Heuristic item determination and model training method, electronic device and computer storage medium

Abstract

The embodiments of the present application provide a method for determining heuristic items and training a model, an electronic device, and a computer storage medium, wherein the method for determining heuristic items includes: obtaining environmental information of an autonomous driving vehicle, driving starting point information of the autonomous driving vehicle, and candidate heuristic item punctuation information to be evaluated, wherein the environmental information includes at least: road boundary information of the area where the autonomous driving vehicle is located, and information about obstacles in the area where the autonomous driving vehicle is located; generating a road boundary subgraph based on the road boundary information, generating an obstacle subgraph based on the obstacle information, and generating a location point subgraph based on the driving starting point information and the candidate heuristic item punctuation information; generating a fully connected graph based on the road boundary subgraph, the obstacle subgraph, and the location point subgraph; inputting the fully connected graph into a graph neural network model for predicting heuristic item punctuation, and determining the heuristic items for driving planning for the autonomous driving vehicle according to the output of the graph neural network model.

Description

Technical Field

[0001] Embodiments of the present application relate to the field of autonomous driving technology, and in particular to a method for determining an heuristic item, a method for training a heuristic item determination model, and corresponding electronic devices and computer storage media. Background Art

[0002] In autonomous driving, driving planning for autonomous vehicles is an essential part. The basic task of driving planning is to calculate the most suitable driving path from the starting point to the end point, which meets multiple requirements such as safety, smoothness, and shortest path, given the perceived environmental information and the starting point and target point, for autonomous driving of autonomous vehicles.

[0003] In order to efficiently solve the driving planning problem, the current mainstream methods in driving planning methods include search-based methods represented by a* and sampling-based methods represented by rrt. These methods all require the setting of a reasonable heuristic function to guide the search direction or sampling distribution of the driving planning algorithm. Among them, the heuristic function is also called the heuristic term, which can quantitatively estimate the proximity of a sampling node in the environment to the most suitable planning path. Through the heuristic function, the search direction or sampling point distribution of the driving planning algorithm can be guided to search or sample in the space where better planning results can be obtained, so as to improve the planning speed and success rate.

[0004] However, the heuristic function of the current driving planning algorithm often fails to obtain accurate results due to factors such as the configuration of the autonomous driving vehicle or the accuracy of information, which leads to the inability to make the most reasonable driving plan. Summary of the invention

[0005] In view of this, an embodiment of the present application provides a solution for determining heuristic items and training a model to at least partially solve the above-mentioned problems.

[0006] According to a first aspect of an embodiment of the present application, a method for determining an heuristic item is provided, comprising: obtaining environmental information of an autonomous driving vehicle, driving starting point information of the autonomous driving vehicle, and candidate heuristic item punctuation information to be evaluated, wherein the environmental information at least includes: road boundary information of the area where the autonomous driving vehicle is located, and information about obstacles in the area where the autonomous driving vehicle is located; generating a road boundary subgraph based on the road boundary information, generating an obstacle subgraph based on the obstacle information, and generating a location point subgraph based on the driving starting point information and the candidate heuristic item punctuation information; generating a fully connected graph based on the road boundary subgraph, the obstacle subgraph, and the location point subgraph; inputting the fully connected graph into a graph neural network model for predicting heuristic item punctuation, and determining the heuristic items for driving planning for the autonomous driving vehicle according to the output of the graph neural network model.

[0007] According to a second aspect of an embodiment of the present application, a method for training an heuristic item determination model is provided, comprising: obtaining road boundary sample information of a preset area where an autonomous driving vehicle is located, obstacle sample information within the area, driving starting point sample information of the autonomous driving vehicle, and candidate heuristic item sample location point information; generating a road boundary sample subgraph based on the road boundary sample information, generating an obstacle sample subgraph based on the obstacle sample information, and generating a location point sample subgraph based on the driving starting point sample information and the candidate heuristic item sample location point information; generating a fully connected sample graph based on the road boundary sample subgraph, the obstacle sample subgraph and the location point sample subgraph; and training a graph neural network model for heuristic item punctuation prediction using the fully connected sample graph as a training sample.

[0008] According to the third aspect of an embodiment of the present application, there is provided an electronic device, comprising: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus; the memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform an operation corresponding to the method described in the first aspect or the second aspect.

[0009] According to a fourth aspect of an embodiment of the present application, a computer storage medium is provided, on which a computer program is stored. When the program is executed by a processor, the method described in the first aspect or the second aspect is implemented.

[0010] According to the scheme provided in the embodiment of the present application, a graph neural network model is used to determine the heuristic items of the driving plan. To this end, the environmental information, driving starting point information and candidate heuristic item punctuation information of the autonomous driving vehicle are first obtained. Through this information, the current environment of the autonomous driving vehicle and the possible target points can be accurately described, so as to provide sufficient and comprehensive information for the determination of subsequent heuristic items. Based on this information, subgraphs of different categories are further generated, namely, road boundary subgraphs for describing roads, obstacle subgraphs for describing obstacles and position point subgraphs for describing various position points. Through these subgraphs, the relationship between information of different categories can be described more accurately and orderly from different category perspectives. Based on these subgraphs, a fully connected graph in which all nodes of the subgraphs are interconnected is generated, and the fully connected graph gives the relationship information between all possible nodes. Based on the fully connected graph, the relationship between all nodes can be predicted and determined through the graph neural network model, so as to find the most reasonable heuristic item punctuation.

[0011] Therefore, no matter what configuration the autonomous driving vehicle has, it will not affect the determination of its heuristic items. Moreover, through various information and the relationship between information (fully connected graph), the environment in which the autonomous driving vehicle is located can be described more accurately and comprehensively, and more accurate and reasonable heuristic items can be determined for it. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of the present application. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0013] Figure 1 A schematic diagram of an exemplary system applicable to the embodiment of the present application;

[0014] Figure 2A A flowchart of the steps of a method for determining a model training method according to a heuristic item in Embodiment 1 of the present application;

[0015] Figure 2B for Figure 2A A schematic diagram of a vectorized representation of various data in the illustrated embodiment;

[0016] Figure 3A A flowchart of a method for determining an inspiration item according to Embodiment 2 of the present application;

[0017] Figure 3B for Figure 3A A schematic diagram of an example scenario in the illustrated embodiment;

[0018] Figure 4 is a schematic diagram of a system for determining an heuristic item according to Embodiment 3 of the present application;

[0019] Figure 5 This is a schematic diagram of the structure of an electronic device according to the fourth embodiment of the present application. DETAILED DESCRIPTION

[0020] In order to enable those skilled in the art to better understand the technical solutions in the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments in the embodiments of the present application should fall within the scope of protection of the embodiments of the present application.

[0021] The specific implementation of the embodiment of the present application is further explained below in conjunction with the accompanying drawings of the embodiment of the present application.

[0022] Figure 1 An exemplary system applicable to the embodiment of the present application is shown. Figure 1 As shown, the system 100 may include a server 102, a communication network 104, and / or one or more autonomous driving vehicles 106. Figure 1 The example is a plurality of autonomous driving vehicles. It should be noted that the solution of the embodiment of the present application can be deployed in the server 102 or in the autonomous driving vehicle 106.

[0023] The server 102 may be any suitable server for storing information, data, programs and / or any other suitable type of content. When the solution of the embodiment of the present application is deployed in the server 102, in some embodiments, the server 102 may receive the environmental information and driving starting point information sent by the autonomous driving vehicle 106, and determine the candidate heuristic item punctuation information for it, generate corresponding subgraphs of multiple categories based on this information, and then generate a fully connected graph based on the subgraphs of multiple categories; and then use the graph neural network model deployed in the server 102 to determine the heuristic items. As an optional example, in some embodiments, the server 102 may also perform driving planning for the autonomous driving vehicle 106 based on the determined heuristic items, and feed back the driving planning results to the autonomous driving vehicle 106. As an optional example, in some embodiments, the subgraphs of multiple categories include a road boundary subgraph for describing road boundary information, an obstacle subgraph for describing obstacle information, and a location point subgraph for describing the driving starting point and candidate heuristic item punctuation of the autonomous driving vehicle.

[0024] In some embodiments, the communication network 104 can be any suitable combination of one or more wired and / or wireless networks. For example, the communication network 104 can include any one or more of the following: the Internet, an intranet, a wide area network (WAN), a local area network (LAN), a wireless network, a digital subscriber line (DSL) network, a frame relay network, an asynchronous transfer mode (ATM) network, a virtual private network (VPN), and / or any other suitable communication network. The autonomous driving vehicle 106 can be connected to the communication network 104 via one or more communication links (e.g., communication link 112), and the communication network 104 can be linked to the server 102 via one or more communication links (e.g., communication link 114). The communication link can be any communication link suitable for transmitting data between the autonomous driving vehicle 106 and the server 102, such as a network link, a dial-up link, a wireless link, a hardwired link, any other suitable communication link, or any suitable combination of such links.

[0025] The autonomous driving vehicle 106 may include a vehicle or other forms of robotic devices with autonomous driving functions. When the solution of the embodiment of the present application is deployed in the server 102, the autonomous driving vehicle 106 may upload the collected corresponding information such as environmental information, driving starting point information, etc. to the server 102, and after receiving the driving plan fed back by the server 102, autonomously control driving based on the driving plan. When the solution of the embodiment of the present application is deployed in the autonomous driving vehicle 106, in some embodiments, the autonomous driving vehicle 106 can determine the candidate heuristic item punctuation information based on the collected environmental information and driving starting point information, generate corresponding subgraphs of multiple categories based on this information, and then generate a fully connected graph based on the subgraphs of multiple categories; and then use the graph neural network model deployed in the autonomous driving vehicle 106 to determine the heuristic items. As an optional example, in some embodiments, the autonomous driving vehicle 106 can also perform driving planning based on the determined heuristic items. As an optional example, in some embodiments, the subgraphs of multiple categories include a road boundary subgraph for describing road boundary information, an obstacle subgraph for describing obstacle information, and a location point subgraph for describing the driving starting point and candidate heuristic item punctuation of the autonomous driving vehicle. As an optional example, in some embodiments, the autonomous driving vehicle 106 may also upload data related to the driving plan to the server 102 for storage.

[0026] Although server 102 is illustrated as one device, in some embodiments, any suitable number of devices may be used to perform the functions performed by server 102. For example, in some embodiments, multiple devices may be used to implement the functions performed by server 102. Alternatively, the functions of server 102 may be implemented using cloud services.

[0027] Based on the above system, the solution of the present application is described below through multiple embodiments. For ease of understanding, the solution of the embodiment of the present application is first described from the perspective of heuristic item determination model training, and then based on the heuristic item determination model completed by the training, the solution of the embodiment of the present application is described from the perspective of model application.

[0028] Embodiment 1

[0029] Reference Figure 2A , shows a flowchart of the steps of a method for determining a model training method according to an inspiration item in Embodiment 1 of the present application.

[0030] The heuristic item determination model training method of this embodiment includes the following steps:

[0031] Step S202: Obtain road boundary sample information of the preset autonomous driving vehicle area, obstacle sample information within the autonomous driving vehicle area, driving starting point sample information of the autonomous driving vehicle, and candidate inspiration item sample location point information.

[0032] Among them, the road boundary sample information can be obtained through the high-precision map of the area where the autonomous driving vehicle is located, which can be used to describe the road conditions in the area; the obstacle sample information can be obtained through the high-precision map of the area where the autonomous driving vehicle is located and / or the data collected by the autonomous driving vehicle through its sensors (including but not limited to cameras, lidar, etc.), which can be used to describe the obstacle conditions on the road in the area where the autonomous driving vehicle is located; the driving starting point sample information of the autonomous driving vehicle can be obtained based on the positioning information of the autonomous driving vehicle. As for the candidate heuristic item sample location point information, in the training stage of the heuristic item determination model, the corresponding simulation environment can be constructed based on the high-precision map and obstacle information, combined with the positioning information of the autonomous driving vehicle and the data collected by the sensor, and then, based on the area in the environment that can be driven and does not collide with obstacles, the candidate heuristic item location points are randomly generated as samples. The candidate heuristic item sample location points may include multiple, which can be considered as multiple target points on the path that the autonomous driving vehicle may travel.

[0033] Based on this, in a feasible way, this step can be implemented as follows: obtaining map sample information of the area where the autonomous driving vehicle is located and obstacle sample information in the area; based on the road boundary sample information and obstacle sample information in the map sample information, as well as the driving start point sample information, generating candidate heuristic item sample location point information for the autonomous driving vehicle. In this way, the cost of obtaining samples is reduced, and the accuracy and generalization of subsequent heuristic item processing in a large number of different environments can be improved. In addition, the obstacle information is integrated to make the sample information more comprehensive and sufficient, providing more accurate and rich information for subsequent heuristic item prediction.

[0034] Further optionally, on the basis of obtaining the above-mentioned various information, the driving planning of the autonomous driving vehicle can be carried out based on the driving starting point corresponding to the driving starting point sample information and the candidate heuristic item position point corresponding to the candidate heuristic item sample position point information to obtain the corresponding driving path; a key point sequence is extracted from the driving path as the expected heuristic item position point corresponding to the candidate heuristic item sample position point information. In this way, a mature driving planning algorithm or manual driving planning can be used to obtain the driving path. The key point sequence extracted from the driving path can be used as the expected heuristic item position point corresponding to the candidate heuristic item sample position point information. The expected heuristic item position point can be used as a reference for the target heuristic item position point predicted by the graph neural network model in the future to evaluate the model training effect.

[0035] For example, the simulation generation and analysis storage process of sample information can be completed through offline data collection. The process includes: first, reading a frame of actual data containing high-precision maps, perception information of autonomous driving vehicles (perception modules of autonomous driving vehicles (at least including data collected by cameras and lidar)), and positioning information of autonomous driving vehicles, and constructing a corresponding simulation environment; then, a group of candidate target points (i.e., candidate heuristic item sample position points) based on the driving starting point are randomly generated in the area of ​​the map that can be driven and does not collide with obstacles in the environment; then, calling the planner to calculate the driving trajectory from the driving starting point to the candidate target point; then, extracting the key point sequence in the driving trajectory, the key point selection method can be sampling at equal intervals on the obtained driving trajectory or selecting a turning point sequence for switching the trajectory direction. The above process is repeated many times, and the environment information, driving starting point information, candidate target point information and result information of each planning process are stored. The key points can be used as the expected heuristic item position points corresponding to the corresponding candidate heuristic item sample position points in the subsequent process.

[0036] It should be noted that in the embodiment of the present application, the candidate heuristic item sample location points include multiple ones. Unless otherwise specified, the numbers related to "multiple", "multiple", etc. in this document all mean two or more.

[0037] Step S204: Generate a road boundary sample subgraph based on the road boundary sample information, generate an obstacle sample subgraph based on the obstacle sample information, and generate a position point sample subgraph based on the driving start point sample information and the candidate heuristic item sample position point information.

[0038] For each category of information, a corresponding subgraph can be generated based on the corresponding positions of the information and the positional relationships between them.

[0039] Specifically, at least one of the following methods may be used to generate the corresponding sub-graph, including:

[0040] (1) Generating a road boundary sample subgraph based on road boundary sample information can be implemented as follows: obtaining the position sequence of location points on the road boundary according to the road boundary sample information; constructing a corresponding road boundary vector according to the position sequence; and constructing and generating a road boundary sample subgraph according to the location points and the road boundary vector.

[0041] (2) Generating an obstacle sample subgraph based on obstacle sample information can be implemented as follows: according to the obstacle sample information, obtaining the outer envelope of the obstacle sample and the arrangement order of the envelope points of the outer envelope; constructing an obstacle vector for the corresponding obstacle sample according to the arrangement order of the envelope points; and constructing and generating an obstacle sample subgraph according to the obstacle vector.

[0042] If there are multiple obstacle samples, an obstacle sample subgraph may be constructed based on the obstacle vector corresponding to each obstacle sample and the positional relationship between each obstacle sample.

[0043] (3) Generating a location point sample subgraph based on the driving starting point sample information and the candidate heuristic item sample location point information can be implemented as follows: according to the configuration (mechanical shape) of the autonomous driving vehicle, corresponding outer envelopes are generated for the driving starting point indicated by the driving starting point sample information and the candidate heuristic item sample location points indicated by the candidate heuristic item sample location information; generating corresponding location point outer envelope vectors according to the arrangement order of the envelope points in each generated outer envelope; and constructing and generating corresponding location point sample subgraphs according to the location point outer envelope vectors.

[0044] In the embodiment of the present application, the aforementioned information, including road boundary sample information, obstacle sample information, driving start point sample information, and candidate heuristic item sample location point information (also referred to as candidate target point information), is described in the form of vectors, such as Figure 2B as shown in .

[0045] Among them, the road boundary selects the front and rear position points in turn according to the position data of the position point to construct the vector; the outer envelope of the obstacle constructs the vector in turn according to the arrangement order of the envelope points; the driving start point and the candidate target point ( Figure 2B Only one example is shown in FIG. 1 , and other candidate target points can be implemented by referring to this example.) According to the mechanical shape of the autonomous driving vehicle, a vector representation of the corresponding outer envelope is constructed. Among them, the road boundary information can correspond to the road boundary class information, and its vector can correspond to the road boundary class vector; the obstacle information can correspond to the obstacle class information, and its vector can correspond to the obstacle class vector; the driving start point and the candidate target point can correspond to the position point class information, and its vector can correspond to the position point class vector.

[0046] Then, each position point in each type of vector is taken as a node, and a subgraph is constructed for the elements of each type of vector, and each node in the subgraph is connected to each other, that is, the road boundary sample subgraph, the obstacle sample subgraph, and the position point sample subgraph.

[0047] Step S206: Generate a fully connected sample graph based on the road boundary sample subgraph, the obstacle sample subgraph and the location point sample subgraph.

[0048] After obtaining the road boundary sample subgraph, obstacle sample subgraph and location point sample subgraph, a fully connected sample graph can be generated based on these subgraphs. That is, each subgraph is connected as a node to form a fully connected complete graph structure, namely, a fully connected sample graph.

[0049] In one feasible approach, all nodes in the road boundary sample subgraph, all nodes in the obstacle sample subgraph, and all nodes in the location point sample subgraph can be connected to each other to generate a fully connected sample graph. That is, all nodes in one subgraph are connected to all nodes in other subgraphs to finally generate a fully connected sample graph. The fully connected sample graph thus generated contains all possible relationships and information between nodes, providing a better foundation for the subsequent training of the graph neural network model.

[0050] Step S208: Using the fully connected sample graph as a training sample, training the graph neural network model for heuristic item punctuation prediction.

[0051] In the fully connected sample graph, at least the weights of the edges within all subgraphs and the weights of the edges between subgraphs need to be used as parameters to be learned by the graph neural network model.

[0052] In addition, in a feasible manner, the graph neural network model is connected to a multi-layer perceptron MLP layer after the vector output layer to output the prediction results efficiently and quickly. Based on this, in a feasible manner, this step can be implemented as follows: input the fully connected sample graph into the graph neural network model to obtain the vector corresponding to the inspiration item sample position point; input the vector corresponding to the inspiration item sample position point into the multi-layer perceptron layer to obtain the information of the target inspiration item position point; and train the graph neural network model according to the difference between the target inspiration item position point and the expected position point.

[0053] Among them, optionally, according to the difference between the target inspiration item position point and the expected position point, training the graph neural network model may include: training the graph neural network model according to the difference between the target inspiration item position point and the expected inspiration item position point, and a preset regression loss function. The method of using a regression loss function is more suitable for the training of models related to autonomous driving tasks, so that the trained model is more in line with the needs of the actual use environment. In the embodiment of the present application, there is no restriction on the specific implementation method of the regression loss function, and the square loss, absolute value or huber loss method can all be applied.

[0054] Exemplarily, after constructing a fully connected graph, the embedding output layer corresponding to multiple candidate target point positions (i.e., candidate inspiration item sample position points) can be connected to a multi-layer perceptron (MLP) layer, and the output results of the multi-layer perceptron layer can be combined into a target position point (i.e., target inspiration item position point), and the distance difference between it and the expected position point is calculated, and the parameters of the graph neural network model are trained using the back-propagation algorithm based on the distance difference.

[0055] It should be noted that in the embodiments of the present application, the graph neural network model can be implemented in any appropriate form, including but not limited to: GNN, GCN, etc., as long as the trained model can realize the prediction of inspired project punctuation. The embodiments of the present application do not limit the specific implementation form of the graph neural network model.

[0056] It can be seen that through this embodiment, multiple corresponding subgraphs are constructed based on multi-category sample information, and then a fully connected graph is constructed based on the multiple subgraphs, and the graph neural network model is trained with the fully connected graph as a training sample. As a result, the trained graph neural network model can comprehensively consider information from all aspects to make more accurate predictions and judgments on possible heuristic item locations, and find the most reasonable heuristic item points. Therefore, no matter what configuration the autonomous driving vehicle has, it does not affect the determination of its heuristic items, and through various information and the relationship between information (fully connected graph), the environment in which the autonomous driving vehicle is located can be more accurately and comprehensively described, and more accurate and reasonable heuristic items can be determined for it.

[0057] Embodiment 2

[0058] Reference Figure 3A , shows a step flow chart of a method for determining an inspiration item according to embodiment 2 of the present application.

[0059] This embodiment is based on the graph neural network model obtained by training in Embodiment 1, and explains the method for determining the heuristic items provided in the embodiment of the present application from the perspective of its application.

[0060] The heuristic item determination method comprises the following steps:

[0061] Step S302: Obtain environmental information of the autonomous driving vehicle, driving starting point information of the autonomous driving vehicle, and punctuation information of candidate heuristic items to be evaluated.

[0062] Among them, the environmental information of the autonomous driving vehicle includes at least: road boundary information of the area where the autonomous driving vehicle is located, and information about obstacles in the area where the autonomous driving vehicle is located. The road boundary information can be obtained based on the high-precision map corresponding to the area where the autonomous driving vehicle is located; the obstacle information can be obtained based on the high-precision map corresponding to the area where the autonomous driving vehicle is located and / or the data collected by the sensors in the autonomous driving vehicle; the driving starting point information can be obtained based on the positioning of the autonomous driving vehicle or manual input; the candidate heuristic project punctuation information can be randomly generated or generated according to preset rules based on the above-mentioned high-precision map, road boundary information, obstacle information and driving starting point information. Among them, the candidate heuristic project punctuation can include multiple, and the preset rules can be appropriately set by those skilled in the art according to actual needs, and the embodiments of the present application do not limit this.

[0063] Step S304: Generate a road boundary subgraph based on the road boundary information, generate an obstacle subgraph based on the obstacle information, and generate a location point subgraph based on the driving start point information and the candidate heuristic item punctuation point information.

[0064] For example, the position sequence of the location points on the road boundary can be obtained according to the road boundary information; the corresponding road boundary vector can be constructed according to the position sequence; and the road boundary subgraph can be constructed according to the location points and the road boundary vector.

[0065] For another example, the outer envelope of the obstacle and the arrangement order of the envelope points of the outer envelope can be obtained according to the information of the obstacle; an obstacle vector is constructed for the corresponding obstacle according to the arrangement order of the envelope points; and an obstacle sub-graph is constructed according to the obstacle vector.

[0066] If there are multiple obstacles, an obstacle sub-graph may be constructed based on the obstacle vectors corresponding to the obstacles and the positional relationship between the obstacles.

[0067] For another example, according to the configuration of the autonomous driving vehicle, corresponding outer envelopes can be generated for the driving starting point indicated by the driving starting point information and the candidate heuristic item punctuation indicated by the candidate heuristic item punctuation information; according to the arrangement order of the envelope points in each generated outer envelope, a corresponding position point outer envelope vector is generated; and according to the position point outer envelope vector, a corresponding position point subgraph is constructed.

[0068] The specific process and implementation of generating each sub-graph as described above can refer to the description of generating the corresponding sample sub-graph in the first embodiment, which will not be repeated here.

[0069] Step S306: Generate a fully connected graph based on the road boundary subgraph, the obstacle subgraph and the location point subgraph.

[0070] That is, each subgraph is used as a node to interconnect the subgraphs. Specifically, all nodes in the road boundary subgraph, all nodes in the obstacle subgraph, and all nodes in the location point subgraph can be interconnected to generate a fully connected graph.

[0071] Step S308: Input the fully connected graph into a graph neural network model for heuristic item punctuation prediction, and determine the heuristic items for driving planning for the autonomous driving vehicle based on the output of the graph neural network model.

[0072] As mentioned earlier, the trained graph neural network model has the function of determining heuristic items for autonomous driving vehicles. Therefore, by inputting the fully connected graph into the graph neural network, corresponding results can be obtained, such as the location point information of specific target heuristic items, or the probability information corresponding to the punctuation points of multiple candidate heuristic items, etc.

[0073] The following is an example of a specific scenario to illustrate the above process. Figure 3B shown.

[0074] Figure 3B In the above, it is assumed that the autonomous driving vehicle is in area X (the specific size and boundary of the area are set by those skilled in the art according to actual needs, such as a certain geographical area, or a certain radius value range centered on the autonomous driving vehicle, etc.), based on the high-precision map of area X, the road boundary information and obstacle information of area X are obtained, and the driving starting point information of the autonomous driving vehicle is determined based on the current positioning of the autonomous driving vehicle. Then, N candidate heuristic project punctuation points (N≥2) are randomly generated based on the above information, and the information of the N candidate heuristic project punctuation points is obtained. Figure 3B In the figure, N is simply represented as 7.

[0075] Next, a corresponding road boundary subgraph is generated based on the road boundary information of area X; a corresponding obstacle subgraph is generated based on the obstacle information within area X; and a corresponding location point subgraph is generated based on the driving starting point information of the autonomous driving vehicle and the punctuation information of N candidate heuristic items.

[0076] Then, the nodes between these subgraphs are fully connected to generate a fully connected graph. The fully connected graph is input into the pre-trained graph neural network model to obtain the model output, for example, the output is the information of 3 position points among the N candidate heuristic project points, and the example in the figure is the candidate target points 2, 3 and 5. Then, the heuristic item can be determined based on this, where these 3 position points are the target heuristic item position points.

[0077] Through this embodiment, a graph neural network model is used to determine the heuristic items of driving planning. To this end, the environmental information, driving starting point information and candidate heuristic item punctuation information of the autonomous driving vehicle are first obtained. Through this information, the current environment of the autonomous driving vehicle and the possible target points can be accurately described, so as to provide sufficient and comprehensive information for the determination of subsequent heuristic items. Based on this information, subgraphs of different categories are further generated, namely, road boundary subgraphs for describing roads, obstacle subgraphs for describing obstacles, and position point subgraphs for describing various position points. Through these subgraphs, the relationship between information of different categories can be described more accurately and orderly from different category perspectives. Based on these subgraphs, a fully connected graph in which all nodes of the subgraphs are interconnected is generated, and the fully connected graph gives the relationship information between all possible nodes. Based on the fully connected graph, the relationship between all nodes can be predicted and determined through the graph neural network model, so as to find the most reasonable heuristic item punctuation.

[0078] Embodiment 3

[0079] Reference Figure 4, showing a schematic diagram of a heuristic item determination system according to embodiment three of the present application. Figure 4 In the system shown, the heuristic item determination scheme provided in the embodiment of the present application is fully explained from three perspectives: offline data collection, network training, and online deployment.

[0080] like Figure 4 As shown, in the offline data collection stage, the simulation generation and analysis storage of relevant information are completed. It includes: first, reading a frame of actual environment information including the high-precision map of the preset area where the autonomous driving vehicle is located, the perception information of the autonomous driving vehicle (information collected by the camera and lidar), the positioning information of the autonomous driving vehicle, etc., and constructing the corresponding simulation environment; then, a group of candidate target points (i.e., candidate heuristic item position points) based on the driving starting point are randomly generated in the area of ​​the map that can be driven and does not collide with obstacles in the environment, and then the planner is called to calculate the driving trajectory from the driving starting point to the candidate target point; then, the key point sequence in the driving trajectory is extracted, and the key point selection method can be sampling at equal intervals on the driving trajectory or selecting a turning point sequence for switching the driving trajectory direction. This process is repeated many times, and the environmental information (including road boundary information, obstacle outer envelope information), driving starting point, candidate target point and result information of each planning process are stored.

[0081] During the network training phase, a graph neural network model is used for training and learning. Including: First, the aforementioned environmental information (including road boundary information, obstacle outer envelope information), as well as driving start point information and candidate target point information are described in the form of vectors. When describing in vector form, the road boundary information selects the front and rear position points in sequence according to the position data of the position points on the road boundary to construct a vector, and the obstacle outer envelope constructs a vector in sequence according to the arrangement order of the envelope points. The driving start point and candidate target points also construct the corresponding outer envelope vector representation according to the mechanical shape of the autonomous driving vehicle, such as Figure 2B As shown. Then, each position point in each category of information is taken as a node, and a subgraph is constructed for each category of information based on their vector representation, and each node in each subgraph is connected in pairs. Subsequently, each subgraph is taken as a node and connected in pairs to form a complete graph structure (i.e., a fully connected graph), in which the weights of the edges within the subgraph and the edge weights between subgraphs are the parameters to be learned by the graph neural network model.

[0082] After constructing a graph structure such as a fully connected graph, it is input into the graph neural network model. The embedding output layer corresponding to the candidate target point position obtained by the graph neural network model is connected to a multi-layer perceptron (MLP) layer. The output results of the multi-layer perceptron layer are combined into the target position point, and the distance difference with the expected position point is calculated as the reward function of the graph neural network model. The parameters of the graph neural network model are trained using the back propagation algorithm.

[0083] When it comes to the online deployment stage, vehicle-side deployment or cloud-side deployment can be used. First, the current environmental information is obtained through high-precision maps, positioning information of autonomous driving vehicles and real-time perception data; then the environmental information is used to construct the input information of the graph neural network model according to the data vectorization method used in training the graph neural network model, and the candidate target points to be evaluated (i.e., candidate heuristic item points) are generated; then, the graph neural network model deployed on the vehicle side can be directly called, or the above input information can be uploaded to the cloud to call the graph neural network model deployed on the cloud, and the heuristic item calculation results of the current candidate target point can be obtained by forward calculation.

[0084] As can be seen from the above, the heuristic item determination system of this embodiment adopts a large number of offline simulations, and trains the simulation results through the graph neural network model, and finally calls the trained graph neural network model online to output the heuristic point position sequence. Among them, in the offline data collection process, firstly, the actual environmental information and other data are imported into the simulation system, and then the target points are randomly set in the drivable range of all maps in the simulation system, and the complete planning method is called to calculate the planning results, and the driving trajectory of each planning result is recorded. Then, the key point sequence in the driving trajectory is extracted, and the key point selection method can be sampling at equal intervals on the driving trajectory or selecting the turning point sequence of the trajectory direction switching. Then, the information including road boundaries, obstacles, vehicle positions (driving starting points), target points, etc. is used as input data, and the key point sequence is used as output data. The input data is modeled in vector form, and the mapping relationship between input and output is trained using the graph neural network model method. After the graph neural network model is trained, the new environmental information and the starting point and target point can be used as input. Through the forward calculation of the graph neural network model, a position point sequence can be quickly obtained, and this position point sequence is used as the heuristic item of the planner to improve the efficiency and success rate of driving planning.

[0085] Among them, the method of training the graph neural network model offline and using the graph neural network model online for heuristic item prediction can improve the accuracy and generalization of heuristic item calculation in a large number of different environments, avoiding the problem that the heuristic items designed based on rules cannot be accurately applied to certain configuration vehicles and cannot consider environmental obstacles. In addition, compared with other methods based on machine learning models, using the graph neural network model for heuristic item calculation and using the vectorized information description method in the graph neural network model to perform high-precision modeling of road boundaries and obstacle information without resolution loss can improve the accuracy and stability of the calculation results in environments where distance is sensitive (such as passing narrow roads, turning around, etc.).

[0086] Embodiment 4

[0087] Reference Figure 5 , shows a schematic diagram of the structure of an electronic device according to the fourth embodiment of the present application. The specific embodiment of the present application does not limit the specific implementation of the electronic device.

[0088] like Figure 5 As shown, the electronic device may include: a processor (processor) 502 , a communication interface (Communications Interface) 504 , a memory (memory) 506 , and a communication bus 508 .

[0089] in:

[0090] The processor 502 , the communication interface 504 , and the memory 506 communicate with each other via a communication bus 508 .

[0091] The communication interface 504 is used to communicate with other electronic devices or servers.

[0092] The processor 502 is used to execute the program 510, and specifically can execute the relevant steps in the above method embodiment.

[0093] Specifically, the program 510 may include program codes, which include computer operation instructions.

[0094] The processor 502 may be a CPU, or an application specific integrated circuit ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present application. The one or more processors included in the smart device may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0095] The memory 506 is used to store the program 510. The memory 506 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

[0096] The program 510 can be specifically used to enable the processor 502 to execute operations corresponding to the method described in any one of the aforementioned method embodiments.

[0097] The specific implementation of each step in program 510 can refer to the corresponding description of the corresponding steps and units in the above method embodiment, and has corresponding beneficial effects, which will not be repeated here. Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process of the above-described devices and modules can refer to the corresponding process description in the above method embodiment, which will not be repeated here.

[0098] An embodiment of the present application also provides a computer program product, including computer instructions, which instruct a computing device to execute operations corresponding to any one of the above-mentioned multiple method embodiments.

[0099] It should be pointed out that, according to the needs of implementation, the various components / steps described in the embodiments of the present application can be split into more components / steps, or two or more components / steps or partial operations of components / steps can be combined into new components / steps to achieve the purpose of the embodiments of the present application.

[0100] The above-mentioned method according to the embodiment of the present application can be implemented in hardware, firmware, or implemented as software or computer code that can be stored in a recording medium (such as a CD ROM, RAM, floppy disk, hard disk or magneto-optical disk), or implemented as a computer code originally stored in a remote recording medium or a non-temporary machine-readable medium downloaded through a network and to be stored in a local recording medium, so that the method described herein can be stored in such software processing on a recording medium using a general-purpose computer, a special-purpose processor or programmable or special-purpose hardware (such as an ASIC or FPGA). It can be understood that a computer, a processor, a microprocessor controller or programmable hardware includes a storage component (e.g., RAM, ROM, flash memory, etc.) that can store or receive software or computer code, and when the software or computer code is accessed and executed by a computer, a processor or hardware, the method described herein is implemented. In addition, when a general-purpose computer accesses the code for implementing the method shown here, the execution of the code converts the general-purpose computer into a special-purpose computer for executing the method shown here.

[0101] Those of ordinary skill in the art will appreciate that the units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the embodiments of the present application.

[0102] The above implementation methods are only used to illustrate the embodiments of the present application, and are not limitations on the embodiments of the present application. Ordinary technicians in the relevant technical field can make various changes and modifications without departing from the spirit and scope of the embodiments of the present application. Therefore, all equivalent technical solutions also belong to the scope of the embodiments of the present application. The scope of patent protection of the embodiments of the present application should be limited by the claims.

Claims

1. A method for determining an heuristic item, comprising: Acquire environmental information of the autonomous driving vehicle, driving starting point information of the autonomous driving vehicle, and candidate heuristic item point information to be evaluated, wherein the environmental information at least includes: road boundary information of the area where the autonomous driving vehicle is located, and information of obstacles in the area where the autonomous driving vehicle is located; Generate a road boundary subgraph based on the road boundary information, generate an obstacle subgraph based on the obstacle information, and generate a location point subgraph based on the driving start point information and the candidate heuristic item punctuation point information; Generate a fully connected graph based on the road boundary subgraph, the obstacle subgraph, and the location point subgraph; The fully connected graph is input into a graph neural network model for predicting punctuation points of inspiration items, and inspiration items for driving planning are determined for the autonomous driving vehicle based on the output of the graph neural network model.

2. The method according to claim 1, wherein: The generating a fully connected graph based on the road boundary subgraph, the obstacle subgraph and the location point subgraph comprises: All nodes in the road boundary subgraph, all nodes in the obstacle subgraph, and all nodes in the location point subgraph are connected to each other to generate a fully connected graph.

3. The method according to claim 1 or 2, wherein: The generating a road boundary subgraph based on the road boundary information includes: According to the road boundary information, obtaining the position sequence of the position points on the road boundary; Constructing a corresponding road boundary vector according to the position sequence; The road boundary subgraph is constructed and generated according to the position point and the road boundary vector.

4. The method according to claim 1 or 2, wherein: The generating an obstacle sub-graph based on the information of the obstacle comprises: According to the information of the obstacle, acquiring the outer envelope of the obstacle and the arrangement order of envelope points of the outer envelope; Constructing an obstacle vector for the corresponding obstacle according to the arrangement order of the envelope points; The obstacle subgraph is constructed and generated according to the obstacle vector.

5. The method according to claim 4, wherein: If the obstacle includes multiple obstacles, then constructing and generating the obstacle subgraph according to the obstacle vector includes: The obstacle subgraph is constructed and generated according to the obstacle vectors corresponding to the obstacles and the positional relationship between the obstacles.

6. The method according to claim 1 or 2, wherein: The generating of the location point sub-graph based on the driving starting point information and the candidate heuristic item punctuation information includes: According to the configuration of the automatic driving vehicle, generating corresponding outer envelopes for the driving starting point indicated by the driving starting point information and the candidate heuristic item punctuation points indicated by the candidate heuristic item punctuation points information respectively; Generate corresponding position point outer envelope vectors according to the arrangement order of envelope points in each generated outer envelope; According to the outer envelope vector of the position point, a corresponding position point subgraph is constructed and generated.

7. A method for training a model for determining an heuristic term, comprising: Obtaining road boundary sample information of a preset area where the autonomous driving vehicle is located, obstacle sample information in the area, driving start point sample information of the autonomous driving vehicle, and candidate heuristic item sample location point information; Generate a road boundary sample subgraph based on the road boundary sample information, generate an obstacle sample subgraph based on the obstacle sample information, and generate a position point sample subgraph based on the driving start point sample information and the candidate heuristic item sample position point information; Generate a fully connected sample graph based on the road boundary sample subgraph, the obstacle sample subgraph and the location point sample subgraph; The fully connected sample graph is used as a training sample to train a graph neural network model for punctuation prediction of inspired items.

8. The method according to claim 7, wherein: The generating a fully connected sample graph based on the road boundary sample subgraph, the obstacle sample subgraph and the location point sample subgraph comprises: All nodes in the road boundary sample subgraph, all nodes in the obstacle sample subgraph, and all nodes in the position point sample subgraph are connected to each other to generate a fully connected sample graph.

9. The method according to claim 7 or 8, wherein: The generating of the road boundary sample subgraph based on the road boundary sample information includes: acquiring the position sequence of the position points on the road boundary according to the road boundary sample information; constructing the corresponding road boundary vector according to the position sequence; constructing and generating the road boundary sample subgraph according to the position points and the road boundary vector; and / or, The generating of the obstacle sample subgraph based on the obstacle sample information includes: acquiring the outer envelope of the obstacle sample and the arrangement order of envelope points of the outer envelope according to the obstacle sample information; constructing an obstacle vector for the corresponding obstacle sample according to the arrangement order of the envelope points; and constructing and generating the obstacle sample subgraph according to the obstacle vector; and / or, The method of generating a location point sample subgraph based on the driving starting point sample information and the candidate heuristic item sample location point information includes: generating corresponding outer envelopes for the driving starting point indicated by the driving starting point sample information and the candidate heuristic item sample location points indicated by the candidate heuristic item sample location information according to the configuration of the autonomous driving vehicle; generating corresponding location point outer envelope vectors according to the arrangement order of envelope points in each generated outer envelope; and constructing and generating a corresponding location point sample subgraph according to the location point outer envelope vectors.

10. The method according to claim 7 or 8, wherein: The method of training a graph neural network model for predicting punctuation points of heuristic items using the fully connected sample graph as a training sample includes: Input the fully connected sample graph into the graph neural network model to obtain the vector corresponding to the inspiration item sample position point; Input the vector corresponding to the inspiration item sample position point into the multi-layer perceptron layer to obtain the information of the target inspiration item position point; The graph neural network model is trained according to the difference between the target inspiration item position point and the expected position point.

11. The method according to claim 10, wherein: The training of the graph neural network model according to the difference between the target inspiration item position point and the expected position point includes: The graph neural network model is trained based on the difference between the target inspiration item position point and the expected inspiration item position point, and a preset regression loss function.

12. The method according to claim 7 or 8, wherein: The obtaining of road boundary sample information of a preset area where the autonomous driving vehicle is located, obstacle sample information in the area, driving start point sample information of the autonomous driving vehicle, and candidate heuristic item sample location point information includes: Obtaining map sample information of the area where the autonomous driving vehicle is located and obstacle sample information in the area; Based on the road boundary sample information and the obstacle sample information in the map sample information, as well as the driving start point sample information, candidate heuristic item sample location point information is generated for the autonomous driving vehicle.

13. The method according to claim 12, wherein: The method further comprises: Based on the driving starting point corresponding to the driving starting point sample information and the candidate heuristic item position point corresponding to the candidate heuristic item sample position point information, performing driving planning of the automatic driving vehicle to obtain a corresponding driving path; A key point sequence is extracted from the driving path as the expected heuristic item position point corresponding to the candidate heuristic item sample position point information.

14. An electronic device comprising: A processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform an operation corresponding to the method according to any one of claims 1-13.

Images