A multi-graph hierarchical road network representation method based on a graph neural network

Abstract

The application discloses a kind of multi-graph hierarchical road network representation methods based on graph neural network, in the method, the structure perception graph neural network of spectral clustering and graph attention network is formed, modeling is carried out to road network representation of different levels, two kinds of virtual nodes, namely structure area and functional area, are introduced, multi-graph mechanism is used to guide virtual node to correspond with the structure area and functional area of real world, structural similarity is established using road type attribute, functional similarity between road segments is defined using city POI information, message sharing is then performed at high level, then updated information is propagated to low level nodes, and functional attributes of road network are supplemented;Using the method, road network representation containing structural features and functional features can be obtained;It is convenient to determine road network structure and functional role;It is helpful to reveal the functional area of city, helpful to route planning and time of arrival estimation and position prediction, and conducive to the construction of intelligent transportation system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent transportation, more particularly, to a multi-graph hierarchical road network representation method based on graph neural network. BACKGROUND

[0002] At present, intelligent transportation system (ITS) has become an indispensable part of people's daily life and plays a crucial role in various transportation applications, such as route planning, arrival time estimation and next location prediction. Road network, as the infrastructure of ITS, plays a very important role in various transportation-related systems and applications.

[0003] Due to its important role, it is necessary to develop a suitable, especially general, method to effectively represent and model road networks. Early studies mainly constrained the road network topology and developed models using standard graph data structures. With the development of deep learning, recent studies have shifted to using graph neural networks to learn the representation of road segments. In this way, the basic features of the road network can be effectively developed and utilized, which is expected to improve the performance of downstream applications.

[0004] However, due to the complexity of the road network, it is not easy to develop an effective representation learning model. In previous studies, at least three major problems have not been well addressed. First, the road network is not "flat". Road segments are naturally divided into different traffic units "clusters", including structural units (such as transportation hubs) or functional units (such as business districts). In addition, the roles of different road segments in the road network are not "equal". Some traffic units play an increasingly important role in various transportation tasks. However, previous studies usually use standard graph neural networks to treat different nodes equally, ignoring their hierarchical characteristics. Second, the road network may not be "small world", it often has a longer average path. For example, the relationship between different road segments in the road network will far exceed the "six degrees of separation" theory. However, in standard graph neural networks, they only aggregate information from nearby nodes, which cannot effectively capture long-distance dependencies between nodes. Third, the road network mainly reflects structural features, and other aspects of information may not be obtained through network structure. For example, it is usually difficult to determine the functional role (such as shopping center) of the traffic unit only according to the road connection of the traffic unit.

[0005] Therefore, it is an urgent problem for those skilled in the art to provide a multi-graph hierarchical road network representation method based on graph neural network. SUMMARY

[0006] In view of the above technical problems, the present application provides a multi-graph hierarchical road network representation method based on a graph neural network, which can effectively obtain a road network representation containing structural features and functional features, facilitate the determination of road network structure and functional role, help reveal the functional areas of a city, facilitate route planning, arrival time estimation and location prediction, and facilitate the construction of an intelligent transportation system.

[0007] To achieve the above object, the present application adopts the following technical solutions:

[0008] A multi-graph hierarchical road network representation method based on a graph neural network comprises the following steps:

[0009] S1, representing a road segment by features;

[0010] S2, establishing the relationship between road segments, constructing a natural connectivity adjacency matrix, a structural similarity matrix and a functional similarity matrix between road segments;

[0011] S3, assuming that each road segment corresponds to a single structural area, different road segments have different importance in a structural area, and a structure perception graph neural network composed of spectral clustering and a graph attention network is used to obtain a structural area assignment matrix representing the important connection between road segments and structural areas; and the structural area is associated with the road segment;

[0012] S4, constructing a functional area based on a structural area with the same function, and constructing a functional area assignment matrix, representing the functional area as a weighted linear combination of the structural area representation based on the functional area assignment matrix, and obtaining the adjacency matrix of the functional area node;

[0013] S5, updating the representation of the functional area, and sequentially transmitting the updated information through the functional area, the structural area and the road segment.

[0014] Further, the step S1 comprises:

[0015] For a given road segment, the road segment serial number ID, the number of lanes LN, the road segment length SL and the longitude and latitude LL are always considered as features for context embedding, for the connection feature, the entire value range is divided into several connection intervals, and interval numbers are used for feature coding, a unique embedded connection is set for each value, the connection vector is concatenated as the context embedding, and the graph node embedding is initialized:

[0016]

[0017] wherein, v ID represents a feature vector corresponding to the road segment serial number ID, v LN represents a feature vector corresponding to the number of lanes LN, v SL represents a feature vector corresponding to the road segment length SL, and vLL denotes the feature vector corresponding to the latitude and longitude LL, denotes the feature vector of the road segment s i , || denotes the vector concatenation operation, and the feature V of all nodes is the initial road segment feature N S .

[0018] Further, the step S2 includes:

[0019] S21, establish the adjacency matrix of the road segment based on the natural connectivity of the road network, denoted as If one road segment is upstream of another road segment, the corresponding value in is 1, otherwise 0. For bidirectional road segments, add two directed connections by reversing their start and end point vertices;

[0020] S22, design a structure code according to the road type. Suppose there are m types of road types, then generate an m-dimensional vector i for each road segment s In , only the position corresponding to the type of road segment s i is 1, and the rest are 0. Then concatenate the structure code of each road segment s i and the sum of the structure codes of the surrounding road segments adjacent to s i , and perform normalization operation:

[0021]

[0022] wherein, denotes the structure code feature of the road segment s i , denotes the unique code vector of the road segment s j , || denotes the concatenation operation, N(s i ) denotes the set of road segments adjacent to the road segment s i , ||x||2 denotes the modulus of the vector x, and a structure matrix is obtained, wherein k S is the number of road segments, and each row in F represents the structure feature of a road segment in the road network. The cosine similarity is used to measure the structural similarity of two road segments:

[0023] SS = F·F T

[0024] Generate a mask matrix to filter some edges with small similarity:

[0025]

[0026] wherein, SS mask is the mask matrix, and SS mask[i,j] represents the element located in the ith row and jth column, θ1 is a similarity threshold, the structural similarity is multiplied by the mask matrix, and the final structural similarity matrix is generated through an activation function

[0027]

[0028] wherein, represents element-wise multiplication of matrices, and σ represents a sigmoid function

[0029] S23, generate a one-hot encoding for each POI according to the type of POI, assume that there are n types of POIs, and generate a one-hot encoding for each POI p j generate an n-dimensional vector In , only the positions of POI p j of the same type are 1, and the remaining positions are 0, and the influence of each POI on its surrounding road segments is considered:

[0030]

[0031] wherein, p j represents a point of interest, dis(p j , s i ) represents the distance from POI p j to road segment s i , θ2 is a distance threshold, and the influence of all POIs on road segment s i forms the functional characteristics of road segment s i :

[0032]

[0033] obtain a functional matrix G, each row of which represents the functional characteristics of a road segment in the road network, R represents a linear space, k S is the number of road segments, and n represents the dimension; then, the cosine similarity is used to measure the functional similarity of two road segments, and the final functional similarity matrix is generated through an activation function

[0034]

[0035] wherein, σ represents a sigmoid function, and the superscript T represents transposition.

[0036] Further, the step S3 comprises:

[0037] S31, performing structural region division by using a spectral clustering algorithm, and giving an adjacency matrix of road segments by subtracting a diagonal matrix DS deriving its graph Laplacian matrix L S , computing the first d' eigenvectors u1,…u d' , obtaining a matrix composed of d' eigenvectors R represents a linear space, represents a linear space composed of k S xd' matrix, a hard mapping M1 from locations to structure regions is obtained by running a standard K-means algorithm on the matrix U;

[0038] an adjacency matrix represents the natural connectivity between road segments, is a structure similarity matrix of road segments, encoding structure similarity as connections between road segments, two graph attention networks are used to model these two graphs respectively, and they are aggregated together to obtain a soft mapping M2, i.e.

[0039]

[0040]

[0041]

[0042] where GAT is a graph attention network, softmax(·) is a row normalization function, N S represents a road segment representation; finally, the hard mapping and the soft mapping are combined together to obtain the final constructed structure zone assignment matrix A SR :

[0043] A SR = softmax(aM1 + (1-a)M2)

[0044] where a is a hyperparameter that controls the fusion of the two mappings;

[0045] S32, the structure zone representation is associated with the road segment representation by A SR :

[0046]

[0047] where N R represents a structure zone feature;

[0048] At the same time, the connection relationship of the structure zone and the connection relationship of the road segment are associated:

[0049]

[0050] where, represents a structure zone connection matrix.

[0051] Further, the step S4 comprises:

[0052] S41, obtaining a functional adjacency matrix From road segments to structure zones:

[0053]

[0054] wherein, denotes a structure zone functional similarity matrix;

[0055] Two graph attention networks are adopted to model the natural connectivity and functional similarity of structure zones respectively, and they are aggregated together to obtain the final functional zone assignment matrix A RZ i.e.

[0056]

[0057]

[0058]

[0059] S42, using the functional zone assignment matrix A RZ , the functional zone representation is expressed as a weighted linear combination of the structure zone representation:

[0060]

[0061] Using N Z , the adjacency matrix of the functional zone node is further derived:

[0062]

[0063] wherein ReLU is a linear rectifier function, and θ is an adjustable hyperparameter.

[0064] Further, the step S5 comprises:

[0065] S51, performing functional zone layer update, updating the functional zone representation, and preparing to pass the information to the next level, using a standard graph convolutional network to update the functional zone embedding:

[0066]

[0067] wherein GCN denotes a graph convolutional network, denotes the functional zone representation in t rounds of iteration, A Z denotes the functional zone adjacency matrix, denotes the functional zone representation in t+1 rounds of iteration;

[0068] Then, the functional zone update information is passed to the next layer to update the structure zone representation:

[0069]

[0070]

[0071] wherein, is the structural zone representation in the t-th iteration, g ZR is the gating unit function from the functional zone to the structural zone, represents the Hadamard product, i.e., element-wise multiplication of matrices, A RZ is the assignment matrix from the structural zone to the functional zone, is the functional zone representation in the t+1-th iteration, w ZR is a learnable parameter;

[0072] Sigmoid represents an activation function, whose formula is:

[0073]

[0074] wherein, e represents a natural constant, and x represents the input of the sigmoid function;

[0075] S52, in the structural zone, first, a standard GCN is used to update the embedding representation of the layer:

[0076]

[0077] Then, the structural zone embedding is forwarded to the next layer to update the link representation:

[0078]

[0079]

[0080] wherein, w RS represents a learnable parameter;

[0081] S53, the updated link representation is combined with the information of the structural zone and the functional zone, and the graph attention network is used to model the relationship between link nodes as

[0082]

[0083] The operation is repeated to obtain the link network representation containing the structural features and the functional features.

[0084] According to the technical solution, compared with the prior art, the present application has at least the following technical effects:

[0085] The application discloses a multi-graph hierarchical road network representation method based on a graph neural network.

[0086] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art upon examination of the following or can be learned by practice of the present application. The objects and other advantages of the present application can be realized and attained by the structure particularly pointed out in the written description and claims hereof.

[0087] The technical solutions of the present application will be further described in detail below with the help of the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0088] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the provided drawings.

[0089] Figure 1 A flowchart of a multi-graph hierarchical road network representation method based on a graph neural network provided by the present application.

[0090] Figure 2 A structural perception graph neural network model structure diagram provided by the present application.

[0091] Figure 3 A structural and functional area diagram generated by HRNR and HRNR+ provided by the present application. DETAILED DESCRIPTION

[0092] The technical solutions in the embodiments of the present application will be described clearly and completely below with the help of the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0093] Referring to Figure 1 The embodiments of the present application disclose a multi-graph hierarchical road network representation method based on a graph neural network, including the following steps:

[0094] S1, a road section is given, and a feature representation thereof is performed;

[0095] S2, a relationship between road sections is established, and three adjacency matrices between road sections are constructed, including a natural connectivity matrix, a structural similarity matrix, and a functional similarity matrix;

[0096] S3, a single structural zone corresponding to each road section is set, different road sections have different importance in a structural zone, and a structural zone distribution matrix A representing the important connection between the road section and the structural zone is obtained by a structure perception graph neural network composed of spectral clustering and a graph attention network SR ; and the structural zone is associated with the road section;

[0097] S4, a functional zone is constructed based on the structural zone with the same function, and a functional zone distribution matrix A RZ is constructed, the functional zone representation is represented as a weighted linear combination of the structural zone representation based on the functional zone distribution matrix, and an adjacency matrix of the functional zone node is obtained;

[0098] S5, the representation of the functional zone is updated, and the updated information is sequentially transmitted through the functional zone, the structural zone, and the road section.

[0099] The above steps will be described in detail as follows:

[0100] Step S1 specifically includes:

[0101] A road section is given, and the road section number ID, the number of lanes LN, the road section length SL, and the latitude and longitude LL are always contextual embedded features. For the connection feature, the entire value range is divided into several connection intervals, and the interval number is used for feature coding. A unique embedded connection is set for each value, and the connection vector is concatenated as a contextual embedding, and the graph node embedding is initialized:

[0102]

[0103] wherein, v ID represents a feature vector corresponding to the road section number ID, v LN represents a feature vector corresponding to the number of lanes LN, v SL represents a feature vector corresponding to the road section length SL, v LL represents a feature vector corresponding to the latitude and longitude LL, represents a feature vector of the road section s i , || represents a vector splicing operation, and the feature V of all nodes is the initial road section feature N S , N S ← V.

[0104] Step S2 specifically includes:

[0105] S21. Establish an adjacency matrix for road segments based on the natural connectivity of the road network, using... This means that if one road segment is upstream of another road segment, then... The corresponding value is 1, otherwise it is 0. For bidirectional segments, simply add two directed connections by reversing their start and end vertices.

[0106] S22. Design a structural code based on road types. Assuming there are m types of roads, then the code for each road segment is s. i Generate an m-dimensional vector exist In the middle, only those that meet the requirements of road segment s i The category is set to 1, and all other positions are 0. Then, for each road segment s... i Structure encoding and s i The structural codes of adjacent surrounding road segments are concatenated and then normalized.

[0107]

[0108] in, Indicates road segment s i The structural coding features, Indicates road segment s j A unique encoded vector, || represents the concatenation operation, N(s i ) indicates the relationship with road segment s i The set of adjacent road segments, where ||x||2 represents the magnitude of vector x, yields a structure matrix. Where k S Given the number of road segments, each row in F represents the structural characteristics of a road segment in the road network. Cosine similarity is used to measure the structural similarity between two road segments:

[0109] SS = F·F T

[0110] Generate a mask matrix to filter edges with low similarity:

[0111]

[0112] Among them, SS mask For the mask matrix, SS mask [i,j] represents the element located in the i-th row and j-th column, θ1 is a similarity threshold, the structural similarity is multiplied by the mask matrix, and then passed through an activation function to generate the final structural similarity matrix.

[0113]

[0114] in, This represents element-wise matrix multiplication, where σ represents the sigmoid function, expressed as:

[0115]

[0116] S23. Generate a one-hot code for each POI (Point of Interest) based on its type. Assume there are n types of POIs, and for each POI p... j Generate an n-dimensional vector exist In the middle, only those that conform to POIp j The type position is 1, and all other positions are 0. Consider the impact of each POI on its surrounding road segments:

[0117]

[0118] Where dis(p) j ,s i ) indicates POI p j To section s i The distance θ2 is a distance threshold, and all POIs are for road segment s. i The impact formed road segment s i Functional features:

[0119]

[0120] A function matrix was obtained. Each row in G represents the functional characteristics of a road segment in the road network. Cosine similarity is then used to measure the functional similarity between two road segments, and an activation function is applied to generate the final functional similarity matrix.

[0121]

[0122] Combined with appendix Figure 2 As shown, step S3 specifically includes:

[0123] S31. Use spectral clustering algorithm to divide structural regions, given the adjacency matrix of road segments. By subtracting the diagonal matrix D S Derive its graphical Laplacian matrix L S Calculate the first d' eigenvectors u1,…u d' This yields a matrix composed of the eigenvectors of d′. R represents a linear space. Indicates by k S The linear space formed by ×d'-dimensional matrices is used to obtain the hard mapping M1 from location to structural regions by running the standard K-means algorithm on matrix U;

[0124] Adjacency matrix representing the natural connectivity between road segments, is a structure similarity matrix for road segments, encoding structure similarity as a new type of connection between road segments, using two graph attention networks to model these two graphs respectively, and aggregating them together to get the soft mapping M2, i.e.

[0125]

[0126]

[0127]

[0128] where GAT is a graph attention network, softmax(·) is a row normalization function, and finally the hard mapping and the soft mapping are combined together to get the final constructed structure zone allocation matrix A SR :

[0129] A SR = softmax(αM1+(1-α)M2)

[0130] where α is a hyperparameter that controls the fusion of the two mappings,

[0131] S32, associate the structure zone representation with the road segment representation through A SR :

[0132]

[0133] where N R represents the structure zone feature,

[0134] At the same time, associate the connection relationship of the structure zone and the connection relationship of the road segment:

[0135]

[0136] where, represents the structure zone connection matrix.

[0137] Step S4 specifically includes:

[0138] S41, construct a functional adjacency matrix from road segments to structure zones:

[0139]

[0140] where, represents the structure zone functional similarity matrix;

[0141] Two other graph attention networks are adopted to model the natural connectivity and functional similarity of the structural regions respectively, and they are aggregated together to obtain the final functional region assignment matrix A RZ , i.e.

[0142]

[0143]

[0144]

[0145] S42, using the functional region assignment matrix A RZ , the functional region representation is expressed as a weighted linear combination of the structural region representation:

[0146]

[0147] Using N Z , the adjacency matrix of the functional region nodes is further derived:

[0148]

[0149] Where ReLU is a linear rectifier function, and theta is an adjustable hyperparameter.

[0150] Step S5 specifically includes:

[0151] S51, update the functional region layer, update the functional region representation, and prepare to pass this information to the next level, using a standard graph convolutional network to update the functional region embedding:

[0152]

[0153] Where GCN represents a graph convolutional network, represents the functional region representation in t rounds of iteration, A Z represents the functional region adjacency matrix, represents the functional region representation in t+1 rounds of iteration

[0154] Then, the functional region update information is passed to the next layer to update the structural region representation:

[0155]

[0156]

[0157] Where, the structural region representation in t rounds of iteration, g ZR is a gating unit function from the functional region to the structural region, and represents the Hadamard product, i.e. the matrix elements are multiplied one by one, A RZA structure-to-function assignment matrix, A function representation in the t+1 iteration,

[0158] Sigmoid is a kind of activation function, whose formula is:

[0159]

[0160] where w ZR is a learnable parameter.

[0161] S52, in the structure area, first update the embedding representation of this layer using the standard GCN:

[0162]

[0163] Then, the structure area embedding is forwarded to the next layer to update the link representation:

[0164]

[0165]

[0166] S53, combine the updated link representation with the information of the structure area and the function area, and use the graph attention network to model the relationship between the link nodes as

[0167]

[0168] Repeat the above steps to obtain the link network representation containing structure features and function features.

[0169] After training is completed, the link-to-structure and structure-to-function assignment matrices A SR and A RZ can be obtained. Using the assignment matrix, the probability distribution of each link belonging to each structure area and the probability distribution of each structure area belonging to each function area can be obtained, and then the link can be clustered, and the links belonging to the same structure area are marked as the same color as the structure area, and further, the links belonging to the same function area are marked as the same color as the function area, that is, the work of high-level semantic clustering of the urban road network can be completed, and then the division method can be analyzed.

[0170] The following verifies the method with a specific embodiment:

[0171] Referring to Figure 3 , as shown in Figure 3 (a) and Figure 3(b) The structural regions constructed by the state-of-the-art method (HRNR) and the proposed method (HRNR+) near the Beijing Zuojiazhuang Bridge are given. The Zuojiazhuang Bridge is the gateway from the city center to the Capital International Airport. It can be seen that in HRNR+, the bridge part is clearly separated from the residential part. While in HRNR, the residential area and the bridge part are mixed together. From the structural similarity, the bridge segment and the residential area are mostly highway and residential road respectively, which helps HRNR+ to distinguish different structural regions.

[0172] Similarly, Figure 3 (c) and Figure 3 (d) are the functional regions generated by HRNR and HRNR+ respectively. The three famous commercial areas in Beijing, Wangfujing, Guomao and Sanlitun, are marked by the icons of stores. It can be seen that in HRNR+, the commercial areas are separated from the surrounding residential areas, while in HRNR, they are mixed together. The functional difference between the commercial area and the residential area helps to divide the functional regions more reasonably. In addition, the three commercial areas belong to the same cluster. This can be explained as follows: using the functional similarity matrix, regional clustering can break through the limitation of geographical connection. This is meaningful for downstream tasks, because similar functional regions may share certain attributes, which makes it easier to distinguish different regions in high-dimensional space.

[0173] From the description of the above embodiments, those skilled in the art can know that the core idea of the present application is to establish a hierarchical graph neural network to model different levels of road network representation. In particular, by aggregating fine-grained road segments, high-level traffic units such as the aforementioned structural clusters and functional clusters are obtained. To this end, two kinds of virtual nodes, namely structural regions and functional regions, are introduced. The structural region is mainly used to represent the spatially connected road segments and plays a certain traffic role, such as an overpass, a crossroads, etc. In addition, the functional region is formed on the basis of the structural region to provide a certain function for the traveling population, such as a shopping area. In addition, a multi-graph mechanism is used to guide the virtual nodes to correspond to the structural regions and functional regions in the real world. In addition to the natural connectivity of the road network, the structural similarity is established by using the road type attribute, and the functional similarity between road segments is defined by using the city POI information. Using such a three-level organization can alleviate the problem of dependence on remote nodes, first performing message sharing at a high level, and then propagating the updated information to low-level nodes. The multi-graph mechanism complements the functional attributes of the road network. In addition, real trajectory data is combined to reveal the functional regions of the city. The trajectory data of users can be used to discover potential functional or lifestyle-related patterns; it is helpful for route planning as well as arrival time estimation and location prediction, and is conducive to the construction of an intelligent transportation system, and is conducive to providing a certain function for the traveling population, such as a shopping area.

[0174] The various embodiments described in this specification are implemented in a progressive manner, each embodiment focusing on the differences from other embodiments, and the same or similar parts between embodiments can be mutually referred to. For the apparatus disclosed by the embodiments, since it corresponds to the method disclosed by the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0175] The above description of disclosed embodiments enables one of ordinary skill in the art to make or use the application. Various modifications to these embodiments will be readily apparent to those of ordinary skill in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Therefore, the application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-graph hierarchical road network representation method based on a graph neural network, characterized in that, The method comprises the following steps: S1, a road section is given, and a feature representation is performed on the road section to obtain an initial road section representation; S2, a relationship between road sections is established, and a natural connectivity adjacency matrix, a structural similarity matrix and a functional similarity matrix are constructed between the road sections; S3, a single structural zone corresponding to each road section is set, different road sections have different importance in a structural zone, a structural zone distribution matrix representing the important connection between the road section and the structural zone is obtained based on the adjacency matrix and the structural similarity matrix through a structure perception graph neural network composed of spectral clustering and a graph attention network; The structural zone representation and the road section representation are associated through the structural zone distribution matrix to obtain a structural zone representation, and the connection relationship of the structural zone and the connection relationship of the road section are associated to obtain a structural zone connection matrix; S4, a functional zone is constructed based on structural zones with the same function, the functional similarity matrix is transferred from the road section to the structural zone to obtain a structural zone functional similarity matrix, and the natural connectivity and the functional similarity of the structural zone are modeled by using two graph attention networks based on the structural zone connection matrix and the structural zone functional similarity matrix to construct a functional zone distribution matrix, the functional zone representation is represented as a weighted linear combination of the structural zone representation based on the functional zone distribution matrix, and the functional zone representation is further derived from the adjacency matrix of the functional zone node; S5, based on the adjacency matrix of the functional zone node, a standard graph convolution network is used to update the functional zone representation, and the updated information is sequentially transmitted among the functional zone, the structural zone and the road section.

2. The multi-graph hierarchical road network representation method based on a graph neural network according to claim 1, characterized in that, The step S1 specifically comprises: Given a road segment, consider the road segment serial number , the number of lanes , the length of the road segment , and the latitude and longitude Always feature context embedding, for contact features, divide the entire value range into several contact intervals, and use interval numbers to encode features, set a unique embedding for each value. Connection, concatenate the connection vector as the context embedding, and initialize the graph node embedding: ； wherein, represents the road segment number corresponding feature vector, represents the number of lanes corresponding feature vector, represents the length of the road segment corresponding feature vector, represents the longitude and latitude corresponding feature vector, represents the road segment feature vector, represents the vector concatenation operation, the features of all nodes is the initial road segment representation .

3. The multi-graph hierarchical road network representation method based on graph neural network according to claim 2, characterized in that, The step S2 specifically comprises: S21, building the adjacency matrix of the road segments based on the natural connectivity of the road network, using to indicate that if one road segment is upstream of another road segment, then the corresponding value is 1, otherwise 0, and for bidirectional road segments, adding two directed connections by reversing their start and end vertexes; S22. Design a structural code based on road type. Assume there are a total of road types. The category is for each road segment. Generate a dimensional vector ,exist Among them, only the sections that meet the requirements are eligible. The category is set to 1, and all other positions are 0. Then, for each road segment... Structure encoding and The structural codes of adjacent surrounding road segments are concatenated and then normalized. ； wherein, denotes the structural encoding feature of a road segment , denotes the unique encoding vector of a road segment , denotes the concatenation operation, denotes the set of road segments adjacent to a road segment , denotes the modulus of a vector , resulting in a structural matrix wherein is the number of road segments, each row in the matrix represents the structural feature of a road segment in the road network, and the structural similarity between two road segments is measured by cosine similarity: ； A mask matrix is generated to filter some edges with small similarity: ； wherein, is a mask matrix, denotes an element located at the row and the column, is a similarity threshold, multiplying the structural similarity and the mask matrix and passing it through an activation function to generate the final structural similarity matrix : ； wherein denotes element-wise multiplication of matrices, denotes a function; S23, according to The type is for each Generate a one-hot code, let Total Class, for each Generate a dimensional vector ,exist In the middle, only those that meet the requirements The type position is 1, and the rest are 0. Consider each Impact on surrounding roads: ； wherein, represents a point of interest, represents a distance to a road segment , is a distance threshold, all road segments having a distance to the point of interest below the distance threshold form a functional characteristic of the road segment . ； get a function matrix , Each row in the function matrix represents the functional characteristics of a road segment in the road network, represents the dimensionality; then the cosine similarity is used to measure the functional similarity between two road segments, and a final functional similarity matrix is generated by an activation function ； wherein denotes function, superscript denotes transpose.

4. The multi-graph hierarchical road network representation method based on graph neural network according to claim 3, characterized in that, The step S3 specifically comprises: S31, using spectral clustering algorithm to divide the structure region, giving the adjacency matrix of road segment , by subtracting the diagonal matrix , deduce its graph Laplacian matrix , calculate the first eigenvectors , get the matrix composed of eigenvectors , represent the linear space composed of dimensional matrix, by running the standard algorithm on the matrix , get the hard mapping from position to structure region ; Adjacency matrix representing the natural connectivity between road segments, is a structure similarity matrix for road segments, encoding structure similarity as connections between road segments, modeling the two graphs with two graph attention networks and aggregating them together to get a soft mapping i.e.: ； wherein, is a graph attention network, is a row normalization function, represents a road segment representation; finally, the hard mapping and the soft mapping are combined to obtain a final constructed structure assignment matrix : ； wherein, is a hyperparameter controlling the fusion of the two mappings; S32, by Linking structure zone characterization to link characterization: ； wherein represents a structural zone representation; The connection relationship of the structural zone and the connection relationship of the road section are associated: ； wherein represents the structure zone connection matrix.

5. The multi-graph hierarchical road network representation method based on graph neural network according to claim 4, characterized in that, The step S4 specifically comprises: S41, the functional similarity matrix is obtained Transition from road segments to structure zones: ； wherein represents a structure zone functional similarity matrix; Two graph attention networks are adopted to model the natural connectivity and functional similarity of the structure regions respectively, and they are aggregated together to obtain the final functional region assignment matrix That is: ； S42, allocating a matrix of functional zones The functional zone representation is expressed as a weighted linear combination of the structural zone representations: ； Utilizing The adjacency matrix of the functional region node is further derived: ； wherein is a linear rectifier function, is an adjustable hyperparameter.

6. The multi-graph hierarchical road network representation method based on graph neural network according to claim 5, characterized in that, The step S5 specifically comprises: S51, the functional zone layer is updated, the functional zone representation is updated, and the information is prepared to be transmitted to the next level, and a standard graph convolution network is used to update the functional zone embedding: ； wherein, denotes a graph convolutional network, denotes a function zone representation in iterations, denotes an adjacency matrix of function zone nodes, denotes a function zone representation in iterations; Then, the functional zone update information is transmitted to the next layer to update the structural zone representation: ； wherein, In the structural zone representation in the k-th iteration, is a gating unit function from the functional zone to the structural zone, denotes the product, i.e. the matrix elements are multiplied one by one, denotes the assignment matrix from the structural zone to the functional zone, denotes the assignment matrix from the structural zone to the functional zone, the functional zone representation in the k-th iteration, are learnable parameters; denotes an activation function, whose formula is ； wherein, represents a natural constant, represents input of the function; S52. In the structural zone, first adopt the standard Update the embedding representation of this layer: ； Then, the structural zone embedding is forwarded to the next layer to update the road section representation: ； wherein, represent learnable parameters; S53, the updated road section representation is combined with the information of the structural zone and the functional zone, and the relationship between the road section nodes is modeled by using a graph attention network: ； The operation is repeated to obtain a road network representation containing structural features and functional features.

Images