: when the prediction accuracy on the validation set reaches or exceeds the prediction accuracy threshold, it is determined that the anomaly shunting model has converged and the training is terminated. For example, the prediction accuracy threshold can be set to 95% in this application.

[0153] It should be noted that the road centerline anomaly type can reflect the nature and repair complexity of the anomaly; the anomaly type weight quantifies the influence degree of the anomaly on the overall geometry, topology or semantic consistency; and the anomaly comprehensive influence score integrates the anomaly distribution and influence range of the candidate road centerline, reflecting the urgency and global importance of repair. When the road centerline anomaly type, anomaly type weight and anomaly comprehensive influence score are used as the input of the anomaly shunting model, the anomaly shunting model can automatically determine whether the anomaly of the candidate road centerline is more suitable for concentrated batch repair (i.e. adjacent road segment consistency adjustment) or independent single-point repair (i.e. local mutation correction) according to the feature mode, thereby realizing dynamic division without manual rule configuration.

[0154] By dividing the batch repair set and the single-point repair set, the static classification method based on a single threshold or fixed rule can be avoided, and the adaptability and accuracy of the repair strategy can be improved. The anomaly shunting model can flexibly select the optimal repair path according to the actual road data distribution and anomaly feature changes, which can improve the efficiency of batch repair when the anomaly distribution is concentrated, and reduce unnecessary global adjustment overhead when the anomaly distribution is scattered. In addition, the repair quality and processing efficiency can be maintained stable in different scenarios.

[0155] As shown in Figure 4 , the method for obtaining the corrected road centerline set comprises:

[0156] For each candidate road centerline in the anomaly priority queue, if the candidate road centerline contains a batch repair set, the spatial positions and anomaly type weights 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 a first parameter set.

[0157] If the candidate road centerline contains a single-point repair set, the spatial positions and abnormal type weights of each abnormal type in the single-point repair set are extracted and constructed into corresponding single-point abnormal data; each single-point abnormal data is input into the second parameter setting model to obtain a second parameter set;

[0158] 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;

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

[0160] The training method of the first parameter setting model comprises:

[0161] A first parameter setting dataset is constructed in advance, the first parameter setting dataset comprising DY group first parameter setting data and a first parameter set corresponding to the DY group first parameter setting data, DY being a positive integer; the first parameter setting data comprises a batch repair subgraph; the first parameter setting dataset is divided into a training set and a validation set, the training set being used for first parameter setting model parameter learning, and the validation set being used for real-time monitoring of the generalization performance and overfitting degree of the first parameter setting model;

[0162] A node output type graph neural network architecture is adopted as the first parameter setting model, the first parameter setting data being input into the node output type graph neural network architecture after standardization and vectorization processing; the node output type graph neural network architecture comprises 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 a probability distribution corresponding to each first parameter set, and finally the first parameter set corresponding to the maximum probability is taken as the prediction result of the first parameter setting model; during the training process, a cross-entropy loss function is used as the optimization objective, a gradient descent type optimization algorithm is used to update the network weights, and an early stopping strategy is set: when the prediction accuracy on the validation set reaches or exceeds a preset threshold, it is determined that the first parameter setting model has converged and the training is terminated.

[0163] The training method of the second parameter setting model comprises:

[0164] A second parameter setting dataset is constructed in advance, the second parameter setting dataset comprising DE group second parameter setting data and a second parameter set corresponding to the DE group second parameter setting data, DE being a positive integer; the second parameter setting data comprises single-point abnormal data; the second parameter setting dataset is divided into a training set and a validation set, the training set being used for second parameter setting model parameter learning, and the validation set being used for real-time monitoring of the generalization performance and overfitting degree of the second parameter setting model;

[0165] The deep neural network with a multilayer perceptron as the core is used as the second parameter setting model. The second parameter setting data is input into the deep neural network after being standardized and vectorized. The deep neural network is composed of an input layer, a hidden layer 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 a probability distribution corresponding to each second parameter set. Finally, the second parameter set corresponding to the maximum probability is taken as the prediction result of the second parameter setting model. In the training process, the cross-entropy loss function is used as the optimization objective, the gradient descent optimization algorithm is used to update the network weights, and the early stopping strategy is set: when the prediction accuracy on the validation set reaches or exceeds the prediction accuracy threshold, it is determined that the second parameter setting model has converged and the training is terminated.

[0166] The first parameter set includes geometric adjustment parameters, topological structure optimization parameters and semantic attribute uniformization parameters. The geometric adjustment parameters include a road centerline transverse displacement amount, a road centerline longitudinal displacement amount and a curvature smoothing coefficient. Specifically, the road centerline transverse displacement amount and the road centerline longitudinal displacement amount are respectively used to control the overall displacement of the batch abnormal nodes in the horizontal direction and the vertical direction in the plane coordinate system, and correct the spatial misplacement caused by acquisition errors or offset accumulation. The curvature smoothing coefficient is used to constrain the curvature variation amplitude of the batch node path, and by adjusting the weight of the curve control point, the smoothing of the large-range curve is realized, and the geometric mutation is eliminated and the transition is ensured to be natural.

[0167] The topological structure optimization parameters include node connection reconstruction weights and a fracture merging threshold. Specifically, the node connection reconstruction weights are used to measure the necessity and priority of reconstructing the connection edges between adjacent batch abnormal nodes. When the connection reconstruction weight is higher than a preset connection reconstruction weight threshold, for example, 0.30, a new connection edge is automatically established or an existing connection edge is replaced, so as to improve the overall path continuity. The fracture merging threshold is used to determine whether the topological fracture section performs a merging operation. When the spatial distance, direction deviation and semantic consistency between the breakpoints all meet the fracture merging threshold, the connection and merging are automatically performed, so as to eliminate the topological fracture and reduce the redundant branches.

[0168] The semantic attribute uniformization parameters include a functional label uniform rule, an attribute conflict priority weight and a node semantic inheritance ratio. Specifically, the functional label uniform rule is used to standardize the core semantic labels such as road function type and lane number during the batch repair process, so as to avoid the semantic inconsistency of adjacent road sections. The attribute conflict priority weight is used to determine the selection order of the conflicting attributes when the multi-source data is fused, so as to ensure the decision consistency of the final semantic result. The node semantic inheritance ratio is used to control the degree of inheritance of the semantic information from the original node after the adjustment of the batch abnormal node, so as to ensure the continuity and reliability of the semantic information while optimizing the geometry and topological structure.

[0169] The first parameter set includes a single-point geometric adjustment offset, a single-point smoothing radius, a node connection weight threshold, a direction correction angle, and a semantic attribute update priority. Specifically, the single-point geometric adjustment offset is used to determine the position fine-tuning amplitude of the single-point anomaly in the plane coordinate system. By applying an 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 smoothness of the curve within the range around the node, and the mutation or sharp corner is eliminated based on the single-point smoothing radius during correction; the node connection weight threshold is used to screen the retention or reconstruction conditions of the node in the topological relationship. When the connection weight with the adjacent node is lower than the node connection weight threshold, the connection operation is disconnected or replaced 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 the adjacent path exceeds the set angle value, the set angle value is, for example, 15°, and 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 functional type to ensure priority matching of key semantics.

[0170] Example 2:

[0171] See also Figure 2 As shown, this embodiment provides a method for determining a road centerline based on a neural network and point cloud spatial features, including:

[0172] Process the original 3D point cloud data of N layers of overlapping roads to obtain the passable road area mask and layer identification map;

[0173] The traversable area within the layer is extracted based on the traversable road mask and layer identification map. A high-confidence skeleton is obtained through non-maximum suppression and 3D refinement. A set of candidate road centerlines is generated through connectivity clustering and length constraints.

[0174] Performing comprehensive anomaly detection 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, including geometric anomaly type, topological anomaly type, and semantic anomaly type;

[0175] An exception 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 according to the exception priority queue; geometric morphology adjustment, topological structure 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 merely describes specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any modification or substitution within the technical range disclosed by the present application can be easily thought of by those skilled in the art, and should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0177] Finally, the above merely describes preferred embodiments of the present application, and is not used to limit the present application, and any modification, equivalent substitution, improvement, etc. within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A method for determining the road centerline based on neural networks and point cloud spatial features, characterized in that: include: Process the original 3D point cloud data of N layers of overlapping roads to obtain the passable road area mask and layer identification map; The traversable area within the layer is extracted based on the traversable road mask and layer identification map. A high-confidence skeleton is obtained through non-maximum suppression and 3D refinement. A set of candidate road centerlines is generated through connectivity clustering and length constraints. Performing comprehensive anomaly detection 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, including geometric anomaly type, topological anomaly type, and semantic anomaly type; Establish an exception priority queue based on the road centerline anomaly types corresponding to the candidate road centerline set; divide the road centerline anomaly types of the candidate road centerlines into a batch repair set and a single-point repair set according to the exception priority queue; Geometric adjustment, topological 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.

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

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

4. The method for determining the road centerline based on neural network and point cloud spatial features according to claim 1, characterized in that: Methods for obtaining the corrected posterior centerline set include: For each candidate road centerline in the exception priority queue, if the candidate road centerline includes a batch repair set, extracting the spatial position and exception type weight corresponding to each exception type in the batch repair set and constructing a corresponding batch exception node; assembling the batch exception nodes into a batch repair subgraph, and inputting the batch repair subgraph into a first parameter setting model to obtain a first parameter set; If the candidate road centerline includes a single-point repair set, the spatial position and the weight of 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 the second parameter setting model to obtain a second parameter set; According to the first parameter set and the second parameter set, geometrically adjust the candidate road centerline, reconstruct the topological structure, and update the semantic attributes to obtain a corrected road centerline; All corrected posterior centerlines were collected into a corrected posterior centerline set.

5. The method for determining the road centerline based on neural network and point cloud spatial features according to claim 1, characterized in that: Methods for determining geometric anomaly types include: Discretize the candidate road centerline into B sequential positioning points according to a preset first spacing. For each sequential positioning point, calculate the shortest Euclidean distance from the sequential positioning point to the boundary of one side of the road, recorded as the first Euclidean distance, and calculate the shortest Euclidean distance from the sequential positioning point to the boundary of the other side of the road, recorded as the second Euclidean distance. Subtract the first Euclidean distance from the second Euclidean distance and take the absolute value to obtain the two-way distance residual corresponding to the sequential positioning point. Sum the first Euclidean distance and the second Euclidean distance to obtain the local road width corresponding to the sequential positioning point. Construct B two-way distance residuals into a two-way distance residual set; construct B local road widths into a local road width set; The symmetry index is calculated based on the two-way 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, a geometric offset anomaly is added to the geometric anomaly type of the candidate road center line; if the road width consistency index is less than the preset road width consistency index threshold, a width anomaly is added to the geometric anomaly type of the candidate road center line.

6. The method for determining the road centerline based on neural network and point cloud spatial features according to claim 1, characterized in that: Methods for determining topological anomaly types include: Discretize the candidate road centerline into a line element sequence according to a preset second spacing, and calculate the spacing between adjacent line elements based on the spatial adjacency relationship. When the spacing between any adjacent line elements is greater than a preset fracture threshold, it is determined that the candidate road centerline has a fracture anomaly, and a fracture anomaly is added to the topological anomaly type of the candidate road centerline. Performing a global intersection search on the candidate road centerlines in three-dimensional space, when it is detected that the candidate road centerlines have geometric intersections at non-adjacent line elements and the spacing is less than a preset self-intersection determination 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 judgment 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 topological anomaly type of the candidate road centerline.

7. The method for determining the road centerline based on neural network and point cloud spatial features according to claim 1, characterized in that: Methods for determining semantic anomaly types include: Pre-constructing a road semantic feature set; the road semantic feature set includes road grade information and road lane number information; the road grade information includes expressways, urban expressways, main roads, secondary roads, and branch roads; Perform spatial superposition calculation on the spatial position of the candidate road centerline and the road vector data or road surface data collected in real time, extract the road object with the highest spatial overlap with the candidate road centerline, and obtain the road grade and lane number corresponding to the road object; If the road grade of the road object is different from the road grade corresponding to the road semantic feature set, it is determined that the candidate road center line has a road grade discrepancy; and a road grade discrepancy is added to the semantic anomaly type of the candidate road center line; If the number of lanes of the road object is different from the number of lanes corresponding to the road semantic feature set, it is determined that the candidate road center line has a lane number discrepancy; and a lane number discrepancy is added to the semantic anomaly type of the candidate road center line.

8. The method for determining the road centerline based on neural network 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 passable road mask and layer identification map, the passable area in the layer corresponding to each candidate road layer number is extracted; for the spatial boundary points of the passable area in the layer, a three-dimensional contour tracking algorithm is used to construct a boundary patch set; Based on the passable area in the layer and the corresponding boundary patch set, a three-dimensional Euclidean distance field and a corresponding distance field gradient are calculated; based on the distance field gradient, a non-maximum suppression operation is performed on the three-dimensional Euclidean distance field along the gradient direction to obtain a skeleton candidate set, and a three-dimensional thinning process is performed on the passable area in the layer to obtain a morphological skeleton; The point cloud data points that exist in both the skeleton candidate set and the morphological skeleton are constructed into a high-confidence skeleton point set, and 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 below a preset minimum skeleton length threshold are eliminated; All retained skeleton line segments are collected to construct a candidate road centerline set.

9. The method for determining the road centerline based on neural network and point cloud spatial features according to claim 1, characterized in that: The method for obtaining the passable road area mask and layer identification map includes: Preprocess the original 3D point cloud data to obtain standard 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 a normal vector, a curvature indicator, an echo intensity mean, and an echo intensity variance; and a neighborhood height histogram of the point cloud data point is statistically calculated 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. Input the point-level spatial feature set into the road domain probability assessment model to obtain the road domain probability score corresponding to each point cloud data point, and label the point cloud data point with a road domain label or a non-road domain label according to the preset road domain probability score threshold; All road and non-road labels are combined with the 3D spatial coordinates of the corresponding point cloud data points to construct a connected domain, and a passable road mask is obtained. All road and non-road labels are combined with the 3D spatial coordinates of the corresponding point cloud data points to construct a connected domain. After connectivity analysis, region elimination and hole filling, a passable road mask with spatial topology continuity is obtained. A layer identification map is obtained by performing layer identification processing based on the pass road area mask.

10. The method for determining the road centerline based on neural network and point cloud spatial features according to claim 9, characterized in that: The method for performing layer identification processing based on a passable road area mask to obtain a layer identification map includes: For each point cloud data point within the road mask, read the point cloud data point's position, neighborhood height histogram, and absolute height value in the gravity direction, and statistically generate height distribution features. Based on the preset height segmentation threshold, the road mask is divided into M height bands. In each height band, the local plane characteristics are analyzed by combining the normal vector and the 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 the candidate road layer and assigned a connected piece number. For highly overlapping or partially overlapping candidate road layers, the horizontal position distribution and the minimum height difference between layers are used to separate them, and the candidate road layer number is determined for each point cloud data point; Map all candidate road layer numbers to the corresponding three-dimensional coordinates to generate a layer identification map that is spatially aligned with the passable road area mask.

11. A road centerline determination system based on a neural network and point cloud spatial features, used to implement the road centerline determination method based on a neural network and point cloud spatial features according to any one of claims 1 to 10, 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 passable road mask and layer identification map; The candidate centerline extraction module extracts the traversable area within a layer based on the traversable road mask and layer identification map, obtains a high-confidence skeleton through non-maximum suppression and 3D refinement, and generates a set of candidate road centerlines through connectivity clustering and length constraints. The centerline verification module performs comprehensive anomaly detection based on the geometric, topological, and semantic features corresponding to the candidate centerline set to obtain the corresponding centerline anomaly type, which includes geometric anomaly type, topological anomaly type, and semantic anomaly type. The centerline hierarchical correction module establishes an exception priority queue based on the centerline anomaly types corresponding to the candidate centerline set; and divides the centerline anomaly types of the candidate centerlines into a batch repair set and a single-point repair set according to the exception priority queue; Geometric adjustment, topological 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.

Images