Map matching method based on dynamic feedback weight optimization of online learning

Abstract

This invention discloses a map matching method based on online learning and dynamic feedback weight optimization, belonging to the field of map matching technology. The method includes: preprocessing road network data for spatial connectivity and gridding; performing standardized gridding processing on vehicle satellite positioning data; calculating the attribute vector feature values ​​of vehicle satellite positioning and candidate road segments within the target grid, wherein the attribute vectors include distance matching degree, cosine similarity, and shape similarity; capturing road network topology and temporal changes using a spatiotemporal graph convolutional network based on a fusion attention mechanism, learning dynamic weights online, and predicting the comprehensive matching degree; evaluating the comprehensive matching degree of vehicle satellite positioning data and candidate road segments, and selecting the optimal matching road segment for the current satellite positioning data. Through the above method, this invention improves the accuracy and efficiency of map matching.

Description

Technical Field

[0001] This invention relates to the field of map matching technology, and more specifically to a map matching method based on dynamic feedback weight optimization using online learning. Background Technology

[0002] With the rapid development of intelligent transportation systems, vehicle-to-everything (V2X) networks, ride-hailing platforms, and logistics management, massive amounts of vehicle satellite positioning data (such as GPS and BeiDou data) are generated and collected in real time. This data forms the core foundation for traffic flow analysis, route planning, travel behavior research, dynamic navigation, and urban traffic management. However, raw satellite positioning data often contains random noise (e.g., positioning drift caused by urban canyons, tunnels, signal blockage, etc.), and its sampling points do not directly correspond to specific road segments in the road network.

[0003] Therefore, accurately and efficiently matching these discrete, error-laden location points to the correct roads—a process known as "map matching"—has become a crucial and challenging key technology.

[0004] Traditional map matching algorithms are mainly divided into two categories: local / incremental algorithms and global algorithms. Local / incremental algorithms (such as point-to-point, point-to-line, and line-to-line matching) typically only consider the geometric relationship (such as distance and direction angle) between the current location point and its neighboring candidate road segments, making a matching decision independently for each point. Their advantages include low computational complexity, fast response speed, and ease of real-time processing. However, their core drawback lies in the "short-sightedness" of their decision-making, ignoring the continuity of vehicle movement and the topology of the road network. In complex scenarios, such as when vehicles are at intersections of elevated bridges and ground-level roads, parallel roads, or dense road networks, relying solely on the geometric information of a single point can easily lead to matching errors. These errors accumulate and propagate with the matching of subsequent points, resulting in low overall matching accuracy and poor robustness.

[0005] Global algorithms (such as those based on Hidden Markov Models (HMMs), Conditional Random Fields (CRFs), or shortest path optimization) treat the entire vehicle trajectory as a whole, finding the optimal matching path by maximizing the posterior probability of the entire trajectory on the road network or minimizing the global cost function. They comprehensively consider both observation probabilities (the matching degree between location points and road segments) and transition probabilities (the connectivity and driving probability between adjacent road segments), thus theoretically achieving higher accuracy than local algorithms. However, despite the improved accuracy of global algorithms, their inherent shortcomings become increasingly apparent in modern applications with large-scale, high real-time requirements, specifically in the following aspects:

[0006] (1) Bottlenecks in computational efficiency and real-time performance:

[0007] Global algorithms require joint solutions for all candidate points along the entire trajectory or within a long sliding window. Their computational complexity increases exponentially or exponentially with the trajectory length. When processing real-time positioning data streams from tens of thousands or even millions of vehicles, the computational overhead of traditional global algorithms becomes enormous, making it difficult to meet millisecond or second-level real-time response requirements. This results in high system latency and an inability to support large-scale online applications.

[0008] (2) The feature dimension is too singular, and the information is not fully utilized:

[0009] Existing global algorithms rely on relatively singular and fixed feature dimensions when defining observation and transition probabilities. They primarily depend on geometric features (such as projection distance and directional consistency) and topological features (such as road segment connectivity and turning angles). However, these features have limited ability to distinguish complex scenarios with "geometric similarity." For example, parallel expressways and auxiliary roads may have highly similar geometric and topological features, but the trajectory behavior characteristics of vehicles on the two roads (such as average speed, speed fluctuation, and acceleration / deceleration frequency) differ significantly. Existing technologies have failed to effectively extract and utilize these multi-dimensional features, such as "trajectory behavior similarity," which can reflect the micro-driving behavior of vehicles. This results in insufficient discriminative ability and limited accuracy of matching models in these "deceptive" scenarios.

[0010] (3) The model is static and lacks dynamic adaptive capability:

[0011] Traditional global algorithms (such as standard Hidden Markov Models) typically use pre-set or offline-trained fixed values ​​for their model parameters (e.g., weights of observation and transition probabilities, standard deviation of Gaussian noise, etc.). This static model assumption makes it unsuitable for adapting to dynamically changing traffic environments in the real world. For example, during peak hours, vehicles may frequently start and stop due to congestion, exhibiting behavior patterns drastically different from off-peak hours; in inclement weather, positioning errors may increase significantly. Static models cannot learn online and dynamically adjust their internal weights based on real-time traffic flow, driving status, and environmental changes, resulting in weak model generalization ability and decreased matching performance in dynamic environments.

[0012] (4) Insufficient ability to model the spatiotemporal dependence of road networks:

[0013] Road networks are complex graph structures with rich spatial dependencies between road segments (such as the mutual influence of traffic flows between upstream and downstream segments). Meanwhile, vehicle trajectories are typical time-series data, with strong correlations between states at different times. Existing algorithms often treat road network topology and trajectory temporal sequence as two independent aspects during modeling, lacking a unified modeling framework that can deeply integrate the spatial topology of the road network and the dynamic changes in trajectory time. For example, they struggle to effectively capture deep spatiotemporal coupling relationships such as "congestion on the road ahead affects the speed change of vehicles on the current road segment," thus limiting further improvements in matching accuracy.

[0014] In summary, while local algorithms are efficient, they suffer from poor accuracy and robustness. Global algorithms, while having relatively high accuracy, suffer from multiple drawbacks, including low computational efficiency, insufficient feature utilization, static models, and inadequate ability to model spatiotemporal dependencies. Consequently, they cannot meet the stringent requirements of real-time matching of large-scale vehicle data while ensuring high accuracy. Summary of the Invention

[0015] In view of the above problems, the present invention is proposed to provide a map matching method based on dynamic feedback weight optimization of online learning to overcome or at least partially solve the above problems.

[0016] To achieve the above objectives, the present invention adopts the following technical solution:

[0017] A map matching method based on dynamic feedback weight optimization using online learning includes the following steps:

[0018] Step 1: Perform spatial connectivity and gridding preprocessing on the road network data, and perform standard gridding processing on the vehicle satellite positioning data;

[0019] Step 2: Calculate the attribute vector feature values ​​of vehicle satellite positioning and candidate road segments within the target grid. The attribute vector includes distance matching degree, cosine similarity, and shape similarity.

[0020] Step 3: Capture the temporal changes of road network topology and attribute vectors using a spatiotemporal graph convolutional network based on a fusion attention mechanism, and learn dynamic weights online to obtain the comprehensive matching degree between vehicle satellite positioning data and candidate road segments;

[0021] Step 4: Evaluate the overall matching degree between vehicle satellite positioning data and candidate road segments, and select the optimal matching road segment for the current satellite positioning data.

[0022] Furthermore, step 1 involves spatial connectivity and gridding preprocessing of the road network data, specifically including:

[0023] Step 1-1: Based on road network data, repair the connectivity of road data;

[0024] Step 1-2: Perform regional gridding processing on the road network data after connectivity restoration: Set the minimum longitude of the road network data region to [value missing]. Minimum latitude Maximum longitude Maximum latitude And the single-side length L of each grid, and establish the spatial grid index of the road network area.

[0025] Furthermore, in steps 1-2, establishing a spatial grid index for the road network area specifically includes:

[0026] With the minimum longitude and minimum latitude The center point of the first grid in the road network area is used as the center point, and the grids in the road network area are numbered by column number and row number;

[0027] Calculate the coordinates of the center point of each grid i in the road network region;

[0028] The vertex coordinate index of each grid is obtained based on the center point coordinates;

[0029] Based on the vertex coordinate indices of all grids in the road network area, a spatial grid index database for the road network area is established.

[0030] Furthermore, in step 1, the vehicle satellite positioning data undergoes standard gridding processing, specifically including:

[0031] Based on the latitude and longitude coordinates of the vehicle satellite positioning data, the grid number corresponding to the vehicle satellite positioning point in the preprocessed road network data is calculated.

[0032] Furthermore, based on the latitude and longitude coordinates of the vehicle satellite positioning data, the grid number corresponding to the vehicle satellite positioning point in the preprocessed road network data is calculated, specifically using the following formula:

[0033] ;

[0034] ;

[0035] Where x represents the longitude coordinates of the vehicle's satellite positioning data, and y represents the latitude coordinates of the vehicle's satellite positioning data. This indicates the grid column number corresponding to the vehicle's satellite positioning point with latitude and longitude coordinates (x, y). This indicates the grid row number corresponding to the vehicle's satellite positioning point with latitude and longitude coordinates (x, y); , , For the Earth's radius, Indicates the minimum longitude of the road network data area. This represents the minimum latitude of the road network data area. Indicates the maximum longitude of the road network data area. L represents the maximum latitude of the road network data area, and L represents the length of one side of each grid in the road network data area.

[0036] Further, step 2 calculates the attribute vector feature values ​​of vehicle satellite positioning and candidate road segments within the target grid, specifically including:

[0037] Step 2-1: Calculate the vertical distance d between the vehicle satellite positioning points within the target grid and the candidate road segments, and map the vertical distance to the distance matching degree value to obtain the distance matching degree feature value δ:

[0038]

[0039] in, This represents the minimum vertical distance threshold for candidate road segments. This represents the maximum vertical distance threshold for the candidate road segment;

[0040] Step 2-2: Calculate the cosine similarity feature value r between the vector direction between the current satellite positioning point and the previous satellite positioning point of the vehicle within the target grid and the direction of the candidate road segment;

[0041] Steps 2-3: Form a positioning sequence data by combining the current satellite positioning point of the vehicle within the target grid with multiple consecutive satellite positioning points, and compress the positioning sequence data using the Douglas-Peucker algorithm;

[0042] Step 2-4: Obtain candidate road segments and form road segment sequence data. Calculate the shape similarity distance s between the compressed positioning sequence data from Step 2-3 and the road segment sequence data using the dynamic time warping algorithm. Perform shape similarity matching mapping based on the shape similarity distance s to obtain the shape similarity feature value β.

[0043]

[0044] in, This represents the minimum value of the shape similarity threshold among candidate road segments. This represents the maximum value of the shape similarity threshold between candidate road segments.

[0045] Furthermore, a spatiotemporal graph convolutional network based on a fusion attention mechanism is used to capture road network topology and temporal changes, and dynamic weights are learned online, specifically including:

[0046] Step 3-1: Construct a topology graph using road network connectivity ;

[0047] in, This represents the attribute vector of n road segments within the corresponding region at time t. for The adjacency matrix, wherein the attribute vector includes distance matching degree, cosine similarity, and shape similarity;

[0048] Step 3-2: Construct a graph convolutional layer to extract spatial structure features from the road network topology graph. The spatial domain is converted to the frequency domain, scaling operations are performed on each dimension, adjacent nodes are aggregated to complete the convolution operation, and then the inverse transformation is performed back to the spatial domain to output the spatial structure features of the road network topology map.

[0049] Step 3-3: Construct a Transformer module for extracting temporal features. The Transformer module takes the road network topology map as input. The time sequence of attribute vectors corresponding to each node T time length is used to output the time sequence characteristics of the road segment matched by the vehicle satellite positioning point;

[0050] Step 3-4: Input the spatial structure features extracted in Step 3-2 and the temporal features extracted in Step 3-3 into the fully connected network layer, and output the comprehensive matching degree value of the vehicle satellite positioning matching road segment.

[0051] Further, in step 3-2, the inverse transformation to the spatial domain is performed using the following transformation formula:

[0052] ;

[0053] Where L is the Laplacian matrix of the road network topology, x is an n-dimensional column vector, L=DA, D is the degree matrix of the road network topology, and A is the adjacency matrix of the road network topology. This represents the attribute vector of vehicle satellite positioning and candidate road segments within the target grid. Indicates the transformation system.

[0054] Furthermore, the Transformer module in step 3-3 outputs the temporal characteristics of the vehicle satellite positioning point matching road segment, specifically including the following steps:

[0055] The attribute vector time series X of length T is input into the embedding layer for encoding, resulting in the encoded embedding matrix E:

[0056]

[0057] in, This is the weight matrix of the embedding layer. This is the bias matrix;

[0058] The embedding matrix E is input into the multi-head attention mechanism module for attention processing;

[0059]

[0060] in This represents the output matrix of the i-th attention head in a multi-head attention mechanism. This is the final weight matrix in the multi-head attention calculation;

[0061] The output of multi-head attention is non-linearly mapped through a feedforward neural network:

[0062]

[0063] in, , Here is the weight matrix of the feedforward neural network. , This is the bias matrix of the feedforward neural network;

[0064] The result of the nonlinear mapping is input into the normalization layer to obtain the temporal features of the road segment matched by the vehicle's satellite positioning point:

[0065]

[0066] in This is the weight matrix of the Transformer output layer. This is the bias matrix of the Transformer output layer.

[0067] Furthermore, the overall matching degree between vehicle satellite positioning data and multiple candidate road segments is evaluated, specifically including:

[0068] The overall matching degree between vehicle satellite positioning data and multiple candidate road segments is calculated. From the multiple candidate road segments with an overall matching degree greater than 0, the candidate road segment with the maximum overall matching degree is selected as the optimal matching road segment for the current vehicle satellite positioning data.

[0069] Furthermore, if the overall matching degree is less than 0, then the current vehicle satellite positioning data has no best matching road segment. Then, the vehicle satellite positioning data without best matching road segment is counted. When the amount of vehicle satellite positioning data without best matching road segment exceeds the threshold, the road network data is updated.

[0070] Specifically, the density clustering DBSCAN algorithm is used to cluster trajectory clusters, extract the center lines of newly added road segments from the trajectory clusters, and combine the center lines of each road segment through topology optimization to update the structure information of newly added roads and update the gridded index; when the amount of vehicle satellite positioning data without the best matching road segment does not reach the threshold, no update operation is performed.

[0071] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a map matching method based on dynamic feedback weight optimization of online learning, which has the following beneficial effects:

[0072] This invention can comprehensively consider multiple dimensions of features, such as the geometric, topological, and trajectory behavior similarity features between vehicle satellite positioning points and candidate road segments. It also uses a spatiotemporal graph convolutional network to capture road network topology and temporal changes, learns dynamic weights online, and predicts the comprehensive matching degree, thereby improving the accuracy and efficiency of map matching and meeting the needs of real-time matching of large-scale vehicle satellite positioning data with roads. Attached Figure Description

[0073] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0074] Figure 1 This is a flowchart of a map matching method based on dynamic feedback weight optimization using online learning, provided in an embodiment of the present invention.

[0075] Figure 2 This is a flowchart of the road network data and satellite positioning data preprocessing process in the map matching method based on online learning and dynamic feedback weight optimization provided in this embodiment of the invention.

[0076] Figure 3 This is a flowchart illustrating the calculation of geometric, topological, and trajectory behavior feature values ​​of vehicle satellite positioning points and multiple grid candidate road segments in the map matching method based on online learning and dynamic feedback weight optimization provided in this embodiment of the invention.

[0077] Figure 4 This is a flowchart of the dynamic weight online learning and comprehensive matching degree prediction process based on spatiotemporal graph convolutional networks in the map matching method based on online learning and dynamic feedback weight optimization provided in the embodiments of the present invention.

[0078] Figure 5 This is a flowchart illustrating the process of evaluating the comprehensive distance matching degree between a vehicle's star positioning point and multiple grid candidate road segments in a map matching method based on online learning and dynamic feedback weight optimization, as provided in an embodiment of the present invention. Detailed Implementation

[0079] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0080] like Figure 1 As shown, this invention discloses a map matching method based on dynamic feedback weight optimization using online learning, comprising the following steps:

[0081] Step 1: Perform spatial connectivity and gridding preprocessing on road network data, and perform standard gridding processing on vehicle satellite positioning data;

[0082] Step 2: Calculate the attribute vector feature values ​​of vehicle satellite positioning and candidate road segments within the target grid. The attribute vector includes geometric distance d, topological direction distance r, and trajectory similarity d feature values.

[0083] Step 3: Use a spatiotemporal graph convolutional network to capture road network topology and temporal changes, learn dynamic weights online, and predict the overall matching degree;

[0084] Step 4: Evaluate the overall matching degree between the vehicle's satellite positioning and multiple candidate road segments to determine the optimal matching road segment.

[0085] refer to Figure 2 In one specific implementation, step 1 performs spatial connectivity and gridding preprocessing on the road network data, and simultaneously performs standard gridding processing on the vehicle satellite positioning data. The specific implementation includes the following steps:

[0086] First, the road network data undergoes spatial connectivity and gridding preprocessing:

[0087] Step 1-1: Based on road network data, repair the connectivity of road data to solve the problem of road network topology disconnection caused by dead-end roads and fragmented road segments;

[0088] Based on the static traffic attribute information of road network segments, such as no left turn or only right turn, analyze the upstream and downstream segments of the road segment, and store the relationship data of the upstream and downstream segments in the database.

[0089] Steps 1-2: Perform gridding processing on the road network data area after connectivity restoration. Assume the minimum longitude of the road network data area is... Minimum latitude Maximum longitude Maximum latitude Each grid cell is Lm*Lm in size. The center point of the first grid cell is determined by its minimum longitude and minimum latitude. Grids are numbered by column and row numbers, starting from 0. Given a row and column number, the coordinates of the center point of each grid cell i are:

[0090]

[0091] ;

[0092] in This represents the vertical coordinate of the center point of grid i. Indicates the column number of grid i. This represents the horizontal coordinate of the center point of grid i. This indicates the row number of grid i. , , L represents the Earth's radius, and L represents the length of one side of the grid.

[0093] Only one vertex needs to be calculated for each grid:

[0094] ;

[0095] The vertices corresponding to the three adjacent grids are as follows:

[0096]

[0097] ;

[0098] .

[0099] This establishes a spatial grid index for the region. In the spatial grid, a square grid has four vertices. The indices of the lower left, lower right, upper left, and upper right vertices of each grid can be obtained from the lower left vertices of the right grid, the upper grid, and the upper right grid, respectively.

[0100] The established spatial grid data is correlated with the road segment spatial data of the road network. The road segment data within each grid is analyzed, and a spatial grid index of the road segments is established and stored in the database.

[0101] Simultaneously, vehicle satellite positioning data undergoes standardized grid processing, specifically including:

[0102] Steps 1-3: Perform standardized gridding processing on the vehicle satellite positioning data. Assume the vehicle satellite positioning data has latitude and longitude coordinates of (x, y), and the gridding unit is Lm*Lm (L is the side length of each grid cell). , Calculate the grid number corresponding to the vehicle's satellite positioning point, where:

[0103] ;

[0104] ;

[0105] This represents the Earth's radius.

[0106] refer to Figure 3 In one specific implementation, step 2, which calculates the geometric, topological, and trajectory behavior feature values ​​of vehicle satellite positioning points and multiple candidate road segments within the target grid, includes the following steps:

[0107] Step 2-1: Calculate the vertical distance d between the vehicle satellite positioning points within the target grid and the candidate road segments, and map the vertical distance to the distance matching degree value to obtain the distance matching degree feature value δ:

[0108]

[0109] in, This represents the minimum vertical distance threshold for candidate road segments. This represents the maximum vertical distance threshold for candidate road segments.

[0110] Step 2-2: Calculate the cosine similarity feature value r between the vector direction between the current satellite positioning point and the previous satellite positioning point of the vehicle within the target grid and the direction of the candidate road segment.

[0111] Steps 2-3: Form a positioning sequence data by combining the current satellite positioning point of the vehicle within the target grid with multiple consecutive satellite positioning points, and compress the positioning sequence data using the Douglas-Peucker algorithm.

[0112] Step 2-4: Obtain candidate road segments and form road segment sequence data. Calculate the shape similarity distance s between the compressed positioning sequence data from Step 2-3 and the road segment sequence data using the dynamic time warping algorithm. Perform shape similarity matching mapping based on the shape similarity distance s to obtain the shape similarity feature value β.

[0113]

[0114] in, This represents the minimum value of the shape similarity threshold among candidate road segments. This represents the maximum value of the shape similarity threshold between candidate road segments.

[0115] It should be noted that in this invention, candidate road segments include not only all candidate road segments located within adjacent grids, but also some candidate road segments located within the target grid. Correspondingly, in some embodiments, the current satellite positioning point of the vehicle within the target grid and the previous consecutive satellite positioning points in steps 2-3 include not only all vehicle satellite positioning points located within the target grid, but also vehicle satellite positioning points located within adjacent grids (i.e., the historical positioning trajectory points of the vehicle satellites in the target grid may originate from adjacent grids). In this invention, adjacent grids include a total of 8 grids: those adjacent to the target grid (top, bottom, left, right, and four corners).

[0116] refer to Figure 4 In one specific implementation, step 3 involves updating the dynamic weights of the online learning network model based on the fusion attention mechanism in real time. The specific implementation includes the following steps.

[0117] Step 3-1: Construct a topology graph through road network connectivity. At time step t, the topology graph... , This represents the attribute vector of n road segments within the corresponding area at time t, where k=3. It includes the matching degree of real-time vehicle satellite positioning data with road segment distance, similarity of direction matching, and similarity of trajectory shape features. For topology The adjacency matrix, Let be the Euclidean distance between road segment i and road segment j.

[0118] The adjacency matrix can be represented as:

[0119] .

[0120] Step 3-2: Construct a graph convolutional layer to extract spatial structure features from the road network topology graph. The spatial domain is transformed into the frequency domain, scaling is performed on each dimension, adjacent nodes are aggregated to complete the convolution operation, and then the domain is inversely transformed back to the spatial domain. The transformation formula is as follows: Where L (L=DA) is the Laplacian matrix of the graph, D is the degree matrix of the graph, A is the adjacency matrix of the graph, and x is an n-dimensional column vector. The characteristics of the nodes are represented by the matching degree between real-time vehicle satellite positioning data and road segment distance, direction matching similarity, and trajectory shape similarity features. This represents the transformation coefficients and outputs the spatial characteristics of the road network topology map.

[0121] Step 3-3: Construct a Transformer layer for extracting temporal features. The input to the Transformer module is the road network topology map. The attribute vectors of each corresponding node, including the distance matching feature value δ, cosine similarity feature value r, and shape similarity feature value β calculated in step 2, are time series of length T. The output is the temporal feature of the vehicle satellite positioning point matching road segment. The calculation process of the Transformer module is as follows.

[0122] Calculate the embedding matrix , in, This is the weight matrix of the embedding layer. This is the bias matrix.

[0123] The output of multi-head attention is calculated as follows:

[0124] ;

[0125] in This represents the output of the i-th attention head in the multi-head attention mechanism. This is the final weight matrix in the multi-head attention calculation.

[0126] Calculate the output after nonlinear mapping through a feedforward neural network.

[0127] ;

[0128] in, , Here is the weight matrix of the feedforward neural network. , This is the bias matrix of the feedforward neural network.

[0129] The final output of the Transformer module is the overall matching score.

[0130] ;

[0131] in, This is the weight matrix of the Transformer output layer. This is the bias matrix of the Transformer output layer.

[0132] Step 3-4: Input the spatial structure features extracted in Step 3-2 and the temporal features extracted in Step 3-3 into the fully connected network layer, and output the comprehensive matching degree value of the vehicle satellite positioning matching road segment.

[0133] refer to Figure 5 In one specific implementation, step 4, which evaluates the comprehensive matching degree between the vehicle's satellite positioning and multiple candidate road segments, specifically includes the following steps.

[0134] Step 4-1: If at least one candidate road segment in the overall matching degree output in Step 3 is greater than 0, the road segment corresponding to the maximum value of the overall distance matching degree is the matching road segment; otherwise, there is no best matching road segment, and continue to Step 4-2.

[0135] Step 4-2: When the amount of unmatched vehicle satellite positioning data exceeds the threshold, use density clustering DBSCAN to cluster trajectory clusters, extract the center lines of new road segments from the trajectory clusters, combine the center lines of each road segment through topology optimization, realize the update of new road structure information, and update the gridded index.

[0136] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0137] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not 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. An online learning based dynamic counterweight optimization map matching method, characterized in that, The method comprises the following steps: Step 1: spatial connectivity and gridding preprocessing of road network data, and standard gridding processing of vehicle satellite positioning data; Step 2: calculating attribute vector characteristic values of vehicle satellite positioning and candidate road segments in the target grid, the attribute vector comprising distance matching degree, cosine similarity and shape similarity; Step 2-1: calculating the vertical distance d between the vehicle satellite positioning point in the target grid and the candidate road segment, and performing distance matching degree value mapping of the vertical distance to obtain the distance matching degree characteristic value δ: Step 2-2: calculating the cosine similarity characteristic value r of the vector direction between the current satellite positioning point and the previous satellite positioning point of the vehicle in the target grid and the direction of the candidate road segment; wherein, is a minimum value of a vertical distance threshold for candidate road segments, is a maximum value of a vertical distance threshold for candidate road segments; Step 2-3: forming positioning sequence data of the current satellite positioning point and the previous continuous satellite positioning points of the vehicle in the target grid, and compressing the positioning sequence data by using a Douglas-Peucker algorithm; Step 2-4: obtaining a candidate road segment and forming road segment sequence data, calculating the shape similarity distance s between the positioning sequence data compressed in step 2-3 and the road segment sequence data by using a dynamic time warping algorithm, and performing shape similarity matching mapping based on the shape similarity distance s to obtain the shape similarity characteristic value β: Step 3: capturing the time sequence changes of the road network topology and the attribute vector by using a spatio-temporal graph convolution network based on a fusion attention mechanism, and learning dynamic weights online to obtain the comprehensive matching degree of the vehicle satellite positioning data and the candidate road segment; specifically comprising: wherein, is a shape similarity threshold minimum value between candidate road segments, is a shape similarity threshold maximum value between candidate road segments; Step 3-4: inputting the spatial structure features extracted in step 3-2 and the time sequence features extracted in step 3-3 into a fully connected network layer to output the comprehensive matching degree value of the vehicle satellite positioning matching road segment; Step 3-1: Constructing a topology graph by road network connectivity ; wherein, represents an attribute vector of an n-path segment in a corresponding region at a time t, is an adjacency matrix of the attribute vector, the attribute vector including a distance matching degree, a cosine similarity, and a shape similarity; Step 3-2: construct the graph convolution layer for extracting the spatial structure features of the road network topology graph The spatial domain is converted into the frequency domain, each dimension is scaled, the convolution operation is completed by aggregating adjacent nodes, and then inverse transformation is performed to the spatial domain to output the spatial structure features of the road network topology graph; inverse transformation to the spatial domain is performed by using the following transformation formula: ； wherein L is a Laplacian matrix of the road network topology graph, x is an n-dimensional column vector, L = D - A, D is a degree matrix of the road network topology graph, and A is an adjacency matrix of the road network topology graph, represents a property vector of a vehicle satellite positioning within a target grid and a candidate road segment of an adjacent grid, represents a transformation coefficient; Step 3-3: build the Transformer module for extracting the time sequence features, which inputs the road network topology graph the attribute vector time sequence corresponding to the time length of each node T, outputting the time sequence features of the vehicle satellite positioning point matching road section; Step 4: evaluating the comprehensive matching degree of the vehicle satellite positioning data and the candidate road segment, and screening out the optimal matching road segment of the current satellite positioning data. In step 1, the spatial connectivity and gridding preprocessing of road network data comprises: 2.The map matching method based on dynamic counterweight optimization of online learning according to claim 1, wherein, Step 1-1: based on the road network data, repairing the connectivity of the road data; In step 1-2, the spatial grid index of the road network region is established, comprising: Step 1-2: The road network data after connectivity repair is subjected to regional gridding processing: the minimum longitude of the road network data region is set as , the minimum latitude , the maximum longitude , the maximum latitude , and the single side length L of each grid, and the spatial grid index of the road network region is established.

3. The map matching method based on dynamic counterweight optimization of online learning according to claim 2, characterized in that, calculating the center point coordinates of each grid i in the road network region; the minimum longitude and the minimum latitude The center point of the first grid of the road network region is taken as the minimum longitude and the minimum latitude, and the grids in the road network region are numbered by column number and row number. obtaining the vertex coordinate index of each grid according to the center point coordinates; establishing a spatial grid index database of the road network region according to the vertex coordinate index of all grids in the road network region. In step 1, the standard gridding processing of vehicle satellite positioning data comprises:

4. The map matching method based on dynamic counterweight optimization of online learning according to claim 1, characterized in that, calculating the grid number corresponding to the vehicle satellite positioning point in the pretreated road network data according to the latitude and longitude coordinates of the vehicle satellite positioning data. The grid number corresponding to the vehicle satellite positioning point in the pretreated road network data is calculated according to the latitude and longitude coordinates of the vehicle satellite positioning data, specifically by the following formula:

5. The online learning based dynamic counterweight optimization map matching method according to claim 4, characterized in that, In step 3-3, the Transformer module outputs the time sequence features of the vehicle satellite positioning point matching road segment, comprising the following steps: ； ； wherein x represents the longitude coordinate of the vehicle satellite positioning data, y represents the latitude coordinate of the vehicle satellite positioning data, represents the grid column number corresponding to the longitude and latitude coordinates (x, y) of the vehicle satellite positioning point, represents the grid row number corresponding to the longitude and latitude coordinates (x, y) of the vehicle satellite positioning point; , , is the radius of the earth, represents the minimum longitude of the road network data region, represents the minimum latitude of the road network data region, represents the maximum longitude of the road network data region, represents the maximum latitude of the road network data region, and L represents the single side length of each grid of the road network data region.

6. The online learning based dynamic counterweight optimization map matching method according to claim 1, wherein, ​ Input the attribute vector time sequence X with a T length into an embedding layer for encoding to obtain an encoded embedding matrix E; Input the embedding matrix E into a multi-head attention mechanism module for attention processing; The output of the multi-head attention is nonlinearly mapped through a feedforward neural network; The result after nonlinear mapping is input into a normalization layer to obtain the time sequence features of the vehicle satellite positioning point matching road section.

7. The online learning based dynamic counterweight optimization map matching method according to claim 1, wherein, The comprehensive matching degree of the vehicle satellite positioning data and the multiple candidate road sections is evaluated, specifically including: The comprehensive matching degree of the vehicle satellite positioning data and the multiple candidate road sections is counted, and from the multiple candidate road sections with a comprehensive total matching degree greater than 0, the candidate road section corresponding to the maximum comprehensive total matching degree is selected as the optimal matching road section of the current vehicle satellite positioning data.

Images