Vehicle behavior decision-making method, equipment, device, storage medium and program product

Abstract

The invention provides a vehicle behavior decision-making method, equipment and device, a storage medium and a program product, and the method comprises the steps: obtaining a constructed driving rule knowledge graph; the driving rule knowledge graph comprises entity elements and relation elements; determining a plurality of sample triads based on the driving rule knowledge graph; each sample triple comprises a head entity element, a tail entity element and a relation element; the plurality of sample triads comprise positive sample triads and negative sample triads; the plurality of sample triads are grouped to obtain a plurality of sample groups, each sample group comprises at least two positive and negative sample pairs, and the positive sample triads and the negative sample triads in each positive and negative sample pair have partially identical elements; taking each sample group in the plurality of sample groups as a sample unit for training the atlas inference network, and training the atlas inference network to obtain a trained atlas inference network; and the map reasoning network is used for driving decision making.

Description

Technical Field

[0001] This disclosure relates to the field of autonomous driving technology, and more specifically, to a vehicle behavior decision-making method, device, apparatus, storage medium, and program product. Background Technology

[0002] In autonomous vehicle planning and control software systems, the behavioral decision-making layer plays the role of the "brain," and the driving decisions made by this layer affect vehicle safety. The decision-making capabilities of the behavioral decision-making layer often rely on deployed neural network models, such as finite state machine models, decision tree models, knowledge-based reasoning decision-making models, and reinforcement learning-based decision-making models.

[0003] However, these models either suffer from poor robustness or require a large amount of training data and computational resources, which affects the efficiency and accuracy of driving behavior decisions. Summary of the Invention

[0004] This disclosure provides at least one vehicle behavior decision-making method, device, apparatus, storage medium, and program product.

[0005] In a first aspect, embodiments of this disclosure provide a vehicle behavior decision-making method, including:

[0006] Obtain the constructed driving rule knowledge graph; the driving rule knowledge graph includes entity elements and relation elements, the entity elements are related to the road environment and / or vehicles, and the relation elements are used to represent the association between different entity elements;

[0007] Based on the driving rule knowledge graph, multiple sample triples are determined; each sample triple includes a head entity element, a tail entity element, and a relation element; the multiple sample triples include positive sample triples and negative sample triples.

[0008] The multiple sample triples are grouped to obtain multiple sample groups, wherein each sample group includes at least two positive and negative sample pairs, and there are some common elements between the positive sample triples and the negative sample triples in each positive and negative sample pair.

[0009] Each of the multiple sample groups is used as a sample unit to train the graph inference network, and the graph inference network is trained to obtain a trained graph inference network. The graph inference network is used to infer the representation vector of the element, and to make driving decisions using the representation vector and the driving rule knowledge graph.

[0010] Optionally, the step of using each of the plurality of sample groups as a sample unit to train the graph inference network to obtain a trained graph inference network includes:

[0011] Each of the multiple sample groups is used as a sample unit for training the graph inference network. The inference loss information under each sample unit is calculated, and the network parameters of the graph inference network are adjusted based on the inference loss information under multiple sample units to obtain the trained graph inference network. The inference loss information is determined based on the distance between the sum of the representation vectors of the head entity element and the relation element and the representation vector of the tail entity element.

[0012] Optionally, the step of calculating the inference loss information for each sample unit and adjusting the network parameters of the graph inference network based on the inference loss information for multiple sample units to obtain the trained graph inference network includes:

[0013] The predicted representation vector of each element under the sample unit is determined based on the graph inference network to be trained.

[0014] For any positive and negative sample pair in the sample unit, a first vector distance is determined based on the sum of the predicted representation vectors of the head entity element and the relation element of the positive sample triplet in the positive sample pair, and the predicted representation vector corresponding to the tail entity element of the positive sample triplet; and a second vector distance is determined based on the sum of the predicted representation vectors of the head entity element and the relation element of the negative sample triplet in the positive and negative sample pair, and the predicted representation vector corresponding to the tail entity element of the negative sample triplet.

[0015] Based on the difference between the first vector distance and the second vector distance corresponding to each pair of positive and negative samples, the inference loss information corresponding to the sample unit is determined;

[0016] With the goal of minimizing the sum of the inference loss information corresponding to each sample unit and the preset positive and negative sample interval values, the network parameters of the graph inference network are adjusted to obtain the trained graph inference network.

[0017] Optionally, the method further includes:

[0018] Acquire real-time vehicle data collected during vehicle operation.

[0019] Data associated with driving rules is extracted from the real-time vehicle data to obtain multiple first elements; the data associated with driving rules includes type and attribute; the type is used to characterize the object information involved in the driving rule, and the attribute is used to describe the relationship between types and / or the value of the type;

[0020] From the first element, at least one target element pair is determined, each target element pair including a first element for use as a head entity element and a first element for use as a relation element;

[0021] Using the trained graph inference network, determine the representation vectors corresponding to the two first elements in each pair of target elements;

[0022] Based on the representation vectors corresponding to the two first elements in each target element pair and the driving rule knowledge graph, driving behavior decisions are made for the vehicle.

[0023] Optionally, the step of making driving behavior decisions for the vehicle based on the representation vectors corresponding to the two first elements in each target element pair and the driving rule knowledge graph includes:

[0024] Based on the representation vectors corresponding to the two first elements, determine the target inference vector corresponding to the target element pair;

[0025] From the driving rule knowledge graph, target entity elements for indicating different driving behaviors are selected, and the similarity between the representation vector corresponding to each target entity element and the target inference vector is calculated.

[0026] If the similarity is greater than a preset threshold, the target driving behavior indicated by the target entity element corresponding to that similarity is taken as the driving behavior to be performed by the vehicle.

[0027] Optionally, based on the representation vectors corresponding to the two first elements, the target inference vector corresponding to the target element pair is determined, including:

[0028] If there are multiple target element pairs, the inference vector that appears most frequently among the inference vectors corresponding to each target element pair shall be taken as the target inference vector.

[0029] Optionally, after calculating the similarity between the representation vector corresponding to each target entity element and the target inference vector, the method further includes:

[0030] If none of the aforementioned similarities are greater than the preset threshold, a driving decision is made to maintain the current driving behavior for the vehicle.

[0031] Optionally, the method further includes:

[0032] Based on the target element pairs corresponding to the target driving behavior, the triples in the driving rule knowledge graph are supplemented.

[0033] Optionally, determining at least one target element pair includes:

[0034] Based on the data source of each of the first elements, at least one of the target element pairs is selected; wherein the data source includes vehicle parameter source and road environment source, and the two first elements in the target element pair correspond to the vehicle parameter source and the road environment source, respectively.

[0035] Optionally, determining at least one target element pair includes:

[0036] From each of the first elements, select the second element related to the type and the third element related to the attribute;

[0037] Based on the type of each second element, determine the first priority of each second element; and based on the attribute of each third element, determine the second priority of each third element.

[0038] Using the first priority and the second priority, at least one target element pair is determined from the second element and the third element.

[0039] Optionally, before determining the representation vectors corresponding to the two first elements in the target element pair using the trained graph inference network, the method further includes:

[0040] From the driving rule knowledge graph, find target tail entity elements that are associated with at least some of the first elements among a plurality of first elements;

[0041] Using the trained graph inference network, the representation vectors corresponding to the two first elements in the target element pair are determined, including:

[0042] After the search fails, the trained graph inference network is used to determine the representation vectors corresponding to the two first elements in the target element pair.

[0043] Optionally, the method further includes:

[0044] If the target tail entity element is found, the driving behavior indicated by the target tail entity element in the driving rule knowledge graph is taken as the driving behavior to be executed by the vehicle.

[0045] Secondly, this disclosure also provides a vehicle behavior decision-making device, including a memory, a transceiver, and a processor;

[0046] Memory is used to store computer programs; transceiver is used to send and receive data under the control of the processor; processor is used to read the computer programs from memory and perform the following operations:

[0047] Obtain the constructed driving rule knowledge graph; the driving rule knowledge graph includes entity elements and relation elements, the entity elements are related to the road environment and / or vehicles, and the relation elements are used to represent the association between different entity elements;

[0048] Based on the driving rule knowledge graph, multiple sample triples are determined; each sample triple includes a head entity element, a tail entity element, and a relation element; the multiple sample triples include positive sample triples and negative sample triples.

[0049] The multiple sample triples are grouped to obtain multiple sample groups, wherein each sample group includes at least two positive and negative sample pairs, and there are some common elements between the positive sample triples and the negative sample triples in each positive and negative sample pair.

[0050] Each of the multiple sample groups is used as a sample unit to train the graph inference network, and the graph inference network is trained to obtain a trained graph inference network. The graph inference network is used to infer the representation vector of the element, and to make driving decisions using the representation vector and the driving rule knowledge graph.

[0051] Optionally, the step of using each of the plurality of sample groups as a sample unit to train the graph inference network to obtain a trained graph inference network includes:

[0052] Each of the multiple sample groups is used as a sample unit for training the graph inference network. The inference loss information under each sample unit is calculated, and the network parameters of the graph inference network are adjusted based on the inference loss information under multiple sample units to obtain the trained graph inference network. The inference loss information is determined based on the distance between the sum of the representation vectors of the head entity element and the relation element and the representation vector of the tail entity element.

[0053] Optionally, the step of calculating the inference loss information for each sample unit and adjusting the network parameters of the graph inference network based on the inference loss information for multiple sample units to obtain the trained graph inference network includes:

[0054] The predicted representation vector of each element under the sample unit is determined based on the graph inference network to be trained.

[0055] For any positive and negative sample pair in the sample unit, a first vector distance is determined based on the sum of the predicted representation vectors of the head entity element and the relation element of the positive sample triplet in the positive sample pair, and the predicted representation vector corresponding to the tail entity element of the positive sample triplet; and a second vector distance is determined based on the sum of the predicted representation vectors of the head entity element and the relation element of the negative sample triplet in the positive and negative sample pair, and the predicted representation vector corresponding to the tail entity element of the negative sample triplet.

[0056] Based on the difference between the first vector distance and the second vector distance corresponding to each pair of positive and negative samples, the inference loss information corresponding to the sample unit is determined;

[0057] With the goal of minimizing the sum of the inference loss information corresponding to each sample unit and the preset positive and negative sample interval values, the network parameters of the graph inference network are adjusted to obtain the trained graph inference network.

[0058] Optionally, the method further includes:

[0059] Acquire real-time vehicle data collected during vehicle operation.

[0060] Data associated with driving rules is extracted from the real-time vehicle data to obtain multiple first elements; the data associated with driving rules includes type and attribute; the type is used to characterize the object information involved in the driving rule, and the attribute is used to describe the relationship between types and / or the value of the type;

[0061] From the first element, at least one target element pair is determined, each target element pair including a first element for use as a head entity element and a first element for use as a relation element;

[0062] Using the trained graph inference network, determine the representation vectors corresponding to the two first elements in each pair of target elements;

[0063] Based on the representation vectors corresponding to the two first elements in each target element pair and the driving rule knowledge graph, driving behavior decisions are made for the vehicle.

[0064] Optionally, the step of making driving behavior decisions for the vehicle based on the representation vectors corresponding to the two first elements in each target element pair and the driving rule knowledge graph includes:

[0065] Based on the representation vectors corresponding to the two first elements, determine the target inference vector corresponding to the target element pair;

[0066] From the driving rule knowledge graph, target entity elements for indicating different driving behaviors are selected, and the similarity between the representation vector corresponding to each target entity element and the target inference vector is calculated.

[0067] If the similarity is greater than a preset threshold, the target driving behavior indicated by the target entity element corresponding to that similarity is taken as the driving behavior to be performed by the vehicle.

[0068] Optionally, based on the representation vectors corresponding to the two first elements, the target inference vector corresponding to the target element pair is determined, including:

[0069] If there are multiple target element pairs, the inference vector that appears most frequently among the inference vectors corresponding to each target element pair shall be taken as the target inference vector.

[0070] Optionally, after calculating the similarity between the representation vector corresponding to each target entity element and the target inference vector, the method further includes:

[0071] If none of the aforementioned similarities are greater than the preset threshold, a driving decision is made to maintain the current driving behavior for the vehicle.

[0072] Optionally, the method further includes:

[0073] Based on the target element pairs corresponding to the target driving behavior, the triples in the driving rule knowledge graph are supplemented.

[0074] Optionally, determining at least one target element pair includes:

[0075] Based on the data source of each of the first elements, at least one of the target element pairs is selected; wherein the data source includes vehicle parameter source and road environment source, and the two first elements in the target element pair correspond to the vehicle parameter source and the road environment source, respectively.

[0076] Optionally, determining at least one target element pair includes:

[0077] From each of the first elements, select the second element related to the type and the third element related to the attribute;

[0078] Based on the type of each second element, determine the first priority of each second element; and based on the attribute of each third element, determine the second priority of each third element.

[0079] Using the first priority and the second priority, at least one target element pair is determined from the second element and the third element.

[0080] Optionally, before determining the representation vectors corresponding to the two first elements in the target element pair using the trained graph inference network, the method further includes:

[0081] From the driving rule knowledge graph, find target tail entity elements that are associated with at least some of the first elements among a plurality of first elements;

[0082] Using the trained graph inference network, the representation vectors corresponding to the two first elements in the target element pair are determined, including:

[0083] After the search fails, the trained graph inference network is used to determine the representation vectors corresponding to the two first elements in the target element pair.

[0084] Optionally, the method further includes:

[0085] If the target tail entity element is found, the driving behavior indicated by the target tail entity element in the driving rule knowledge graph is taken as the driving behavior to be executed by the vehicle.

[0086] Thirdly, embodiments of this disclosure also provide a vehicle behavior decision-making device, comprising:

[0087] The graph acquisition module is used to acquire the constructed driving rule knowledge graph; the driving rule knowledge graph includes entity elements and relation elements, the entity elements are related to the road environment and / or vehicles, and the relation elements are used to represent the association between different entity elements;

[0088] The sample determination module is used to determine multiple sample triples based on the driving rule knowledge graph; each sample triple includes a head entity element, a tail entity element, and a relation element; the multiple sample triples include positive sample triples and negative sample triples;

[0089] The sample grouping module is used to group the multiple sample triples to obtain multiple sample groups, wherein each sample group includes at least two positive and negative sample pairs, and there are some common elements between the positive sample triples and the negative sample triples in each positive and negative sample pair.

[0090] The training inference module is used to train the graph inference network by using each of the multiple sample groups as a sample unit to train the graph inference network, thereby obtaining a trained graph inference network. The graph inference network is used to infer the representation vectors of elements and to make driving decisions using the representation vectors and the driving rule knowledge graph.

[0091] Fourthly, embodiments of this disclosure also provide a processor-readable storage medium storing a program for causing a processor to perform the steps of the vehicle behavior decision-making method as described in the first aspect above.

[0092] Fifthly, embodiments of this disclosure also provide a computer program product, which, when invoked by a computer, causes the computer to execute the steps of the vehicle behavior decision-making method of the first aspect described above.

[0093] The vehicle behavior decision-making method, device, apparatus, storage medium, and program product provided in this disclosure, because the graph network uses fewer samples and resources during training, utilizes a small number of sample triples selected from the driving rule knowledge graph to train the graph inference network. This reduces memory resources and training time, resulting in a graph inference network with reliable accuracy. During training, using a sample group comprising at least two positive and negative sample pairs as a single sample unit allows each training sample unit to contain richer and more reliable information, leading to more accurate network learning results and improved robustness of the trained graph inference network. Furthermore, after obtaining the trained graph inference network, driving decisions can be quickly inferred using the representation vectors inferred by the network and the driving rule knowledge graph, reducing the time consumption of matching inference and improving inference efficiency.

[0094] In summary, the vehicle behavior decision-making method provided in this disclosure has advantages in terms of memory consumption for constructing the graph inference network and time consumption for matching inference. It can solve the shortcomings of traditional algorithms and improve the efficiency and accuracy of vehicle behavior decision-making.

[0095] To make the above-mentioned objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0096] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to this disclosure and, together with the specification, serve to explain the technical solutions of this disclosure. It should be understood that the following drawings only show some embodiments of this disclosure and should not be considered as limiting the scope. Those skilled in the art can obtain other related drawings based on these drawings without creative effort.

[0097] Figure 1 A schematic diagram of a system structure provided by an embodiment of this disclosure is shown;

[0098] Figure 2 A flowchart of a vehicle behavior decision-making method provided by an embodiment of this disclosure is shown;

[0099] Figure 3 A schematic diagram of a type of obstacle provided by an embodiment of this disclosure is shown;

[0100] Figure 4 This illustration shows a schematic diagram of a road network type provided in an embodiment of the present disclosure;

[0101] Figure 5 This diagram illustrates a structural schematic of a behavioral decision-making type provided by an embodiment of the present disclosure;

[0102] Figure 6 A schematic diagram of a driving rules knowledge graph provided in an embodiment of this disclosure is shown;

[0103] Figure 7 A schematic diagram of a vehicle behavior decision-making device provided in an embodiment of this disclosure is shown;

[0104] Figure 8 A schematic diagram of a vehicle behavior decision-making device provided in an embodiment of this disclosure is shown. Detailed Implementation

[0105] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. The components of the embodiments of this disclosure described and shown herein can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this disclosure is not intended to limit the scope of the claimed disclosure, but merely represents selected embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of protection of this disclosure.

[0106] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0107] In this disclosure, the term "multiple" refers to two or more objects, and other quantifiers are similar. In this disclosure, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0108] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein.

[0109] Research has found that the role of the vehicle's behavioral decision-making layer is to formulate correct driving strategies using perceived road environment information and the vehicle's own driving parameters (such as speed, orientation, and position) to ensure safe driving on the road. Models deployed at the behavioral decision-making layer can include finite state machine models, decision tree models, knowledge-based reasoning decision-making models, and reinforcement learning-based decision-making models. However, each of these models has its own drawbacks. For example, finite state machine models are suitable for driving decisions in simple scenarios but struggle with behavioral decision-making tasks in complex traffic scenarios (such as urban road environments with rich structured features). Decision tree models require defining separate decision networks for various driving scenarios; in complex scenarios, the control logic becomes more complex, and decision accuracy decreases. Knowledge-based reasoning decision-making models, while possessing strong decision-making capabilities, suffer from high computational complexity and resource requirements. Reinforcement learning-based decision-making models, while offering reliable decision accuracy, require large amounts of training data and computational resources, resulting in relatively long training times. Therefore, it can be seen that traditional behavioral decision-making algorithms have some drawbacks to varying degrees, affecting the decision-making efficiency and accuracy of the behavioral decision-making layer.

[0110] Based on the above research, this disclosure provides a vehicle behavior decision-making method, device, apparatus, storage medium, and program product. Since graph networks use fewer samples and resources during training, training the graph inference network with a small number of sample triplets selected from the driving rule knowledge graph can consume less memory resources and training time, resulting in a graph inference network with reliable accuracy. During training, using a sample group containing at least two positive and negative sample pairs as a single sample unit allows each training sample unit to contain richer and more reliable information, thereby making the network learning results more accurate and improving the robustness of the trained graph inference network. Furthermore, after obtaining the trained graph inference network, driving decisions can be quickly inferred using the representation vectors inferred by the network and the driving rule knowledge graph, reducing the time consumption of matching inference and improving inference efficiency. In summary, the vehicle behavior decision-making method provided by the embodiments of this disclosure has advantages in terms of memory consumption in constructing graph inference networks and time consumption of matching inference, overcoming the shortcomings of traditional algorithms and improving the efficiency and accuracy of vehicle behavior decisions.

[0111] The shortcomings of the above solutions are the result of the inventor's practical experience and careful research. Therefore, the discovery process of the above problems and the solutions proposed in this disclosure below should be considered as the inventor's contribution to this disclosure.

[0112] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0113] It should be noted that the specific terms mentioned in the embodiments of this disclosure include:

[0114] TransE Algorithm: An algorithm proposed to solve relational data in knowledge graphs. Its core idea is to vectorize the relation elements and entities in the knowledge graph. By continuously learning and adjusting the vector representations of the head entity element, relation element, and tail entity element, the vector of the head entity element plus the vector of the relation element is made as equal as possible to the vector of the tail entity element.

[0115] The technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this disclosure.

[0116] This disclosure provides a vehicle behavior decision-making method, device, apparatus, storage medium, and program product to address the problems of low decision-making efficiency and poor robustness when making behavior decisions for vehicles.

[0117] The method and apparatus are based on the same inventive concept. Since the methods and apparatus solve problems in similar ways, the implementation of the apparatus and methods can refer to each other, and the repeated parts will not be described again.

[0118] The vehicle-mounted server involved in the embodiments of this disclosure may refer to a device with certain data processing capabilities, a device with wireless connection function, or other processing devices connected to a wireless modem, etc.

[0119] The technical solutions provided in this disclosure can be applied to autonomous vehicle planning and control software systems. For example... Figure 1The diagram illustrates a system architecture according to an embodiment of this disclosure, including a cloud-based vehicle-mounted server 11 and an autonomous vehicle 12. The vehicle-mounted server 11 and the autonomous vehicle 12 are connected via a communication network. The vehicle-mounted server 11 can acquire road environment data and vehicle driving parameter data from the vehicle 12, and make decisions regarding vehicle driving behavior based on this data, resulting in a decision-making behavior. This decision-making behavior can then be used to control vehicle driving, thereby improving the safety and rationality of autonomous driving.

[0120] The following describes the vehicle behavior decision-making method provided in this embodiment, taking the vehicle system server as the execution subject as an example.

[0121] like Figure 2 The flowchart shown is a vehicle behavior decision-making method provided in an embodiment of this disclosure, which may include the following steps:

[0122] S201: Obtain the constructed driving rule knowledge graph; the driving rule knowledge graph includes entity elements and relation elements. Entity elements are related to the road environment and / or vehicles, and relation elements are used to represent the association between different entity elements.

[0123] Here, the driving rule knowledge graph can include multiple entity elements and multiple relationship elements (hereinafter referred to as "relationship"). Entity elements can be further divided into head entity elements (hereinafter referred to as "head") and tail entity elements (hereinafter referred to as "tail"). Entity elements are related to parameters of the road environment and / or the vehicle itself, while relationship elements are used to represent the associations between different entity elements. For example, head entity elements, tail entity elements, and relationship elements can form a triple. If the sum of the representation vectors corresponding to the head entity element and the relationship elements equals the representation vector corresponding to the tail entity element, then the triple is considered a correct triple. Entity elements and relationship elements can be determined based on collected historical driving data. Historical driving data can include various historical road environment data, historical driving parameter data, and historical driving behavior collected by the vehicle. Historical road environment data can include data such as various road objects, lanes, traffic signs, traffic lights, and driving scenarios. Road objects can include both static and dynamic objects; traffic signs can include speed limit signs, pedestrian crossing signs, straight-ahead signs, and turning signs; driving scenarios can include highway scenarios, urban road scenarios, intersection scenarios, and urban roundabout scenarios. Historical driving parameter data can include vehicle speed, vehicle battery level, whether turn signals are activated, and vehicle acceleration. Historical driving behavior is used to characterize the vehicle's actual driving decisions, such as left / right turns, U-turns, straight-ahead driving, lane changes, and acceleration.

[0124] Specifically, a driving rule knowledge graph can be constructed using data related to driving rules extracted from historical driving data. For example, a driving rule can include two parts: type and attributes. The type is used to characterize the object information involved in the driving rule. It can be used to describe the type to which the data information in the driving rule belongs. Specifically, it can describe the name of the object involved, the superior type and subordinate type to which the object belongs, etc. It can be mainly divided into intelligent vehicle type, obstacle type, road network type, and behavior decision type.

[0125] The `EgoVehicle` type indicates the various autonomous vehicles appearing on the road. The `Obstacle` type can be defined as traffic participants or road entities that affect the autonomous vehicle during its operation. The structure of the `Obstacle` type can be as follows: Figure 3 As shown, obstacle types can specifically include static obstacle types and dynamic obstacle types. Static obstacle types can include roadblocks, buildings, etc. Buildings can be houses, walls, office buildings, etc. Roadblocks can be fixed roadblocks and retractable roadblocks, etc. Dynamic obstacle types can include pedestrians, vehicles, animals, etc.

[0126] The RoadNetwoek type is primarily used to reflect the driving scenarios and road elements of intelligent vehicles. Specifically, the RoadNetwoek type can be categorized as follows: Figure 4 As shown. In Figure 4 In the context of RoadNetworkek, the specific types can include driving scenario types and road element types. Driving scenario types can include highway types, urban road types, intersection types, and urban roundabout types. Intersection types can include intersections without traffic lights and intersections with traffic lights. Road element types can include lane types, lane line types, pedestrian crossing types, road sign types, stop line types, and traffic light types. Road sign types can include speed limit sign types, pedestrian crossing sign types, straight-ahead sign types, left-turn sign types, right-turn sign types, U-turn sign types, etc.

[0127] For the Dehavior type, elements under this type can typically serve as tail entity elements in a driving rule knowledge graph, representing posture behavior influenced by head entity elements and relational elements. Specifically, the structure of the Dehavior type can be as follows: Figure 5 As shown, in Figure 5In this context, the Dehavior type can specifically include lateral behavior decision types and longitudinal behavior decision types. Lateral behavior decision types can include constant speed driving type, acceleration driving type, deceleration driving type, start type, and stop type, etc., while longitudinal behavior decision types can include lane keeping type and left / right lane changing type, etc.

[0128] Driving rule attributes describe the relationships between elements of different types, and / or the values ​​of elements within each type. Specifically, driving rule attributes can be divided into relational attributes and data attributes. Relational attributes indicate the relationships between elements of different types. For example, relational attributes can describe the topological relationships between roads, the relative positions of intelligent vehicles and obstacles, the positional relationships between intelligent vehicles and lanes, the relationships between intelligent vehicles and road signs, and the driving scenario in which the intelligent vehicle is located. For instance, the relational attribute isonLane(EgoVehicle, Lane1) indicates that the intelligent vehicle is currently traveling in lane Lane1, hasLaneMarker(Lane1, StraightSign) indicates that there is a straight-ahead road sign ahead of lane Lane1, hasFrontObstacle(EgoVehicle, Vehicle1) indicates that there is a vehicle Vehicle1 ahead of the intelligent vehicle, and hasRoadCondition(expressway, greenlight) indicates that the current highway scenario is a green light, allowing passage.

[0129] Data attributes are used to indicate the values ​​of elements of various types. The value types can include integer variables (int), string variables (string), floating-point variables (double), etc. For example, `distToObstacle(Vehicle1, x1)` indicates that the distance between the intelligent vehicle and the vehicle ahead (Vehicle1) is x1; `hasLaneNumber(Highway, 3)` indicates that the number of lanes in a highway scenario is 3; and `hasLightColor(TrafficLight, 'red')` indicates that the traffic light is red. For ease of understanding, Table 1 below shows some relational and data attribute names, domains, and value ranges:

[0130]

[0131] (Table 1)

[0132] In practical implementation, after obtaining a certain amount of historical driving data, driving rules can be extracted from the historical driving data based on the attributes and types associated with driving rules and relevant driving regulations, resulting in multiple attribute elements and type elements. Then, a driving rule knowledge graph can be constructed based on these attribute and type elements. This driving rule knowledge graph not only includes correct triples, but also allows for rich interactions between different triples, enabling the deduction of new data from existing data, thereby achieving a higher level of data recognition, reasoning, and decision management. Optionally, the driving rule knowledge graph can be constructed manually or using any existing graph construction method. Figure 6 The diagram shown is a schematic of a driving rule knowledge graph provided in an embodiment of this disclosure, including entity elements 1 to 15 and relation elements 1 to 12. The elements in entity elements 1 to 15 can be different or there can be some overlap. The entity elements and relation elements can be different or there can be some overlap.

[0133] For any given attribute element and any element of any type, within different triples, an attribute element can function as either an entity element of a different type or a relation element. For example, it can be the head entity element in triple 1, the tail entity element in triple 2, and a relation element in triple 3. Similarly, an element of a certain type can function as either an entity element of a different type or a relation element. For example, it can be the head entity element in triple 4, the tail entity element in triple 5, and a relation element in triple 6. In other words, each element obtained from historical driving data can, under the condition of complying with driving regulations, function as either an entity element or a relation element of a different type within different correct triples. For example, the element isonLane(EgoVehilec, Lane1) can be used as a head entity element, hasLaneMarker(Lane1, StraightSign) can be used as a relation element, and the element Driving at a constant speed can be used as a tail entity element; or, hasLaneMarker(Lane1, StraightSign) can be used as a head entity element, the element has Trafficsigns(Highspeedscenes, Speed ​​limitof 120) can be used as a relation element, and the element Driving at a constant speed can be used as a tail entity element.

[0134] For example, in the constructed driving rule knowledge graph, an entity element can correspond to a node in the graph. Nodes corresponding to two entity elements with an association relationship can be connected using a directional connection line, on which the relationship element can be labeled. For instance, the head entity element 1 (vehicle 1 ahead of lane 1) corresponds to a node 1, the left lane can be changed to a relationship element 1, and the tail entity element 1 (change lanes to the left) corresponds to a node 2. There is a connection relationship between node 1 and node 2, where the head entity element 1 points to the tail entity element 1, and this connection relationship corresponds to relationship element 1.

[0135] S202: Based on the driving rule knowledge graph, determine multiple sample triples; each sample triple includes a head entity element, a tail entity element, and a relation element; the multiple sample triples include positive sample triples and negative sample triples.

[0136] The vehicle behavior decision-making method provided in this disclosure uses the TransE graph inference algorithm for behavior decision-making. The TransE graph inference algorithm utilizes the translation invariance of word vectors to represent entities and relations using vectors. In this disclosure, entities are entity elements, and relations are relation elements. In a correct representation, based on the translation invariance of word vectors, the sum of the representation vector of the head entity element (hereinafter denoted by h) and the representation vector of the relation element (hereinafter denoted by r) equals the representation vector of the tail entity element (hereinafter denoted by t), that is, h + r = t. Here, the representation vector can be an embedding vector. In an incorrect representation, h + r ≠ t.

[0137] Sample triples are selected from entity and relation elements in the driving rules knowledge graph. A sample triple can include a head entity element, a tail entity element, and a relation element. A positive sample triple is a triple where each element has the relation h + r = t, and a negative sample triple is a triple where each element has the relation h + r ≠ t.

[0138] In practical implementation, after obtaining the driving rule knowledge graph, multiple positive sample triples and multiple negative sample triples can be selected from it based on the relationships between entity elements in the graph. For example, in Figure 6In this context, we can select entity element 1, relation element 1, and entity element 2 as positive sample triple 1, where entity element 1 is the head entity element and entity element 2 is the tail entity element; select entity element 1, relation element 1, and entity element 5 as negative sample triple 1, where entity element 1 is the head entity element and entity element 5 is the tail entity element; select entity element 1, relation element 3, and entity element 4 as positive sample triple 2, where entity element 1 is the head entity element and entity element 4 is the tail entity element; select entity element 1, relation element 2, and entity element 7 as negative sample triple 2, where entity element 1 is the head entity element and entity element 7 is the tail entity element; and select entity element 6, relation element 5, and entity element 7 as positive sample triple 3, where entity element 6 is the head entity element and entity element 7 is the tail entity element.

[0139] S203: Divide multiple sample triples into multiple sample groups, wherein each sample group includes at least two positive and negative sample pairs, and there are some common elements between the positive sample triples and the negative sample triples in each positive and negative sample pair.

[0140] In practice, after selecting multiple sample triples, the sample groups can be divided based on the elements included in the positive and negative sample triples. During this division, for any positive sample triple, a negative sample triple with one or two identical elements can be selected from the negative sample triples to obtain a corresponding sample group. This positive sample triple and the selected negative sample triple form a positive-negative sample pair. Understandably, a positive sample triple can appear in different positive-negative sample pairs, and a negative sample triple can also appear in different positive-negative sample pairs. At most, one element in each positive or negative sample triple can be identical across different positive and negative sample pairs.

[0141] For example, consider positive triplet 1 and negative triplet 1 in the above text as one positive-negative sample pair, consider positive triplet 1 and negative triplet 2 in the above text as another positive-negative sample pair, consider positive triplet 2 and negative triplet 2 as one positive-negative sample pair, and consider positive triplet 3 and negative triplet 2 as one positive-negative sample pair.

[0142] After obtaining multiple pairs of positive and negative samples, at least two pairs can be grouped together to form sample groups containing at least two pairs of positive and negative samples. Different sample groups must contain at least one pair of positive and negative samples that are different from each other.

[0143] S204: Using each sample group in multiple sample groups as a sample unit to train the graph inference network, the graph inference network is trained to obtain a trained graph inference network; the graph inference network is used to infer the representation vector of the element, and the representation vector and the driving rule knowledge graph are used to make driving decisions.

[0144] Here, the graph inference network can specifically be the TransE graph network.

[0145] In practice, a sample group can be considered as a sample unit. The positive and negative sample pairs within each sample unit are then input into the TransE graph network to be trained. The TransE graph network outputs the representation vector corresponding to each element in the positive and negative sample pairs. For example, sample unit 1 is sample group 1, which includes positive and negative sample pair 1 and positive and negative sample pair 2. Positive and negative sample pair 1 includes positive triplet 1 and negative triplet 1, and positive and negative sample pair 2 includes positive triplet 2 and negative triplet 2. The TransE graph network to be trained can output the head entity representation vector h1 corresponding to the head entity element, the relation representation vector r1 corresponding to the relation element, and the tail entity representation vector t1 corresponding to the tail entity element in positive triplet 1, and output the head entity representation vector h1 corresponding to the head entity element in negative triplet 1. ′ The relation representation vector r1 corresponding to each relation element ′ The tail entity representation vector t1 corresponding to the tail entity element ′ Similarly, the TransE graph network to be trained can output the head entity representation vector h2 corresponding to the head entity element, the relation representation vector r2 corresponding to the relation element, and the tail entity representation vector t2 corresponding to the tail entity element in the positive sample triple 2, as well as the head entity representation vector h2 corresponding to the head entity element in the negative sample triple 2. ′ The relation representation vector r2 corresponding to each relation element ′ The tail entity representation vector t2 corresponding to the tail entity element ′ .

[0146] After obtaining the representation vectors corresponding to each element in the positive and negative triplet pairs in each sample unit, the distance between these representation vectors can be used to calculate the loss of the TransE graph network to be trained. This loss is then used to iteratively train the TransE graph network, resulting in a trained network. Since a sample unit contains at least two positive and negative sample pairs, more sample data can be referenced when calculating the loss using sample units, thus improving the reasonableness of the calculated loss. Training with a reasonable loss can improve the robustness of the trained TransE graph network.

[0147] After obtaining the trained TransE graph network, the TransE graph network can be used to infer the elements corresponding to the driving data collected in real time by the vehicle, thereby obtaining the representation vectors corresponding to the elements. Then, the representation vectors and the representation vectors corresponding to the tail entity elements related to decision-making behavior in the driving rule knowledge graph can be used to make real-time decisions on the driving behavior of the vehicle.

[0148] Thus, since the graph network uses fewer samples and resources during training, training the graph inference network with a small number of sample triples selected from the driving rule knowledge graph reduces memory consumption, resource consumption, and training time, while achieving a graph inference network with reliable accuracy. During training, using sample groups containing at least two positive and negative sample pairs as a single sample unit provides the network with richer sample data for learning, thereby improving the robustness of the trained graph inference network. Furthermore, after obtaining the trained graph inference network, driving decisions can be quickly inferred using the representation vectors inferred by the network and the driving rule knowledge graph, reducing the time consumption of matching inference and improving inference efficiency. In summary, the vehicle behavior decision-making method provided in this disclosure has advantages in terms of memory consumption for constructing the graph inference network and time consumption for matching inference, overcoming the shortcomings of traditional algorithms and improving the efficiency and accuracy of vehicle behavior decisions.

[0149] In one embodiment, the sum of the representation vectors of the head entity element and the relation element in a positive sample triple is equal to the representation vector of the tail entity element, while the sum of the representation vectors of the head entity element and the relation element in a negative sample triple is not equal to the representation vector of the tail entity element.

[0150] Based on the translation invariance of word vectors, if the representation vectors of each element output by the TransE graph network to be trained are accurate, then the distance between the representation vector h of the head entity element plus the representation vector r of the relation element in the positive sample triple output by the network, and the representation vector t of the tail entity element will be 0. Similarly, the distance h of the representation vector h of the head entity element in the negative sample triple output by the network will also be 0. ′ Add the representation vector r of the relation elements ′ The representation vector t of the tail entity element ′The distance between them must not be equal to 0. In this embodiment of the disclosure, the distance between the representation vector h of the head entity element plus the representation vector r of the relation element, and the representation vector t of the tail entity element, is represented by L2 (also known as the Euclidean norm), and is denoted by d(h+r,t). Therefore, it can be seen that for a positive sample triple (head, relationship, tail), the smaller the distance d(h+r,t), the better; for a negative sample triple (head... ′ relationship ′ , tail ′ In terms of distance d(h) ′ +r ′ , t ′ The larger the better. Specifically, the head, relationship, and tail of a positive sample triplet, and the head of a negative sample triplet... ′ relationship ′ and tail ′ Some parts are the same.

[0151] Therefore, the steps for training the graph inference network in S204 above can be implemented as follows:

[0152] Each sample group in multiple sample groups is used as a sample unit to train the graph inference network. The inference loss information under each sample unit is calculated, and the network parameters of the graph inference network are adjusted based on the inference loss information under multiple sample units to obtain the trained graph inference network. The inference loss information is determined by the distance between the sum of the representation vectors of the head entity elements and relation elements and the representation vector of the tail entity elements.

[0153] Here, inference loss information is used to characterize the difference between the representation vector corresponding to the sample unit output by the graph inference network and the true representation vector corresponding to the sample unit. One sample unit can correspond to one inference loss information, which is determined by the vector distance between the sum of the representation vectors of the head entity elements and relation elements in the sample unit and the representation vector of the tail entity elements.

[0154] In practice, positive and negative sample pairs from each sample unit can be input into the graph inference network to be trained, yielding predicted representation vectors for each element in the positive sample triplet and for each element in the negative sample triplet. Then, using these predicted representation vectors, the vector distance between sample units can be calculated, and the inference loss information can be determined based on this distance. After obtaining the inference loss information for multiple sample units, this information can be used to adjust the network parameters of the graph inference network to be trained, resulting in a trained graph inference network.

[0155] In this way, training based on the inference loss information of each sample unit can fully learn the commonalities and differences between each positive and negative sample pair in the sample unit during the training process, thereby improving the robustness of the trained graph inference network.

[0156] In one embodiment, inference loss information and graph inference network training can be performed according to the following steps:

[0157] Step 1: Determine the predicted representation vector of each element under the sample unit based on the graph inference network to be trained.

[0158] Here, each element under a sample unit refers to the element within each triplet of the positive and negative sample pairs included in that sample unit. For example, in a sample unit comprising two positive and negative sample pairs, this sample unit includes 12 elements: the head entity element, tail entity element, and relation element corresponding to the two positive sample triples, and the head entity element, tail entity element, and relation element corresponding to the two negative sample triples. The predicted representation vector is the representation vector predicted by the graph inference network to be trained for a given element.

[0159] For example, for any positive and negative sample pair in any sample unit, the graph inference network to be trained can be used to output the predicted representation vectors corresponding to each element of the positive sample triplet in the positive and negative sample pair, and output the predicted representation vectors corresponding to each element of the negative sample triplet in the positive and negative sample pair.

[0160] Step 2: For any positive and negative sample pair in the sample unit, determine the first vector distance corresponding to the positive sample triplet based on the sum of the predicted representation vectors of the head entity element and relation element of the positive sample triplet and the predicted representation vector corresponding to the tail entity element of the positive sample triplet; and determine the second vector distance corresponding to the negative sample triplet based on the sum of the predicted representation vectors of the head entity element and relation element of the negative sample triplet and the predicted representation vector corresponding to the tail entity element of the negative sample triplet.

[0161] Here, the first vector distance can be the L2 norm distance between the sum of the predicted representation vectors of the head entity element and the relation element in the positive sample triple and the predicted representation vector corresponding to the tail entity element. The second vector distance can be the L2 norm distance between the sum of the predicted representation vectors of the head entity element and the relation element in the negative sample triple and the predicted representation vector corresponding to the tail entity element.

[0162] In practice, for any positive and negative sample pair in any sample unit, the first vector distance d(h+r,t) corresponding to the positive sample triple can be calculated using the predicted representation vectors corresponding to the head entity element, relation element, and tail entity element in the positive sample triple. Similarly, the second vector distance d(h′+r′,t′) corresponding to the negative sample triple can be calculated using the predicted representation vectors corresponding to the head entity element, relation element, and tail entity element in the negative sample triple.

[0163] Step 3: Determine the inference loss information corresponding to each sample unit based on the difference between the first vector distance and the second vector distance for each pair of positive and negative samples.

[0164] Taking a sample unit containing two positive and negative sample pairs as an example, for each positive and negative sample pair, the difference between the first vector distance corresponding to the positive sample triplet and the second vector distance corresponding to the negative sample triplet can be determined. Then, by using the sum of the distance differences corresponding to the two positive and negative sample pairs, the inference loss information corresponding to the sample unit can be determined.

[0165] Step 4: With the goal of minimizing the sum of the inference loss information corresponding to each sample unit and the preset positive and negative sample interval values, adjust the network parameters of the graph inference network to obtain the trained graph inference network.

[0166] In practice, after obtaining the inference loss information corresponding to each sample unit, the target loss can be determined by using the preset positive and negative sample interval value and the sum of the inference loss information corresponding to each sample unit. Then, the network parameters of the graph inference network are adjusted with the goal of minimizing the target loss to obtain the trained graph inference network.

[0167] Specifically, the target loss can be expressed using Formula 1:

[0168] min∑ (h，r，t)∈S ∑ (h′，r′，t′)∈S′ [γ+d(h1+r1,t1)-d(h1′+r1′,t1′)+d(h2+r2,t2)-d(h2′+r2′,t2′)] + (Formula 1)

[0169] Where S represents the set of positive sample triples, including each positive sample triple in each sample unit; S′ represents the set of negative sample triples, including each negative sample triple in each sample unit; (h, r, t) represents the representation vectors corresponding to the head entity element, the relation element, and the tail entity element in the positive sample triples; (h′, r′, t′) represents the representation vectors corresponding to the head entity element, the relation element, and the tail entity element in the negative sample triples; and γ represents the preset positive-negative sample interval value. d(h1+r1, t1) and d(h1′+r1′, t1′) represent the first and second vector distances corresponding to a pair of positive and negative samples in a sample unit, respectively; d(h2+r2, t2) and d(h2′+r2′, t2′) represent the first and second vector distances corresponding to another pair of positive and negative samples in a sample unit, respectively; d(h1+r1, t1)-d(h1′+r1′, t1′)+d(h2+r2, t2)-d(h2′+r2′, t2′) represents the inference loss information corresponding to a sample unit; [x] + Max(0, x) represents the maximum value between 0 and x; ∑ (h，r，t)∈S ∑ (h′，r′，t′)∈S′ [γ+d(h1+r1,t1)-d(h1′+r1′,t1′)+d(h2+r2,t2)-d(h2′+r2′,t2′)] + Indicates the target loss.

[0170] For example, after obtaining the predicted representation vectors of each element under each sample unit, each predicted representation vector can be substituted into Formula 1 above to obtain the target loss of the graph inference network. Then, the target loss is used to iteratively train the graph inference network to achieve reasonable adjustment of the network parameters and obtain a trained graph inference network.

[0171] Understandably, gradient backpropagation can be used to train a graph inference network. Thus, for the representation vector *hi* corresponding to the head entity element in any positive sample triplet, its gradient can be determined according to the following formula:

[0172]

[0173] Where hi represents the representation vector corresponding to the i-th head entity element. Let represent the gradient of hi, and let loss represent the partial derivative of the target loss. Let represent the partial derivative of hi, where hi1 represents the representation vector corresponding to the head entity element in the first positive triplet in the i-th sample unit, hi1′ represents the representation vector corresponding to the head entity element in the first negative triplet in the i-th sample unit, hi2 represents the representation vector corresponding to the head entity element in the second positive triplet in the i-th sample unit, and hi2′ represents the representation vector corresponding to the head entity element in the second negative triplet in the i-th sample unit.

[0174] Similarly, for the representation vector ri corresponding to the relation entity element in any positive sample triplet, its gradient can be determined according to the following formula:

[0175]

[0176] Where ri represents the representation vector corresponding to the i-th relation element. Represents the gradient of ri. Let ri1 ​​represent the representation vector corresponding to the relation element in the first positive triplet in the i-th sample unit, ri1′ represent the representation vector corresponding to the relation element in the first negative triplet in the i-th sample unit, ri2 represent the representation vector corresponding to the relation element in the second positive triplet in the i-th sample unit, and ri2′ represent the representation vector corresponding to the relation element in the second negative triplet in the i-th sample unit.

[0177] Similarly, for the representation vector *ti* corresponding to the relation entity element in any positive sample triplet, its gradient can be determined according to the following formula:

[0178]

[0179] Where ti represents the representation vector corresponding to the i-th tail entity element. Denotes the gradient of ti. Let ti represent the partial derivative of ti, where ti1 represents the representation vector corresponding to the tail entity element in the first positive triplet in the i-th sample unit, ti1′ represents the representation vector corresponding to the tail entity element in the first negative triplet in the i-th sample unit, ti2 represents the representation vector corresponding to the tail entity element in the second positive triplet in the i-th sample unit, and ti2′ represents the representation vector corresponding to the tail entity element in the second negative triplet in the i-th sample unit.

[0180] Thus, based on formulas two through four above, the predicted representation vectors of each element in the sample unit determined by the graph inference network to be trained, and the target loss corresponding to formula one, the gradients corresponding to the head entity element, relation element, and tail entity element can be determined respectively. Then, the solved gradients can be used to perform backpropagation training on the graph inference network to achieve reasonable adjustment of the network parameters and obtain a trained graph inference network.

[0181] Thus, the sum of the representation vectors corresponding to the head entity element and relation element in a positive sample triple, and the distance between them and the representation vector of the tail entity element, reflect the network's first inference loss when inferring from positive sample triples. Similarly, the sum of the representation vectors corresponding to the head entity element and relation element in a negative sample triple, and the distance between them and the representation vector of the tail entity element, reflect the network's second inference loss when inferring from negative sample triples. A smaller first inference loss indicates better inference performance, while a larger second inference loss also indicates better inference performance. Therefore, using the first and second inference losses to determine the target loss and then training the network can improve the accuracy of the trained graph inference network.

[0182] In one embodiment, after training the graph inference network, it can be deployed in the vehicle's infotainment server. Then, during vehicle operation, the vehicle's infotainment server and the graph inference network are used to make driving behavior decisions. Specifically, driving behavior decisions can be made according to the following steps S1 to S5:

[0183] S1: Acquire real-time vehicle data collected during vehicle operation.

[0184] Here, real-time vehicle data refers to various data collected in real time during the operation of an autonomous vehicle. Specifically, real-time vehicle data can include vehicle driving parameter data and road environment data. Vehicle driving parameter data characterizes real-time operating parameters of the vehicle, such as speed, acceleration, torque, horsepower, orientation, battery / fuel level, engine speed, and various instrument parameters. Road environment data characterizes the real-time environment of the road the vehicle is traveling on, such as static / dynamic obstacles, lanes, lane markings, traffic signs, driving scenarios, and the speed and orientation of dynamic obstacles. Optionally, real-time vehicle data does not include data related to driving decisions.

[0185] For example, autonomous vehicles can use sensors with various functions to collect data in real time while the vehicle is in motion, obtain real-time vehicle data, and send the real-time vehicle data to the vehicle's onboard server. In this way, the onboard server can also obtain the real-time vehicle data.

[0186] S2: Extract data associated with driving rules from real-time vehicle data to obtain multiple first elements; the data associated with driving rules includes type and attribute; the type is used to characterize the object information involved in the driving rule, and the attribute is used to describe the relationship between types and / or the value of the type.

[0187] Here, the first element can include attribute elements and / or type elements, and the number of first elements can be determined based on the amount of real-time vehicle data. The richer the real-time vehicle data collected, the more first elements can be proposed.

[0188] In practice, the vehicle server can classify real-time vehicle data by type and attribute based on driving rules, obtain data associated with driving rules, and use these data as the first element.

[0189] Optionally, the step of determining the first element based on real-time vehicle data is similar to the step of extracting attribute elements and type elements from historical driving data to construct a driving rule knowledge graph, and will not be described again in this embodiment.

[0190] S3: From the first element, determine at least one target element pair, each target element pair including a first element for use as a head entity element and a first element for use as a relation element.

[0191] Here, the target element pair may include two first elements, which are considered as a selected set of head entity elements and relation elements, used as input to the graph inference network for vector inference. If the selection is accurate, the two first elements in the target element pair may indeed be the correct set of head entity elements and relation elements in the actual application.

[0192] If the filtering is accurate, the two first elements in the target element pair may be two head entity elements or two relation elements. However, in application, one of them is still used as the head entity element and the other as the tail entity element. In this case, after the network infers the representation vector, the representation vector corresponding to the tail entity element obtained using the representation vector is often inaccurate and will be identified and discarded instead of being used, thus avoiding the problem of incorrect driving decisions.

[0193] In practice, after obtaining multiple first elements, two of them can be selected as a target element pair. Furthermore, the number of target element pairs to be selected can be determined according to pre-specified selection rules. For example, if the selection rule is to select only one pair, then only one target element pair can be selected. If the selection rule requires the target element pair to cover all first elements, multiple target element pairs can be obtained through iterative combination.

[0194] S4: Using the trained graph inference network, determine the representation vectors corresponding to the two first elements in each target element pair.

[0195] In practice, the two first elements of each target element pair can be input into the trained graph inference network to obtain the representation vectors output by the graph inference network for the two first elements respectively.

[0196] S5: Make driving behavior decisions for the vehicle based on the representation vectors corresponding to the two first elements in each target element pair and the driving rule knowledge graph.

[0197] In practice, for each target element pair, link prediction can be performed using the representation vectors corresponding to the two first elements in the pair to obtain the prediction vector for that target element pair. If there are multiple target element pairs, the mean, mode, etc., of the prediction vectors corresponding to each pair can be used as the final prediction vector. Then, the prediction vector can be compared with the tail entity representation vectors corresponding to each tail entity element related to driving decision in the driving rule knowledge graph, and driving decisions can be made based on the comparison results. The representation vectors corresponding to each tail entity element in the driving rule knowledge graph can be pre-inferred using a trained graph inference network.

[0198] For example, when the comparison result indicates that there is a tail entity representation vector that is consistent with the prediction vector, the decision behavior corresponding to the tail entity representation vector is taken as the decided vehicle driving behavior; when the comparison result indicates that there is no tail entity representation vector that is consistent with the prediction vector, the decided vehicle driving behavior can be to maintain the current state.

[0199] Thus, since the trained graph inference network has reliable inference accuracy, using it to infer the target element pairs corresponding to real-time vehicle data can yield relatively accurate representation vectors. Then, based on these inferred representation vectors, behavioral decisions can be made, resulting in more accurate decision-making actions.

[0200] In one embodiment, S5 described above can be implemented according to the following steps:

[0201] S5-1: Determine the target inference vector corresponding to the target element pair based on the representation vectors corresponding to the two first elements.

[0202] In practice, for each pair of target elements, the representation vectors corresponding to the two first elements in the pair can be added together to obtain an inference vector.

[0203] If a target element pair contains only one element, then the inference vector corresponding to that target element pair can be used as the target inference vector. If there are multiple target element pairs, after obtaining the inference vectors corresponding to each target element pair, a final target inference vector is obtained by taking the average, weighting, taking the mode, and randomly selecting one of the target inference vectors.

[0204] S5-2: From the driving rule knowledge graph, select target entity elements that indicate different driving behaviors, and calculate the similarity between the representation vector corresponding to each target entity element and the target inference vector.

[0205] Here, the target entity element is an entity element in the constructed driving rule knowledge graph used to indicate driving behavior (i.e., behavioral decisions). Typically, the target entity element can be a tail entity element in the driving rule knowledge graph.

[0206] For example, target entity elements that indicate different driving behaviors can be selected from the driving rule knowledge graph. Then, the vector similarity between the representation vector and the target inference vector corresponding to each target entity element can be calculated separately.

[0207] S5-3: If there is a similarity greater than a preset threshold, the target driving behavior indicated by the target entity element corresponding to the similarity is taken as the driving behavior to be executed by the vehicle.

[0208] Here, the preset threshold can be set based on experience, and this embodiment of the disclosure does not impose specific limitations. For example, the preset threshold can be 80%, 90%, 95%, etc. The target entity element corresponding to a similarity is the entity element corresponding to the representation vector used to calculate that similarity.

[0209] For example, each calculated similarity can be compared with a preset threshold to determine if there is a similarity greater than the preset threshold. If so, the target driving behavior (i.e., target decision behavior) indicated by the target entity element corresponding to the similarity greater than the preset threshold can be used as the driving behavior to be executed by the vehicle. If multiple similarities exist that are greater than the preset threshold, the target driving behavior (i.e., target decision behavior) indicated by the target entity element corresponding to the highest similarity can be used as the driving behavior to be executed by the vehicle.

[0210] In this way, since the driving behaviors indicated by the target entity elements in the driving rule knowledge graph are the safe driving behaviors indicated in the collected historical driving data, the decision on driving behavior can be made by comparing the similarity between vectors, so that the driving decisions made meet the driving safety requirements, thereby improving the rationality and safety of driving decisions.

[0211] Conversely, if all similarities are no greater than a preset threshold, the vehicle will make a driving decision to maintain its current driving behavior.

[0212] Here, if the similarity between the inferred target inference vector and each target entity element vector is not greater than a preset threshold, it can be concluded that the inferred target inference vector cannot indicate the correct driving decision, and the driving behavior to be performed by the vehicle can be determined to be to maintain the current driving behavior.

[0213] In this way, when reasonable driving behavior cannot be predicted, maintaining the vehicle's current driving behavior can avoid the problem of driving danger caused by unreasonable driving behavior intervention, thereby improving driving safety and decision-making rationality.

[0214] In one embodiment, to improve the rationality of the determined target inference vector, the above S5-1 can also be implemented according to the following steps:

[0215] If there are multiple target element pairs, the inference vector that appears most frequently among the inference vectors corresponding to each target element pair shall be taken as the target inference vector.

[0216] Here, the inference vector corresponding to the target element pair is obtained by adding the representation vectors corresponding to the two first elements in the target element pair.

[0217] For example, if there are 10 target element pairs, the inference vectors corresponding to target element pairs 1 to 6 are all inference vector 1, and the inference vectors corresponding to target element pairs 7 to 10 are inference vectors 2 to 5 respectively, then inference vector 1 can be determined as the target inference vector.

[0218] Thus, if multiple target elements show repeated occurrences of the inference vectors obtained from the corresponding inferences, it indicates that the decision behavior indicated by the inference vector may be the most reasonable decision behavior. Therefore, by using this inference vector as the target inference vector for driving behavior decision-making, the rationality and accuracy of the final determined decision behavior can be improved.

[0219] In one embodiment, to improve the rationality of the determined driving behavior, before executing S4, a driving behavior decision can be made using a driving rule knowledge graph and each first element. If no decision can be made, a graph reasoning network can then be used for decision-making. Specifically, this can be implemented according to the following steps:

[0220] From the driving rules knowledge graph, find target tail entity elements that are associated with at least some of the first elements among multiple first elements.

[0221] In a driving rule knowledge graph, a tail entity element may be associated with multiple different sets of head entity elements and relational elements. That is, using different head entity elements and relational elements may yield the same tail entity element. Therefore, the more head entity elements and relational elements a tail entity element is associated with, the more accurate the driving behavior it indicates. Thus, after obtaining multiple first elements, tail entity elements can be searched using these first elements and various head entity elements and relational elements in the driving rule knowledge graph to determine which target tail entity elements are associated with at least some of the multiple first elements.

[0222] At least some of the elements can be determined based on a preset ratio or a preset quantity. For example, if there are 10 first elements and the preset ratio is 50%, then if it is determined that there is a tail entity element that has a correct association with at least 5 first elements, then that tail entity element can be used as the target tail entity element. Alternatively, if there are 10 first elements and the preset quantity is 4, then if it is determined that there is a tail entity element that has a correct association with at least 4 first elements, then that tail entity element can be used as the target tail entity element.

[0223] Optionally, if multiple target tail entity elements are found, the target tail entity element with the most associated first elements can be used as the final target tail entity element found.

[0224] Understandably, if multiple first elements fail to identify any target tail entity element from the driving rule knowledge graph, a graph inference network can be used for decision-making. That is, after a failed search, steps S4 and S5 can be executed to utilize the trained graph inference network to determine the representation vectors corresponding to the two first elements in each target element pair, and then the network-inferred representation vectors are used to make driving behavior decisions.

[0225] Conversely, if a target tail entity element is found in the driving rule knowledge graph, the driving behavior indicated by the target tail entity element in the driving rule knowledge graph can be directly used as the driving behavior to be executed by the vehicle.

[0226] For example, if the first element is a three-lane road with a green traffic light for straight-ahead traffic, the vehicle is currently traveling in the first lane, there are road maintenance personnel ahead of the first vehicle, the second lane on the left has lane-changing conditions and is the execution lane, and the third lane on the right does not have lane-changing conditions, if a target tail entity element indicating a left lane change is found in the driving rules knowledge graph, then the driving behavior to be performed by the vehicle can be directly determined as a left lane change.

[0227] In this way, before using the graph reasoning network, decisions can be made by querying the driving rule knowledge graph. If a matching driving behavior can be found directly, and this driving behavior is used as the driving behavior to be performed by the vehicle, the computational cost of using the network for reasoning can be saved, and the rationality and accuracy of the decision results can be guaranteed.

[0228] Optionally, if a target tail entity element is found in the driving rule knowledge graph, a secondary decision can be made using a graph inference network. The similarity between the representation vector corresponding to the target tail entity element and the target inference vector determined by the network is then calculated. If the similarity is greater than a threshold, it indicates that the driving behavior found in the driving rule knowledge graph is consistent with the decision behavior inferred by the network. The target driving behavior indicated by the target tail entity element is then taken as the driving behavior to be executed by the vehicle. In this way, using both graph lookup and network inference for driving behavior decision-making allows for mutual verification between the two methods, thereby improving the rationality and accuracy of the decided driving behavior. Understandably, if the similarity between the representation vector corresponding to the target tail entity element and the target inference vector determined by the network is not greater than a preset threshold, the target driving behavior indicated by the target tail entity element can be taken as the driving behavior to be executed by the vehicle, or the vehicle's driving behavior can be determined to be maintaining its current driving behavior.

[0229] In one embodiment, if the target tail entity element cannot be found, after inferring the representation vector using the graph inference network and determining the target inference vector using the representation vector, if the decided target driving behavior is not the vehicle maintaining its current driving behavior, the existing driving rule knowledge graph can be supplemented using the two first elements corresponding to the target inference vector and the decided driving behavior. Specifically, the triples in the driving rule knowledge graph can be supplemented based on the target element pairs corresponding to the target driving behavior.

[0230] Here, the target driving behavior is the driving behavior indicated by the matching entity element, and the similarity between the representation vector of the matching entity element and the target inference vector is greater than a preset threshold. The target element pair corresponding to the target driving behavior is the matching target element pair corresponding to the two target representation vectors used to determine the target inference vector. The target representation vector is the representation vector corresponding to the two first elements in the matching target element pair. For example, when the target inference vector is determined based on the first element 1 and the first element 2 in target element pair 1, then the target element pair corresponding to the target driving behavior is target element pair 1.

[0231] In practice, after obtaining the target driving behavior, the two first elements in the target element pair corresponding to the target driving behavior can be used as the head entity element and the relation element, respectively, and the matching entity element can be used as the tail entity element to obtain a new triple. This new triple is then added to the driving rule knowledge graph to update the driving rule knowledge graph.

[0232] In this way, after the target driving behavior is inferred from the network, the driving rule knowledge graph is updated using the target element pair corresponding to the target driving behavior. This allows the target driving behavior to be directly retrieved from the driving rule knowledge graph the next time a target element pair is obtained, without having to use the network for inference again, thereby improving decision-making efficiency.

[0233] In one embodiment, to improve the accuracy of the selected target element pairs, the above-mentioned S3 can be implemented according to the following steps:

[0234] Based on the data source of each first element, at least one target element pair is selected; wherein, the data source includes vehicle parameter source and road environment source, and the two first elements in the target element pair correspond to the vehicle parameter source and the road environment source, respectively.

[0235] Here, the data source of the first element indicates the source from which the data was collected. Specifically, the data source can include vehicle parameter sources and road environment sources; among them, the vehicle parameter source indicates that the data originates from the vehicle's own driving parameters. Typically, the data source of the first element corresponding to vehicle driving parameter data can be the vehicle parameter source. For example, the data source of the first element corresponding to vehicle speed, acceleration, torque, horsepower, orientation, battery / fuel level, RPM, and various instrument parameters is the vehicle parameter source.

[0236] The "road environment source" parameter is used to characterize that the data originates from the road environment. Typically, the data source of the first element corresponding to the road environment data can be considered the road environment source. For example, the data source of the first element corresponding to data on static / dynamic obstacles, lanes, lane lines, traffic signs, and driving scenarios is the road environment source.

[0237] In practice, after obtaining multiple first elements, the data source of each first element can be determined. Then, a target first element can be selected from the first elements whose data source is vehicle parameters, and another target first element can be selected from the first elements whose data source is road environment. These two target first elements are considered as a pair of target elements. In this way, at least one pair of target elements can be selected based on the data source.

[0238] In this way, since the first element under the vehicle parameter source and the first element under the road environment source can usually be used as a head entity element and a relation element, the probability of selecting target element pairs that include two head entity elements or two relation elements can be reduced by using data sources to filter target element pairs, thereby improving the rationality of the selected target element pairs.

[0239] In another embodiment, S3 can also be implemented according to P1 to P3 as follows:

[0240] P1: From each first element, select the second element related to the type and the third element related to the attribute.

[0241] Here, the second element is the type data related to the data associated with driving rules in the first element, and the third element is the attribute data related to the data associated with driving rules in the first element.

[0242] For example, since the first element is obtained based on the classification of real-time vehicle data by type and attribute, the first element can include attribute elements and type elements. Therefore, a second element belonging to type elements and a third element belonging to attribute elements can be distinguished from the first element.

[0243] P2: Determine the first priority of each second element based on its type, and determine the second priority of each third element based on its attribute.

[0244] In practical implementation, different priorities can be pre-set for various types. For example, dynamic obstacles have the highest priority, followed by static obstacles, then road elements, and finally driving scenarios. Similarly, different priorities can be pre-set for various attributes. For instance, the positional relationship between the intelligent vehicle and obstacles has the highest priority, followed by the topological relationship of the road, and the numerical values ​​between road components have the lowest priority. Alternatively, the priorities of various types and attributes can be determined based on their impact on driving safety. For example, types and attributes with the highest impact on driving safety have the highest priority; those with a moderate impact have the next highest priority; and those with the lowest impact have the lowest priority.

[0245] After determining the second and third elements, the first priority of each second element can be determined based on its type and the priority of each type; at the same time, the second priority of each third element can be determined based on its attributes and the priority of each attribute.

[0246] P3: Using the first priority and the second priority, determine at least one target element pair from the second and third elements.

[0247] In practice, the highest priority among the first priorities of each second element can be selected, and this second element with the highest first priority can be used as a target second element. Similarly, the highest second priority among the second priorities of each third element can be selected, and this third element with the highest second priority can be used as a target third element. Then, the target second element and the target third element can be considered as a single target element pair.

[0248] Optionally, if multiple target element pairs need to be selected, they can be selected in descending order of priority, with one target second element selected from the second element and one target third element selected from the third element each time, to obtain a target element pair.

[0249] Thus, different priorities have different degrees of impact on driving safety. By using the priorities corresponding to the second and third elements respectively to select target element pairs, the rationality of the selected target element pairs can be improved.

[0250] Based on the above embodiments, this disclosure proposes a cognitive reasoning-based intelligent connected vehicle driving behavior decision-making method. First, driving rules are extracted and a driving rule knowledge graph is established, specifically by storing the extracted and summarized driving rules in the driving rule knowledge graph. Then, a knowledge-based graph reasoning network is designed. This graph reasoning network improves the TransE graph reasoning algorithm to reason out representation vectors corresponding to entity elements and relation elements. Then, link prediction reasoning is performed using the reasoned representation vectors. Once the link prediction is successful, a target tail entity element can be selected from the tail entity elements of the behavior decision class, and the driving behavior indicated by the target tail entity element is used to make intelligent driving decisions for the current driving scenario. The graph reasoning algorithm of this disclosure has advantages in terms of memory consumption for network construction and time consumption for matching reasoning, solving the shortcomings of traditional decision-making algorithms and improving the efficiency of intelligent connected vehicle driving behavior decision-making.

[0251] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0252] Reference Figure 7The diagram shown is a schematic representation of a vehicle behavior decision-making device according to an embodiment of this disclosure. This device can be deployed with an in-vehicle server and may include: a memory 710, a transceiver 720, and a processor 730.

[0253] The memory 710 is used to store computer programs; the transceiver 720 is used to receive and send data under the control of the processor 730.

[0254] Among them, Figure 7 In this context, the bus architecture can include any number of interconnected buses and bridges, specifically linking various circuits together, represented by one or more processors (processor 730) and memory (memory 710). The bus architecture can also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. The bus interface provides an interface. The transceiver 720 can be multiple elements, including transmitters and receivers, providing a unit for communicating with various other devices over transmission media, including wireless channels, wired channels, optical fibers, etc. The processor 730 is responsible for managing the bus architecture and general processing, and the memory 710 can store data used by the processor 730 during operation.

[0255] The processor 730 can be a central processing unit (CPU), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or a complex programmable logic device (CPLD). The processor can also adopt a multi-core architecture.

[0256] The processor 730 invokes a computer program stored in the memory 710 to execute the steps of any of the methods provided in the embodiments of this disclosure according to the obtained executable instructions, for example:

[0257] Obtain the constructed driving rule knowledge graph; the driving rule knowledge graph includes entity elements and relation elements, the entity elements are related to the road environment and / or vehicles, and the relation elements are used to represent the association between different entity elements;

[0258] Based on the driving rule knowledge graph, multiple sample triples are determined; each sample triple includes a head entity element, a tail entity element, and a relation element; the multiple sample triples include positive sample triples and negative sample triples.

[0259] The multiple sample triples are grouped to obtain multiple sample groups, wherein each sample group includes at least two positive and negative sample pairs, and there are some common elements between the positive sample triples and the negative sample triples in each positive and negative sample pair.

[0260] Each of the multiple sample groups is used as a sample unit to train the graph inference network, and the graph inference network is trained to obtain a trained graph inference network. The graph inference network is used to infer the representation vector of the element, and to make driving decisions using the representation vector and the driving rule knowledge graph.

[0261] Optionally, the step of using each of the plurality of sample groups as a sample unit to train the graph inference network to obtain a trained graph inference network includes:

[0262] Each of the multiple sample groups is used as a sample unit for training the graph inference network. The inference loss information under each sample unit is calculated, and the network parameters of the graph inference network are adjusted based on the inference loss information under multiple sample units to obtain the trained graph inference network. The inference loss information is determined based on the distance between the sum of the representation vectors of the head entity element and the relation element and the representation vector of the tail entity element.

[0263] Optionally, the step of calculating the inference loss information for each sample unit and adjusting the network parameters of the graph inference network based on the inference loss information for multiple sample units to obtain the trained graph inference network includes:

[0264] The predicted representation vector of each element under the sample unit is determined based on the graph inference network to be trained.

[0265] For any positive and negative sample pair in the sample unit, a first vector distance is determined based on the sum of the predicted representation vectors of the head entity element and the relation element of the positive sample triplet in the positive sample pair, and the predicted representation vector corresponding to the tail entity element of the positive sample triplet; and a second vector distance is determined based on the sum of the predicted representation vectors of the head entity element and the relation element of the negative sample triplet in the positive and negative sample pair, and the predicted representation vector corresponding to the tail entity element of the negative sample triplet.

[0266] Based on the difference between the first vector distance and the second vector distance corresponding to each pair of positive and negative samples, the inference loss information corresponding to the sample unit is determined;

[0267] With the goal of minimizing the sum of the inference loss information corresponding to each sample unit and the preset positive and negative sample interval values, the network parameters of the graph inference network are adjusted to obtain the trained graph inference network.

[0268] Optionally, the method further includes:

[0269] Acquire real-time vehicle data collected during vehicle operation.

[0270] Data associated with driving rules is extracted from the real-time vehicle data to obtain multiple first elements; the data associated with driving rules includes type and attribute; the type is used to characterize the object information involved in the driving rule, and the attribute is used to describe the relationship between types and / or the value of the type;

[0271] From the first element, at least one target element pair is determined, each target element pair including a first element for use as a head entity element and a first element for use as a relation element;

[0272] Using the trained graph inference network, determine the representation vectors corresponding to the two first elements in each pair of target elements;

[0273] Based on the representation vectors corresponding to the two first elements in each target element pair and the driving rule knowledge graph, driving behavior decisions are made for the vehicle.

[0274] Optionally, the step of making driving behavior decisions for the vehicle based on the representation vectors corresponding to the two first elements in each target element pair and the driving rule knowledge graph includes:

[0275] Based on the representation vectors corresponding to the two first elements, determine the target inference vector corresponding to the target element pair;

[0276] From the driving rule knowledge graph, target entity elements for indicating different driving behaviors are selected, and the similarity between the representation vector corresponding to each target entity element and the target inference vector is calculated.

[0277] If the similarity is greater than a preset threshold, the target driving behavior indicated by the target entity element corresponding to that similarity is taken as the driving behavior to be performed by the vehicle.

[0278] Optionally, based on the representation vectors corresponding to the two first elements, the target inference vector corresponding to the target element pair is determined, including:

[0279] If there are multiple target element pairs, the inference vector that appears most frequently among the inference vectors corresponding to each target element pair shall be taken as the target inference vector.

[0280] Optionally, after calculating the similarity between the representation vector corresponding to each target entity element and the target inference vector, the method further includes:

[0281] If none of the aforementioned similarities are greater than the preset threshold, a driving decision is made to maintain the current driving behavior for the vehicle.

[0282] Optionally, the method further includes:

[0283] Based on the target element pairs corresponding to the target driving behavior, the triples in the driving rule knowledge graph are supplemented.

[0284] Optionally, determining at least one target element pair includes:

[0285] Based on the data source of each of the first elements, at least one of the target element pairs is selected; wherein the data source includes vehicle parameter source and road environment source, and the two first elements in the target element pair correspond to the vehicle parameter source and the road environment source, respectively.

[0286] Optionally, determining at least one target element pair includes:

[0287] From each of the first elements, select the second element related to the type and the third element related to the attribute;

[0288] Based on the type of each second element, determine the first priority of each second element; and based on the attribute of each third element, determine the second priority of each third element.

[0289] Using the first priority and the second priority, at least one target element pair is determined from the second element and the third element.

[0290] Optionally, before determining the representation vectors corresponding to the two first elements in the target element pair using the trained graph inference network, the method further includes:

[0291] From the driving rule knowledge graph, find target tail entity elements that are associated with at least some of the first elements among a plurality of first elements;

[0292] Using the trained graph inference network, the representation vectors corresponding to the two first elements in the target element pair are determined, including:

[0293] After the search fails, the trained graph inference network is used to determine the representation vectors corresponding to the two first elements in the target element pair.

[0294] Optionally, the method further includes:

[0295] If the target tail entity element is found, the driving behavior indicated by the target tail entity element in the driving rule knowledge graph is taken as the driving behavior to be executed by the vehicle.

[0296] Based on the same inventive concept, this disclosure also provides a vehicle behavior decision-making device corresponding to the vehicle behavior decision-making method. Since the principle of the device in this disclosure for solving the problem is similar to the above-mentioned vehicle behavior decision-making method in this disclosure, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.

[0297] like Figure 8 The diagram shown is a schematic representation of a vehicle behavior decision-making device provided in an embodiment of this disclosure, comprising:

[0298] The graph acquisition module 801 is used to acquire the constructed driving rule knowledge graph; the driving rule knowledge graph includes entity elements and relation elements, the entity elements are related to the road environment and / or vehicles, and the relation elements are used to represent the association between different entity elements;

[0299] The sample determination module 802 is used to determine multiple sample triples based on the driving rule knowledge graph; each sample triple includes a head entity element, a tail entity element, and a relation element; the multiple sample triples include positive sample triples and negative sample triples;

[0300] The sample grouping module 803 is used to group the plurality of sample triples to obtain a plurality of sample groups, wherein each sample group includes at least two positive and negative sample pairs, and there are some common elements between the positive sample triples and the negative sample triples in each positive and negative sample pair.

[0301] The training inference module 804 is used to train the graph inference network by using each of the multiple sample groups as a sample unit for training the graph inference network, thereby obtaining a trained graph inference network. The graph inference network is used to infer the representation vector of the element and to make driving decisions using the representation vector and the driving rule knowledge graph.

[0302] Optionally, the training inference module 804, when training the graph inference network using each of the plurality of sample groups as a sample unit to obtain the trained graph inference network, is used to:

[0303] Each of the multiple sample groups is used as a sample unit for training the graph inference network. The inference loss information under each sample unit is calculated, and the network parameters of the graph inference network are adjusted based on the inference loss information under multiple sample units to obtain the trained graph inference network. The inference loss information is determined based on the distance between the sum of the representation vectors of the head entity element and the relation element and the representation vector of the tail entity element.

[0304] Optionally, the training inference module 804, when calculating the inference loss information for each sample unit and adjusting the network parameters of the graph inference network based on the inference loss information for multiple sample units to obtain the trained graph inference network, is used to:

[0305] The predicted representation vector of each element under the sample unit is determined based on the graph inference network to be trained.

[0306] For any positive and negative sample pair in the sample unit, a first vector distance is determined based on the sum of the predicted representation vectors of the head entity element and the relation element of the positive sample triplet in the positive sample pair, and the predicted representation vector corresponding to the tail entity element of the positive sample triplet; and a second vector distance is determined based on the sum of the predicted representation vectors of the head entity element and the relation element of the negative sample triplet in the positive and negative sample pair, and the predicted representation vector corresponding to the tail entity element of the negative sample triplet.

[0307] Based on the difference between the first vector distance and the second vector distance corresponding to each pair of positive and negative samples, the inference loss information corresponding to the sample unit is determined;

[0308] With the goal of minimizing the sum of the inference loss information corresponding to each sample unit and the preset positive and negative sample interval values, the network parameters of the graph inference network are adjusted to obtain the trained graph inference network.

[0309] Optionally, the device further includes:

[0310] Behavioral decision module 805 is used for:

[0311] Acquire real-time vehicle data collected during vehicle operation.

[0312] Data associated with driving rules is extracted from the real-time vehicle data to obtain multiple first elements; the data associated with driving rules includes type and attribute; the type is used to characterize the object information involved in the driving rule, and the attribute is used to describe the relationship between types and / or the value of the type;

[0313] From the first element, at least one target element pair is determined, each target element pair including a first element for use as a head entity element and a first element for use as a relation element;

[0314] Using the trained graph inference network, determine the representation vectors corresponding to the two first elements in each pair of target elements;

[0315] Based on the representation vectors corresponding to the two first elements in each target element pair and the driving rule knowledge graph, driving behavior decisions are made for the vehicle.

[0316] Optionally, the behavior decision module 805, when making driving behavior decisions for the vehicle based on the representation vectors corresponding to the two first elements in each target element pair and the driving rule knowledge graph, is used to:

[0317] Based on the representation vectors corresponding to the two first elements, determine the target inference vector corresponding to the target element pair;

[0318] From the driving rule knowledge graph, target entity elements for indicating different driving behaviors are selected, and the similarity between the representation vector corresponding to each target entity element and the target inference vector is calculated.

[0319] If the similarity is greater than a preset threshold, the target driving behavior indicated by the target entity element corresponding to that similarity is taken as the driving behavior to be performed by the vehicle.

[0320] Optionally, the behavior decision module 805, when determining the target inference vector corresponding to the target element pair based on the representation vectors corresponding to the two first elements, is used to:

[0321] If there are multiple target element pairs, the inference vector that appears most frequently among the inference vectors corresponding to each target element pair shall be taken as the target inference vector.

[0322] Optionally, after calculating the similarity between the representation vector corresponding to each target entity element and the target inference vector, the behavior decision module 805 is further configured to:

[0323] If none of the aforementioned similarities are greater than the preset threshold, a driving decision is made to maintain the current driving behavior for the vehicle.

[0324] Optionally, the device further includes:

[0325] Module 806 for map completion is used for:

[0326] Based on the target element pairs corresponding to the target driving behavior, the triples in the driving rule knowledge graph are supplemented.

[0327] Optionally, the behavior decision module 805, when determining at least one target element pair, is used to:

[0328] Based on the data source of each of the first elements, at least one of the target element pairs is selected; wherein the data source includes vehicle parameter source and road environment source, and the two first elements in the target element pair correspond to the vehicle parameter source and the road environment source, respectively.

[0329] Optionally, the behavior decision module 805, when determining at least one target element pair, is used to:

[0330] From each of the first elements, select the second element related to the type and the third element related to the attribute;

[0331] Based on the type of each second element, determine the first priority of each second element; and based on the attribute of each third element, determine the second priority of each third element.

[0332] Using the first priority and the second priority, at least one target element pair is determined from the second element and the third element.

[0333] Optionally, before determining the representation vectors corresponding to the two first elements in the target element pair using the trained graph inference network, the behavior decision module 805 is further configured to:

[0334] From the driving rule knowledge graph, find target tail entity elements that are associated with at least some of the first elements among a plurality of first elements;

[0335] When the behavior decision module 805 uses the trained graph inference network to determine the representation vectors corresponding to the two first elements in the target element pair, it is used for:

[0336] After the search fails, the trained graph inference network is used to determine the representation vectors corresponding to the two first elements in the target element pair.

[0337] Optionally, the behavior decision module 805 is further configured to:

[0338] If the target tail entity element is found, the driving behavior indicated by the target tail entity element in the driving rule knowledge graph is taken as the driving behavior to be executed by the vehicle.

[0339] It should be noted that the division of units in the embodiments of this disclosure is illustrative and only represents one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units.

[0340] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a processor-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0341] It should be noted that the apparatus provided in this embodiment can implement all the method steps implemented in the above method embodiment and can achieve the same technical effect. Therefore, the parts and beneficial effects that are the same as those in the method embodiment will not be described in detail here.

[0342] On the other hand, embodiments of this disclosure also provide a processor-readable storage medium storing a program for causing a processor to execute the vehicle behavior decision-making methods provided in the above embodiments.

[0343] It should be noted that the processor-readable storage medium provided in this embodiment can implement all the method steps implemented in the above method embodiments and achieve the same technical effect. Therefore, the parts and beneficial effects that are the same as those in the method embodiments will not be described in detail here.

[0344] Processor-readable storage media can be any available medium or data storage device that the processor can access, including but not limited to magnetic storage (e.g., floppy disks, hard disks, magnetic tapes, magneto-optical disks (MOs), etc.), optical storage (e.g., CDs, DVDs, BDs, HVDs, etc.), and semiconductor storage (e.g., ROMs, EPROMs, EEPROMs, non-volatile memory (NAND flash), solid-state drives (SSDs)).

[0345] This disclosure also provides a computer program product that, when invoked by a computer, causes the computer to execute the steps of the vehicle behavior decision-making method described above.

[0346] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0347] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-executable instructions. These computer-executable instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0348] These processor-executable instructions may also be stored in a processor-readable memory that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the processor-readable memory produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0349] These processors can execute instructions that can also be loaded onto a computer or other programmable data processing device, causing a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable device for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0350] Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from its spirit and scope. Therefore, if such modifications and variations fall within the scope of the claims of this disclosure and their equivalents, this disclosure is also intended to include such modifications and variations.

Claims

1. A vehicle behavior decision-making method, characterized in that, include: Obtain the constructed driving rule knowledge graph; The driving rule knowledge graph includes entity elements and relational elements. The entity elements are related to the road environment and / or vehicles, and the relational elements are used to represent the association between different entity elements. Based on the driving rule knowledge graph, multiple sample triples are determined; each sample triple includes a head entity element, a tail entity element, and a relation element. The plurality of sample triplets includes positive sample triplets and negative sample triplets; The multiple sample triples are grouped to obtain multiple sample groups, wherein each sample group includes at least two positive and negative sample pairs, and there are some common elements between the positive sample triples and the negative sample triples in each positive and negative sample pair. Each of the multiple sample groups is used as a sample unit to train the graph inference network, and the graph inference network is trained to obtain a trained graph inference network. The graph reasoning network is used to reason and obtain the representation vector of the element, and to make driving decisions using the representation vector and the driving rule knowledge graph.

2. The method according to claim 1, characterized in that, The step of using each of the plurality of sample groups as a sample unit to train the graph inference network, and obtaining a trained graph inference network, includes: Each of the multiple sample groups is used as a sample unit for training the graph inference network. The inference loss information under each sample unit is calculated, and the network parameters of the graph inference network are adjusted based on the inference loss information under multiple sample units to obtain the trained graph inference network. The inference loss information is determined based on the distance between the sum of the representation vectors of the head entity element and the relation element and the representation vector of the tail entity element.

3. The method according to claim 2, characterized in that, The process of calculating the inference loss information for each sample unit and adjusting the network parameters of the graph inference network based on the inference loss information for multiple sample units to obtain the trained graph inference network includes: The predicted representation vector of each element under the sample unit is determined based on the graph inference network to be trained. For any positive and negative sample pair in the sample unit, a first vector distance is determined based on the sum of the predicted representation vectors of the head entity element and the relation element of the positive sample triplet in the positive sample pair, and the predicted representation vector corresponding to the tail entity element of the positive sample triplet; and a second vector distance is determined based on the sum of the predicted representation vectors of the head entity element and the relation element of the negative sample triplet in the positive and negative sample pair, and the predicted representation vector corresponding to the tail entity element of the negative sample triplet. Based on the difference between the first vector distance and the second vector distance corresponding to each pair of positive and negative samples, the inference loss information corresponding to the sample unit is determined; With the goal of minimizing the sum of the inference loss information corresponding to each sample unit and the preset positive and negative sample interval values, the network parameters of the graph inference network are adjusted to obtain the trained graph inference network.

4. The method according to claim 1, characterized in that, The method further includes: Acquire real-time vehicle data collected during vehicle operation. Data associated with driving rules is extracted from the real-time vehicle data to obtain multiple first elements; the data associated with driving rules includes type and attribute; the type is used to characterize the object information involved in the driving rule, and the attribute is used to describe the relationship between types and / or the value of the type; From the first element, at least one target element pair is determined, each target element pair including a first element for use as a head entity element and a first element for use as a relation element; Using the trained graph inference network, determine the representation vectors corresponding to the two first elements in each pair of target elements; Based on the representation vectors corresponding to the two first elements in each target element pair and the driving rule knowledge graph, driving behavior decisions are made for the vehicle.

5. The method according to claim 4, characterized in that, The step of making driving behavior decisions for the vehicle based on the representation vectors corresponding to the two first elements in each target element pair and the driving rule knowledge graph includes: Based on the representation vectors corresponding to the two first elements, determine the target inference vector corresponding to the target element pair; From the driving rule knowledge graph, target entity elements for indicating different driving behaviors are selected, and the similarity between the representation vector corresponding to each target entity element and the target inference vector is calculated. If the similarity is greater than a preset threshold, the target driving behavior indicated by the target entity element corresponding to that similarity is taken as the driving behavior to be performed by the vehicle.

6. The method according to claim 5, characterized in that, Based on the representation vectors corresponding to the two first elements, the target inference vector corresponding to the target element pair is determined, including: If there are multiple target element pairs, the inference vector that appears most frequently among the inference vectors corresponding to each target element pair shall be taken as the target inference vector.

7. The method according to claim 5, characterized in that, After calculating the similarity between the representation vector corresponding to each target entity element and the target inference vector, the method further includes: If none of the aforementioned similarities are greater than the preset threshold, a driving decision is made to maintain the current driving behavior for the vehicle.

8. The method according to claim 5, characterized in that, The method further includes: Based on the target element pairs corresponding to the target driving behavior, the triples in the driving rule knowledge graph are supplemented.

9. The method according to claim 4, characterized in that, Determining at least one target element pair includes: Based on the data source of each of the first elements, at least one of the target element pairs is selected; wherein the data source includes vehicle parameter source and road environment source, and the two first elements in the target element pair correspond to the vehicle parameter source and the road environment source, respectively.

10. The method according to claim 4, characterized in that, Determining at least one target element pair includes: From each of the first elements, select the second element related to the type and the third element related to the attribute; Based on the type of each second element, determine the first priority of each second element; and based on the attribute of each third element, determine the second priority of each third element. Using the first priority and the second priority, at least one target element pair is determined from the second element and the third element.

11. The method according to claim 4, characterized in that, Before using the trained graph inference network to determine the representation vectors corresponding to the two first elements in the target element pair, the process further includes: From the driving rule knowledge graph, find target tail entity elements that are associated with at least some of the first elements among a plurality of first elements; Using the trained graph inference network, the representation vectors corresponding to the two first elements in the target element pair are determined, including: After the search fails, the trained graph inference network is used to determine the representation vectors corresponding to the two first elements in the target element pair.

12. The method according to claim 11, characterized in that, The method further includes: If the target tail entity element is found, the driving behavior indicated by the target tail entity element in the driving rule knowledge graph is taken as the driving behavior to be executed by the vehicle.

13. A vehicle behavior decision-making device, characterized in that, Includes memory, transceiver, and processor; The memory is used to store a computer program; the transceiver is used to send and receive data under the control of the processor; the processor is used to read the computer program in the memory and execute the steps of the vehicle behavior decision method as described in any one of claims 1 to 12.

14. A vehicle behavior decision-making device, characterized in that, include: The graph acquisition module is used to acquire the constructed driving rule knowledge graph; The driving rule knowledge graph includes entity elements and relational elements. The entity elements are related to the road environment and / or vehicles, and the relational elements are used to represent the association between different entity elements. The sample determination module is used to determine multiple sample triples based on the driving rule knowledge graph; each sample triple includes a head entity element, a tail entity element, and a relation element; The plurality of sample triplets includes positive sample triplets and negative sample triplets; The sample grouping module is used to group the multiple sample triples to obtain multiple sample groups, wherein each sample group includes at least two positive and negative sample pairs, and there are some common elements between the positive sample triples and the negative sample triples in each positive and negative sample pair. The training inference module is used to train the graph inference network by using each of the multiple sample groups as a sample unit for training the graph inference network, and to obtain the trained graph inference network. The graph reasoning network is used to reason and obtain the representation vector of the element, and to make driving decisions using the representation vector and the driving rule knowledge graph.

15. A processor-readable storage medium, characterized in that, The processor-readable storage medium stores a program for causing the processor to execute the vehicle behavior decision method as described in any one of claims 1 to 12.

16. A computer program product, characterized in that, When the computer program product is invoked by a computer, it causes the computer to execute the vehicle behavior decision-making method as described in any one of claims 1 to 12.

Images