Overpass identification method under support of graph attention network

Abstract

The invention discloses an overpass identification method under the support of a graph attention network. According to the method, road data of an open street map (OSM) is utilized to carry out an experiment, and the method comprises the following steps: firstly, carrying out data set extraction on an obtained road network of a city center through a point density clustering method to obtain a road data set needing to be identified; secondly, modeling a road network into a dual graph, and representing related nodes in the graph by combining local features and global features of the road; and finally, carrying out dichotomy in a supervised learning mode by using a graph attention network to realize the recognition of the overpass. Experimental results show that various classification indexes, such as the accuracy rate, the recall rate and the F1 value, all exceed 91%, the method is superior in the aspect of identifying complex and atypical overpasses, and in addition, the model is good in generalization ability on different city data sets and high in adaptability.

Description

[0002] Overpass Recognition Method Supported by Graph Attention Network Technical Field

[0003] The present invention relates to the field of road network pattern recognition, and more specifically to an overpass recognition method supported by a graph attention network. Background Art

[0004] Interchange structures are the main pattern types of road networks. Their spatial structures play a crucial role in urban traffic networks by diverting traffic flows in different directions and reducing traffic conflicts. The recognition of overpasses is beneficial for cartographic generalization of road networks, vehicle navigation, traffic flow analysis, etc. It is somewhat difficult to accurately recognize overpasses in road networks, and multiple factors need to be considered, including various external factors, complex internal structures, geometric features of each section of the road, and topological relationships between roads.

[0005] Traditional overpass recognition methods mainly include point feature-based and line feature-based recognition. Point feature-based recognition methods utilize the node distribution characteristics of roads and retrieve in areas where point density aggregates to achieve overpass recognition. Such methods are simple and effective, but due to the overly single extracted features and the lack of geometric and topological features of roads, they are prone to misrecognition. Therefore, this method can only be used for auxiliary recognition and is difficult to recognize complex overpasses. Line feature-based methods mainly establish a template library of typical overpasses based on the spatial relationships between roads and use matching methods for overpass recognition. Such recognition methods take into account the geometric and topological features between roads and express overpasses more fully. However, the recognition process depends on the establishment of the template library, and the generalization ability is not strong, and non-typical overpasses outside the template library cannot be effectively recognized.

[0006] Graph Attention Networks (GAT) can generalize convolutional operations to the graph domain, enabling convolutional processing of graph-structured data with non-Euclidean structures to capture high-dimensional features. In addition, GAT uses an attention mechanism to express the connection relationships between nodes with continuous numbers and learns node features according to different importance levels. This characteristic is particularly applicable to overpass structures with a large number of roads and complex road intersections and can effectively extract their features.

[0007] In summary, in view of the deficiencies of traditional methods in insufficient feature expression and dependence on template libraries, this paper proposes an overpass recognition method based on graph attention networks. Summary of the Invention

[0008] The present invention proposes an overpass recognition method supported by a graph attention network. The road network data to be recognized is converted into a dual graph, and the roads are constructed into graph-structured data using the dual graph. Whether the data is an overpass is used as a label and input into the GAT model for training to complete the recognition of overpasses. The present invention mainly includes four parts: the preliminary positioning of overpasses, the construction and annotation of the dual graph, the extraction of graph node features, and the overpass recognition based on GAT.

[0009] (1) Preliminary positioning of overpasses: After preprocessing the data, the nodes in the road network are extracted, and the distribution range of overpasses is obtained using the method of point density clustering. The vector roads in the node-dense area are used as the roads to be recognized and made into a dataset.

[0010] (2) Construction and annotation of the dual graph: In the way of the dual graph, points are used to represent roads, and edges are used to represent the connection relationships between roads. The dataset is constructed into a graph structure, and then the dataset is divided into two categories: overpass and non-overpass by manual annotation.

[0011] (3) Extraction of graph node features: Two categories of a total of five features, namely local road features and global road features, are extracted to describe the roads, and they are used as the dual graph node features and input into the graph attention network.

[0012] (4) Overpass recognition based on GAT: Through the calculation of the attention layer and the graph pooling operation, graph-level features are obtained, which can identify the classification result of the graph and realize the overall recognition of overpasses.

[0013] The overpass recognition method combining the graph attention network proposed by the present invention can automatically calculate the weight assignment of feature factors, effectively reduce the recognition errors caused by subjective assignment, and improve the recognition accuracy. And it can automatically perform recognition, avoiding the operation steps of artificially setting the recognition threshold and continuously debugging the threshold size, and improving the recognition efficiency. Description of the Drawings

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the schematic diagrams of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0015] Figure 1 It is the flow chart of the overpass recognition based on the graph attention network provided by the present invention.

[0016] Figure 2 It is the schematic diagram of point density clustering provided by the present invention.

[0017] Figure 3Schematic diagram of dual graph construction and icon annotation provided by the present invention.

[0018] Figure 4 Schematic diagram of polar coordinate division provided by the present invention.

[0019] Figure 5 Model diagram of overpass recognition based on GAT provided by the present invention.

[0020] Figure 6 Accuracy and loss values of the training set and validation set provided by the present invention.

[0021] Figure 7 Recognition result diagram of overpass provided by the present invention.

[0022] Table 1 shows the buffer construction algorithm

[0023] Table 2 shows the training results of different models

[0024] Table 1 Buffer construction algorithm

[0025]

[0026]

[0027] Table 2 Training results of different models

[0028] Specific implementation manners

[0029] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0030] An overpass recognition method combining a graph attention network is disclosed in an embodiment of the present invention. The flowchart is as Figure 1 shown, and includes:

[0031] 1. Data preprocessing and extraction of the data set:

[0032] Step1: Use ArcGIS to perform topological detection and attribute table inspection on the acquired data, and delete duplicate sections and isolated sections. For the road extraction of the road to be recognized to meet the subsequent experimental conditions, it is necessary to screen the road network without affecting the overall structure of the road, select the main road in the attribute table and the auxiliary roads connected to the main road, and delete other roads.

[0033] Step 3: Use the point density clustering method to determine the dataset of the road to be recognized. The clustering process is carried out on the nodes of the road. The process of point density clustering is as Figure 2 shown. The buffer construction algorithm is shown in Table 1. First, by statistically analyzing the distribution characteristics of road nodes, a node is randomly marked and a buffer is established for the marked node. If there are other nodes in the buffer, buffers of the same size are established with these nodes as the centers, and at the same time, these intersecting buffers are merged until no new nodes fall into the buffer. A new node is reselected for buffer establishment until all nodes are traversed. Calculate the number of road nodes in the buffer and compare it with a pre-set threshold. Extract the roads in the buffer where the number is greater than the threshold to obtain the dataset to be recognized.

[0034] Step 4: Construct a dual graph: Convert the data to be recognized into a dual graph, with the roads as the nodes of the graph and the connection relationships between the roads as the edges of the graph. As Figure 3 shown, convert the roads into a dual graph. By comparing with the remote sensing image, perform a tagging operation on the dataset, where 1 represents an overpass and 0 represents a non-overpass.

[0035] 2. Selection of geometric features:

[0036] Step 5: In this paper, the local features of the road and the global features describing the overall structure of the road are extracted as the feature information of the nodes. For the local features of a single road, the length, curvature, and connectivity of the road are used as indicators for description.

[0037] (1) Road length: The sum of the lengths of all road segments that make up the road. The calculation formula is:

[0038]

[0039] In the formula, length_seg i represents the length of the i-th road segment, and n represents the total number of road segments.

[0040] (2) Road curvature: Represents the angular change value of the road in space and is used to evaluate the shape of the road. The calculation formula is:

[0041]

[0042] In the formula, a i is the angular change of the road at point p i , and L is the road length.

[0043] (3) Road connectivity: Represents the number of roads connected to the road and is used to evaluate the degree of closeness of the current road to other roads. The calculation formula is:

[0044]

[0045] In the formula, E is the edge set of the dual graph, i is the current road, and j ∈ M represents other roads in the graph structure except i.

[0046] Step6: For the global features of the road, use the coordinates of the road and the proportion of the road line density as indicators for description.

[0047] (1) Road coordinates: Represent the direction and distance of the road relative to the data center, and are used to describe the position information of the road. The calculation method is as follows:

[0048] First, taking the center of the data set as the origin and the farthest distance as the radius, establish a polar coordinate system, and divide the coordinates into 4 angular regions and 5 distance regions, as Figure 4 shown;

[0049] Then, according to the established polar coordinate system, assign a code to each region, and use the region code where the midpoint of the road is located as the road coordinate feature. The coding formula is:

[0050]

[0051] In the formula, θ is the angular coordinate of the point, ρ is the radius coordinate of the point, and γ is the farthest distance.

[0052] (2) Proportion of road line density: Represents the ratio of the road length to the total length of the original graph, and represents the proportion of the road line density in the region. The calculation formula is:

[0053]

[0054] In the formula, length_origin i is the total length of the roads in the region, i ∈ O is the set of regions where the roads exist, and L is the road length.

[0055] Step7: Finally, combine the local features and global features of the road to generate a feature vector, which is used as the input of the GAT model to support subsequent analysis.

[0056] 3. Identification of overpasses by GAT

[0057] Step:8: This experimental model is divided into three parts: an input layer, a hidden layer, and an output layer. The model is as Figure 5 shown.

[0058] Step: 9: Comprehensively analyze the recognition accuracy and model loss obtained by selecting different numbers of convolutional layers and different convolutional kernels. The hidden layer of the GAT model used in this experiment consists of 3 convolutional layers and 1 pooling layer. Each convolutional layer is a graph attention layer containing 64 convolutional kernels. ReLU is used as the activation function, and the Adam optimizer is used for parameter update. The learning rate is set to 0.005, and the L2 regularization parameter is used. The cross-entropy loss function is used to measure the error between the model prediction result and the true label, and the backpropagation algorithm is used to update the model parameters. The accuracy and loss values of the model training set and validation set are as Figure 6 shown.

[0059] Step 10: Use the transformed dual graph, labels, and extracted features as the input of the model. The input and output of overpass recognition in this experiment are as follows:

[0060] Input: Convert the dataset into a dual graph, use the road segments as the nodes of the graph, and the connection relationship as the edges of the graph. Finally, the obtained road network graph model G = (V, E, A). Each graph model contains N nodes, and each node has 5 calculated feature values {f1, f2, f3, f4, f5}, corresponding to 5 local and global features of the road respectively. All nodes form an N×5 feature vector and an N×N adjacency matrix.

[0061] Output: The recognition result is as Figure 7 shown, and finally the predicted probability of the dataset to be recognized is obtained, and the overpasses in the road network are recognized.

[0062] Step 11: Input the test data into the trained model for overpass recognition, calculate the accuracy, recall rate, and the harmonic mean F1 of the two, as shown in Table 2. Analyze the above three indicators to finally evaluate the overpass recognition quality.

[0063] Step 12: End.

[0064] This paper proposes an overpass recognition method supported by a graph attention network. The road dataset is screened by point density clustering, and then combined with the local and global features of the road, and the GAT is used to recognize the overpasses.

[0065] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined by this solution can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown in this solution, but will conform to the widest scope consistent with the principles and novel features disclosed by the present invention.

Claims

1. A method for overpass recognition supported by a graph attention network, including preliminary positioning of overpasses, construction and annotation of dual graphs, extraction of graph node features, and an overpass recognition model based on GAT. The initial positioning steps for the overpass are as follows: S1: Optimize the acquired data and use ArcGIS to remove unimportant sections, leaving only the main roads and their connected auxiliary roads. Perform topology detection and attribute table check on the data to delete duplicate and isolated sections. S2: extracting road nodes on the optimized data, establishing a buffer zone on the data according to the node threshold and the distance threshold to obtain the distribution range of the overpass, and using the roads extracted in the buffer zone as the data to be identified; The steps for modeling and labeling the dual graph are as follows: S3: Take road segments as nodes and road nodes as edges to construct a dual graph of data; S4: Use manual methods to label the data by comparing it with high-resolution remote sensing images; The steps for extracting graph node features are as follows: S5: Select local features of the road and global features of the road representing the overall structure as feature information of the node. Local road features include three features: road length, curvature, and road connectivity, and global road features include two features: road coordinates and road line density ratio. The steps of interchange recognition in the graph attention network are as follows: S6: Use the dual graph after graph data conversion as input data, divide the data into training set, validation set and test set samples, and input the training set and validation set data into the model for supervised training; S7: Analyze the recognition accuracy obtained by different numbers of convolution layers and different convolution kernels, and adjust to the optimal parameters; S8: Input the test data into the trained model to identify the overpass, calculate the accuracy, recall rate and F1 value of the recognition result, analyze the above indicators, and make a final evaluation on the overpass recognition quality; S9: End.

2. The overpass recognition method supported by a graph attention network according to claim 1 is characterized in that: In step S5, local road features including road length, curvature, and road connectivity, and global road features including road coordinates and road line density ratio are selected.

3. The overpass recognition method supported by a graph attention network according to claim 1 or claim 2, characterized in that: In steps S6 to S8, the graph attention network is combined to identify overpasses, and high-dimensional learning of road features is performed through supervised classification to obtain deeper features of the road. In the final result, the method in this paper can identify complex and atypical overpasses, has good generalization ability on different urban data sets, and shows strong adaptability.

Images