CN115424435A

CN115424435A - Cross-link road identification method and cross-link road identification method

Info

Publication number: CN115424435A
Application number: CN202210956434.3A
Authority: CN
Inventors: 郭翊麟; 杜舒; 唐俊杰
Original assignee: Alibaba China Co Ltd
Current assignee: Alibaba China Co Ltd
Priority date: 2022-08-10
Filing date: 2022-08-10
Publication date: 2022-12-02
Anticipated expiration: 2042-08-10
Also published as: CN115424435B

Abstract

The specification provides a training method for a cross-link road recognition network and a method for recognizing a cross-link road, wherein a road data set is obtained and comprises a training data subset and a test data subset; training a cross-link road recognition network through the training data subset; the input of the cross-link road identification network is a road relation graph, the road relation graph comprises a plurality of nodes and edges connecting the nodes, each node is used for representing a triple, and each edge is used for representing a road relation between two triples; predicting the test data subset according to the cross-link road identification network, and determining error data in the test data subset, wherein the identification result of the cross-link road identification network does not accord with the label; and acquiring the update information of the road relation graph aiming at the error data, and retraining the cross-link road recognition network based on the update information.

Description

Cross-link road identification method and cross-link road identification method

Technical Field

One or more embodiments of the present disclosure relate to the field of computer application technologies, and in particular, to a cross-link road identification method and a cross-link road identification method.

Background

In the process of navigating a vehicle for a navigation application, the following problems may exist: as shown in fig. 1, the navigation route is a _ link (link represents a section of road) to b _ link and then to c _ link, but the b _ link is short, so that the user cannot immediately respond when the vehicle walks to the b _ link and hears the guidance voice of "turning right to c _ link", and the vehicle of the user may go straight from the b _ link to the c' -link, so that the user misses the correct navigation route, which causes inconvenience to the user. The triplet that has the above problem (i.e., one set of three consecutive links, such as a _ link, b _ link, and c _ link) is called a link-crossing link. And a method for accurately identifying the existence of the cross-link road is lacked in the related art.

Disclosure of Invention

In view of the above, one or more embodiments of the present disclosure provide a cross-link road identification method and a cross-link road identification method.

According to a first aspect of one or more embodiments of the present specification, there is provided a training method across a link road recognition network, the method comprising:

acquiring a road data set, wherein the road data set comprises road information of a plurality of triples and a label corresponding to each triplet, and the label is used for indicating whether the triplet is link-crossing road data or not; the road data set comprises a training data subset and a testing data subset; the triplet comprises three links connected together;

training a cross-link road recognition network through the training data subset; the input of the cross-link road identification network is a road relation graph, and the output is an identification result of the cross-link road; the road relation graph comprises a plurality of nodes and edges connecting the nodes, each node is used for representing a triple, and each edge is used for representing the road relation between two triples; the road relation is used for representing whether the two triples have the same road information or representing the position relation of links in the two triples;

predicting the test data subset according to the cross-link road identification network, and determining error data in the test data subset, wherein the identification result of the cross-link road identification network does not accord with the label;

acquiring update information of the road relation graph aiming at the error data, wherein the update information comprises: node data corresponding to the error data and edge data corresponding to the error data;

retraining the cross-link road recognition network based on the updated information.

According to a second aspect of one or more embodiments herein, there is provided a method of identifying a link-crossing road, comprising:

acquiring data to be detected, and constructing a road relation graph to be detected according to the data to be detected;

identifying and processing the road relation graph to be detected through a link-crossing road identification network to obtain an identification result of whether a node in the road relation graph to be detected is link-crossing road data or not; the cross-link road recognition network is obtained by training through the training method of the cross-link road recognition network.

According to a third aspect of embodiments herein, there is provided a training apparatus across a link road recognition network, comprising:

the road data set acquisition unit is used for acquiring a road data set, wherein the road data set comprises road information of a plurality of triples and a label corresponding to each triplet, and the label is used for indicating whether the triples are link-crossing road data or not; the road data set comprises a training data subset and a testing data subset; the triplet comprises three links connected together;

the cross-link road recognition network training unit is used for training a cross-link road recognition network through the training data subset; the input of the cross-link road identification network is a road relation graph, and the output is an identification result of the cross-link road; the road relation graph comprises a plurality of nodes and edges connecting the nodes, each node is used for representing a triple, and each edge is used for representing the road relation between two triples; the road relation is used for representing whether the two triples have the same road information or representing the position relation of links in the two triples;

the test data subset prediction unit is used for predicting the test data subset according to the cross-link road identification network and determining error data in the test data subset, wherein the identification result of the cross-link road identification network does not accord with the label;

an update information acquisition unit configured to acquire update information of the road map for the error data, the update information including: node data corresponding to the error data and edge data corresponding to the error data;

and the cross-link road recognition network retraining unit is used for retraining the cross-link road recognition network based on the updated information.

According to a fourth aspect of embodiments herein, there is provided an apparatus for identifying a link-crossing road, comprising:

the to-be-detected road relation graph building unit is used for obtaining to-be-detected data and building a to-be-detected road relation graph according to the to-be-detected data;

the identification result acquisition unit is used for identifying the road relation graph to be detected through a link-crossing road identification network to obtain an identification result of whether a node in the road relation graph to be detected is link-crossing road data; the cross-link road recognition network is obtained by training through the training method of the cross-link road recognition network.

According to a fifth aspect of embodiments herein, there is provided a computer readable storage medium having stored thereon computer instructions, which when executed by a processor, implement the aforementioned training method for a cross-link road recognition network or the cross-link road recognition method.

According to a sixth aspect of embodiments herein, there is provided a computer apparatus comprising:

a processor;

a memory for storing processor-executable instructions;

the processor executes the executable instructions to realize the training method of the cross-link road recognition network or the cross-link road recognition method.

The specification provides a training method of a cross-link road identification network and a method for identifying a cross-link road, wherein a road data set is obtained, the road data set comprises road information of a plurality of triples and a label corresponding to each triplet, and the label is used for indicating whether the triples are cross-link road data or not; the road data set comprises a training data subset and a test data subset; the triplet comprises three links connected together; training a cross-link road recognition network through the training data subset; the input of the cross-link road identification network is a road relation graph, and the output of the cross-link road identification network is an identification result of a cross-link road; the road relation graph comprises a plurality of nodes and edges connecting the nodes, each node is used for representing a triple, and each edge is used for representing the road relation between two triples; the road relation is used for representing whether the two triples have the same road information or representing the position relation of links in the two triples; predicting the test data subset according to the cross-link road identification network, and determining error data in the test data subset, wherein the identification result of the cross-link road identification network does not accord with the label; acquiring update information of the road relation graph aiming at the error data, wherein the update information comprises: node data corresponding to the error data and edge data corresponding to the error data; retraining the cross-link road recognition network based on the updated information.

By processing the wrong data (the data which is wrongly classified by the cross-link road recognition network), correct road information of the data which is wrongly classified in the wrong data due to inaccurate road information can be determined, and the cross-link road recognition network is updated through the correct road information, so that the network can learn more accurate knowledge iteratively through the self-adaptive feedback mechanism, the network training can eliminate the influence of inaccurate road data set, the cross-link road recognition network is more accurate, the cross-link recognition result is more accurate, the navigation of the cross-link road data can be corrected, the yaw loss of a user is reduced, and the use experience of the user is improved.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the specification.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present specification and together with the description, serve to explain the principles of the specification.

FIG. 1 is a schematic diagram of a cross link problem shown in this specification.

FIG. 2 is a flow chart illustrating a method of training across a link road recognition network according to an exemplary embodiment of the present description.

FIG. 3 is a road map shown in accordance with an exemplary embodiment of the present description.

FIG. 4 is a flow chart illustrating a method of identifying a cross-link road according to an exemplary embodiment of the present description.

FIG. 5 is a block diagram illustrating a method of training across a link road recognition network in accordance with one embodiment of the present disclosure.

FIG. 6 is a block diagram of a training apparatus across a link road recognition network, shown in accordance with an exemplary embodiment of the present specification.

FIG. 7 is a block diagram illustrating an apparatus for identifying a link-crossing road according to an exemplary embodiment.

FIG. 8 is a hardware block diagram of a computer device in which a training apparatus for a cross-link road recognition network or an apparatus for recognizing a cross-link road is shown according to an exemplary embodiment of the present specification.

Detailed Description

Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, like numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with one or more embodiments of the specification. Rather, they are merely examples of apparatus and methods consistent with certain aspects of one or more embodiments of the specification, as detailed in the claims that follow.

It should be noted that: in other embodiments, the steps of the corresponding methods are not necessarily performed in the order shown and described herein. In some other embodiments, the method may include more or fewer steps than those described herein. Moreover, a single step described in this specification may be broken down into multiple steps for description in other embodiments; multiple steps described in this specification may be combined into a single step in other embodiments.

Navigation applications typically have vehicle navigation functionality for assisting a user in driving a vehicle by voice guidance information or other forms of guidance action (such as on-screen displayed logos and the like) to drive a shorter/faster/more convenient road to a destination.

In map application, a road network is divided into multiple sections according to a certain rule, each section of road can be regarded as a link, a navigation route is provided for a user according to a corresponding link according to a starting place and a destination of the user during navigation planning, the user is reminded to walk according to the navigation route in the driving process, and corresponding loss caused by the fact that the user walks by mistake is avoided.

In some cases, the navigation route includes cross-link road data, which may cause the user to drift. Specifically, as shown in fig. 1, when there is a triplet including three consecutive links, a _ link, b _ link, and c _ link, in the navigation route, the triplet is the road data across the links, which may cause the following problems: when the vehicle passes through the a _ link, the user is reminded of driving to the right front, and when the vehicle goes from the b _ link to the c _ link, the user is reminded of turning to the right, and because the length of the b _ link is short, when the vehicle goes to the b _ link and the guide action (which can be voice or other forms) of turning to the right is heard by the user, the vehicle speed of the user is too fast, or the user cannot react to the guide action immediately due to chatting and the like, the vehicle driven by the user may move from the b _ link to the c' -link, so that the user deviates from a navigation route (namely, is inconvenient to yaw), and the user is caused.

Therefore, it is necessary to identify the cross-link road data in the map to give the user a better use experience.

Related art there are the following methods of identifying link-crossing road data: in the related art, the cross-link road data may be screened out through a certain strategy, for example, when the length of the triple intermediate link is smaller than a certain length threshold (for example, 50 meters), and the minimum included angle between the triple intermediate link and any of the remaining links is larger than a certain value, the triple is determined to be the cross-link road data, but the screened cross-link road data may be inaccurate, for example, some screened data is not the cross-link road data, and some cross-link road data may not be identified.

It can be seen that a method for accurately identifying the cross-link road data is lacked in the related art.

In order to solve the above problem, it is considered that the road data corresponding to the triplets is unstructured data, and therefore the road data across links can be identified by the neural network. Further, the result of identifying the cross-link road data only through the graph neural network is not accurate, and the reason for this is found that, in some cases, the actual condition of the road is updated, but the stored data used for training is not updated (for example, after a certain road is repaired, the width of the road is changed, or the road composition of the road is changed, but the stored data used for training is not updated with this information), or the originally collected information of the road is not accurate, which causes that the trained graph neural network is trained based on inaccurate data, and the accuracy of the trained graph neural network is affected.

If each training data is corrected manually, much time is wasted. In order to solve the above problem, it is considered that the training data can be updated by an adaptive feedback method, so that the data used for training can be closer to the actual situation, and the trained graph neural network is more accurate. After the training of the graph neural network is completed, data which is wrongly classified by the graph neural network (hereinafter referred to as wrong score data, namely data which is wrong in the recognition result of the cross-link road in the cross-link road recognition process) can be found, the wrong score data is analyzed, and update information, namely information which is closer to reality of the wrong score data, is obtained, so that the obtained updated wrong score data is used as training data to update (retrain) the graph neural network, and the graph neural network can learn a correct graph structure.

In other words, the specification provides a method and a device for training a cross-link road identification network, which are used for acquiring a road data set, wherein the road data set comprises road information of a plurality of triples and a label corresponding to each triplet respectively, and the label is used for indicating whether the triplet is cross-link road data or not; the road data set comprises a training data subset and a testing data subset; the triplet comprises three links connected together; training a cross-link road recognition network through the training data subset; the input of the cross-link road identification network is a road relation graph, and the output of the cross-link road identification network is an identification result of a cross-link road; the road relation graph comprises a plurality of nodes and edges connecting the nodes, each node is used for representing a triple, and each edge is used for representing the road relation between two triples; the road relation is used for representing whether the two triples have the same road information or representing the position relation of links in the two triples; predicting the test data subset according to the cross-link road identification network, and determining error data in the test data subset, wherein the identification result of the cross-link road identification network is inconsistent with the label; acquiring update information of the road relation graph aiming at the error data, wherein the update information comprises: node data corresponding to the error data and edge data corresponding to the error data; retraining the cross-link road recognition network based on the updated information.

By processing the wrong score data (the data which is wrongly classified by the cross-link road recognition network), the correct road information of the data which is wrongly classified due to the inaccurate road information in the wrong score data can be determined, and the cross-link road recognition network is updated through the correct road information, so that the network can learn more accurate knowledge iteratively through the self-adaptive feedback mechanism, the influence of inaccurate road data set can be eliminated through network training, the cross-link road recognition network is more accurate, the cross-link recognition result is more accurate, the navigation of the cross-link road data can be corrected, the yaw loss of a user is reduced, and the use experience of the user is improved.

Next, a method for training a cross-link road recognition network shown in this specification will be described in detail.

As shown in fig. 2, fig. 2 is a flowchart of a training method for a cross-link road recognition network shown in this specification, including the following steps:

step 201, a road data set is obtained.

The road data set comprises road information of a plurality of triples and a label corresponding to each triplet, wherein the label is used for indicating whether the triplet is link-crossing road data or not; the road data set comprises a training data subset and a testing data subset; the triplet comprises three links connected.

And step 203, training a cross-link road recognition network through the training data subset.

The input of the cross-link road identification network is a road relation graph, and the output is an identification result of the cross-link road; the road relation graph comprises a plurality of nodes and edges connecting the nodes, each node is used for representing a triple, and each edge is used for representing the road relation between two triples; the road relation is used for representing whether the two triples have the same road information or representing the position relation of links in the two triples.

Next, step 201 and step 203 will be explained.

First, the nouns in step 201 and step 203 will be explained.

Link is also a segment of road, and in map navigation application, a real road network is often divided according to attribute information of the road (such as length, name, number of lanes, road grade, road composition, speed limit, direction, city to which the road belongs, and the like) to obtain different Link segments. For a specific dividing method, reference may be made to methods in the related art, and details are not described herein.

The triple is a road data combination including three links connected in sequence, and in this specification, the triple is used as an identification unit of the cross-link problem according to the reason of the cross-link problem. It should be noted that, in this specification, the reason why the triple is selected and the middle link of the triple or the last two links of the triple are not selected as the unit for identifying the cross-link problem is that whether the cross-link problem exists cannot be determined only by these two links without other information, and the cross-link problem exists is related to all three links of the triple, and the absence of any information cannot accurately determine whether the cross-link problem exists.

The cross-link road recognition network in the present specification includes at least one graph neural network. For a general graph neural network, an input is a graph including a plurality of nodes and edges (for example, as shown in fig. 3), and in the process of processing the input by the graph neural network, the nodes and edges in the graph are continuously updated in a multi-layer process (the update is generally obtained by convolution multiplication of other nodes and edges around the nodes and weights), and finally a node vector in the output graph is obtained. In addition, sometimes, to achieve the corresponding purpose, a classification sub-network follows the graph neural network, for example, in the context of this specification, the sub-network takes a node vector as an input, and takes a result whether the triplet corresponding to the node is a cross-link road as an output.

For purposes of this specification, the cross-link road recognition network may include a graph neural network (a graph neural network includes a sub-network for updating node vectors and a classification sub-network, the two sub-networks being trained together as a whole during the training process). The cross-link road recognition network may also be a graph neural network including a node vector for updating the road relationship graph, and a classification network (both networks trained separately). Of course, the cross-link road identification network may have other structures, and the specific structure of the cross-link road identification network is not limited in this specification.

For training of the graph neural network, the graph neural network is trained, that is, a graph is constructed through training data, and a process of obtaining weights of the graph neural network is trained through the constructed graph. It should be noted that, in this specification, a corresponding road relationship graph may be constructed according to road data, and supervised learning may be performed according to the labels of the training data subset, so that the network may learn the relationship between graph structures and road interaction information (attention). The specific training method can be referred to in the related art, and the specific training method of the graph neural network is not limited in this specification.

It should be noted that the Graph neural Network may be any common Graph neural Network, for example, a Graph Convolutional Network (GCN), a Graph Attention Network (GAT), or a Graph Auto-encoder Network (GAE), which is only an example of some Graph neural networks, and the specific form of the Graph neural Network is not limited in this specification.

The road relationship graph is a graph constructed by a road data set or a training data subset, and the specific method for constructing the graph may refer to the method in the related art, which is not described in detail herein. In order to solve the problems existing in the navigation scene, in the constructed road relationship diagram, a node represents a triplet, a feature vector of the node is a vector used for representing road information of the triplet, an edge represents a road relationship between the triplet and the triplet, and a feature vector of the edge is used for representing a road relationship between two triplets. Specific meanings of the road information and the road relationship will be described in detail below, and will not be described in detail herein.

In addition, the reason why the above-described problem exists in the present specification is that the road data in the road data set does not match the actual situation, in other words, the road data set in the present specification has a problem of inaccuracy, or the road data set in the present specification cannot completely reflect the actual road situation.

It should be further noted that the training data subset is only used as training data in this iteration, the test data is also only used as test data in this iteration, after step 209 is completed, the data in the test data subset may also be used as training data to re-train the neural network of the graph, and the data in the training data subset may also be used as test data again. The test data subsets and training data subsets in this specification do not mean that the data is used for both test data or always training data. Further, the division of the training data and the test data may be random.

After the nouns in step 201 and step 203 are explained, a specific implementation of the above steps will be explained.

For the specific acquisition of the road data, the acquisition of the road data in step 201 may be to randomly select a part of the data from a database storing the road data as the road data, or to use the whole amount of data in the database as the road data set.

The specific acquisition of the road data may be, in addition to the above method: since for the network, if the data is data which is greatly different from the cross-link data, for example, the directions of three links in the triad are completely the same, the network can easily distinguish the triad from the triad corresponding to the cross-link road. In other words, it is more desirable in this specification for the network to have the ability to screen triples across link roads and triples that are very similar across link roads. Therefore, in the training process, in order to enable more training data to be data enabling the model to learn more graph structures, further reduce training cost and avoid noise influence, and achieve the optimal training effect with the least data, the acquired road data may be data including a triple of a link-crossing road and a triple of a link-crossing road which is very similar to the link-crossing road.

Based on the above analysis, in step 201, the road data may be first screened according to a strategy for screening a link-crossing road in the related art, and the road data obtained by screening is used as training data.

In other words, step 201 includes: acquiring three links connected with each other; taking the three links as the triples in the road data set when the three links meet a preset road condition, wherein the preset road condition comprises: the road length of a middle link of the three links is smaller than a length threshold (for example, may be 50 m), and an included angle between the middle link and any other link of the three links is greater than a preset included angle threshold (for example, may be 30 degrees).

It should be noted that three links connected are also referred to as three links connected in sequence, and the included angle refers to the minimum included angle between two links. In addition, in addition to the screening according to the above conditions, the road data meeting the requirements may be further screened according to the following manner when the above conditions are satisfied, for example, the road data may be screened according to information about road attributes such as road type, road grade, road composition, and the like.

Among them, the road types may include: one-way lanes, two-way lanes, etc.; the road grade may include: national road, provincial road, city express, major/minor road, etc.; the road configuration may include: upper and lower line separation, roundabout, service area, main road, auxiliary road, etc. The road data needs to be screened by the information about the road attributes because some attributes do not have the cross-link problem, for example, the cross-link problem does not exist in the small roads and the service areas.

In addition, after the road data set is obtained, if the road data set is obtained by means of screening, in order to enrich the road data so that the link-crossing road recognition network can learn richer knowledge during training, the triplets of links before and after the road data set (i.e., the triplets consisting of links connected to any link in the triplets) can also be obtained after the road data set is obtained. Therefore, in the process of constructing the road relation graph, the triplets and the triplets connected with the triplets exist a connection relation, and the information of the connected triplets can be considered in the updating of the screened triplets, so that the cross-link road identification network can learn richer graph structures in the process of updating the road relation graph corresponding to the training data.

In other words, after step 201, the method further comprises: acquiring an adjacent triple with the position adjacent to the position of the link according to the position of the link in the triple in the road data set; adding the adjacent triplet to the road data set.

Finally, it should be noted that, in order to enable the graph neural network to learn a richer graph structure, the road relationship graph can be constructed not only by using the training data subset but also by using the whole amount of road data, so that the relationship in the constructed graph is more comprehensive and richer, and the network learning result can be more accurate.

In other words, step 203 comprises: and constructing a road relation graph according to the training data subset and the testing data subset, and training to obtain the cross-link road recognition network according to the labels of the triples in the training data subset of the road relation graph. Of course, it is also feasible to construct a graph neural network from only a subset of the training data.

The cross-link road recognition network is obtained through training by acquiring road data comprising a training data subset and a testing data subset and performing supervised learning through the training data subset until the network converges. But the identification effect of the network is not good, and the identification effect of the network needs to be further improved through the steps 205 to 209.

Step 205, predicting the test data subset according to the cross-link road identification network, and determining error data in the test data subset, wherein the identification result of the cross-link road identification network does not accord with the label.

And step 207, acquiring the update information of the road relation map according to the error data.

Wherein the update information includes: node data corresponding to the erroneous data, and edge data corresponding to the erroneous data.

Step 205 and step 207 are feedback processes of closed loop feedback (adaptive feedback). Next, step 205 and step 207 will be collectively described.

First, the nouns of step 205 and step 207 will be explained.

The wrong data is the data which is identified wrongly by the cross-link road identification network in the test data subset, and the test data also has labels, so the wrong data can be found out by comparing the labels of the test data in the test data subset with the identification result of the cross-link road identification network.

Because the error data is classified incorrectly, the cross-link road identification network may not learn the correct relationship between the nodes, or may learn an error with respect to the relationship between the edges and the nodes, so that in order to achieve a better identification effect for the cross-link road identification network, the updated data in step 205 includes the node data and the edge data. That is, the updated information is used to update the feature vectors of the nodes and edges in the road relationship graph, and in some cases, the structure of the graph may also change due to the fact that the feature vectors of the edges become larger.

And predicting the test data subset according to the cross-link road identification network, namely updating the node characteristic vectors of the road relation graph corresponding to the test data through the cross-link road identification network, and obtaining an identification result of whether each node is a cross-link road, wherein wrong data is data with inconsistent identification results and test data labels.

After the terms related to step 205 and step 207 are explained, the specific implementation of step 205 and step 207 will be explained below.

For step 205, the data used for prediction may be all data in the test data subset, and when there is more data in the test data subset, a better test effect may be achieved only with a part of the test data, so to improve the overall efficiency, the data used for testing may also be a part of the data in the test data subset.

For the latter, step 205 comprises: selecting predictive data from the subset of test data; and predicting the prediction data through the cross-link road identification network, and determining error data in the prediction data, wherein the recognition result of the cross-link road identification network is inconsistent with the label.

For the specific method of selecting the prediction data described above, it may be that a part of the data is randomly selected from the test data subset as the prediction data.

In addition, it is considered that if the prediction data are randomly screened, the prediction result may be biased, for example, the network may only have a good recognition effect on part of the road data, but may not have a good recognition effect on other road data. For example, when the road relationship includes multiple relationships (the condition that the road relationship includes multiple relationships will be described in detail below, and will not be described herein for the moment), the network may only have a better prediction effect on data with a part of the road relationship, and therefore, in order to accurately determine the prediction effect of the network, it should be ensured that the data is unbiased in the aspect of the road relationship, that is, screening prediction data should screen a part of data from each relationship, so that it can be determined whether the network has a better prediction effect on all the road relationships.

In other words, in the case where the edges of the road relationship graph are used to characterize a plurality of road relationships between two triples, the selecting the predicted data from the test data subset includes: and aiming at the test data subset, extracting partial road data from the road data corresponding to each road relationship, and forming the extracted road data into prediction data.

In addition, in order to determine that the network has a good prediction effect on all traffic flow triples, it should be ensured that the prediction data is unbiased in traffic flow hierarchy, that is, each traffic flow has corresponding prediction data.

In other words, the selecting of the prediction data from the test data subset includes: counting the traffic flow of the links included in each triple aiming at the test data subset, and determining the traffic flow range to which the average value of the traffic flow of the three links included in each triple belongs; and extracting part of road data from the road data corresponding to each traffic flow range, and forming the extracted road data into prediction data.

The traffic flow can be counted based on daily flow, or can be counted based on monthly flow, seasonal flow or annual flow. Of course, statistics may be performed based on daily traffic, monthly traffic, seasonal traffic, and annual traffic, and then partial data may be extracted from all traffic ranges corresponding to each statistical method. For example, partial data may be extracted from a plurality of traffic flow ranges corresponding to annual traffic flow, and partial data may be extracted from a plurality of traffic flow ranges corresponding to seasonal traffic flow, and so on.

For the execution timing of step 207, updating the error data may be updating as long as there is error data. In addition, in order to improve the network training efficiency, the updating can be performed only when the proportion of the error data to the prediction data (the data used for prediction, all data in the test data subset, or part of data in the test data subset) is greater than a certain value, so that under the condition that the error data is not large, the training of the task network can reach the usable precision, and the network training efficiency can be improved by reducing the updating.

In other words, step 207 includes: determining the error division ratio of error data to prediction data; the prediction data is part or all of the test data subset; and acquiring the update information of the road relation graph under the condition that the error ratio is greater than a preset ratio threshold value.

In the above situation, considering the situation of different traffic flow, if the number of the wrong-separation data is the same, the wrong-separation data will cause more vehicle yaw for the data with larger traffic flow range, and the application aims to reduce the number of the vehicles yaw as much as possible. Therefore, under the condition of judging whether the wrong-distribution ratio is larger than the ratio threshold value or not, the wrong-distribution data can be divided into a plurality of groups of data according to the traffic flow range, the wrong-distribution ratio is lower when the traffic flow is larger, and therefore the network can have a better cross-link road identification effect on the group with the larger traffic flow, and therefore yaw of most vehicles is prevented.

In other words, the method further comprises: and for the test data subset, counting the traffic flow of the links included in each triple, and determining the traffic flow range to which the average value of the traffic flow of the three links included in each triple belongs. Determining the proportion of error data to the predicted data; when the proportion of the error is larger than a preset proportion threshold, acquiring the update information of the road relation graph comprises the following steps: for each traffic flow range, performing: determining the fraction error ratio of the fraction error data corresponding to the traffic flow range to the prediction data; acquiring the update information of the road relation graph under the condition that the proportion of the error is larger than the proportion threshold value corresponding to the traffic flow range; the larger the average value of the traffic flow of the road data corresponding to the traffic flow range is, the lower the proportional threshold value corresponding to the traffic flow range is.

After the execution timing of step 207 is described in detail, a specific method for updating error data will be described next.

The updating of the error data may be performed by manually analyzing the error data to determine correct edge data and node data (the node data is used to represent the road information of the triples corresponding to the nodes, and the edge data is used to represent the road relationships between the triples and the triples).

In addition to identification by a manual analysis method, it is considered that, for road information or road relationships, if there are a plurality of road information or a plurality of road relationships, a change in any one of the road information or road relationships is associated with the remaining road information or road relationships. For example, if a road is changed from a one-way road to a two-way road, the road may be a two-way road due to the fact that the road is widened, and therefore, when part of information is accurate and part of information is inaccurate in road information or road relationships, interaction between different road information or different road relationships may be learned through the relationship network, and whether each road information or road relationship is correct or not may be determined through the relationship network. In other words, step 207 may also be implemented by a relational network.

In other words, step 207 includes: acquiring the updating information of the road relation graph through a relation network; the input of the relation network is a characteristic vector of a node or an edge in the road relation graph, and the output is updated node data or edge data.

The relationship network will be further explained below. The output of the relationship network may be: whether each bit of feature value in the feature vector of the node or the edge is accurate or not. Under the condition that a plurality of road relationships exist in the road relationship graph or a plurality of road information exists in the node, the eigenvalue of the eigenvector represents one road relationship or one road information, and then the output of the relationship network can be understood as: whether each road relationship or each road information is accurate.

For the road relationship, because the edge usually represents whether the relationship exists between the two, in the feature vector of the edge, the value of each feature value is generally not 0, that is, 1, so that not only can whether the feature value is accurate be output through the relationship network, but also what the feature vector of the accurate edge is can be determined through the identification result of the relationship network.

In addition, for the nodes, the road information of the nodes is attribute information of the nodes, each attribute information generally includes several categories, for example, for the road information of the road type, which generally includes a unidirectional road and a bidirectional road, so the feature vectors of the nodes can also be updated through the relationship network. Specifically, still following the above example, the output of the relationship network may be whether the road type of the node is a one-way road, whether the road type of the node is a two-way road, and so on. The visible relation network can well update the feature vectors of the nodes and edges.

It should be noted that, it may also be determined whether the cause of the misclassification of the misclassified data is caused by inaccuracy of the feature vectors of the nodes or the edges through the relationship network, and if so, the feature vectors of the nodes or the edges may not be updated through the relationship network, and may be updated through a manual analysis method, so that a better updating effect may also be achieved.

In addition, in order to enable the relational network to better realize the update of the error data, the relational network can be updated (retrained) through the recognition result of the error data before the feature vectors of the nodes or edges are updated through the relational network every time, so that the relational network can be more suitable for the current scene.

In other words, the obtaining the update information of the road relationship diagram through the relationship network includes: selecting partial data from the error data, and acquiring a relation label of the selected partial data, wherein the relation label is used for representing accurate edge or node data corresponding to the error data; taking the selected partial data as relationship training data, and updating a pre-trained relationship network through the relationship training data; and acquiring updating information corresponding to the residual data through the updated relation network aiming at the residual data except the relation training data in the error data.

The accurate edge or node data may be characterized by the result of whether each feature value is accurate, or may be characterized by whether the feature value is a or B (i.e., whether the road type corresponding to the node is a one-way road or not, and whether the road type corresponding to the node is a two-way road or not).

Step 209, retraining the cross-link road recognition network based on the updated information.

Specifically, the updated error data is added into the training data subset (meanwhile, the part of data is removed from the test data subset), the constructed road relation diagram is updated based on the updated error data, and the cross-link road recognition network is retrained based on the new road relation diagram, so that the network achieves a better cross-link recognition effect.

In addition, it should be noted that, after the steps 201 to 209 are performed, in order to enable the network to have a better recognition effect, the steps 201 to 209 may also be repeatedly performed, so that the network may update the model through a continuous feedback process, and learn richer information to further improve the recognition effect of the model.

After the above method is generally described, a description will be given next of a method that can improve the accuracy of network identification.

For the identification of the link-crossing road through the neural network, the road relationship may be one road relationship or a plurality of road relationships, and similarly, the road information may include one kind of road information or a plurality of kinds of road information.

If various road information and various road relations are adopted, a graph with more complex relations can be constructed, because the neural network actually learns the combination relation among different data, the more the types of the data in the training data are, the richer the contents of edges and nodes are, and the better the recognition effect of the cross-link road recognition network after training is, the more abundant the road relation graph can be generated by the road relation graph with various road relations, and the graph neural network can learn the richer structural information of the graph through the road relation graph, so that the graph neural network obtained through training has a better recognition effect on the cross-link road.

In other words, the road relationship in the present specification may include a plurality of relationships, and of course, the road information may include a plurality of information at the same time. That is, the road relationship graph includes a plurality of road relationship sub-graphs, each road relationship sub-graph includes a plurality of nodes and edges connecting the nodes, and the edges connecting the nodes in different road relationship sub-graphs represent different road relationships. The road relationship diagram having the above relationship is shown in fig. 3.

It should be noted that the boxes in fig. 3 represent a subgraph corresponding to a relationship, a connection line between nodes in different boxes represents that two nodes are the same node, and the connection of the nodes in fig. three is a full-connection relationship, that is, if the node a is connected to the node B and the node B is connected to the node C, the connection line is not drawn between the node C and the node a in the graph, but the node C and the node a are actually connected. The road relation subgraph is the area represented by the frame in fig. 3.

Other relationships in fig. 3, i.e., combinations of two or more road relationships in the above relationships, such as the road types between two triplets are consistent and the road grades are consistent, may be considered as having one other relationship between the two triplets. The other relationship may be preset, or may be obtained by the foregoing obtaining of the update information. For example, if a combination of two or more road relationships is found to be more relevant across link roads by manual analysis, other relationships may be added to the road relationship graph.

It should be noted that fig. 3 is not a diagram actually used by a computer, but a diagram drawn to facilitate a user to read and view training results. Meanwhile, although the same node in fig. 3 is represented by a plurality of nodes in the graph, in the training process of the neural network of the actual graph, the plurality of nodes are the same in feature vector, and the change is also synchronous.

Next, description will be made first of all on the road relationship and the road information, which may be information for characterizing basic attributes of a road, and may include, for example, at least one of a road grade, a road type, a road composition, a city to which the road belongs, and an upstream-downstream relationship with other triples, a road length, a road width, a number of lanes, a road out-degree, a road in-degree, a road direction, a road angle, and the like.

The road grade, road type, road composition have been described above and will not be described herein. The city to which the road belongs is the city in which the position of the road is represented. The road length is the length of each link in the representation triad, the road width is the width of each road in the representation triad, and the road out-degree and the road in-degree represent the number of roads which can pass through at the downstream of one road and the number of roads which can enter the current road at the upstream of one road. The road direction and the road angle are similar, namely, the azimuth information representing the road. It should be noted that if the information is a numerical value, the numerical value may be directly used as a feature value in a feature vector of a node, and if the information is not a numerical value (for example, a category feature), the information needs to be encoded and then processed.

Similar to the road information, the road relationship indicates whether the two triplets have the same specific road information, for example, if all links of the two triplets are bidirectional roads, the two triplets are considered to have the road relationship of the road category. The road relationship may also characterize the positional relationship between the triplets and the triplets. Two triples may be considered to have an upstream-downstream relationship, for example, if they have the same link between them.

The road relationship may include at least one of: whether it is the same road type, whether it is the same road composition, whether it is the same road type, whether it belongs to the same city, or whether it has an upstream-downstream relationship.

It should be noted that the upstream and downstream relationship not only indicates that two triples include the same link, but also includes that two triples are both within the same area, and can go from one triplet to another, which is also called that there is an upstream and downstream relationship between two triples.

It should be noted that the graph neural network learns the relationship between the combination of various information and the output, and the combination of various information and the cross-link road have a certain relationship, so that the graph neural network for identifying the cross-link road can be obtained through the training of the data including various relationships and/or including various road information.

After a method for training a cross-link road recognition network is described, a method for recognizing a cross-link road shown in this specification will be described below.

FIG. 4 is a flowchart of a method of identifying a cross-link road shown in accordance with one exemplary embodiment of the present description, including:

step 401, acquiring data to be detected, and constructing a road relation graph to be detected according to the data to be detected.

The data to be detected can be manually input, and the method for constructing the road relation graph to be detected according to the data to be detected can be similar to the method for constructing the graph in the related art, and is not described herein again. The data to be detected can be all data of the map, and therefore the link-crossing road can be marked in advance. The data to be detected can also be data to be detected, wherein when the navigation route is generated, the combination of three links which are sequentially connected in the navigation route is regarded as a triple, and all the triples are regarded as the data to be detected, so that the cross-link road in the navigation route can be identified rapidly in a targeted manner, and the guidance action corresponding to the cross-link road can be prompted in a targeted manner.

And step 403, identifying the road relation graph to be detected through the cross-link road identification network to obtain an identification result of whether a node in the road relation graph to be detected is the cross-link road data.

The cross-link road recognition network is obtained by training the training method of the cross-link road recognition network.

And inputting the constructed road relation graph to be detected into the network to obtain an identification result of whether each node in the road relation graph to be detected is the link-crossing road data, so that the identification of the link-crossing road can be completed.

When the link-crossing road data exists in the navigation route, the guide actions of the triplets corresponding to the link-crossing road need to be processed, for example, two guide actions of the triplets are simultaneously broadcasted or displayed, so that the user can be prevented from deviating, and the use experience of the user is improved.

A method for training a cross-link road recognition network shown in the present specification will be described below with a specific embodiment.

The overall architecture of the training method of the graph neural network is shown in fig. 5, and comprises a data acquisition module, a relationship construction module, a deep learning model module and a relationship updating module. The data acquisition module is used for acquiring training data and test data; the relationship building module is used for building a road relationship graph comprising various road relationships. In addition, when the judgment result of the relationship updating in the closed-loop feedback mechanism is 'yes', a new relationship graph is constructed and the training data is updated; the deep learning model module is used for learning information from different relational graphs to improve the model identification effect; the relation updating module consists of an expert experience module and a relation generating module, the expert experience module is used for extracting part of test data and judging the quality of the part of data (judging whether the wrong score proportion is smaller than a proportion threshold value), the relation generating module is used for re-determining a new road relation from the wrong score data, and finally, the new road relation is fed back to the relation building module based on whether the new relation exists or not, so that the closed loop is realized, the model learning is richer, and the model identification effect is improved.

The specific functions of the above-described modules will be described in detail below.

A data acquisition module: road data are obtained from a database, cross-link data are screened out according to a certain strategy, the triple representation shows that the screening strategy is generally generated according to the methods such as road type, road grade and road composition and the like and front and back link data of the data are obtained at the same time. And providing manual labeling for the data, wherein the label means whether the data cross the link, and the rest data is used as test data for prediction, so that a road data set, a test data subset and a training data subset are obtained. After a batch of data is obtained by screening by using the strategy, real link-crossing road data and non-link-crossing road data exist, and the real link-crossing road needs to be identified through the following modules.

A relationship building module: constructing different road relationship subgraphs according to different road relationship types through a road data set, wherein a point represents a triple, and the type of an edge mainly comprises the following relationships: upstream and downstream relationships, the same road type, the same road composition, the same city, etc. Each node also contains basic information of each road link in the triples, such as information of road length, road width, lane number, road out degree, road in degree, road direction, road angle and the like, and the information is used for processing in a subsequent deep learning model module, wherein numerical features are directly used, and category features are correspondingly coded. Therefore, the road relation subgraph comprising various road relations can be constructed through the relation construction module, and the road relation graph is formed by a plurality of subgraphs together.

A deep learning model module: after the relationship building module builds the whole relationship graph, some graph models can be used, such as: GCN, GAT, GAE and the like learn the structural relationship of the whole graph and road interaction information, and then perform supervised learning on training data until the model converges.

The relation updating module mainly comprises screening test data, an expert experience submodule and a relation generating submodule, wherein the expert experience submodule is mainly used for judging the quality of the test data, and the relation generating module is mainly used for analyzing and generating a new relation according to a quality judging result and feeding the new relation back to the relation building module.

And after the model learning in the deep learning model module is converged, screening whether the test data is the test data or not to obtain a test data subset in the road data set and obtain a classification result of the test data.

The expert experience sub-module is used for extracting partial data from the test data subset as prediction data and carrying out quality judgment on the prediction data, wherein the extraction method comprises the following steps:

step one, according to a relation construction module, randomly extracting a part of data from different relations, wherein the part of data is assumed to be data _ A;

and secondly, counting the traffic flow of each triad in the data _ A, such as daily flow, monthly flow, seasonal flow, annual flow and the like, and randomly extracting a part of data from the traffic flow, which is assumed to be data _ B.

And then evaluating the data _ B data, and comparing the classification result of the dataB with the label to find out error data. And judging whether the fraction error ratio corresponding to the fraction error data is smaller than a preset ratio threshold, if so, updating the fraction error data, and if so, updating the fraction error data through the relationship generation submodule.

And the relation generation submodule is used for analyzing the error data to obtain accurate characteristic vectors of edges and nodes of the error data. The analysis may be performed manually or by a relational network. After the analysis is completed, the road relation graph needs to be updated through updated information, the data are added to training data, and the cross-link road recognition network is retrained through the training data, so that the recognition effect of the network can be improved.

Corresponding to the embodiments of the method, the present specification also provides embodiments of the apparatus and the terminal applied thereto.

As shown in fig. 6, fig. 6 is a block diagram of a training apparatus across a link road recognition network according to an exemplary embodiment shown in the present specification, the apparatus including:

the road data set acquiring unit 610 is configured to acquire a road data set, where the road data set includes road information of multiple triplets and a label corresponding to each triplet, and the label is used to indicate whether the triplet is link-crossing road data; the road data set comprises a training data subset and a testing data subset; the triplet comprises three links connected together;

a network training unit 620, configured to train a link-crossing road recognition network through the training data subset; the input of the cross-link road identification network is a road relation graph, and the output is an identification result of the cross-link road; the road relation graph comprises a plurality of nodes and edges connecting the nodes, each node is used for representing a triple, and each edge is used for representing the road relation between two triples; the road relation is used for representing whether the two triples have the same road information or representing the position relation of links in the two triples;

a test data subset prediction unit 630, configured to predict the test data subset according to the cross-link road identification network, and determine error data in which an identification result of the cross-link road identification network in the test data subset does not match a tag;

an update information acquisition unit 640 configured to acquire update information of the road map for the error data, the update information including: node data corresponding to the error data and edge data corresponding to the error data;

a network retraining unit 650 configured to retrain the cross-link road recognition network based on the updated information.

In an optional embodiment, the road data set acquiring unit 610 is configured to acquire three links connected; taking the three links as the triples in the road data set when the three links meet a preset road condition, wherein the preset road condition comprises: the road length of the middle link of the three links is smaller than a length threshold, and an included angle between the middle link and any other link of the three links is larger than a preset included angle threshold.

In an optional embodiment, the road data set adding unit 660 (not shown in the figure) is further included, configured to obtain, according to a position of a link in a triplet in the road data set, an adjacent triplet whose position is adjacent to the position of the link; adding the adjacent triplet to the road data set.

In an alternative embodiment, the road relationship graph comprises a plurality of road relationship sub-graphs, each road relationship sub-graph comprising a plurality of nodes and edges connecting the nodes, the edges connecting the nodes in different road relationship sub-graphs representing different road relationships.

In an optional embodiment, the road relationship comprises at least one of: whether the roads are of the same road type, whether the roads are of the same road composition, whether the roads are of the same road type, whether the roads belong to the same city or have an upstream-downstream relationship; and/or; the road information includes at least one of: road grade, road type, road composition, city to which the road belongs, and upstream and downstream relation of other triples, road length, road width, lane number, road departure, road entrance, road direction and road angle.

In an alternative embodiment, the test data subset prediction unit 630 includes: a prediction data screening subunit 631 (not shown in the figure) for selecting prediction data from the test data subset; a prediction data predictor 632 (not shown) for: predicting the prediction data through the cross-link road recognition network, and determining error data in the prediction data, wherein the recognition result of the cross-link road recognition network does not accord with the label.

In an optional embodiment, the edges of the road relationship graph are used for characterizing various road relationships between two triples; the predicted data filtering subunit 631 is configured to extract, for the test data subset, part of the road data from the road data corresponding to each road relationship, and combine the extracted road data into predicted data.

In an alternative embodiment, the predictive data filtering subunit 631 is configured to count the traffic flow of the links included in each triplet with respect to the test data subset, and determine the traffic flow range to which the average of the traffic flow of the three links included in each triplet belongs; and extracting part of road data from the road data corresponding to each traffic flow range, and forming the extracted road data into prediction data.

In an optional embodiment, the update information obtaining unit 640 is configured to determine an error ratio of error data to prediction data; the prediction data is part or all of the test data subset; and acquiring the update information of the road relation graph under the condition that the error ratio is greater than a preset ratio threshold value.

In an optional embodiment, the vehicle flow rate statistics module 670 (not shown in the drawings) is further included, configured to count the vehicle flow rates of the links included in each triplet and determine a vehicle flow rate range to which the average of the vehicle flow rates of the three links included in each triplet belongs, for the test data subset; an update information acquisition unit 640 configured to, for each of the traffic volume ranges: determining the fraction error ratio of the fraction error data corresponding to the vehicle flow range to the prediction data; acquiring the update information of the road relation graph under the condition that the proportion of the error is greater than the proportion threshold value corresponding to the traffic flow range; the larger the average value of the traffic flow of the road data corresponding to the traffic flow range is, the lower the proportional threshold value corresponding to the traffic flow range is.

In an optional embodiment, the update information obtaining unit 640 is configured to obtain update information of the road relationship graph through a relationship network; the input of the relation network is a characteristic vector of a node or an edge in the road relation graph, and the output is updated node data or edge data.

In an optional embodiment, the update information obtaining unit 640 is configured to use the selected partial data as relationship training data, and update a pre-trained relationship network through the relationship training data; and acquiring updating information corresponding to the residual data through the updated relation network aiming at the residual data except the relation training data in the error data.

As shown in fig. 7, fig. 7 is a block diagram of an apparatus for identifying a link-crossing road according to an exemplary embodiment shown in the present specification, the apparatus including:

the road relation graph building unit 710 is used for acquiring data to be detected and building a road relation graph to be detected according to the data to be detected;

the identification result obtaining unit 720 is configured to perform identification processing on the road relation graph to be detected through a link-crossing road identification network to obtain an identification result of whether a node in the road relation graph to be detected is link-crossing road data; the cross-link road recognition network is obtained by training the training method of the cross-link road recognition network.

The implementation process of the functions and actions of each module in the above device is specifically described in the implementation process of the corresponding step in the above method, and is not described herein again.

For the device embodiment, since it basically corresponds to the method embodiment, reference may be made to the partial description of the method embodiment for relevant points. The above-described embodiments of the apparatus are merely illustrative, wherein the modules described as separate parts may or may not be physically separate, and the parts displayed as modules may or may not be physical modules, may be located in one place, or may be distributed on a plurality of network modules. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution in the specification. One of ordinary skill in the art can understand and implement it without inventive effort.

As shown in fig. 8, fig. 8 is a hardware structure diagram of a computer device in which a training apparatus of the cross-link road recognition network or an apparatus for recognizing a cross-link road according to an embodiment is located, and the device may include: a processor 1010, a memory 1020, an input/output interface 1030, a communication interface 1040, and a bus 1050. Wherein the processor 1010, memory 1020, input/output interface 1030, and communication interface 1040 are communicatively coupled to each other within the device via a bus 1050.

The processor 1010 may be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an Application Specific Integrated Circuit (ASIC), or one or more Integrated circuits, and is configured to execute related programs to implement the technical solutions provided in the embodiments of the present disclosure.

The Memory 1020 may be implemented in the form of a ROM (Read Only Memory), a RAM (Random Access Memory), a static storage device, a dynamic storage device, or the like. The memory 1020 may store an operating system and other application programs, and when the technical solution provided by the embodiments of the present specification is implemented by software or firmware, the relevant program codes are stored in the memory 1020 and called to be executed by the processor 1010.

The input/output interface 1030 is used for connecting an input/output module to input and output information. The i/o module may be configured as a component in a device (not shown) or may be external to the device to provide a corresponding function. The input devices may include a keyboard, a mouse, a touch screen, a microphone, various sensors, etc., and the output devices may include a display, a speaker, a vibrator, an indicator light, etc.

The communication interface 1040 is used for connecting a communication module (not shown in the drawings) to implement communication interaction between the present device and other devices. The communication module can realize communication in a wired mode (such as USB, network cable and the like) and also can realize communication in a wireless mode (such as mobile network, WIFI, bluetooth and the like).

Bus 1050 includes a path that transfers information between various components of the device, such as processor 1010, memory 1020, input/output interface 1030, and communication interface 1040.

It should be noted that although the above-mentioned device only shows the processor 1010, the memory 1020, the input/output interface 1030, the communication interface 1040 and the bus 1050, in a specific implementation, the device may also include other components necessary for normal operation. In addition, those skilled in the art will appreciate that the above-described apparatus may also include only those components necessary to implement the embodiments of the present description, and not necessarily all of the components shown in the figures.

Embodiments of the present specification further provide a computer-readable storage medium, on which a computer program is stored, where the computer program, when executed by a processor, implements the aforementioned training method for a cross-link road recognition network or the cross-link road recognition method.

Computer-readable media, including both permanent and non-permanent, removable and non-removable media, may implement the information storage by any method or technology. The information may be computer readable instructions, data structures, modules of a program, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static Random Access Memory (SRAM), dynamic Random Access Memory (DRAM), other types of Random Access Memory (RAM), read Only Memory (ROM), electrically Erasable Programmable Read Only Memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital Versatile Discs (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device. As defined herein, a computer readable medium does not include a transitory computer readable medium such as a modulated data signal and a carrier wave.

The present specification also provides a computer program for implementing the aforementioned training method for a cross-link road recognition network or the cross-link road recognition method.

It should also be noted that the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrases "comprising one of 8230; \8230;" 8230; "does not exclude the presence of additional like elements in a process, method, article, or apparatus that comprises that element.

The foregoing description has been directed to specific embodiments of this disclosure. Other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims may be performed in a different order than in the embodiments and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing may also be possible or may be advantageous.

Claims

1. A method of training across a link road recognition network, the method comprising:

predicting the test data subset according to the cross-link road identification network, and determining error data in the test data subset, wherein the identification result of the cross-link road identification network is inconsistent with the label;

2. The method of claim 1, the obtaining a set of road data, comprising:

acquiring three links connected with each other;

taking the three links as the triples in the road data set when the three links meet a preset road condition, wherein the preset road condition comprises: the road length of the middle link of the three links is smaller than a length threshold, and an included angle between the middle link and any other link of the three links is larger than a preset included angle threshold.

3. The method of claim 1, the road relationship graph comprising a plurality of road relationship subgraphs, each road relationship subgraph comprising a plurality of nodes and edges connecting the nodes, the edges connecting the nodes in different road relationship subgraphs representing different road relationships.

4. The method of claim 3, wherein the predicting the test data subset according to the cross-link road identification network, and determining error data in the test data subset for which the identification result of the cross-link road identification network does not conform to the tag, comprises:

extracting part of road data from the road data corresponding to each road relation aiming at the test data subset, and forming the extracted road data into prediction data;

and predicting the prediction data according to the cross-link road recognition network, and determining error data in the prediction data, wherein the recognition result of the cross-link road recognition network is inconsistent with the label.

5. The method as set forth in claim 1, wherein,

the road relationship includes at least one of: whether the roads are of the same road type, whether the roads are of the same road composition, whether the roads are of the same road type, whether the roads belong to the same city or have an upstream-downstream relationship; and/or;

the road information includes at least one of: road grade, road type, road composition, upstream and downstream relation between the city to which the road belongs and other triples, road length, road width, number of lanes, road out-degree, road in-degree, road direction and road angle.

6. The method according to claim 1, wherein the acquiring, for the error data, update information of the road map includes:

determining the error division ratio of error data to prediction data; the prediction data is part or all of the test data subset;

and acquiring the update information of the road relation graph under the condition that the error ratio is greater than a preset ratio threshold value.

7. The method as set forth in claim 1, wherein,

the method further comprises the following steps: counting the traffic flow of the links included in each triple aiming at the test data subset, and determining the traffic flow range to which the average value of the traffic flow of the three links included in each triple belongs;

the acquiring, for the error data, the update information of the road relationship map includes:

for each traffic flow range, performing:

determining the fraction error ratio of the fraction error data corresponding to the vehicle flow range to the prediction data;

acquiring the update information of the road relation graph under the condition that the proportion of the error is larger than the proportion threshold value corresponding to the traffic flow range;

the larger the traffic flow average value of the road data corresponding to the traffic flow range is, the lower the proportional threshold value corresponding to the traffic flow range is.

8. The method of claim 1, wherein the obtaining of the updated information of the road relationship map comprises:

acquiring the updating information of the road relation graph through a relation network; the input of the relation network is a feature vector of a node or an edge in the road relation graph, and the output of the relation network is updated node data or edge data.

9. The method of claim 8, wherein the obtaining the updated information of the road relationship graph through the relationship network comprises:

selecting partial data from the error data, and acquiring a relationship label of the selected partial data, wherein the relationship label is used for representing accurate edge or node data;

taking the selected partial data as relationship training data, and updating a pre-trained relationship network through the relationship training data;

and acquiring updating information corresponding to the residual data through the updated relation network aiming at the residual data except the relation training data in the error data.

10. A method of identifying a link-crossing road, comprising:

identifying and processing the road relation graph to be detected through a link-crossing road identification network to obtain an identification result of whether a node in the road relation graph to be detected is link-crossing road data or not; wherein the cross-link road recognition network is trained by the method of any one of claims 1 to 9.

11. A training apparatus across a link road recognition network, comprising:

the cross-link road recognition network training unit is used for training a cross-link road recognition network through the training data subset; the input of the cross-link road identification network is a road relation graph, and the output of the cross-link road identification network is an identification result of a cross-link road; the road relation graph comprises a plurality of nodes and edges connecting the nodes, each node is used for representing a triple, and each edge is used for representing the road relation between two triples; the road relation is used for representing whether the two triples have the same road information or representing the position relation of links in the two triples;

and the cross-link road recognition network retraining unit is used for retraining the cross-link road recognition network based on the updating information.

12. An apparatus for identifying a link-crossing road, comprising:

the identification result acquisition unit is used for identifying the road relation graph to be detected through a link-crossing road identification network to obtain an identification result of whether a node in the road relation graph to be detected is link-crossing road data or not; wherein the cross-link road recognition network is trained by the method of any one of claims 1 to 9.

13. A computer device, comprising:

a processor;

a memory for storing processor-executable instructions;

wherein the processor implements the method of any one of claims 1-10 by executing the executable instructions.

14. A computer readable storage medium having stored thereon computer instructions which, when executed by a processor, implement the method of any one of claims 1-10.