DEVICE AND METHOD FOR ROUTE FINDING WITH ETA PREDICTION BASED ON A NEURAL GRAPH NETWORK

Abstract

A device and method for searching for a route using an estimated time of arrival (ETA) prediction based on a neural graph network are provided, comprising a memory 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 search request, wherein the ETA prediction model is based on a neural graph network formed by converting road information into a graph.

Description

BACKGROUND AREA

[0001] Example embodiments of the present disclosure relate to a route search apparatus and method using an estimated time of arrival (ETA) prediction based on a graph neural network, and more particularly to a route search apparatus and method that perform route search by expressing a road network as a graph neural network and using the road network to predict an ETA. DISCUSSION OF THE BACKGROUND

[0002] Since most vehicles are now equipped with navigation devices, and due to the increasing number of vehicles, traffic jams on roads and intersections are common, even when driving on a road that is already familiar, it is common practice to search for a route in advance using a navigation device and be guided so that sections such as roads or intersections where traffic jams occur can be identified in advance and avoided.

[0003] In addition, users often use estimated time of arrival (ETA) information when searching for a route. For example, the departure time or the time of an appointment is determined based on the ETA information. Since the ETA information is not accurate, the user adds an extra time to determine the departure time or the appointment time. However, because the ETA information is accurate, the extra time can be reduced, allowing the user to save time without losing time on the road.

[0004] For example, during a route search process in navigation, an optimal route, a recommended route, etc. are calculated using the ETA information.

[0005] As deep learning-based information prediction technology continues to evolve, active research is currently underway to explore deep learning-based technology for more accurate estimated time of arrival prediction.

[0006] Therefore, an efficient method for representing road data is required so that the road data can be used as input for a deep learning model. SUMMARY

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

[0008] A device for searching for 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 investigation 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 a link of road information into a node of a graph and converting the connectivity between links into edges of the graph.

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

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

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

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

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

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

[0016] A method for generating an estimated time of arrival (ETA) prediction model according to the present disclosure includes converting, by a processor, links of road information into nodes and converting the connectivity between the links into edges of a graph to convert the road information in the graph, including data related to it, by the processor into a global attribute of the graph, a node attribute of the graph, and an edge attribute of the graph, and learning, by the processor, an ETA prediction model using the global attribute of the graph, the node attribute of the graph, and the edge attribute of the graph as input values.

[0017] In the present disclosure, the global attribute of the graph may include information about the previous arrival time for the entire route, the node attribute of the graph may include information about a previous speed or a transit time of a link corresponding to a node of the graph, and the edge attribute of the graph may include data such as the connectivity between the nodes and whether a road type has changed.

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

[0019] Within the scope of the present disclosure, the ETA prediction model may output ETA values ​​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, by a processor, a route search request; performing, by the processor, a route search based on an ETA prediction model; and providing, by the processor, a route search result, 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 the route search may further include calculating, by the processor, a plurality of candidate routes in response to the route search request; and calculating, by the processor, an ETA of each candidate route using the ETA prediction model.

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

[0023] In the present disclosure, the calculation of the ETA of each candidate route by the ETA prediction model may include, by the processor, converting links of the candidate routes into nodes of the graph and converting the connectivity between the links into edges of the graph, thereby converting each candidate route into a corresponding graph; and, by the processor, calculating an ETA of each candidate route by the ETA prediction model by inputting data regarding the global attribute of the graph, the node attribute of the graph, and the edge attribute of the graph as input values.

[0024] In the present disclosure, the global attribute of the graph may include previous ETA information along the entire route, the node attribute of the graph may include information about a previous speed or traversal time of a link corresponding to a node of the graph, and the edge attribute of the graph may include data such as connectivity between nodes and changes in road type.

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

[0026] In the present disclosure, a weighting may be applied to the ETA prediction model that increases the influence of a preset time zone on the output value of the ETA prediction model. BRIEF DESCRIPTION OF THE DRAWINGS Fig. 1 is an exemplary diagram showing a schematic configuration of a device for searching for a route using the estimated time of arrival (ETA) based on a graphic neural network according to an embodiment of the present disclosure. Fig. 2 and Fig. 3 are exemplary diagrams for describing a graph transformation method of the apparatus for searching a route using ETA prediction based on a graph neural network according to an embodiment of the present disclosure. Fig. 4 is an exemplary diagram for describing an ETA prediction model of the apparatus for searching a route using a graphic neural network-based ETA prediction according to an embodiment of the present disclosure. Fig. 5 is a flowchart for describing a method for searching for a route using graph neural network-based ETA prediction according to an embodiment of the present disclosure. Fig. 6 is a flowchart for describing the use of an ETA prediction model in the method for searching for a route using the ETA prediction based on a graphic neural network according to an embodiment of the present disclosure. Fig. 7 is a flowchart for describing an ETA prediction process of the method for searching for a route using a graph neural network-based ETA prediction according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE ILLUSTRATED EMBODIMENTS

[0027] Embodiments of a route search apparatus and method using estimated time of arrival (ETA) prediction based on a graphic neural network according to the present disclosure will be described below with reference to the accompanying drawings. Incidentally, the thickness of lines and the size of components shown in the drawings may be exaggerated for clarity and simplicity of description. Furthermore, the terms used below are defined in light of their functions in the present disclosure and may vary depending on the intention of a user or operator or common practice. Therefore, these terms should be defined in the context of the present description.

[0028] Fig. 1 is an example diagram illustrating a schematic configuration of an apparatus for searching for a route using ETA prediction based on a graph neural network according to an embodiment of the present disclosure.

[0029] As in Fig. 1, the apparatus for searching a route using ETA prediction based on a graph neural network according to the present embodiment includes a GPS module 110, a storage module 120, a processor 130, and a communication module 140.

[0030] The GPS module 110 receives a GPS signal to determine the current position of a vehicle and also receives current position information from a navigation device that uses a GPS module.

[0031] In addition, although not shown in the drawing, the GPS module 110 can receive traffic information (traffic volume) and vehicle volume information for each link, lane information for each road, and traffic situation information (e.g., accident information, incident information, construction site information, etc.) via a real-time traffic information receiving module (not shown) or the communication module 140.

[0032] An internal memory (or database) (not shown) of the processor 130 stores digital map data (e.g., precision map data, a high-resolution map, and the like), and the digital map data includes geographic coordinates indicating latitude and longitude in units of degrees / minutes / seconds.

[0033] The storage module 120 may store connection information based on the digital map data and may store a traffic information history for each connection received via the communication module 140.

[0034] The storage module 120 can store a real-time traffic information history for each link in units of seasons, days, and times.

[0035] In this case, the storage module 120 may store the real-time traffic information history for each connection in the server 200 or store the real-time traffic information history for a certain period of time (eg, a certain period of time, ie, one year) by locking with the server 200.

[0036] Therefore, the information stored in the storage module 120 should be understood to include the information stored on the server 200.

[0037] The processor 130 can learn, e.g., perform deep learning, by reflecting the information stored in the memory module 120 (e.g., the real-time traffic information for each link and information about road characteristics (e.g., the number of lanes, the presence of signals, the number of turns, the collection trajectory, the traffic information history, and the like)).

[0038] The processor 130 is a concept that includes a path search engine (e.g., a path search algorithm) and an ETA prediction model (e.g., a deep learning model for ETA prediction).

[0039] The processor 130 can calculate a plurality of candidate routes through the route search engine and calculate (compute) an optimal route among the candidate routes through the ETA prediction model.

[0040] The processor 130 can calculate a more accurate optimal route during route search through an ETA prediction model based on a neural graph network.

[0041] The communication module 140 communicates with the server 200 (e.g., an ETA prediction server, a cloud server, a navigation server, and the like).

[0042] The apparatus for searching for 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 the vehicle and at least one external server 200 connected to the navigation terminal (not shown) by communication (e.g., a navigation server, a cloud server, an ETA prediction server, and the like).

[0043] For example, in a state where communication is connected to the server 200, the apparatus for searching for a route using the ETA prediction based on a graph neural network according to the present embodiment may calculate (compute) a plurality of candidate paths and guide an optimal route (ie, an optimal candidate path with the lowest cost) among the plurality of candidate paths using one or more pieces of information provided to or from the server 200 (e.g., the real-time traffic information, the ETA prediction information, cost information, and the like).

[0044] Alternatively, in some embodiments, the server 200 may be configured to receive a route search request from a navigation terminal installed in the vehicle, calculate an optimal route by considering the estimated time of arrival (ETA) for a corresponding route, and then transmit the optimal route to the navigation terminal.

[0045] Since the server 200 is also equipped with a computing device such as a processor to perform the above route search, it can be said that the route search by the server 200 is performed by a processor.

[0046] Fig. 2 and Fig. 3 are exemplary diagrams for describing a graph transformation method of the apparatus for searching for a route using ETA prediction based on a graph neural network according to an embodiment of the present disclosure; and Fig. 4 is an exemplary diagram for describing an ETA prediction model of the apparatus for searching a route using a graph neural network-based ETA prediction according to an embodiment of the present disclosure.

[0047] The ETA prediction model in the apparatus for route search using ETA prediction based on a graph neural network according to the present disclosure uses a graph neural network, and the graph neural network is a deep learning model that learns data with a graph structure. By learning the connectivity between the data and a topological structure of the data, embeddings expressing an attribute of a graph can be generated.

[0048] The graph has a combinatorial structure with node and edge attributes. If two nodes are adjacent, the connection between the nodes is expressed as an edge. If the edge is directed, it is called a directed graph.

[0049] Road information typically consists of nodes and links, where nodes refer to road junctions (e.g. intersections) and links refer to road lines that connect these junctions.

[0050] Therefore, road information, as in the Fig. shown, can be converted into a graph by expressing the links that form the road as nodes of the graph and the connectivity and change information between the links as edges.

[0051] The graph used in the present disclosure may be a directed graph with global properties that represent the graph itself.

[0052] In this case, the characteristics of the entire route can be expressed globally.

[0053] So, if node, edge, and global attributes are represented as V, E, and u, respectively, a graph with attributes V, E, and u can be represented as G(V, E, u).

[0054] Furthermore, the graph represents the directionality of the edges as sender / receiver nodes according to the direction of travel of the road. The sender node can contain information about a departure link and the receiver node can contain information about an arrival link.

[0055] The graph's node attribute can contain data such as a link attribute, a past speed of a corresponding link (e.g., a speed during the last 40 minutes), and a link transit time. The graph's edge attribute can contain data such as connectivity between nodes and changes in road type. The global attribute can contain data such as past ETA information on the corresponding link (e.g., ETA information over the last 40 minutes), information about the day of the week (weekday and weekend), and accident information (accident and construction).

[0056] As in Fig. As shown in Figure 4, according to the present disclosure, a road network is expressed as data of a graph structure to construct an ETA prediction model based on a graph neural network. Specifically, a graph network (GN) block structure is used to consider not only a spatial feature of the road network but also a temporal feature of the road network. Therefore, the ETA prediction model according to the present disclosure has the advantage of reflecting complex road network features and minimizing the loss of spatio-temporal data.

[0057] The structure of the ETA prediction model according to the present disclosure is a GN block-based structure (see Fig. 4). This allows learning about the relationship between a node block, an edge block, and a global block.

[0058] 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 interactions between neighboring nodes and edges; feature transformation updates the features of nodes and edges into new features through linear or nonlinear transformation of the features; and feature aggregation aggregates the features of neighboring nodes and edges to generate new features of neighboring nodes and edges.

[0059] In the present disclosure, the model structure is designed such that the features of a predicted target connection and neighboring connections in a node block of the GN block are reflected to the node block, the edge block, and the global block, and then the neighboring connections are excluded from learning, thereby enabling the focus on the predicted target connection.

[0060] Furthermore, the discriminatory value between a link and a neighboring link within a route and the change in road type can be expressed as a binary value, and the binary values ​​can be used as additional features. When learning is performed by adding the additional features, the connectivity of the edges can be emphasized.

[0061] If a GN block core of Fig. Meanwhile, when the feature pyramid network is repeated k times, features can be reflected over a wider range of connections than k-hop (a group of nodes that can be reached via k edges from a reference node). In this case, a bidirectional feature pyramid network (BiFPN) is used to accurately reflect the features of different scales obtained by k-times repetition. The BiFPN is designed to distinguish the degrees to which input features with different resolutions contribute to the output features and is a technique that adds a bottom-up method to the existing top-down FPN so that resolution information for different scales can be extracted.

[0062] Furthermore, in terms of a loss function, a driver tends to have lower service satisfaction when an actual arrival time exceeds an ETA. According to the present disclosure, to complement the above problem, a prediction weight smaller than a correct answer is applied during model learning. That is, in order to detect a change in traffic conditions, weights are also applied to peak hours (e.g., 5:40 to 11:30 and 15:30 to 19:30) when traffic volume is high. That is, in the ETA prediction model according to the present disclosure, a weight may be applied to predict a value larger than a predicted value smaller than a correct answer, and a weight may be applied to increase the influence of a preset time zone on an output value of the ETA prediction model.

[0063] For example, the Huber loss is a technique that combines the advantage of the mean square error (MSE), which is differentiable at all points, with the advantage of the mean absolute error (MAE), which is insensitive to an abnormal value.

[0064] Furthermore, an exponential moving average (EMA) may be used in the context of this disclosure. The EMA is a technique that gives high weight to recent data and low weight to past data to achieve low influence. This technique enables learning by considering trends.

[0065] Meanwhile, regarding an output value of the ETA prediction model, both node and global attributes can be used as correct labels (destination node and destination global) in the present disclosure. The destination node refers to a node to be predicted by the model and may include link transit time information for up to 30 minutes in units of 5 minutes. The global destination refers to a global destination to be predicted by the model and may include ETA information for the entire route for up to 30 minutes in units of 5 minutes.

[0066] Like a learning process of a general deep learning model, the learning process is performed to derive a result value using an initial artificial intelligence (AI) model based on the input values ​​described above (the global, node, and edge attributes), derive an error value between the derived value and the correct value (label value) through a loss function, and perform AI model learning (parameter update) based on the derived error value.

[0067] Fig. 5 is a flowchart for describing a method for searching for a route using a graph neural network-based ETA prediction according to an embodiment of the present disclosure, Fig. 6 is a flowchart for describing the use of an ETA prediction model in the method for searching for a route using an ETA prediction based on a graph neural network according to an embodiment of the present disclosure; and Fig. 7 is a flowchart for describing an ETA prediction process of the method for searching a route using graph neural network-based ETA prediction according to an embodiment of the present disclosure.

[0068] As in Fig. As shown in Figure 5, processor 130 receives a route search request (S200). For example, processor 130 may receive a route search request including information about the departure location, destination, and departure time from a user via an input module of a navigation device.

[0069] Then, the processor 130 performs a route search based on the ETA prediction model (S210).

[0070] A route search operation is generally performed using a route search engine. The route search engine can search for a route from a starting point to a destination based on information such as starting point information, destination information, and departure time information, and select and output a candidate route suitable for various conditions (e.g., shortest distance, shortest time, priority on toll-free roads). However, since the route search engine is a technology already widely used in the technical field of the present disclosure, a detailed description is omitted here.

[0071] Accordingly, the processor 130 selects, as in Fig. 6, the route search engine selects a plurality of candidate routes based on the route search engine (S211). In this case, the route search engine may use a method to select candidate routes by calculating the cost for each link and ranking the candidate routes in a low order of summed cost. Here, link cost is a concept for quantifying the cost of traversing the link and can be calculated by considering the length of the link and the expected transit time of the link. Furthermore, the estimated time of arrival (ETA) of a candidate route may be predicted and included in the cost to determine the final cost of the corresponding route.

[0072] However, the specific method for calculating the cost of a navigation route may vary depending on the user's intention and the design of the navigation system.

[0073] In this case, since the route search engine often uses the ETA also in calculating the candidate routes, in some embodiments the route search engine may be configured to use the above-described ETA prediction model according to the present embodiment to calculate the candidate routes.

[0074] Alternatively, as described below, the route search engine may select a candidate route according to the existing method, but apply a cost re-estimation method by additionally performing ETA prediction using the ETA prediction model described above according to the present disclosure.

[0075] Accordingly, the processor 130 calculates ETAs of the candidate routes using the above-described ETA prediction model according to the present embodiment of the disclosure (S212).

[0076] As in Fig. For example, as shown in Figure 7, 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 expressing the links constituting the candidate routes as nodes of the graph and expressing the connectivity and change information between the links as edges.

[0077] Thereafter, the processor 130 inserts related data into the global attribute, the node attribute of the graph, and the edge attribute of the graph (S2222) and then performs ETA prediction using a graph neural network-based learning model (S2223). That is, the ETA prediction is performed by inputting data related to the candidate routes in the same way as attribute data input for learning in Fig. 4 can be carried out.

[0078] As described above, the output of an ETA prediction model can be an ETA value and / or a link transit time for the entire route. That is, the ETA value for the entire route can be directly predicted, or the transit time of each link on the entire route can be predicted, and in the latter case, a sum of the link transit times can be used as the ETA prediction value.

[0079] Subsequently, the processor 130 calculates the costs of the candidate routes based on the calculated ETA (S213).

[0080] Next, the processor 130 derives a route search result based on the calculated cost (S214). That is, the processor 130 may recalculate and sort the costs based on the ETAs of the estimated candidate routes to sort the candidate routes in order of lower (smaller) cost.

[0081] Subsequently, the processor 130 provides the result of the route search (S220). This means that the processor 130 can provide the user with a list of possible routes and their estimated arrival times as a route search result.

[0082] According to an apparatus and method for searching a route using an estimated time of arrival (ETA) based on a neural network of the present disclosure, the efficiency of time and resource utilization is increased by providing a driver with accurate ETA information by expressing a road network as a neural graph network to predict an ETA.

[0083] Furthermore, the apparatus and method for searching a route using the ETA prediction based on a graphic neural network and the method according to the present disclosure have the effect of sensitively detecting a change in traffic information (speed) in the peak hour by applying a weight with respect to the ETA prediction and improving a driver's perceived service satisfaction with respect to the ETA prediction.

[0084] While the present disclosure has been described with reference to the embodiments shown in the drawings, these embodiments are merely illustrative, and it should be understood that various modifications and equivalent other embodiments can be derived by those skilled in the art based on the embodiments. Therefore, the technical scope of the present disclosure should be defined by the appended claims.

Claims

[1] Device for searching a route using an estimated time of arrival (ETA) prediction based on a neural graph network, the device comprising: a memory module (120) configured to store digital map data; and a processor (130) configured to perform a route search based on an ETA prediction model in accordance with a route exploration request, where the ETA prediction model is generated based on a graph neural network formed by converting road information into a graph. [2] The device according to claim 1, wherein the ETA prediction model is generated by converting a link of road information into a node of a graph and by converting the connectivity between links into edges of the graph. [3] The device according to claim 1 or 2, wherein the ETA prediction model is generated based on a neural graph network that includes features of the route as global attributes. [4] Device according to one of the preceding claims, wherein a feature of the node of the graph contains information about a past speed or a transit time of a link corresponding to a node of the graph. [5] The device of any preceding claim, wherein a weight is configured to predict a value greater than a predicted value less than a correct answer, and wherein the weight is configured to be applied to the ETA prediction model. [6] The device of any preceding claim, wherein a weight is configured to increase the influence of a preset time zone on an output value of the ETA prediction model, and wherein the weight is configured to be applied to the ETA prediction model. [7] The device of any preceding claim, wherein the processor (130) is further configured to: calculate a plurality of candidate routes according to a route search query and calculate an ETA of each candidate route through the ETA prediction model. [8] The apparatus of any preceding claim, wherein the processor (130) is further configured to calculate the cost of each candidate route based on the calculated ETA of each candidate route. [9] A method for generating an estimated time of arrival (ETA) prediction model, the method comprising: Converting, by a processor (130), links of the road information into nodes and converting the connectivity between the links into edges of a graph to convert the road information into the graph; Inserting related data by the processor (130) into a global attribute of the graph, a node attribute of the graph, and an edge attribute of the graph; and Learning an ETA prediction model by the processor (130) 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 of claim 9, wherein: the global attribute of the diagram contains information about previous arrival times for a route; the node attribute of the graph contains information about a past speed or a transit time of a link corresponding to a node of the graph; and the edge attribute of the graph contains information about the connectivity between the nodes and information about whether a road type has been changed.

Images