A method and system for calculating estimated arrival time based on graph network

Abstract

The embodiment of the present invention provides a method and system for calculating the estimated arrival time based on a graph network, and simultaneously mines the vehicle type and weather information of the historical trajectory data. In terms of the deep integration of time and space information, the most natural graph neural network is used for modeling, and the mutual influence of the traffic speed of the connected sections in the local area is fully considered, and the road topology is modeled with the most appropriate graph structure. By projecting the trajectory data on the graph structure, the time information and the spatial structure information are fully integrated through feature modeling; in the feature modeling process, the road attribute information that affects the vehicle flow in the high-precision map is introduced, such as the road type of the section, the maximum speed limit of the section, the number of lanes of the section, whether it is an intersection, etc. This information is introduced into the static graph network modeling in the form of information embedding; first, the cache machine is used to ensure the real-time query, and secondly, the world is divided into local areas to ensure the stability and convergence speed of the graph network training.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of high-precision map application in trajectory data mining, and in particular to a method and system for calculating an estimated arrival time based on a graph network. Background Art

[0002] The problem of Estimated time of arrival (ETA) prediction is a key business problem that map companies and travel companies need to solve. In the context of a fast-paced society, the accuracy of ETA calculation is particularly important. Currently, most map companies and travel companies predict ETA by mining historical trajectory data. Recurrent Neural Network (RNN) and Convolutional Neural Network (CNN) are used for modeling, but the common method cannot deeply integrate time information and spatial information. ETA is in four-dimensional space, and time and spatial position affect each other. In addition, road attribute information, weather information, vehicle type information, holiday information, etc. are all key influencing factors for ETA estimation. Summary of the invention

[0003] The embodiments of the present invention provide a method and system for calculating the estimated arrival time based on a graph network, which fully utilizes the road attribute information of high-precision maps and mines the vehicle type and weather information of historical trajectory data. The trajectory data is projected onto the graph structure and the time information and spatial structure information are fully integrated through feature modeling.

[0004] In a first aspect, an embodiment of the present invention provides a method for calculating an estimated arrival time based on a graph network, comprising:

[0005] Step S1, dividing the high-precision map into a number of fixed-size areas, and constructing a super section based on all road sections in each area;

[0006] Step S2: obtaining a given query path, projecting the given query path onto a super-segment of a high-precision map, and determining a super-segment sequence corresponding to the given query path;

[0007] Step S3, querying the passing time of each super-segment in the super-segment sequence within different future fixed time ranges based on the cache table obtained by pre-training to determine the passing time of the super-segment sequence; determining the estimated arrival time ETA based on the passing time of the super-segment sequence.

[0008] Preferably, the step S1 specifically includes:

[0009] The high-precision map is divided into a number of fixed-size areas, and all road segments contained in each area form a super road segment; wherein the road segment is the smallest road expression unit in the high-precision map;

[0010] A graph structure is defined for each super segment, wherein the segment is a vertex of the graph structure; if any two segments are adjacent and satisfy upward connectivity or downward connectivity, the two adjacent segments constitute an edge.

[0011] Preferably, in step S3, before querying the passing time of each super-segment in the super-segment sequence within different future fixed time ranges based on the cache table obtained through pre-training, the step further includes:

[0012] Capturing the real-time traffic flow and historical traffic flow in the area where the graph structure is located to determine the historical trajectory sequence and real-time trajectory sequence of each road section, wherein the historical trajectory sequence includes the real-time average passing time sequence and the real-time average passing speed sequence of the road section at any time within the cycle day, and the real-time trajectory sequence is the historical average passing time sequence and the historical average passing speed sequence of the road section at any time within the cycle day;

[0013] Projecting the historical trajectory sequence and the real-time trajectory sequence onto the corresponding super-segment to obtain the node feature vector, edge feature vector, and graph feature vector of the super-segment; the node feature vector includes the real-time average passing time series, the real-time average passing speed series, the historical average passing time series, the historical average passing speed series, and segment attribute information on the segment; the graph feature vector is the real-time passing time vector of the super-segment;

[0014] A graph neural network is constructed, and graph neural network training is performed based on the node feature vector and the graph feature vector to obtain a predicted passing time list, and a cache table is updated based on the predicted passing time list.

[0015] Preferably, the real-time traffic flow and historical traffic flow in the area where the graph structure is located are captured to determine the historical trajectory sequence and real-time trajectory sequence of each road section, specifically including:

[0016] All the trajectory segments on the road section passed at t0 are intercepted, and the time of the last trajectory point of each intercepted trajectory segment minus the time of the first point of the trajectory segment is used to obtain a passing time t of the road section; determine the mean passing time d0 of all trajectory segments; calculate t0-i*Δt r The travel time is denoted as d i ,Δt r The preset fixed time window is i=1,2,3,…,m; determine the real-time average passing time series vec-time at time t0 on the road section t0 For m , ctm-1 ,…, ct2,ct1, ct0]; calculate the real-time average passing speed sequence vec-speed at time t0 on the road section t0 [v m , v m-1 ,…, v2,v1, v0];

[0017] Aggregate all trajectories at the same time in the same cycle day to determine the p Δt before the specified time h The historical average of the moment through the time series, q Δt h The historical average passing speed sequence at time Δt h , Δt h The preset time window is a fixed time window.

[0018] Preferably, constructing a graph neural network specifically includes:

[0019] Construct the graph structure encoding module, graph network update module and graph network decoding prediction module of the graph neural network;

[0020] The graph structure encoding module is used to encode the node feature vector, edge feature vector and graph feature vector of the super section to generate a node feature potential representation vector, an edge feature potential representation vector and a graph feature potential representation vector;

[0021] The graph network update module is used to update the node feature potential representation vector, the edge feature potential representation vector and the graph feature potential representation vector respectively;

[0022] The graph network decoding prediction module is used to predict a predicted passing time list based on the node feature potential representation vector, the edge feature potential representation vector and the graph feature potential representation vector.

[0023] Preferably, the graph network decoding prediction module includes an update function for updating the node feature potential representation vector Φ v , an update function used to update the edge feature potential representation vector Φ e , an update function used to update the latent representation vector of graph features Φ u ;

[0024] For any edge, the update function Φ e The independent variables are the corresponding edge feature vector, the edge source node feature vector, the edge target node feature vector, and the graph feature vector u; the edge feature vector set is recorded as E′, and the edge feature vector set connected to node i is recorded as Aggregate the edges of node i to obtain the representation vector of the edges of node i, and perform super-segment aggregation on the updated edges to obtain the representation vector of the aggregated edges.

[0025] As a preferred embodiment, it also includes:

[0026] Constructing the loss function of the graph neural network, the loss function includes the loss function on the super segment , the loss function on the road segment , the cumulative loss function on all road sections ;

[0027]

[0028] In the above formula, n Indicates the total number of super sections, m （i） The number of segments in super segment i; represents the passing time of super section i, represents the predicted passing time of super section i; is the cumulative passing time of all sections on the super section i, is the cumulative predicted passing time of all sections on the super section i; The passing time of section j in super section i, The predicted passing time of segment j in super segment i; Free passing time of super section i, free passing time refers to the passing time when the number of vehicles on the road is less than the preset value or there are no vehicles; δ To adjust the parameters.

[0029] In a second aspect, an embodiment of the present invention provides an estimated arrival time calculation system based on a graph network, comprising:

[0030] The super-segment division module divides the high-precision map into several fixed-size areas and constructs a super-segment based on all the road segments in each area;

[0031] A traffic flow fusion module obtains a given query path, projects the given query path to a super-segment of a high-precision map, and determines a super-segment sequence corresponding to the given query path;

[0032] The graph neural network prediction module queries the passing time of each super-segment in the super-segment sequence within different fixed time ranges in the future based on the cache table obtained through pre-training to determine the passing time of the super-segment sequence; and determines the estimated arrival time ETA based on the passing time of the super-segment sequence.

[0033] In a third aspect, an embodiment of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of the method for calculating an estimated arrival time based on a graph network as described in the embodiment of the first aspect of the present invention are implemented.

[0034] In a fourth aspect, an embodiment of the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for calculating the estimated arrival time based on a graph network as described in the embodiment of the first aspect of the present invention.

[0035] The embodiment of the present invention provides a method and system for calculating the estimated arrival time based on a graph network, and at the same time mines the vehicle type and weather information of the historical trajectory data. In terms of the deep integration of time and space information, the most natural graph neural network is used for modeling, and the mutual influence of the traffic speed of the connected road sections in the local area is fully considered, and the road topology is modeled with the most appropriate graph structure. By projecting the trajectory data on the graph structure, the time information and the spatial structure information are fully integrated through feature modeling; secondly, in the feature modeling process, the road attribute information that affects the vehicle flow in the high-precision map is introduced, such as the road type of the section, the maximum speed limit of the section, the number of lanes of the section, whether it is an intersection, etc. This information is introduced into the static graph network modeling in the form of information embedding; thirdly, in terms of engineering, the cache machine is first used to ensure the real-time query. Secondly, the world is divided into local areas to ensure the stability and convergence speed of graph network training. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0037] Figure 1 A flowchart of a method for calculating an estimated arrival time based on a graph network according to an embodiment of the present invention;

[0038] Figure 2 An example diagram of modeling a super-segment graph structure according to an embodiment of the present invention;

[0039] Figure 3 A flowchart of a graph neural network processing according to an embodiment of the present invention;

[0040] Figure 4 A flowchart of extracting features from a super-segment graph according to an embodiment of the present invention;

[0041] Figure 5 A flowchart of a map ETA application according to an embodiment of the present invention;

[0042] Figure 6 is a flow chart of ETA calculation according to an embodiment of the present invention;

[0043] Figure 7 FIG. 4 is a schematic diagram of a physical structure according to an embodiment of the present invention. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] The problem of Estimated time of arrival (ETA) prediction is a key business problem that map companies and travel companies need to solve. In the context of a fast-paced society, the accuracy of ETA calculation is particularly important. Currently, most map companies and travel companies predict ETA by mining historical trajectory data. Recurrent neural networks and convolutional neural networks are used for modeling, but the common methods cannot deeply integrate time information and spatial information. ETA is in four-dimensional space, and time and spatial position affect each other. In addition, road attribute information, weather information, vehicle type information, holiday information, etc. are all key influencing factors for ETA estimation.

[0046] Therefore, the embodiments of the present invention provide an estimated arrival time calculation method and system based on a graph network, which fully utilizes the road attribute information of high-precision maps, and at the same time mines the vehicle type and weather information of historical trajectory data; in terms of the deep fusion of time and space information, the most natural graph neural network is used for modeling; and the caching mechanism is used in engineering to construct a graph structure of a local area to ensure fast training and convergence on the graph. The following will be explained and introduced through multiple embodiments.

[0047] Figure 1 A method for calculating an estimated arrival time based on a graph network is provided for an embodiment of the present invention, comprising:

[0048] Step S1, dividing the high-precision map into a number of fixed-size areas, and constructing a super section based on all road sections in each area;

[0049] In this embodiment, the road topology is divided into local areas, and the road traffic conditions at different times are learned in the local areas. Because driving behaviors often have great similarities in local areas, the road topology structure is modeled in the local area using a graph network (GNN). A local area is a graph network. The size of the area is determined by fine-tuning the experimental data.

[0050] The high-precision map is divided into a number of fixed-size areas, and all road segments contained in each area form a super road segment; wherein the road segment is the smallest road expression unit in the high-precision map;

[0051] A graph structure is defined for each super segment, wherein the segment is a vertex of the graph structure; if any two segments are adjacent and satisfy upward connectivity or downward connectivity, the two adjacent segments constitute an edge.

[0052] If two segments are adjacent and satisfy the vehicle flow rules (upward or downward connectivity), there is an edge between the two vertices. Graph S=(S,E), S is the set of segments in the supersegment, and E is the set of edges in the supersegment. Any segment rs in the supersegment i ∈S. Any two segments rs in the super segment i and rs j , if rs i and rs j If they are physically adjacent and vehicles can flow directly on these two road segments, then rs i and rs j The edge e ij ∈E. The corresponding road topology map is as follows Figure 2 ,The corresponding super segment on the left side is the road topology graph, and the corresponding one on the right side is the graph structure.

[0053] Step S2: obtaining a given query path, projecting the given query path onto a super-segment of a high-precision map, and determining a super-segment sequence corresponding to the given query path;

[0054] The significant characteristics of traffic flow in a region are manifested in the following three aspects: first, the real-time traffic flow pattern (the traffic flow conditions of the previous moment in time series will affect the later ones, lag). Secondly, the historical traffic flow pattern, which mainly shows the periodicity of daily (for example, the similarity of the daily working hours / off-get off work time periods) and weekly (for example, the similarity of traffic flow on different days of the week). Therefore, dynamic traffic flow modeling needs to capture both real-time and historical traffic flow patterns. In order to increase the dimension of data, real-time and historical time are expressed in the form of time series.

[0055] Capture the real-time traffic flow and historical traffic flow in the area where the graph structure is located to determine the historical trajectory sequence and real-time trajectory sequence of each road section. The historical trajectory sequence includes the real-time average passing time sequence and real-time average passing speed sequence of the road section at any time during the cycle day. The real-time trajectory sequence is the historical average passing time sequence and historical average passing speed sequence of the road section at any time during the cycle day; the real-time and historical time series are represented as follows. Take time t0 as a reference, and use the m fixed time windows Δt before time t0 for real-time r Time series Tseq real To express, Tseq real That is [t0- m*Δt r , 。。。, t0-2Δt r , t0-Δt r , t0]. The historical time is represented by the time series of the corresponding time period and the corresponding day of the week x weeks ago. The first p and last q fixed time windows Δt corresponding to the ith week t0 in the day h Time series Tseq his , Tseq his That is [t0-Δt h -10080i, t0-2Δt h -10080i, ..., t0- p*Δt h -10080i, t0, t0+Δt h -10080i, t0+2Δt h -10080i, ..., t0+q*Δt h -10080i ]. (There are 60*24*7=10080 minutes in a week, i ).

[0056] All the trajectory segments on the road section passed at t0 are intercepted (assuming there are c trajectory segments), and the time of the last trajectory point of each intercepted trajectory segment minus the time of the first point of the trajectory segment is used to obtain the time t of passing the road section; determine the mean value d0 of the passing time of all trajectory segments; calculate t0-i*Δt r The travel time is denoted as d i ,Δt r The preset fixed time window is i=1,2,3,…,m; determine the real-time average passing time series vec-time at time t0 on the road section t0 For m , ct m-1 ,…, ct2, ct1, ct0]; in the same way, calculate the real-time average passing speed sequence vec-speed at time t0 on the road section t0 [v m , v m-1,…, v2, v1, v0]; the speed is obtained by dividing the length of the trajectory segment by the passing time, and the trajectory length is obtained by accumulating the Euclidean distance between two adjacent trajectory points (after converting the longitude and latitude coordinates to the UTM coordinates of the 3-degree zone, the plane Euclidean distance is calculated).

[0057] Aggregate all trajectories at the same time in the same cycle day (days with the same number of measurements in a cycle) to determine the p Δt before the specified time h The historical average of the moment through the time series, q Δt h The historical average passing speed sequence at time Δt h , Δt h The preset fixed time window is: For example, the time at 8 am on Monday corresponds to: Aggregate all the trajectories at 8 am on Monday in the previous x weeks, calculate the average passing time hct0 and the average passing speed at 8 am. Then calculate the p Δt before 8 am h time and q Δt h A sequence of the average transit time and average transit speed. vec-time-his t0 for [hct p ,hct p-1 ,…, hct1, hct0, hct1,…, hct q-1 , hct q ], vec-speed-his t0 [hv p ,hv p-1 ,…, hv1, hv0, hv1,…, hv q-1 , hv q ].

[0058] Projecting the historical trajectory sequence and the real-time trajectory sequence onto the corresponding super-segment to obtain the node feature vector, edge feature vector, and graph feature vector of the super-segment; the node feature vector includes the real-time average passing time series, the real-time average passing speed series, the historical average passing time series, the historical average passing speed series, and segment attribute information on the segment; the graph feature vector is the real-time passing time vector of the super-segment;

[0059] The static features of the graph structure are represented by the structure of the graph. The graph is composed of nodes and edges. The features of the graph structure are represented by node features and edge features through complex aggregation. The dynamic traffic flow features on the graph structure are represented by the features of the dynamic trajectory by projecting the historical trajectory data on the corresponding graph structure. There may be multiple projections of historical trajectory sequences on a unique super segment, and each trajectory sequence may pass through the super segment multiple times.

[0060] Node-level features represent: real-time passing time and average speed on the road segment, historical average passing time and average speed, and road segment attribute information. Road attribute information includes section length, road grade, road type, maximum speed limit, number of lanes, whether it is an intersection, etc. These road attributes are embedded in a single attribute to obtain an attribute vector, and then all attribute vectors are concat to form a vec-attr. drep seg [vec-time t0 ,vec-speed t0 , vec-time-his t0 , vec-speed-his t0 , vec-attr]

[0061] Graph feature representation: real-time passing time vector of a super segment. Because a super segment contains several modes of travel, each mode of travel has a passing time. It is denoted as drep g [vec-time0, vec-time1,…, vec-time k ]

[0062] In addition, embedding vectors emb-vec are used for road segments and super-road segments. seg and emb-vec g Respectively. The corresponding representation vector emb-vec seg and emb-vec g The purpose of information sharing can be achieved through learning on the neural network. Because there may be a situation where a road section appears in different super sections, in this case, the information of the road section needs to be shared. At the same time, the same super section may appear in different paths, in this case, the information of the super section also needs to be shared.

[0063] Node feature vector vec i is: [vec-time t0 ,vec-speed t0 ,vec-time-his t0 , vec-speed-his t0 , vec-attr, emb-vec seg ].

[0064] Graph feature vector vec g For: [vec-time0, vec-time1,…, vec-time k , emb-vec s-seg ].

[0065] The specific graph feature calculation process is as follows Figure 4 shown.

[0066] A graph neural network is constructed, and graph neural network training is performed based on the node feature vector and the graph feature vector to obtain a predicted passing time list, and a cache table is updated based on the predicted passing time list.

[0067] A graph structure encoding module, a graph network updating module and a graph network decoding prediction module are constructed for the graph neural network; the graph structure encoding module is used to encode the node feature vector, the edge feature vector and the graph feature vector of the super section to generate the node feature potential representation vector, the edge feature potential representation vector and the graph feature potential representation vector; the graph network updating module is used to update the node feature potential representation vector, the edge feature potential representation vector and the graph feature potential representation vector respectively; the graph network decoding prediction module is used to predict a predicted passing time list based on the node feature potential representation vector, the edge feature potential representation vector and the graph feature potential representation vector.

[0068] In this embodiment, the graph neural network is composed of three graph network modules, such as Figure 3 First, the graph is encoded by the graph structure module to encode the super segment S The original predicted dynamic and static information (node ​​v feature vector v i , the eigenvector e of edge e j , the super-segment global feature vector u) is encoded to generate the node v feature potential representation vector hv i (i∈V), edge e feature potential representation vector he j (j∈E) and the global latent representation vector h of the hyper-segment u Then the network update module is updated to hv i , he j The graph network update module designs three different update functions for nodes, edges and super links, respectively, denoted as Φ v , Φ e , Φ u . Three update functions Φ v , Φ e , Φ u Both use the structure of MultiLayer Perceptrons.

[0069] For any edge j(j∈E), the update function Φ e The independent variable is the eigenvector he corresponding to the edge j , the source node feature vector hv of edge j sj, the target node feature vector hv of edge j tj And the feature vector u of the super segment. The representation vector of edge j is updated and recorded as he′ j That is: he′ j =Φ e (he j , hv sj , hv tj , u). When all edges are updated, the set of edges formed is recorded as E', and the set of edges connected to node i is recorded as E' i (Because a node can be connected to multiple nodes, there are multiple edges. Figure 1 The nodes C and BDHI in the figure are connected to each other, and there are 4 edges. It is necessary to aggregate the edges of node i to obtain the representation vector of the edges of node i. And perform super-segment aggregation on the updated edges to obtain the representation vector of the aggregated edges

[0070] For any node i(i∈S), the update function Φ v The independent variable is the feature potential table hv of node i i , and u. The fixed point set obtained after updating all nodes is recorded as V′. V′ is needed to update the super segment aggregation to obtain the representation vector of the vertex aggregation

[0071] After completing the above-mentioned edge and node updates and edge aggregation results, the representation vector u of the super segment is updated to obtain u′.

[0072] In the training phase, for a given start time, each super segment s is learned i The passing time within different fixed time ranges in the future, and record them in the cache table Tab si .

[0073] The training of graph neural network is to train the nodes vec i The representation vector vec of the graph ss The graph representation vector contains the list of times to be predicted.

[0074] The training of graph neural network adopts the weighted method of multiple loss functions, because the passing time of the section and the cumulative passing time on the super section can both reflect the passing time on the super section. The main goal of training is to predict the passing time of the super section, but the passing time at the section level (including the passing time of each section and the cumulative passing time of each section) can assist the rapid convergence of the model.

[0075] The true value of graph neural network training is the time it takes to pass the road section and time to pass the super section For a given time L = l ss + l s + l sc , l ss refers to the loss function on the super section, l s is the loss function on the road segment, l sc is the cumulative loss function on all road segments. Free passing time of super section i The free time of segment k on super segment i. Free time refers to the time when there are few or no vehicles on the road. The smoothed mean absolute error is used. Compared with the squared error loss, the smoothed mean absolute error loss is less sensitive to outliers in the data. The smoothed mean absolute error is adjusted by adjusting the parameter The size of makes the loss function tend to Mean Absolute Error or Mean Square Error.

[0076]

[0077] n Indicates the total number of super sections, m （i） The number of segments in super segment i; represents the passing time of super section i, represents the predicted passing time of super section i; is the cumulative passing time of all sections on the super section i, is the cumulative predicted passing time of all sections on the super section i; The passing time of section j in super section i, The predicted passing time of segment j in super segment i; Free passing time of super section i, free passing time refers to the passing time when the number of vehicles on the road is less than the preset value or there are no vehicles; δ To adjust the parameters.

[0078] Step S3, querying the passing time of each super-segment in the super-segment sequence within different future fixed time ranges based on the cache table obtained by pre-training to determine the passing time of the super-segment sequence; determining the estimated arrival time ETA based on the passing time of the super-segment sequence.

[0079] In the ETA query phase, for a given query path S t and departure time t, and S t Projected onto the hyper-segment of the high-precision map, calculate S t The corresponding super segment sequence rseq st Use [s1, s2, …, s n ] represents. [s1,s2,…, s n ]sequences are queried in chronological order. i The departure time on is recorded as st i , in s i The time of passing through is t i , then st i =t+t1+t 2+…+ +t i-1 .t i Is through the Tab si Find the departure time in i The ETA calculation process is as follows: Figure 6 shown.

[0080] In the application process Figure 5 As shown: After receiving the navigation request, the map engine calls the map engine path planning module to calculate several optional paths. After the user selects a path through manual interaction, the map engine starts the ETA time calculation module. Finally, the ETA calculation result is returned to the map engine module for display.

[0081] The embodiment of the present invention further provides a system for calculating an estimated arrival time based on a graph network, based on the method for calculating an estimated arrival time based on a graph network in the above embodiments, including:

[0082] The super-segment division module divides the high-precision map into several fixed-size areas and constructs a super-segment based on all the road segments in each area;

[0083] A traffic flow fusion module obtains a given query path, projects the given query path to a super-segment of a high-precision map, and determines a super-segment sequence corresponding to the given query path;

[0084] The graph neural network prediction module queries the passing time of each super-segment in the super-segment sequence within different fixed time ranges in the future based on the cache table obtained through pre-training to determine the passing time of the super-segment sequence; and determines the estimated arrival time ETA based on the passing time of the super-segment sequence.

[0085] Based on the same concept, the embodiment of the present invention also provides a schematic diagram of an entity structure, such as Figure 7 As shown, the server may include: a processor 810, a communication interface 820, a memory 830 and a communication bus 840, wherein the processor 810, the communication interface 820 and the memory 830 communicate with each other through the communication bus 840. The processor 810 may call the logic instructions in the memory 830 to execute the steps of the estimated arrival time calculation method based on the graph network as described in the above embodiments. For example, it includes:

[0086] Step S1, dividing the high-precision map into a number of fixed-size areas, and constructing a super section based on all road sections in each area;

[0087] Step S2: obtaining a given query path, projecting the given query path onto a super-segment of a high-precision map, and determining a super-segment sequence corresponding to the given query path;

[0088] Step S3, querying the passing time of each super-segment in the super-segment sequence within different future fixed time ranges based on the cache table obtained by pre-training to determine the passing time of the super-segment sequence; determining the estimated arrival time ETA based on the passing time of the super-segment sequence.

[0089] Based on the same concept, an embodiment of the present invention further provides a non-transitory computer-readable storage medium, which stores a computer program, which includes at least one code segment, and the at least one code segment can be executed by a main control device to control the main control device to implement the steps of the estimated arrival time calculation method based on a graph network as described in the above embodiments. For example, it includes:

[0090] Step S1, dividing the high-precision map into a number of fixed-size areas, and constructing a super section based on all road sections in each area;

[0091] Step S2: obtaining a given query path, projecting the given query path onto a super-segment of a high-precision map, and determining a super-segment sequence corresponding to the given query path;

[0092] Step S3, querying the passing time of each super-segment in the super-segment sequence within different future fixed time ranges based on the cache table obtained by pre-training to determine the passing time of the super-segment sequence; determining the estimated arrival time ETA based on the passing time of the super-segment sequence.

[0093] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and for parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0094] It should be understood by those skilled in the art that embodiments of the present invention may be provided as methods, devices, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0095] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (apparatus), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0096] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0097] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0098] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0099] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for calculating estimated arrival time based on graph network, characterized in that: include: Step S1, dividing the high-precision map into a number of fixed-size areas, and constructing a super section based on all road sections in each area; Step S2: obtaining a given query path, projecting the given query path to a super-segment of a high-precision map, and determining a super-segment sequence corresponding to the given query path; Step S3, querying the passing time of each super-section in the super-section sequence within different future fixed time ranges based on the cache table obtained by pre-training, so as to determine the passing time of the super-section sequence; Determine the estimated time of arrival (ETA) based on the passing time of the super-section sequence; The step S1 specifically includes: The high-precision map is divided into a number of fixed-size areas, and all road segments contained in each area form a super road segment; wherein the road segment is the smallest road expression unit in the high-precision map; A graph structure is defined for each super segment, wherein the segment is a vertex of the graph structure; if any two segments are adjacent and satisfy upstream connectivity or downstream connectivity, the two adjacent segments constitute an edge; In the step S3, before querying the passing time of each super-segment in the super-segment sequence within different future fixed time ranges based on the cache table obtained by pre-training, the step further includes: Capturing the real-time traffic flow and historical traffic flow in the area where the graph structure is located to determine the historical trajectory sequence and real-time trajectory sequence of each road section, wherein the historical trajectory sequence includes the real-time average passing time sequence and the real-time average passing speed sequence of the road section at any time within the cycle day, and the real-time trajectory sequence is the historical average passing time sequence and the historical average passing speed sequence of the road section at any time within the cycle day; Projecting the historical trajectory sequence and the real-time trajectory sequence onto the corresponding super-segment to obtain the node feature vector, edge feature vector, and graph feature vector of the super-segment; the node feature vector includes the real-time average passing time series, the real-time average passing speed series, the historical average passing time series, the historical average passing speed series, and segment attribute information on the segment; the graph feature vector is the real-time passing time vector of the super-segment; A graph neural network is constructed, and graph neural network training is performed based on the node feature vector and the graph feature vector to obtain a predicted passing time list, and a cache table is updated based on the predicted passing time list.

2. The estimated arrival time calculation method based on graph network according to claim 1 is characterized in that: Capture the real-time and historical traffic flows in the area where the graph structure is located to determine the historical and real-time trajectory sequences of each road segment, including: All the trajectory segments on the road section passed at t0 are intercepted, and the time of the last trajectory point of each intercepted trajectory segment minus the time of the first point of the trajectory segment is used to obtain a passing time t of the road section; determine the mean passing time d0 of all trajectory segments; calculate t0-i*Δt r The travel time is denoted as d i ,Δt r The preset fixed time window is i=1,2,3,…,m; determine the real-time average passing time series vec-time at time t0 on the road section t0 For m , ct m-1 ,…, ct2,ct1, ct0]; calculate the real-time average passing speed sequence vec-speed at time t0 on the road section t0 [v m , v m-1 ,…, v2,v1, v0]; Aggregate all trajectories at the same time in the same cycle day to determine the p Δt before the specified time h The historical average of the moment through the time series, q Δt h The historical average passing speed sequence at time Δt h , Δt h The preset time window is a fixed time window.

3. The estimated arrival time calculation method based on graph network according to claim 1 is characterized in that: Constructing a graph neural network, including: Construct the graph structure encoding module, graph network update module and graph network decoding prediction module of the graph neural network; The graph structure encoding module is used to encode the node feature vector, edge feature vector and graph feature vector of the super section to generate a node feature potential representation vector, an edge feature potential representation vector and a graph feature potential representation vector; The graph network update module is used to update the node feature potential representation vector, the edge feature potential representation vector and the graph feature potential representation vector respectively; The graph network decoding prediction module is used to predict a predicted passing time list based on the node feature potential representation vector, the edge feature potential representation vector and the graph feature potential representation vector.

4. The estimated arrival time calculation method based on graph network according to claim 3 is characterized in that: The graph network decoding prediction module includes an update function for updating the node feature potential representation vector Φ v , an update function used to update the edge feature potential representation vector Φ e , an update function used to update the latent representation vector of graph features Φ u ; For any edge, the update function Φ e The independent variables are the corresponding edge feature vector, the edge source node feature vector, the edge target node feature vector, and the graph feature vector u; the edge feature vector set is recorded as E′, and the edge feature vector set connected to node i is recorded as Aggregate the edges of node i to obtain the representation vector of the edges of node i, and perform super-segment aggregation on the updated edges to obtain the representation vector of the aggregated edges.

5. The estimated arrival time calculation method based on graph network according to claim 3 is characterized in that: Also includes: Constructing the loss function of the graph neural network, the loss function includes the loss function on the super segment l ss , the loss function on the road segment l s , the cumulative loss function on all road sections l sc ; In the above formula, n Indicates the total number of super sections, m （i） The number of segments in super segment i; represents the passing time of super section i, represents the predicted passing time of super section i; is the cumulative passing time of all sections on the super section i, is the cumulative predicted passing time of all sections on the super section i; The passing time of section j in super section i, The predicted passing time of segment j in super segment i; Free passing time of super section i, free passing time refers to the passing time when the number of vehicles on the road is less than the preset value or there are no vehicles; δ To adjust the parameters.

6. A system for calculating estimated arrival time based on graph network, characterized in that: include: The super-segment division module divides the high-precision map into several fixed-size areas and constructs a super-segment based on all the road segments in each area; A traffic flow fusion module obtains a given query path, projects the given query path to a super-segment of a high-precision map, and determines a super-segment sequence corresponding to the given query path; A graph neural network prediction module queries the passing time of each super-segment in the super-segment sequence within different future fixed time ranges based on a cache table obtained through pre-training to determine the passing time of the super-segment sequence; and determines an estimated arrival time ETA based on the passing time of the super-segment sequence; The super-segment division module is also used to: divide the high-precision map into a number of fixed-size areas, and all the road segments contained in each area form a super-segment; wherein the road segment is the smallest road expression unit in the high-precision map; define a graph structure for each super-segment, wherein the road segment is the vertex of the graph structure; if any two road segments are adjacent and satisfy the upstream connectivity or the downstream connectivity, the two adjacent road segments constitute an edge; The graph neural network prediction module is also used to: capture the real-time traffic flow and historical traffic flow in the area where the graph structure is located to determine the historical trajectory sequence and real-time trajectory sequence of each road section, the historical trajectory sequence includes the real-time average passing time series and the real-time average passing speed series of the road section at any time within the cycle day, and the real-time trajectory sequence is the historical average passing time series and the historical average passing speed series of the road section at any time within the cycle day; project the historical trajectory sequence and the real-time trajectory sequence onto the corresponding super-road section to obtain the node feature vector, edge feature vector, and graph feature vector of the super-road section; the node feature vector includes the real-time average passing time series, the real-time average passing speed series, the historical average passing time series, the historical average passing speed series and the road section attribute information on the road section; the graph feature vector is the real-time passing time vector of the super-road section; construct a graph neural network, perform graph neural network training based on the node feature vector and the graph feature vector, obtain a predicted passing time list, and update the cache table based on the predicted passing time list.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the method for calculating the estimated arrival time based on a graph network as described in any one of claims 1 to 5 are implemented.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for calculating the estimated arrival time based on a graph network as described in any one of claims 1 to 5 are implemented.

Images