Apparatus and method for predicting search route using ETA based on graph neural network

Abstract

The present invention provides an apparatus and method for predicting a search route using an ETA based on a graph neural network, the apparatus comprising: a storage module configured to store digital map data; and a processor configured to perform a route search based on an ETA prediction model based on a map neural network formed by converting the road information into a map according to the route exploration request.

Description

Technical Field

[0001] Exemplary embodiments of the present disclosure relate to an apparatus and method for searching a route using an estimated time of arrival (ETA) prediction based on a graph neural network, and more specifically, to an apparatus and method for searching a route by representing a road network as a graph neural network and predicting ETA using the road network to perform a route search. Background Art

[0002] Recently, since most vehicles are equipped with navigation devices and due to the increase in vehicles, traffic congestion often occurs on roads and intersections even when traveling on roads that users already know. It is common to perform route searches and receive guidance in advance through navigation devices so that congested sections such as roads or intersections can be identified in advance and avoided.

[0003] In addition, when searching for routes, users often use estimated time of arrival (ETA) information. For example, the departure time or appointment time is determined on the basis of the ETA information. Therefore, when the ETA information is inaccurate, the user adds extra time to determine the departure time or appointment time, and when the ETA information is accurate, the extra time can be reduced, so that the user can save time without wasting time on the road.

[0004] In this way, during the route search process of navigation, ETA information is used to calculate the best route, recommended route, etc.

[0005] Meanwhile, with the development of deep learning-based information prediction technology, research on deep learning-based technology for more accurately predicting the estimated arrival time is being actively conducted.

[0006] Therefore, an efficient technique is needed to represent road data in order to allow the road data to be used as input to deep learning models. Summary of the invention

[0007] Various embodiments of the present disclosure are directed to an apparatus and method for searching a route using estimated time of arrival (ETA) prediction based on a graph neural network, which represents a road network as a graph neural network to achieve ETA prediction.

[0008] A device for searching a route using an estimated time of arrival (ETA) prediction based on a graph neural network according to the present disclosure includes: a storage module configured to store digital map data; and a processor configured to perform a route search based on an ETA prediction model according to a route exploration request, wherein the ETA prediction model is based on a graph neural network formed by converting road information into a graph.

[0009] In the present disclosure, the ETA prediction model may be based on converting road information into links of nodes of a graph and converting connections between the links into edges of the graph.

[0010] In the present disclosure, the ETA prediction model may be based on a graph neural network that includes features of the entire route as global attributes.

[0011] In the present disclosure, a node attribute of a graph may include information about a past speed or a passing time of a link corresponding to the node of the graph.

[0012] In the present disclosure, a weight may be applied to the ETA prediction model for predicting a value greater than a predicted value, where the predicted value is less than the correct answer.

[0013] In the present disclosure, a weight that increases the influence of a preset time zone on an output value of the ETA prediction model may be applied to the ETA prediction model.

[0014] In the present disclosure, the processor may calculate a plurality of candidate routes according to a route search request, and calculate the ETA of each candidate route through an ETA prediction model.

[0015] In the present disclosure, the processor may calculate the cost of each candidate route based on the calculated ETA of each candidate route.

[0016] A method for generating an estimated time of arrival (ETA) prediction model according to the present disclosure includes: a processor converting links of road information into nodes, and converting connections between links into edges of a graph, thereby converting road information into a graph, which includes: the processor converting relevant data into global attributes of the graph, node attributes of the graph, and edge attributes of the graph; and the processor using the global attributes of the graph, the node attributes of the graph, and the edge attributes of the graph as input values ​​to learn the ETA prediction model.

[0017] In the present disclosure, global attributes of a graph may include past ETA information for the entire route, node attributes of a graph may include information about past speeds or transit times of links corresponding to nodes of the graph, and edge attributes of a graph may include data such as connections between nodes and whether road types have changed.

[0018] In the present disclosure, the ETA prediction model may output link transit time as an output value.

[0019] In the present disclosure, the ETA prediction model may output an ETA value for the entire route.

[0020] A method for searching a route using an estimated time of arrival (ETA) prediction based on a graph neural network according to the present disclosure includes: receiving a route search request by a processor; performing a route search based on an ETA prediction model by the processor; and providing a route search result by the processor, wherein the ETA prediction model is based on a graph neural network formed by converting road information into a graph.

[0021] In the present disclosure, performing a route search may further include: in response to a route search request, calculating, by a processor, a plurality of candidate routes; and calculating, by the processor, an ETA of each candidate route by using an ETA prediction model.

[0022] In the present disclosure, performing the route search may further include calculating, by the processor, a cost of each candidate route based on the calculated ETA of each candidate route.

[0023] In the present disclosure, calculating the ETA of each candidate route through the ETA prediction model may include: a processor converting the links of the candidate route into nodes of a graph, and converting the connections between the links into edges of the graph, thereby converting each candidate route into a corresponding graph; and the processor calculating the ETA of each candidate route through the ETA prediction model that takes data related to the global properties of the graph, the node properties of the graph, and the edge properties of the graph as input values.

[0024] In the present disclosure, global attributes of a graph may include past ETA information about the entire route, node attributes of a graph may include information about past speeds or transit times of links corresponding to the nodes of the graph, and edge attributes of a graph may include data such as connections between nodes and whether the road type has changed.

[0025] In the present disclosure, a weight may be applied to the ETA prediction model for predicting a value greater than a predicted value, where the predicted value is less than the correct answer.

[0026] In the present disclosure, a weight that increases the influence of a preset time zone on an output value of the ETA prediction model may be applied to the ETA prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is an exemplary diagram showing a schematic construction of a device for searching a route using estimated time of arrival (ETA) prediction based on a graph neural network according to one embodiment of the present disclosure.

[0028] Figure 2 and Figure 3 is an example diagram for describing a graph transformation method for a device for searching a route using ETA prediction based on a graph neural network according to one embodiment of the present disclosure.

[0029] Figure 4 is an example diagram for describing an ETA prediction model of a device for searching a route using a graph neural network-based ETA prediction according to one embodiment of the present disclosure.

[0030] Figure 5 is a flowchart for describing a method for searching a route using ETA prediction based on a graph neural network according to one embodiment of the present disclosure.

[0031] Figure 6 is a flowchart for describing the use of an ETA prediction model in a method for searching a route using a graph neural network-based ETA prediction according to one embodiment of the present disclosure.

[0032] Figure 7 is a flowchart for describing an ETA prediction operation of a method for searching a route using a graph neural network-based ETA prediction according to one embodiment of the present disclosure. DETAILED DESCRIPTION

[0033] An embodiment of an apparatus and method for searching a route using an estimated time of arrival (ETA) prediction based on a graph neural network according to the present disclosure will be described below with reference to the accompanying drawings. In the process, the thickness of the lines and the sizes of the components shown in the accompanying drawings may be exaggerated for the sake of clarity and convenience of description. In addition, the terms used below are defined in consideration of their functions in the present disclosure and may vary depending on the intention or convention of the user or operator. Therefore, these terms should be defined according to the context of this specification.

[0034] Figure 1 is an example diagram showing a schematic construction of a device for searching a route using ETA prediction based on a graph neural network according to one embodiment of the present disclosure.

[0035] like Figure 1 As shown, the device for searching a route using ETA prediction based on a graph neural network according to the present embodiment includes a global positioning system (GPS) module 110 , a storage module 120 , a processor 130 , and a communication module 140 .

[0036] The GPS module 110 receives a GPS signal to detect a current location of a vehicle, and also receives current location information of a navigation device using the GPS module.

[0037] In addition, although not shown in the drawings, the GPS module 110 can receive traffic (traffic volume) information and vehicle traffic volume information of each link, lane information of each road, and traffic condition information (e.g., accident information, event information, construction information, etc.) through a real-time traffic information receiving module (not shown) or a communication module 140.

[0038] The internal memory (or database) (not shown) of the processor 130 stores digital map data (eg, precise map data, high definition (HD) map, etc.), and the digital map data includes geographic coordinates indicating latitude and longitude in degrees / minutes / seconds.

[0039] The storage module 120 may store link information based on the digital map data, and may store a traffic information history for each link received through the communication module 140 .

[0040] The storage module 120 may store the real-time traffic information history of each link in units of season, day, and time.

[0041] In this case, the storage module 120 may store the real-time traffic information history of each link in the server 200 , or store the real-time traffic information history for a certain period of time (eg, a designated period of time, ie, one year) by interlocking with the server 200 .

[0042] Therefore, the information stored in the storage module 120 below should be understood to include the information stored in the server 200 .

[0043] The processor 130 may learn, for example, by performing deep learning by reflecting the information stored in the storage module 120 (e.g., real-time traffic information for each link and information about road characteristics (e.g., the number of lanes, the presence of traffic lights, the number of turns, the collection trajectory, the history of traffic information, etc.)).

[0044] The processor 130 is a concept including a path search engine (eg, a path search algorithm) and an ETA prediction model (eg, a deep learning model for ETA prediction).

[0045] The processor 130 may calculate (calculate) a plurality of candidate routes through a route search engine, and calculate (calculate) an optimal route among the candidate routes through an ETA prediction model.

[0046] The processor 130 may calculate a more accurate optimal route during route search through a graph neural network-based ETA prediction model.

[0047] The communication module 140 communicates with the server 200 (eg, an ETA prediction server, a cloud server, a navigation server, etc.).

[0048] The device for searching a route using ETA prediction based on a graph neural network according to the present embodiment is a concept including a navigation terminal (not shown) installed in a vehicle and at least one external server 200 (e.g., a navigation server, a cloud server, an ETA prediction server, etc.) connected to the navigation terminal (not shown) via communication.

[0049] For example, in a state of being communicatively connected to the server 200, the device for searching routes using ETA prediction based on a graph neural network according to the present embodiment can calculate (calculate) multiple candidate paths, and use one or more pieces of information provided to or provided from the server 200 (for example, real-time traffic information, ETA prediction information, cost information, etc.) to guide the optimal route among the multiple candidate paths (i.e., the best candidate path with the lowest cost).

[0050] Alternatively, in some embodiments, the server 200 may be configured to receive a route search request from a navigation terminal installed in a vehicle, calculate an optimal route by reflecting an ETA of a corresponding route, and then provide the optimal route to the navigation terminal.

[0051] In addition, since the server 200 is also equipped with a computing device (such as a processor) to perform the above-mentioned route search operation, it can be described as the route search operation performed by the server 200 being performed by the processor.

[0052] Figure 2 and Figure 3 is an example diagram for describing a graph transformation method of a device for searching a route using ETA prediction based on a graph neural network according to one embodiment of the present disclosure; Figure 4 is an example diagram for describing an ETA prediction model of a device for searching a route using a graph neural network-based ETA prediction according to one embodiment of the present disclosure.

[0053] The ETA prediction model in the device for searching a route using ETA prediction based on a graph neural network according to the present disclosure utilizes a graph neural network, and the graph neural network is a deep learning model for learning graph-structured data. By learning the connection between the data and the topological structure of the data, an embedding representing the attributes of the graph can be generated.

[0054] A graph has a combined structure with node and edge properties, and when two nodes are adjacent to each other, the connection between the nodes is represented as an edge. When there is directionality in the edges, it is called a directed graph.

[0055] Meanwhile, road information is generally composed of nodes and links, where nodes refer to road nodes (eg, intersections) and links refer to road lines connecting these nodes.

[0056] Therefore, if Figure 2 and Figure 3 As shown, the road information can be converted into a graph by representing the links constituting the road as nodes of the graph and representing the connection and change information between the links as edges.

[0057] A graph used in the present disclosure may be a directed graph including global properties representing the graph itself.

[0058] In this case, the characteristics of the entire route can be represented globally.

[0059] Therefore, when the node, edge, and global attributes are denoted as V, E, and u respectively, a graph with attributes V, E, and u can be denoted as G(V, E, u).

[0060] Additionally, the graph represents the directionality of the edges as sender / receiver nodes according to the road travel direction. The sender node may include information about the outgoing link, while the receiver node may include information about the incoming link.

[0061] The node attributes of the graph may include data such as link attributes, past speed of the corresponding link (e.g., speed within the past 40 minutes), and link passing time. The edge attributes of the graph may include data such as the connection between nodes and whether the road type has changed. Global attributes may include data such as past ETA information about the corresponding route (e.g., ETA information within the past 40 minutes), day of the week information (weekdays and weekends), and accident information (collision and construction).

[0062] like Figure 4 As shown, according to the present disclosure, the road network is represented as graph-structured data to generate an ETA prediction model based on a graph neural network. Specifically, the graph network (GN) block structure is used not only to implement the application of the spatial characteristics of the road network, but also to implement the application of the temporal characteristics of the road network. Therefore, the ETA prediction model according to the present disclosure has the advantages of reflecting complex road network characteristics and minimizing the loss of spatiotemporal data.

[0063] refer to Figure 4 The ETA prediction model structure according to the present disclosure is based on the GN block structure. This allows learning the relationship between node blocks, edge blocks and global blocks.

[0064] The features of the GN block-based structure are feature propagation, feature transformation and feature aggregation. Feature propagation propagates the features of nodes and edges to model the interactions between adjacent nodes and edges. Feature transformation updates the features of nodes and edges into new features by performing linear or nonlinear transformations on the features. Feature aggregation aggregates the features of adjacent nodes and edges to generate new features for adjacent nodes and edges.

[0065] In the present disclosure, the model structure is designed so that the features of the predicted target link and adjacent links are reflected to the global blocks in the node blocks, edge blocks, and node blocks of the GN block, and then the adjacent links are excluded from learning, allowing focus on the predicted target link.

[0066] In addition, the distinction value between a link within a route and an adjacent link and whether the road type has changed can be represented as a binary value, and the binary value can be used as an additional feature. When learning is performed by adding additional features, the connection of edges can be emphasized.

[0067] At the same time, when Figure 4 When the GN block core is repeated k times, the features in the wider range of links can be reflected as k-hops (a set of nodes that can be reached from the reference node through k edges). In this case, in order to correctly reflect the features of various scales obtained by k repetitions, the present disclosure applies a bidirectional feature pyramid network (BiFPN). BiFPN is to distinguish the contribution of input features with different resolutions to output features, and is a technique that adds a bottom-up approach to the existing top-down FPN to allow the extraction of resolution information of various scales.

[0068] In addition, for the loss function, when the actual arrival time exceeds the ETA, the driver tends to have a lower service satisfaction. According to the present disclosure, in order to compensate for the above-mentioned problem, a weight that is less than the correct answer is applied to the prediction during model learning. That is, in order to detect changes in traffic conditions, weights are also applied to peak hours with heavy traffic (e.g., 05:40 to 11:30 and 15:30 to 19:30). That is, in the ETA prediction model according to the present disclosure, weights can be applied to predict a value greater than the predicted value (which is less than the correct answer), and weights can be applied to increase the influence of a preset time zone on the output value of the ETA prediction model.

[0069] For example, Huber loss may be used in the present disclosure, and Huber loss is a technique that combines the advantages of mean square error (MSE), which is differentiable at all points, and mean absolute error (MAE), which is insensitive to outliers.

[0070] In addition, in the present disclosure, an exponential moving average (EMA) may be applied. EMA is a technique that gives a higher weight to recent data and a lower weight to past data to achieve a lower impact. This technique enables learning to take trends into account.

[0071] Meanwhile, regarding the output value of the ETA prediction model, in the present disclosure, both the attributes of the node and the global can be used as the correct labels (target node and target global). The target node refers to the node to be predicted by the model, and can include the link passing time information in units of 5 minutes up to 30 minutes in the future, and the target global refers to the global to be predicted by the model, and can include the ETA information of the entire route in units of 5 minutes up to 30 minutes in the future.

[0072] Similar to the learning process of a general deep learning model, this learning process is performed to infer the result value using the initial artificial intelligence (AI) model based on the above input values ​​(global, node, and edge attributes), derive the error value between the inferred value and the correct value (label value) through the loss function, and perform AI model learning (parameter update) based on the derived error value.

[0073] Figure 5 is a flowchart for describing a method for searching a route using ETA prediction based on a graph neural network according to one embodiment of the present disclosure, Figure 6 is a flowchart for describing using an ETA prediction model in a method for searching a route using ETA prediction based on a graph neural network according to one embodiment of the present disclosure, Figure 7 is a flowchart for describing an ETA prediction operation of a method for searching a route using a graph neural network-based ETA prediction according to one embodiment of the present disclosure.

[0074] like Figure 5 As shown, the processor 130 receives a route search request (S200). For example, the processor 130 may receive a route search request including departure point information, destination information, and departure time information from a user through an input module of the navigation device.

[0075] Then, the processor 130 performs a route search using the ETA prediction model ( S210 ).

[0076] The operation of route search is generally performed based on a route search engine. The route search engine can search for a route from a departure point to a destination based on information such as departure point information, destination information, and departure time information, and select and output candidate routes that meet various conditions (e.g., shortest distance, shortest time, free road priority). However, since the route search engine is a technology that has been widely used in the technical field of the present disclosure, a detailed description thereof will be omitted here.

[0077] Therefore, if Figure 6As shown, the processor 130 selects a plurality of candidate routes based on a route search engine (S211). In this case, the route search engine may use a method of selecting candidate routes by calculating the cost of each link and calculating the candidate routes in order of low total cost. Here, the cost of a link is a concept of quantifying the cost for passing through a link, and may be calculated by considering the length of the link and the expected passing time of the link. In addition, the ETA of the candidate route may be predicted and reflected in the cost to calculate the final cost of the corresponding route.

[0078] However, a specific calculation method of the cost of the navigation route may vary depending on the user's intention and the design of the navigation system.

[0079] In this case, since the route search engine often uses ETA even when calculating candidate routes, in some embodiments, the route search engine may be configured to include the above-mentioned ETA prediction model according to an embodiment of the present disclosure to calculate candidate routes.

[0080] Alternatively, as described below, the route search engine may select candidate routes according to an existing method, but may adopt a method of re-estimating costs by additionally performing ETA prediction using the above-mentioned ETA prediction model according to an embodiment of the present disclosure.

[0081] Therefore, the processor 130 calculates the ETA of the candidate routes using the above-mentioned ETA prediction model according to an embodiment of the present disclosure ( S212 ).

[0082] For example, Figure 7 As shown, the processor 130 converts the candidate route data into a directed graph (S2221). That is, the candidate route can be converted into a directed graph by representing the links constituting the candidate route as nodes of the graph and representing the connection and change information between the links as edges.

[0083] Thereafter, the processor 130 inserts the relevant data into the global attributes, the node attributes of the graph, and the edge attributes of the graph (S2222), and then performs ETA prediction using a learning model based on a graph neural network (S2223). Figure 4 The attribute data in the input is used for learning in the same way as the data related to the candidate routes are input to perform ETA prediction.

[0084] As described above, the ETA prediction model output value can be the ETA value of the entire route and / or the link passing time. That is, the ETA value of the entire route can be directly predicted, or the passing time of each link of the entire route can be predicted, and in the latter case, the sum of the link passing times can be used as the ETA prediction value.

[0085] Subsequently, the processor 130 calculates the cost of the candidate routes based on the calculated ETA ( S213 ).

[0086] Thereafter, the processor 130 derives the route search result based on the calculated cost (S214). That is, the processor 130 may recalculate the cost and sort based on the estimated ETA of the candidate routes to sort the candidate routes in order of low (small) cost.

[0087] Subsequently, the processor 130 provides a route search result (S220). That is, the processor 130 may provide a list of candidate routes and their ETAs to the user as a route search result.

[0088] According to the device and method for searching a route using estimated time of arrival (ETA) prediction based on a graph neural network disclosed in the present invention, accurate ETA information is provided to the driver by predicting the ETA by representing the road network as a graph neural network, which has the effect of improving the efficiency of time and resource utilization.

[0089] In addition, the device and method for searching routes using ETA prediction based on a graph neural network according to the present disclosure have the effect of sensitively detecting changes in traffic information (speed) during peak hours by applying weights related to ETA and improving drivers' service satisfaction with ETA prediction.

[0090] Although the present disclosure has been described with reference to the embodiments shown in the accompanying drawings, these embodiments are merely illustrative, and it should be understood that those skilled in the art can derive various modifications and other equivalent embodiments based on these embodiments. Therefore, the technical scope of the present disclosure should be defined by the appended claims.

Claims

1. A device for searching a route using estimated time of arrival prediction based on a graph neural network, the device comprising: a storage module configured to store digital map data; as well as a processor configured to perform a route search based on an estimated arrival time prediction model in response to a route exploration request, Among them, the estimated arrival time prediction model is based on a graph neural network formed by converting road information into a graph.

2. The device according to claim 1, wherein: The estimated arrival time prediction model is based on converting links of road information into nodes of a graph and converting connections between links into edges of the graph.

3. The device according to claim 2, wherein: The estimated time of arrival prediction model is based on a graph neural network, which includes the characteristics of the entire route as global attributes.

4. The device according to claim 2, wherein: The characteristics of the node of the graph include information about past speeds or transit times of links corresponding to the node of the graph.

5. The device according to claim 1, wherein: A weight is applied to the estimated time of arrival prediction model for values ​​predicted to be greater than a predicted value, wherein the predicted value is less than a correct answer.

6. The device according to claim 1, wherein: A weight for increasing the influence of a preset time zone on an output value of the estimated arrival time prediction model is applied to the estimated arrival time prediction model.

7. The device according to claim 1, wherein: The processor calculates a plurality of candidate routes according to the route search request, and calculates an estimated arrival time of each candidate route by using the estimated arrival time prediction model.

8. The device according to claim 7, wherein: The processor calculates a cost for each candidate route based on the calculated estimated time of arrival for each candidate route.

9. A method for generating an estimated time of arrival prediction model, the method comprising: The processor converts the links of the road information into nodes, and converts the connections between the links into edges of a graph, thereby converting the road information into the graph; Including, by the processor, relevant data into global properties of the graph, node properties of the graph, and edge properties of the graph; as well as An estimated time of arrival prediction model is learned by the processor using the global attribute of the graph, the node attribute of the graph, and the edge attribute of the graph as input values.

10. The method according to claim 9, wherein: The global attributes of the graph include past estimated time of arrival information for the entire route; The node attributes of the graph include information about past speeds or transit times of links corresponding to nodes of the graph; and The edge attributes of the graph include data including connections between the nodes and whether a road type has changed.

11. The method according to claim 9, wherein: The estimated arrival time prediction model outputs the link transit time as an output value.

12. The method according to claim 9, wherein: The estimated time of arrival prediction model outputs an estimated time of arrival value for the entire route.

13. A method for searching a route using estimated time of arrival prediction based on a graph neural network, the method comprising: Receiving, by the processor, a route search request; performing, by the processor, a route search based on an estimated time of arrival prediction model; as well as providing, by the processor, route search results, Among them, the estimated arrival time prediction model is based on a graph neural network formed by converting road information into a graph.

14. The method according to claim 13, wherein: Performing a route search involves: calculating, by the processor, a plurality of candidate routes in response to the route search request; and The processor calculates an estimated arrival time of each candidate route using the estimated arrival time prediction model.

15. The method according to claim 14, wherein: Performing the route search further includes calculating, by the processor, a cost for each candidate route based on the calculated estimated time of arrival for each candidate route.

16. The method according to claim 14, wherein: Calculating the estimated arrival time of each candidate route by the estimated arrival time prediction model includes: The processor converts the links of the candidate routes into nodes of the graph, and converts the connections between the links into edges of the graph, thereby converting each candidate route into a corresponding graph; and The processor calculates the estimated arrival time of each candidate route through the estimated arrival time prediction model by inputting data related to the global attributes of the graph, the node attributes of the graph, and the edge attributes of the graph as input values.

17. The method of claim 16, wherein: The global attributes of the graph include past estimated arrival time information for the entire route; The node attributes of the graph include information about past speeds or transit times of links corresponding to nodes of the graph; and The edge attribute data of the graph includes the connection between the nodes and whether the road type has changed.

18. The method of claim 14, wherein: A weight is applied to the estimated time of arrival prediction model for values ​​predicted to be greater than a predicted value, wherein the predicted value is less than a correct answer.

19. The method according to claim 14, wherein: A weight for increasing the influence of a preset time zone on an output value of the estimated arrival time prediction model is applied to the estimated arrival time prediction model.

Images