A road network data matching method based on a graph neural network

Abstract

The application discloses a road network data matching method based on a graph neural network, and comprises the following steps: pre-processing and stroke extraction of road network data, constructing a graph model for describing road section relations by using the road network data, using nodes of the graph to represent stroke road sections, and using edges of the graph to represent connection relations between adjacent road sections; automatically encoding geometric elements of the road sections as road section geometric features, using attribute data of the road sections as road section semantic features, and using coordinates and azimuth angles of the road sections as road section position features; based on a GraphSAGE framework, using initial features of road graph nodes as input, and learning graph node embedding of the road sections by using a spatial, attribute or mixed neighborhood aggregation function; learning graph node embedding of road network data of different sources respectively, calculating road section similarity between different data sources by using the graph node embedding, determining a road section matching relation, and realizing road network matching; and the application improves the performance of road network matching based on the graph neural network.

Description

Technical Field

[0001] This invention relates to the field of road network matching technology, and in particular to a road network data matching method based on graph neural networks. Background Technology

[0002] As a crucial transportation infrastructure, timely updates to road network data are essential for traffic management, emergency response, and urban planning. Determining the correspondence between elements in road network data from different sources through road network matching is not only a vital technical approach for road network updates but also provides technical support for tasks such as crowdsourced road network data quality assessment, thus attracting significant attention from scholars in the field of geographic information. Traditional road network matching methods primarily measure the similarity of road network structures through geometric and topological attributes of road network data to determine the matching relationships of road network elements. However, manually designed features and thresholds are easily limited by expert experience, leading to performance degradation under complex road network structures. In recent years, road network data modeling based on graph neural networks has become a research hotspot, achieving excellent performance in multiple road network modeling tasks. However, existing methods often employ a direct neighborhood aggregation approach on the graph topology to learn the embedded representation of the road network structure, failing to consider the spatial relationships of road network elements in this crucial step and thus not fully utilizing the representation learning capabilities of graph neural networks. Summary of the Invention

[0003] This invention addresses road network matching tasks and proposes a road network data matching method based on graph neural networks, employing the concept of explicit spatial modeling. First, a road graph model of the road network data is constructed, and geometric, semantic, and location features are extracted. Then, based on the GraphSAGE framework, three neighborhood aggregation operators—spatial, attribute, and hybrid—are proposed, incorporating the calculation of spatial relationships and attribute types of road network elements during the neighborhood aggregation operation. Finally, the matching relationships of road network elements are determined using the similarity of graph node embeddings. To verify the effectiveness of this method, extensive experiments were conducted using real road network data. The proposed method achieved optimal performance across all indicators on the experimental data, improving the matching accuracy by more than 11% and the recall rate by more than 6.8% compared to the baseline graph neural network method.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] A road network data matching method based on graph neural networks includes:

[0006] Step 1: Preprocess the road network data and extract strokes. Use the road network data to construct a graph model describing the relationship between road segments. Use the nodes of the graph to represent stroke road segments and the edges of the graph to represent the connection relationship between adjacent road segments.

[0007] Step 2: Automatically encode the geometric elements of the road segment as the geometric features of the road segment, use the attribute data of the road segment as the semantic features of the road segment, and use the coordinates and orientation angles of the road segment as the location features of the road segment. Combine the three types of features—geometric, semantic, and location—as the initial feature description of the road map node.

[0008] Step 3: Based on the GraphSAGE framework, using the initial features of the road graph nodes as input, learn the graph node embeddings of the road segments using spatial neighborhood aggregation functions, attribute neighborhood aggregation functions, or hybrid neighborhood aggregation functions;

[0009] Step 4: For road network data from different sources, learn the graph node embedding of the road network data respectively, and use the graph node embedding to calculate the road segment similarity between different data sources, determine the road segment matching relationship, and realize road network matching.

[0010] Further, step 1 includes:

[0011] First, the road network data is preprocessed to filter out noisy data and then stroke extraction is performed on the road network data.

[0012] Then, based on the processed road network data, the nodes and edges of the graph are determined, a road graph is constructed, the midpoint of the stroke segment is selected as the node of the road graph, the neighboring nodes of each node are determined according to the distance relationship between road segments, and the adjacent nodes are connected to construct the edges of the graph.

[0013] Furthermore, the neighboring nodes of each node are determined in the following manner:

[0014] Set a neighborhood search radius, and then search for road segment nodes in the neighborhood whose Hausdorf distance is higher than a certain threshold for each road segment node.

[0015] Furthermore, in step 2, when extracting the geometric features of the road segment, each road segment in the road network data is traversed, the latitude and longitude of the road segment endpoints are compared, and the one with the smaller longitude or latitude is taken as the starting point of the road segment.

[0016] Furthermore, the semantic features of the road segment include road type and road name.

[0017] Furthermore, the coordinates of the road segment are represented as the normalized starting point coordinates within the geographical range of the road network tiles, and the azimuth angle of the road segment is determined by the angle between the starting point and the boundary of the road network tiles.

[0018] Furthermore, the spatial neighborhood aggregation function is:

[0019]

[0020] in p is the embedding vector of node v at the k-th layer in the case of spatial aggregation. u and p v These are the coordinates of the midpoint of the road segment represented by nodes u and v, respectively; p u -p v The spatial relationship between nodes is represented by the Euclidean distance between their locations; W (k) σ is a learnable parameter; N(v) is the set of neighboring nodes of node v; |N(v)| is the number of neighboring nodes of node v; σ is the activation function.

[0021] Furthermore, the attribute neighborhood aggregation function is:

[0022]

[0023] in It is the embedding vector of node v at the k-th layer in the case of attribute aggregation. W is the embedding vector of the (k-1)th layer of a neighboring node u that has the same road type as node v; (k) These are the trainable parameters; N(v) is the set of neighboring nodes of node v; |N(v)| is the number of neighboring nodes of node v; σ is the activation function.

[0024] Furthermore, the hybrid neighborhood aggregation function is:

[0025]

[0026] in p is the embedding vector of node v at the k-th layer in the case of hybrid aggregation. u and p v These are the coordinates of the midpoint of the road segment represented by nodes u and v, respectively; p u -p v The spatial relationship between nodes is represented by the Euclidean distance between their locations; W (k) σ is a learnable parameter; N(v) is the set of neighboring nodes of node v; |N(v)| is the number of neighboring nodes of node v; σ is the activation function.

[0027] Further, step 4 includes:

[0028] For the road network data involved in road network matching, each road map is constructed and its features are extracted, and then road map node embedding learning is performed;

[0029] Calculate the similarity of road segments from different sources, and find matching road segments by comparing the magnitude of road segment similarity. This includes: First, using the nearest neighbor weighted Manhattan distance to calculate the distance d between the high-dimensional vectors of node embeddings from different sources of road network data. veThen, the node weight coefficient w is calculated using the node connection distance d, and w is normalized to obtain w′, thereby obtaining the similarity score of the road segment corresponding to the node:

[0030] w = d / d (1 / d)

[0031] score=(1 / (1+d ve ))w′

[0032] Where d is the Hausdorff distance between roads.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] This invention employs the concept of explicit spatial modeling and designs a novel neighborhood aggregation function for graph neural networks. It introduces spatial relationship operations into road network embedding learning, enhancing the learning representation of road network data by graph embedding, improving the application level of graph neural networks in road network matching, and verifying the effectiveness of the method using real road network data.

[0035] (1) This invention proposes a road network data matching method based on graph neural networks. By constructing a road map and learning graph node embeddings, road network matching based on graph node embeddings is achieved. Three neighborhood aggregation functions, namely spatial, attribute, and hybrid, are designed to improve the ability of graph neural networks to represent the spatial structure of road networks.

[0036] (2) Road network matching experiments were conducted using real road network data. The performance of classic road network matching algorithms and multiple graph neural network road network matching algorithms was compared and studied. The experiments verified the effectiveness of the method of the present invention. Compared with the baseline graph neural network method, the method of the present invention improved the matching accuracy by more than 11% and the recall by more than 6.8% on the experimental area data. Attached Figure Description

[0037] Figure 1 This is a flowchart of a road network data matching method based on graph neural network according to an embodiment of the present invention; in step (2), type, name, and maxspeed represent road attributes, (x,y) represent the starting coordinates of the road segment, and angle represents the angle with the vertical axis; in step (3), P2, P3, and P4 represent the neighbor nodes of P1;

[0038] Figure 2 This is the road map construction process in an embodiment of the present invention; v a v b v c The numbers e represent the corresponding nodes in the graph for each road segment. ab Represents node v a and v b The connected edge, e acRepresents node v a and v c The connected edge, e bc Represents node v b and v c Connecting edges;

[0039] Figure 3 This is an example diagram showing the features of a node road segment in an embodiment of the present invention;

[0040] Figure 4 This is a schematic diagram of the location features of the node road segment in the embodiment of the present invention; the four coordinates (minX,minY) are the corner coordinates of the current tile, (x,y) represents the starting coordinate of the road segment, and angle represents the angle between the line segment connecting (minX,minY) and (x,y) and the vertical axis of the current tile;

[0041] Figure 5 This invention provides a comparison of OSM (red) and TeleAtlas (blue) road network data in the test area of ​​Ansbach in this embodiment of the invention.

[0042] Figure 6 The matching accuracy and recall of different methods under different thresholds in the embodiments of the present invention;

[0043] Figure 7 The matching accuracy of the hybrid aggregation method (GS+SP) under different neighborhood range and neighborhood sampling depth settings in the embodiments of the present invention;

[0044] Figure 8 The OSM and TeleAtlas road network matching results in the test area of ​​this invention are visualized in this embodiment; the red line represents OSM road network data, the blue line represents TeleAtlas road network data, and the black arrows represent the matching correspondence between road networks; (b) and (c) show the matching results of complex road segments, and (d) shows road segments with large differences in road network structure;

[0045] Figure 9 This invention provides a comparison of the road network embedding matching algorithm and the baseline algorithm in typical road segments.

[0046] Figure 10 These are the ablation experiment results of three types of road network features in this embodiment of the invention. Detailed Implementation

[0047] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments:

[0048] like Figure 1 As shown, a road network data matching method based on graph neural networks includes:

[0049] This invention utilizes graph neural networks to learn representations of road network data, calculates road segment similarity based on graph node embedding, and infers road segment matching relationships from different sources, thereby achieving overall road network matching. This invention not only improves road map construction methods to address the differences in multi-source road network data (including introducing stroke extraction and adding location features to road map feature extraction), but also proposes a graph node embedding method based on improved neighborhood aggregation, which can better capture road network structure information and significantly improve the performance of road network matching tasks. The basic process of this invention is as follows: Figure 1 As shown, it mainly includes 4 steps:

[0050] (1) Road graph model construction. The road network data is preprocessed and stroke is extracted. A graph model describing the relationship between road segments is constructed using the road network data. The nodes of the graph represent stroke road segments, and the edges of the graph represent the connection relationship between adjacent road segments.

[0051] (2) Road map feature extraction. The geometric elements of the road segment are automatically encoded as the geometric features of the road segment, the attribute data of the road segment is used as the semantic features of the road segment, and the coordinates and orientation of the road segment are used as the location features of the road segment. The three types of features, namely geometric, semantic and location, are combined as the initial feature description of the road map node.

[0052] (3) Improved neighborhood aggregation road graph node embedding. Three types of neighborhood aggregation functions, namely spatial, attribute, and hybrid, are proposed to learn the graph node embedding of road segments. These functions can better capture the complex neighborhood relationships of the road network and provide a more effective road segment embedding representation for road network matching.

[0053] (4) Node embedding similarity measurement. For road network data from different sources, the graph node embedding of the road network data is learned respectively, and the similarity of road segments between different data sources is calculated using the graph node embedding to determine the road segment matching relationship and realize road network matching.

[0054] 1. Road map model construction

[0055] Road network data is typically represented using graph models, which describe the topological relationships of the road network and provide a basic data structure for analytical tasks such as route planning and traffic prediction. Considering that road network data from different sources does not always contain road travel direction information, this invention uses undirected graphs to construct road maps of the road network data. There are two common methods for constructing road maps: ① treating road segments as graph nodes and the connections between road segments as graph edges; ② also treating the connection points and intersections of road segments as graph nodes and the road segments as graph edges. This invention uses preprocessed road segments as units and employs graph node embedding to represent the characteristics of the road segment itself and its neighborhood. This allows for the calculation of the similarity of road segments in different road network data, thereby determining the road network matching relationship. Therefore, the method of using road segments as graph nodes is selected. The road map construction process is as follows: Figure 2As shown. For a given road network data R, this invention constructs a corresponding road graph G through road network data preprocessing and stroke extraction, road graph node extraction and neighbor node connection. R = {V, E}, where V and E are the set of nodes and the set of edges of the road graph, respectively.

[0056] First, the road network data is preprocessed. Noisy data, such as road segments that are too short or cannot be topologically connected, is filtered out. A stroke extraction operation is then performed on the road network data. Stroke extraction aims to unify the differences in road segment segmentation modeling between road network data from different sources, improving the success rate of road network matching. For example, different sources of road network data may provide different numbers of road segments for the same road entity, leading to the inability to establish road segment matching relationships due to differences in segment representation. Therefore, the original road segments of the road network data need to be processed by breaking them down according to the length of road network elements, intersections, and corners, and the results are used as input for subsequent graph model construction to eliminate inconsistencies in the topological relationship descriptions between different data sources and increase the possibility of finding road matching pairs.

[0057] Then, based on the processed road network data, the nodes and edges of the graph are determined, and a road graph is constructed. The midpoint of the stroke segment is selected as a node in the road graph. Figure 2 (c) Extracting the geometric, semantic, and location features of road segments as initial feature inputs for graph nodes. Specific node feature extraction will be discussed in the next section. After extracting graph nodes, the neighboring nodes of each node are determined based on the distance relationships between road segments, and adjacent nodes are connected to construct graph edges. The neighborhood of a road segment node can be determined directly based on the road segment topological adjacency relationship, or it can be determined by searching for neighboring road segments in geographic space that satisfy a certain distance relationship. This invention adopts the latter, more general method, setting a neighborhood search radius, and then searching for road segment nodes in the neighborhood of each road segment node whose Hausdorff distance is higher than a certain threshold. For example, in Figure 2 (d) Middle road segments a, b, and c (each mapped to road map node v) a v b v c The Hausdorff distances between v and v are all within the set threshold, therefore v a v b and v c Should be in accordance with v a -v b -v c -v a Connect the edges e to form the road graph ab e bc e ac This method brings flexibility to handling parallel roads and complex road segments in road network matching tasks.

[0058] As the scale of road network data (especially the number of road segments) increases, the computational overhead of neighborhood node search increases rapidly. In practical implementations, it is advisable to consider building spatial indexes such as KD-Trees for the road network to improve the efficiency of neighborhood search for road segment nodes.

[0059] 2. Road map feature extraction

[0060] After the road map is constructed, road network data features need to be extracted for the graph nodes as input for graph embedding learning. For road network matching, it is often necessary to comprehensively determine the matching relationship between road networks using both geometric and semantic features. Geometric features of roads include attributes such as the shape, length, and width of the road centerline, while semantic features include road type, grade, and speed limits. Furthermore, this invention adds road location features to represent the spatial relationships of road segments, finally obtaining the feature input for each graph node, such as... Figure 3 As shown.

[0061] 2.1 Geometric Features

[0062] This invention employs automatic encoding of line features to represent the geometric characteristics of road segments. Compared to traditional manual feature design methods, it can better capture the geometric representation of road segments and provides a more flexible and universal representation. However, unlike trajectory data arranged according to the displacement time sequence of moving objects, road line features do not necessarily possess clear or unique directional information (e.g., missing driving direction, two-way roads), and different road network data sources may also use opposite directional representations for the same road segment. This will lead to different embedding vectors calculated by the sequence data encoder, which requires specifying the data order. Therefore, the coordinate order of road features is determined by comparing the latitude and longitude of the two endpoints of the road features, ensuring consistency across multiple road network data sets, thereby making the embedding results of the sequence encoder comparable. To this end, each road segment in the road network data is traversed, and the latitude and longitude of the segment endpoints are compared, with the smaller longitude or latitude being taken as the starting point of the road segment.

[0063] This invention employs an LSTM autoencoder to encode road segments, avoiding discrepancies in the encoder's knowledge of geometry and structure caused by varying sampling rates across different road networks. Through training, the encoder acquires fixed-dimensional vectors of road segment line features, unaffected by data length, structure, or the number of coordinate points. Two thousand road features are selected, and five sets of data are re-collected at intervals of 1m, 10m, 15m, 20m, and 25m as training data input, generating corresponding labels. To ensure the generated embedding vectors are closer to the original data, random noise is added to the original input. This invention uses an encoder to encode the road segment line features of all graph nodes, converting variable-length line features into fixed-length 128-dimensional vectors. For the loss function, this invention uses a contrastive loss function.

[27] To meet the needs of data similarity modeling, the contrastive loss function is defined as:

[0064]

[0065] In the formula: N≤n≤1; L contrast The comparison loss function is used to train the model; N is the number of samples; T is the sample matching index function (it takes a value of 1 when two samples match, and 0 otherwise); D = ||a n -b n ||2 represents the Euclidean distance between the two sample embedding vectors; R is a pre-set threshold representing the radius around the sample embedding space.

[0066] 2.2 Semantic Features

[0067] The same road attribute data from different road network data sources may differ. This is related to factors such as the definition of attribute fields, collection time, and even data quality. When selecting semantic features, road attributes can be selected as semantic feature inputs from the perspectives of attribute universality, consistency, and comparability by comparing the attribute data of the road network data involved in the matching. Based on the data characteristics of the research objects OSM and TeleAtlas, this invention selects road type and road name, normalizes them, and obtains corresponding enumeration values ​​as semantic features of node road segments.

[0068] 2.3 Location Features

[0069] To represent the spatial relationships such as location and direction between road segments, road segment coordinates and azimuth angles are used as location features. For large-scale road network data, the road network can be tiled (e.g., HERE tile partitioning) to improve computational efficiency and reduce computational complexity. Road segment coordinates are represented as the normalized starting point coordinates (x, y) within the geographical range of the road network tiles. The road segment azimuth angle can be determined by the angle between the starting point and the boundary of the road network tiles, such as... Figure 4 As shown.

[0070] 3. Improved Road Graph Node Embedding for Neighborhood Aggregation

[0071] Road map construction and feature extraction transform road network data into effective input data for graph neural networks. To obtain effective road network data embedding representations, a graph embedding learning framework must be selected, and embedding generation algorithms must be designed around the road network matching task. Based on the applicability of graph models to graph data not seen during model training, graph embedding learning frameworks can be divided into transductive and inductive types. Transductive graph models cannot directly apply to new graph topologies, while inductive graph models, represented by GraphSAGE, can generalize extracted patterns to graph data structures not seen during training. For road network matching tasks, the model needs to be applicable to road network data from different regions and with different topologies, making inductive frameworks more practical. Therefore, this invention selects GraphSAGE as the basic architecture of the graph neural network and improves the representation performance of the graph neural network for road network data by improving the neighborhood aggregation function.

[0072] 3.1 GraphSAGE Framework

[0073] The primary goal of the GraphSAGE framework is to learn a generative function for graph node embeddings. This function generates embedding vectors for graph nodes by sampling and aggregating neighborhood features. Sampling involves randomly selecting a certain number of nodes from the target node's neighborhood using a random walk algorithm. Aggregation, on the other hand, utilizes a predefined function to compute features from the node's neighborhood samples, progressively updating the target node's embedding vector until a specified aggregation depth (network layer number) is reached. Overall, the aggregation operation in GraphSAGE allows the model to gather information from local neighborhoods, improving its understanding of the entire graph structure and thus better capturing the complex relationships between nodes.

[0074] The formula for updating the graph node embedding in GraphSAGE can be defined as:

[0075]

[0076] In the formula: It is the embedding vector of node v; f is the embedding vector of node v at the k-th layer; N(v) is the set of neighboring nodes of node v; (k) It is an activation function; W (k) is a learnable parameter; k is the number of network layers (it can also represent the number of aggregators and the number of weight matrices); AGG is the aggregation function, including four basic implementations: Mean, GCN, Pooling, and LSTM. Analysis and experiments revealed that using Mean aggregation is more beneficial for the fusion of adjacent road information because the determination of road network matching relationships is influenced by the common features of adjacent roads. Without affecting generality, different network layers use the same activation function, i.e., f. (k)=σ. The node embedding update formula using Mean aggregation is:

[0077]

[0078] In the formula: |N(v)| is the number of neighboring nodes of node v.

[0079] GraphSAGE supports both unsupervised and supervised training. Before training, a road graph needs to be constructed, and the training and validation sets need to be determined based on the task. Then, neighbor node sampling and feature aggregation operations are performed to update node information. During training, a loss function needs to be calculated to measure the model's performance. Afterwards, the model parameters are updated based on the gradient of the loss function using the backpropagation algorithm. Finally, the model is evaluated and iterated to gradually approach the optimal result.

[0080] GraphSAGE's unsupervised learning is based on the following assumption: node v has similar embeddings to its neighbor u, but is dissimilar to node n, which has no intersection with it. The loss function is:

[0081]

[0082] In the formula: e v The embedding is generated by node v; node u is a neighbor node of node v; u n ~p n (u) indicates negative sampling, meaning node u n It is the negative sampling distribution P from node u. n The samples were obtained from sampling; N is the number of samples.

[0083] 3.2 Improvement of Neighborhood Aggregation Function

[0084] Neighborhood aggregation, a crucial step in the GraphSAGE framework for calculating the embedding vector of target nodes, is an important direction for improving the GraphSAGE algorithm and enhancing the performance of its downstream tasks. For road network data, features such as the spatial layout and road grade of road segments provide important references for node-road segment neighborhood aggregation. However, traditional neighborhood aggregation functions cannot effectively utilize these features, resulting in insufficient learning of the spatial relationships of road network structural elements by the GNN, thus hindering performance improvement in matching tasks. Therefore, this invention improves the GNN neighborhood aggregation function, enhancing its characterization of the spatial and typological relationships between graph nodes and their neighbors, thereby improving the performance of road network structure graph representation. Specifically, this invention improves the neighborhood aggregation function, enhancing the aggregation operation's characterization of the spatial and typological relationships between target nodes and their neighbors, thus improving the performance of road segment representation.

[0085] (1) Spatial aggregation

[0086] Spatial aggregation explicitly considers the spatial positional relationship between the target node road segment and its neighboring road segments in the neighborhood aggregation function to improve the representation performance of road segment embedding on its neighborhood spatial layout. Since GraphSAGE does not consider the spatial positional order of data when processing spatial data, some scholars have compensated for the model's shortcomings in describing such structural features by manually designing spatial features as graph node feature inputs. This modeling method requires relevant expert knowledge and also has the problems of high data dependence and weak model generalization ability. In molecular structure analysis, some scholars have proposed the Spatial Graph Convolution (SGCN) method, which allows the model to directly learn the spatial weights between nodes to adjust the graph convolution operation. Based on GraphSAGE neighborhood aggregation (Equation (2)), this invention introduces a description of the spatial positional relationship between the target node and its neighboring nodes to better capture the spatial relationship between nodes. The spatial aggregation function can be written as:

[0087]

[0088] In the formula: p u and p v These are the coordinates of the midpoint of the road segment represented by nodes u and v, respectively; p u -p v It represents the spatial relationship between nodes, calculated from the Euclidean distance between node positions.

[0089] (2) Attribute aggregation

[0090] Attribute aggregation is a neighborhood aggregation function that selects neighboring nodes of the same type to update the embedding representation of the target node based on the road type relationship between the target node and the corresponding road segments. Considering that the road attribute classification of road network data from different sources may not be the same, this invention divides road types into three categories: Type I roads, Type II roads, and Type III roads, thereby improving the applicability of the attribute aggregation function. Among them, Type I roads include highways, main roads, secondary roads, and tertiary roads; Type II roads include residential roads, community streets, and service roads; Type III roads include pedestrian roads, sidewalks, and bicycle lanes. Based on GraphSAGE neighborhood aggregation (Equation (2)), this invention filters road segments of the same type in the node's neighborhood according to these three road types for updating the target node's embedding. Attribute aggregation can be written as:

[0091]

[0092] In the formula: W is the embedding vector of the (k-1)th layer of a neighboring node u that has the same road type as node v; (k) These are parameters that can be trained.

[0093] (3) Mixed polymerization

[0094] Hybrid aggregation combines spatial aggregation and attribute aggregation operations. First, the spatial relationship between the target node's road segment and its neighboring road segments is explicitly considered in the neighborhood aggregation function. Then, neighboring nodes of the same type are selected based on road type relationships to update the node's embedding representation. Hybrid aggregation can be written as:

[0095]

[0096] 4-node embedding similarity measure

[0097] For the road network data involved in road network matching, separate road maps are constructed and features are extracted, followed by road map node embedding learning. Road segment embedding vectors are used to calculate the similarity of road segments from different sources; matching road segments are found by comparing these similarities. First, the distance d between the high-dimensional node embedding vectors of road network data from different sources is calculated. ve Because the embedding vector has a large dimension, the nearest neighbor weighted Manhattan distance is used for calculation. Then, the node weight coefficient w (Equation (8)) is calculated using the node connection distance d (such as Hausdorff distance) and normalized to obtain w′. This is because road segments that are spatially closer may have stronger correlation, while those that are farther apart have lower correlation. After comparison, linear normalization can obtain better results. Finally, the similarity measure of the road segments corresponding to the nodes is obtained (Equation (9)).

[0098] w = d / d (1 / d) #(8)

[0099] score=(1 / (1+d ve ))w′#(9)

[0100] In the formula, d represents the Hausdorff distance between roads. After obtaining the similarity between nodes, the matching relationship of road segments can be determined by comparing the similarity metric with a pre-set threshold. If the similarity between nodes from two different road network data sources is higher than the set matching threshold, the road segments corresponding to the two nodes are considered a match; otherwise, they are not a match. For the entire road network data, all its road segments are traversed for similarity analysis to complete the road network matching task.

[0101] To verify the effectiveness of the present invention, the following experiment was conducted:

[0102] a. Experimental Data and Experimental Setup

[0103] a.1 Experimental Data

[0104] The road network matching experiment selected Ansbach, Germany, as the test area, covering approximately 99.92 square kilometers. Road network data sources included OpenStreetMap (OSM) and TeleAtlas commercial electronic navigation data. OSM is an open-source, editable global geographic information database containing vector data of elements such as roads, rivers, and buildings. Within the test area, OSM contained 6258 roads, which, after preprocessing, resulted in 12785 roads; TeleAtlas contained 5024 roads, which, after preprocessing, resulted in 7523 roads. To evaluate the performance of different algorithms in road network matching, this invention constructed ground truth values ​​for the Ansbach OSM-TeleAtlas road network matching by combining manual comparison of vector data with comparison of reference high-resolution satellite remote sensing imagery. Figure 5 The test area data is displayed using OSM and TeleAtlas. Overlaying the two road network data shows that the OSM data contains more road segment elements than the TeleAtlas data.

[0105] a.2 Experimental Setup

[0106] This invention implements road graph construction and graph embedding algorithms using Python and TensorFlow, and completes road network matching experiments on a computer platform configured with an AMD EPYC 73522 4-core processor and an NVIDIA RTX A6000 graphics processing unit. In road graph construction, the neighborhood query threshold for graph nodes is set to 100 meters. In feature extraction, after multiple experimental comparisons, the following settings were selected for the LSTM: batch size BS = 128, sequence length SL = 256, hidden layer size HS = 128, and learning rate LR = 0.001. In graph embedding learning, the GraphSAGE unsupervised training method is used, employing a random walk algorithm to generate training data samples for training the graph embedding model. After multiple experimental comparisons, the following settings were selected for GraphSAGE: hidden layer channel number HS = 256, mini-batch size MBS = 512, learning rate LR = 0.00001, neighborhood sampling depth NSD = 3, random walk step size WL = 3, and random walk number of steps NW = 7.

[0107] b Experimental Results

[0108] To verify the effectiveness of the method proposed in this invention, experiments were conducted comparing the following methods: the classic road network matching algorithm Limited Stroke Algorithm (DSO), the Graph Convolutional Network (GCN) method directly applying graph neural networks, the Geospatial Graph Embedding (GraphSAGE, GS) method, and the improved neighborhood aggregation graph embedding methods GS+S, GS+P, and GS+SP proposed in this invention (representing the use of spatial aggregation functions, attribute aggregation functions, and hybrid aggregation functions on the GS framework, respectively). In the experiments, both the graph neural network methods (including the baseline methods GS and GCN) and the method of this invention used the same initial features for the graph nodes.

[0109] b.1 Experimental Evaluation Indicators

[0110] The experimental metrics used are accuracy, recall, and match success rate, which can be calculated as follows:

[0111] accuracy=1-[b / (ac)]#(10)

[0112] recall=(abc) / d#(11)

[0113] match=(ac) / a#(12)

[0114] Where a represents the total number of road segment elements in the experimental area, b represents the number of road segment elements that were incorrectly matched, c represents the number of road segment elements that failed to match (i.e., no matching road segment could be found), and d represents the actual number of road segment elements that should have been matched.

[0115] b.2 Model Parameter Selection

[0116] When performing road network embedding matching, the matching threshold parameter for determining road segment similarity affects the final road network matching result. Therefore, the optimal matching threshold setting is first selected through experimental comparison. For example... Figure 6 As shown, increasing the matching threshold increases the model's matching accuracy but decreases the recall; conversely, decreasing it decreases the accuracy. Considering the calculation metrics under different thresholds, setting the node matching similarity threshold to 0.6 not only yields high accuracy and recall but also comprehensively outperforms the DSO method.

[0117] Furthermore, the neighborhood range of nodes in road graph construction and the neighborhood sampling depth in graph embedding learning also affect model performance. This invention compares and analyzes the settings of these two parameters. Experimental results are as follows: Figure 7As shown, for the neighborhood range, both excessively large and excessively small ranges are detrimental to the acquisition of node features, leading to a decrease in the accuracy of the matching results. The setting of the neighborhood range is affected by the road length in the data; setting the neighborhood range to 100 meters on the experimental data in the test area yields optimal performance. Regarding the neighborhood sampling depth, while increasing the sampling depth expands the range of representation learning and the depth of information transmission, it also increases computation time. Therefore, a balance must be struck between these two factors; both excessively large and excessively small ranges are detrimental to performance improvement. It can be seen that setting the neighborhood sampling depth to 3 on the experimental data in the test area results in the best performance.

[0118] b.3 Analysis of Experimental Results

[0119] After 1153 training epochs, the loss and accuracy of the GS+SP method converged on both the training and validation sets. The accuracy on the training and test sets were 93.63% and 87.27%, respectively, and the recall was 92.02% and 84.37%, respectively. The confusion matrices are shown in Table 1. It can be seen that the trained network has strong generalization ability and good practicality.

[0120] Table 1. Road network matching confusion matrix based on the hybrid aggregation method (GS+SP)

[0121]

[0122] Table 2 presents an experimental comparison between the proposed methods (GS+S, GS+P, GS+SP) and three baseline methods: DSO, GCN, and GS. It can be seen that compared to the GS and GCN methods, the DSO method exhibits the best performance in both accuracy and recall, thus directly applying the graph neural network algorithm cannot achieve optimal results. The proposed GS+S and GS+P methods show significant performance improvements over GS, and their accuracy surpasses that of the DSO method, but their recall and matching success rates are slightly lower. The GS+SP method, however, surpasses all three baseline methods in all metrics, fully demonstrating its effectiveness. By comparing the three improved neighborhood aggregation road network embedding matching methods (GS+S, GS+P, and GS+SP), it can be seen that different aggregation methods contribute differently to the experimental results. GS+S significantly improves accuracy compared to the GS method, but its recall does not change significantly, thus verifying that spatial aggregation incorporating road segment spatial relationships can effectively improve matching accuracy. While GS+P improves accuracy compared to the GS method, its contribution to recall is greater, effectively verifying the performance improvement, especially the contribution to recall, of attribute aggregation. The GS+SP method improves both precision and recall compared to other methods, combining the advantages of the previous two methods, which fully demonstrates the contribution of hybrid aggregation to performance improvement.

[0123] Table 2 Performance comparison between the road network embedding matching algorithm and the baseline algorithm of this invention.

[0124]

[0125] To understand the effectiveness of GS+SP, this invention visualizes the road network matching results. For example... Figure 8 As shown, the algorithm of this invention can effectively match road network data from different sources. In particular, the algorithm can successfully achieve road network matching in more complex road sections such as intersections and grid sections. Figure 8 As shown in (b) and (c), the black arrows between different colored road networks represent the matching relationships between road segments calculated by the method of this invention. However, the algorithm failed to effectively match road segments with large differences in road network structure, such as... Figure 8 (d) is shown in the circle.

[0126] To analyze the advantages and characteristics of the road network embedding matching method of the present invention compared with the baseline method, the present invention compares the differences in road network matching results under typical road network structures. Figure 9 The visualization results of road network matching relationships under different matching algorithms are shown. Black arrows between different colored road networks represent correct road segment matches found by the corresponding method, while green arrows represent incorrect matches, marked with red circles. It can be seen that sample road 1 is a highway with many parallel road segments. Except for GS+SP and GS+S, other methods show incorrect matches because the parallel road segments are close together, leading to the incorrect identification of closely spaced road segments in different datasets as matched road segments, without fully considering the spatial relationships between road segments. Sample roads 2 and 3 are rural roads with many intersecting road segments. The DSO, GCN, and GS methods have many problems, not only showing incorrectly matched road segments but also missing matches. The three methods proposed in this invention all show good results for this sample road segment. Sample roads 4 and 5 are community neighborhood roads with varying degrees of grid structure. Except for GS+SP and GS+S, other methods show a certain number of incorrect matches. This is because there are closely spaced non-matching road segments in different datasets, affecting the judgment of matching results. In summary, the DSO, GCN, GS, and GS+P methods are prone to mismatches in complex road conditions, especially when there are closely spaced road segments, which are easily misidentified as matching roads. The GCN, GS, and GS+S methods incorrectly determine that road segments that are far apart cannot be matched. The GS+S and GS+SP methods can effectively distinguish matching relationships in complex road conditions, and the GS+SP method can also successfully match some far-distance road segments.

[0127] In summary, the method of this invention can find roads with the same name from different data sources on road segments with complex road network structures, such as township roads and community streets, and performs better in terms of matching accuracy and recall. Compared with traditional road network matching methods, the method of this invention only uses the basic attribute information of road network data and achieves performance improvement in road network matching tasks through road network structure features automatically learned by GNN; compared with GNN road network matching methods, the method of this invention achieves a significant improvement in matching accuracy by improving the neighborhood aggregation function under the condition that the initial features of GNN are the same.

[0128] b.4 Characteristic Ablation Experiment

[0129] This invention utilizes geometric, semantic, and location features of road segments in road network matching GNN modeling (including baseline GCN, GS method, and the GS+S, GS+P, and GS+SP methods proposed in this invention). To understand the impact of these three types of features on the performance of the GNN road network matching algorithm, feature ablation experiments were conducted, i.e., location, semantic, and geometric features were removed sequentially, and the matching accuracy and recall of the GNN road network matching algorithm were compared. Figure 10 As shown.

[0130] Experiments show that all GNN methods experience varying degrees of decrease in matching accuracy and recall when not using all features. Among the GCN, GS, and GS+P methods, the accuracy decrease is smallest when using only "geometric + semantic" features, and largest when using "semantic + location" features. Therefore, it can be seen that road location features have a smaller impact on GNN matching performance, while geometric features have a larger impact. The GS+S and GS+SP methods show the smallest accuracy decrease when using only "geometric + location" features, indicating that location features significantly improve accuracy in these two methods. This is because both methods use spatial neighborhood aggregation, utilizing the spatial relationships of roads during node embedding computation. From the impact of features on recall, we can see that the methods exhibit roughly the same performance trend, but the GS+P and GS+SP methods show significantly better performance than the other methods when using only "geometric + semantic" and only "semantic + location" features. This is because these two methods use attribute neighborhood aggregation, utilizing the semantic features of roads during node embedding computation.

[0131] In summary, this invention proposes a road network data matching method based on graph neural networks (GCNs) for road network matching. This method employs an explicit spatial modeling approach, utilizing graph node embedding updates based on road space and type information. By improving the neighborhood aggregation function of the GCN, it enhances the processing and utilization of spatial relationships and road types during the graph embedding learning stage, thereby improving the GCN's ability to represent complex road network structures and ultimately enhancing the performance of GCN-based road network matching. Multi-source road network matching experiments were conducted using real road network data. The results show that the proposed method outperforms classic stroke-limited road network matching algorithms and baseline graph neural network (GCN and GraphSAGE) road network matching algorithms in terms of matching accuracy, recall, and success rate. Compared to baseline graph neural network methods, the proposed method improves matching accuracy by over 11% and recall by over 6.8% on experimental data. The proposed method also performs well in complex environments such as grid sections and parallel sections. This invention provides an effective graph embedding learning framework for road network matching tasks, significantly improving road network matching performance.

[0132] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A road network data matching method based on graph neural networks, characterized in that, include: Step 1: Preprocess the road network data and extract strokes. Use the road network data to construct a graph model describing the relationship between road segments. Use the nodes of the graph to represent stroke road segments and the edges of the graph to represent the connection relationship between adjacent road segments. Step 2: Automatically encode the geometric elements of the road segment as the geometric features of the road segment, use the attribute data of the road segment as the semantic features of the road segment, and use the coordinates and orientation angles of the road segment as the location features of the road segment. Combine the three types of features—geometric, semantic, and location—as the initial feature description of the road map node. Step 3: Based on the GraphSAGE framework, using the initial features of the road graph nodes as input, learn the graph node embeddings of the road segments using spatial neighborhood aggregation functions, attribute neighborhood aggregation functions, or hybrid neighborhood aggregation functions; Step 4: For road network data from different sources, learn the graph node embedding of the road network data respectively, and use the graph node embedding to calculate the road segment similarity between different data sources, determine the road segment matching relationship, and realize road network matching; The spatial neighborhood aggregation function is: in Nodes in the case of spatial aggregation No. Layer embedding vectors, and These are nodes and The coordinates of the midpoint of the road segment are indicated. It represents the spatial relationship between nodes, calculated from the Euclidean distance between node positions; These are learnable parameters; It is a node The set of neighboring nodes; It is a node The number of neighboring nodes; It is an activation function; The attribute neighborhood aggregation function is: in Nodes in the case of attribute aggregation No. Layer embedding vectors, It is with nodes Neighboring nodes with the same road type The Layer embedding vector; These are parameters that can be trained; It is a node The set of neighboring nodes; It is a node The number of neighboring nodes; It is an activation function; The hybrid neighborhood aggregation function is: in Nodes in the case of hybrid polymerization No. Layer embedding vectors, and These are nodes and The coordinates of the midpoint of the road segment are indicated. It represents the spatial relationship between nodes, calculated from the Euclidean distance between node positions; These are learnable parameters; It is a node The set of neighboring nodes; It is a node The number of neighboring nodes; It is an activation function.

2. The road network data matching method based on graph neural networks according to claim 1, characterized in that, Step 1 includes: First, the road network data is preprocessed to filter out noisy data and then stroke extraction is performed on the road network data. Then, based on the processed road network data, the nodes and edges of the graph are determined, a road graph is constructed, the midpoint of the stroke segment is selected as the node of the road graph, the neighboring nodes of each node are determined according to the distance relationship between road segments, and the adjacent nodes are connected to construct the edges of the graph.

3. The road network data matching method based on graph neural networks according to claim 2, characterized in that, The neighboring nodes of each node are determined in the following way: Set a neighborhood search radius, and then search for road segment nodes in the neighborhood whose Hausdorf distance is higher than a certain threshold for each road segment node.

4. The road network data matching method based on graph neural networks according to claim 1, characterized in that, In step 2, when extracting the geometric features of the road segment, each road segment in the road network data is traversed, the latitude and longitude of the road segment endpoints are compared, and the one with the smaller longitude or latitude is taken as the starting point of the road segment.

5. The road network data matching method based on graph neural networks according to claim 1, characterized in that, The semantic features of the road segment include road type and road name.

6. The road network data matching method based on graph neural networks according to claim 1, characterized in that, The coordinates of the road segment are represented as the normalized starting point coordinates within the geographical range of the road network tiles, and the azimuth angle of the road segment is determined by the angle between the starting point and the boundary of the road network tiles.

7. The road network data matching method based on graph neural networks according to claim 1, characterized in that, Step 4 includes: For the road network data involved in road network matching, each road map is constructed and its features are extracted, and then road map node embedding learning is performed; Calculate the road segment similarity from different source road network data, and find matching road segments by comparing the magnitude of road segment similarity. This includes: First, using the nearest neighbor weighted Manhattan distance to calculate the distance between the high-dimensional vectors of node embeddings from different source road network data. Then, using the connection distance of the nodes Calculate node weight coefficients and to Normalization is performed to obtain This leads to the similarity score of the road segment corresponding to the node. : in The distance between the roads is the Hausdorf distance.

Images