Automatic driving knowledge graph scene node construction method and system

Abstract

The invention discloses an automatic driving knowledge graph scene node construction method and system, and belongs to the technical field of automatic driving. The method comprises the steps of dividing an observation area for scene information representation, defining vehicle key scene parameters in the area, and constructing scene nodes based on a regional danger level (TTC). According to the method, the richness of scene information and the number of scenes are considered through efficient scene expression, the efficiency of automatic driving scene data knowledge summarization and retrieval is improved, and the auxiliary effect of driving experience knowledge in the LLMs decision making process is enhanced; the prediction and risk assessment capability of the system on diversified driving scenes is improved, and a solid foundation is laid for realizing safer, more efficient and more intelligent automatic driving.

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous driving, and particularly relates to a method and system for constructing scene nodes of an autonomous driving knowledge graph. Background Art

[0002] Large language models (LLMs) possess powerful information processing and natural language understanding capabilities, providing important support for the analysis and decision-making of complex tasks. However, directly applying LLMs to autonomous driving systems has significant limitations, especially the impact of the hallucination problem. Due to the probabilistic nature of LLMs, the opacity of the operation mechanism, and the fuzzy boundaries, the model may generate decisions that do not conform to reality. Without the assistance of professional driving knowledge, the hallucination problem will further amplify safety hazards and affect the reliability and controllability of the system.

[0003] In autonomous driving, driving experience knowledge is the key to alleviating the hallucination problem, and this knowledge often needs to be obtained through the summary, extraction, and retrieval of specific scenarios. However, complex scene representation methods and a large number of scenes have greatly increased the difficulty of collecting, storing, and efficiently retrieving the set of driving scenes, restricting the effective utilization of knowledge. For example, OpenSCENARIO, as an internationally authoritative standard for autonomous driving simulation scenarios, its scene description involves external states such as roads, traffic facilities, weather conditions, and traffic participants, as well as information such as the driving tasks and states of autonomous vehicles. It divides the scene model into six layers and further divides the scenes into three abstract levels: functional scenarios, logical scenarios, and specific scenarios. Although this multi-level and multi-level scene description method can comprehensively capture key driving information, it also brings the problem of an explosion in the number of scene nodes, resulting in a significant increase in the complexity of scene representation. In addition, Scenic, as Google's autonomous driving scene generation language, its declarative syntax and randomness support make it easier to write test scenes, but at the same time increase the complexity of scene representation. Scenic allows the introduction of random factors in the scene, such as the speed, position, and orientation of vehicles, which helps to generate diverse test cases and cover more edge cases, thus better evaluating the robustness of autonomous driving systems. However, this requirement for randomness and diversity also means that the scene representation method needs to handle more parameters and variables, further increasing the complexity of scene representation.

[0004] It can be seen that with the development of technology, driving scenes have become increasingly complex, and the number of scene nodes has shown an explosive growth. This growth not only increases the difficulty of scene matching but also causes many difficulties in the collection, storage, and retrieval of scene nodes.

[0005] First, the traffic conditions that autonomous driving systems need to handle are increasing day by day. Each scene node contains a large amount of information, such as vehicle position, speed, pedestrian behavior, etc. As the number of scene nodes increases, the resources and time required to collect this data also increase, and it becomes extremely difficult to collect scene nodes accurately and comprehensively. This poses a huge challenge to autonomous driving systems with extremely high requirements for real-time performance and accuracy. Second, a huge number of scene nodes require a large amount of storage space, and it is necessary to ensure that the storage structure is convenient for subsequent query and use. However, the capacity and cost of storage devices are limited. Facing the explosion in the number of scene nodes, existing storage methods face the dual challenges of capacity and structure optimization. How to effectively store and manage massive data with limited resources has become an urgent problem to be solved. In addition, in an autonomous driving system, quickly and accurately retrieving relevant scene nodes is crucial for the system to make correct decisions. However, with the huge number of scene nodes, how to quickly and accurately retrieve the required scene nodes has become a major problem. Existing retrieval algorithms will have greatly reduced efficiency and accuracy when dealing with such a large number of scene nodes, which seriously affects the timely response and correct decision-making ability of autonomous driving systems to driving scenarios.

[0006] Therefore, it is crucial to design a simple and efficient method for representing autonomous driving scenarios. An ideal scenario representation method should achieve a balance between the number of scenarios and information expression, being able to comprehensively capture key driving information while reducing the complexity of matching and retrieval. Summary of the Invention

[0007] The present invention discloses a method and system for constructing scene nodes of an autonomous driving knowledge graph. By means of efficient scene expression, it takes into account the richness of scene information and the number of scenarios, improves the efficiency of knowledge summary and retrieval of autonomous driving scene data, enhances the auxiliary role of driving experience knowledge in the decision-making process of LLMs, and improves the system's prediction and risk assessment capabilities for diverse driving scenarios, laying a solid foundation for more secure, efficient, and intelligent autonomous driving.

[0008] To achieve the above object, the technical solution of the present invention includes the following contents.

[0009] A method for constructing scene nodes of an autonomous driving knowledge graph, the method comprising:

[0010] Defining the observation area of the host vehicle in the current scene and dividing the observation area into several sub-areas;

[0011] Extracting key scene parameters within the observation area;

[0012] Based on the key scenario parameters, calculate the regional hazard level of each sub-region and the overall risk of the scenario node, and use the regional hazard level and the overall risk value as the attributes of the scenario node in the knowledge graph.

[0013] Further, the sub-regions include: the left adjacent lane sub-region, the rear sub-region, the front sub-region, and the right adjacent lane sub-region.

[0014] Further, the key scenario parameters include: road information and the positions, lane IDs, speeds, and accelerations of the background vehicles and the host vehicle.

[0015] Further, calculating the regional hazard level of each sub-region based on the key scenario parameters includes:

[0016] According to the positions of the background vehicles and the host vehicle, obtain the distance between each background vehicle and the host vehicle;

[0017] Based on the distance between each background vehicle and the host vehicle, the speed of the background vehicle, and the speed of the host vehicle, obtain the TTC value between the background vehicle and the host vehicle;

[0018] Based on the road information and the lane IDs of the background vehicles, obtain the distribution of the background vehicles in each sub-region;

[0019] When there is one background vehicle in a sub-region, the TTC value of the sub-region is the TTC value between the background vehicle and the host vehicle;

[0020] When there is no background vehicle in a sub-region, the TTC value of the sub-region is a set value;

[0021] When there are multiple background vehicles in a sub-region, the TTC value of the sub-region is the minimum TTC value between the host vehicle and the background vehicles;

[0022] Based on the set value and a set threshold, discretize the TTC values of each sub-region to obtain the regional hazard level of each sub-region.

[0023] Further, the process of calculating the overall risk value of the scenario node includes:

[0024] Obtain the weights of each sub-region;

[0025] Based on the weights of the sub-regions, weight the regional hazard levels of each sub-region to obtain the overall risk value of the scenario node.

[0026] A method for constructing an autonomous driving knowledge graph, the method includes:

[0027] Define the observation area of the host vehicle in each scenario, and divide the observation area into several sub-areas;

[0028] Construct a scenario node, and the attributes of the scenario node are generated based on the method for constructing an autonomous driving knowledge graph scenario node according to any one of claims 1 to 5;

[0029] Generate entity nodes based on the sub-areas in different scenarios;

[0030] After connecting the entity node with the corresponding scenario node, an autonomous driving knowledge graph is obtained.

[0031] Furthermore, store the autonomous driving knowledge graph in a graph database.

[0032] An autonomous driving knowledge graph scenario node construction system, the system includes:

[0033] An area definition module, configured to define the observation area of the host vehicle in the current scenario, and divide the observation area into several sub-areas;

[0034] A parameter extraction module, configured to extract key scenario parameters in the observation area;

[0035] A node construction module, configured to calculate the regional hazard level of each sub-area and the overall risk value of the scenario node based on the key scenario parameters, and use the regional hazard level and the overall risk value as the attributes of the scenario node in the knowledge graph.

[0036] An autonomous driving knowledge graph construction system, the system includes:

[0037] An area definition module, which defines the observation area of the host vehicle in each scenario, and divides the observation area into several sub-areas;

[0038] A scenario node construction module, configured to construct a scenario node, and the attributes of the scenario node are generated based on the method for constructing an autonomous driving knowledge graph scenario node according to any one of claims 1 to 5;

[0039] An entity node construction module, configured to generate entity nodes based on the sub-areas in different scenarios;

[0040] A knowledge graph generation module, configured to obtain an autonomous driving knowledge graph after connecting the entity node with the corresponding scenario node.

[0041] An electronic device, characterized in that the electronic device includes: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the method for constructing an autonomous driving knowledge graph scenario node according to any one of the above or the method for constructing an autonomous driving knowledge graph construction system according to any one of the above is implemented.

[0042] Compared with the prior art, the present invention has at least the following beneficial effects.

[0043] The present invention proposes an innovative method for constructing scene nodes of an autonomous driving knowledge graph, aiming to improve the safety and efficiency of autonomous driving systems. The effectiveness of the present invention has been verified through a series of comprehensive evaluation methods, ensuring its reliability and practicality in actual applications.

[0044] Compared with scene representation methods such as OpenSCENARIO and Scenic, the present invention innovatively introduces the division of the observation area and the hierarchical representation of the time to collision (TTC), significantly reducing the complexity of the scene and the number of nodes. In the present invention, the driving scene is divided into four sub-areas: front, rear, left, and right, and only the background vehicles that may pose a threat to the host vehicle in each area are concerned, thus effectively reducing the interference of irrelevant information. In addition, by calculating and discretizing the TTC values of each sub-area (divided into fixed threat levels: -1, 0, 1, 2, 3, 4), the complexity of scene representation is further reduced, avoiding the risk of explosion in the number of nodes in traditional methods. In contrast, OpenSCENARIO uses structured languages such as XML to define scenes. Although it can describe the behaviors and events of traffic participants in detail, as the complexity of the scene increases, the number of objects and interactions that need to be explicitly defined increases sharply, resulting in a rapid expansion of the number of nodes; Scenic uses high-level descriptive languages and probability sampling mechanisms to generate scenes. Although it can simplify scene generation, it relies on complex probability constraints for modeling, and the generated scenes may be too dependent on the quality of the rules. When the scene requires precise quantification of threats (TTC), the processing is relatively complex. The present invention not only focuses on key points through area division and threat grading, but also comprehensively evaluates scene risks through weight calculation, achieving efficient and accurate scene representation, demonstrating obvious advantages in driving scene processing. Brief Description of the Drawings

[0045] Figure 1 is a flowchart of a method for constructing scene nodes of an autonomous driving knowledge graph shown according to an exemplary embodiment.

[0046] Figure 2 is a schematic diagram of calculating the TTC value of a sub-area shown according to an exemplary embodiment. Detailed Embodiments

[0047] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below through specific embodiments in combination with the accompanying drawings.

[0048] The present invention provides a method for constructing scenario nodes of an autonomous driving knowledge graph, which reduces the complexity of scenario representation and the number of nodes by dividing the scenario area and the level of regional danger (TTC), avoids the explosion of the number of scenario nodes, and improves the retrieval efficiency of scenario nodes.

[0049] 1) Divide the observation area to reduce scenario complexity

[0050] According to the actual driving requirements, the ego vehicle only needs to focus on the vehicles that may pose a threat to safety within a specific range, without comprehensively perceiving all background vehicles. The present invention designs a rectangular observation area and divides it into four sub-areas: left, rear, front, and right. This division method can reduce the interference of background information and only retain the vehicles closely related to the safety of the ego vehicle, thereby reducing the data processing burden. Secondly, by screening the most threatening vehicle in each sub-area, the explosive growth of the number of scenario nodes is avoided. This method can effectively control the number of nodes and make the representation of scenario information more efficient.

[0051] 2) Construct scenario nodes based on TTC values

[0052] To reduce the number of scenario nodes and ensure accurate perception of threats at the same time, the present invention evaluates the threats of sub-areas based on TTC values and constructs scenario nodes accordingly. The TTC value represents the potential time of collision between vehicles, and the smaller the value, the greater the threat. In the scenario of a single background vehicle, the TTC value with this vehicle is directly calculated; in the scenario of multiple vehicles, the minimum TTC value is taken to lock the maximum threat. For sub-areas without vehicle interaction, the TTC value is set to -1 for marking. This method only extracts the key threat information of each sub-area, avoids the complexity of processing all vehicle data in full, thereby significantly reducing the number of nodes, and provides a reliable basis for risk assessment at the same time.

[0053] 3) Discretize TTC values to simplify scenario representation

[0054] To simplify the scenario representation, the present invention reduces the storage and calculation pressure of data dimensions by discretizing continuous TTC values into classification levels. The TTC values are divided into 6 categories: -1, 0, 1, 2, 3, 4. Among them, -1 represents no vehicle interaction, 0 represents extremely low threat, and the remaining levels represent increasing threats in turn. The discretization is based on setting thresholds and level upper limits. Vehicles above the threshold are considered to have negligible threats. This method avoids the explosion of the number of node combinations, reduces the complexity of scenario nodes, and at the same time retains sufficient description accuracy of threat information, providing support for efficient storage and retrieval.

[0055] 4) Define the overall risk value of scenario nodes

[0056] After calculating the TTC level of the sub-region, the present invention further defines the overall risk value of the scenario node to comprehensively evaluate the global threat. The overall risk value is based on the threat levels and weights of each sub-region. The weights are set according to the importance of the sub-region to driving safety. For example, the front region has a higher weight because it directly affects the driving path. This weighted evaluation method highlights the characteristics of high-risk regions and weakens the influence of secondary regions, making the overall risk value more comprehensively reflect the safety status of the driving scenario. In addition, the overall risk value can also provide a basis for sorting the node priorities, making the information storage and retrieval more efficient. Finally, the scenario node combines the sub-region TTC level and the overall risk value to construct a refined and effective scenario representation, supporting the accurate decision-making of the autonomous driving system.

[0057] The present invention provides a method for constructing a scenario node of an autonomous driving knowledge graph, which reduces the complexity of the scenario representation and the number of nodes by dividing the scenario area and the level of regional danger degree (TTC), avoids the explosion of the number of scenario nodes, and improves the retrieval efficiency of the scenario nodes.

[0058] The method for constructing a scenario node of an autonomous driving knowledge graph in the present invention includes three key steps, as Figure 1 shown, including: observation area division, scenario parameter extraction, and scenario node construction based on TTC.

[0059] Step 1: Observation area division for scenario information representation.

[0060] In the actual driving scenario, the ego vehicle usually only needs to focus on the background vehicles within a specific range that may pose a safety threat. Based on this discovery, the present invention defines an observation rectangular area and only focuses on the vehicles within this area. As Figure 2 shown, the observation area is divided into 4 different sub-regions: the left adjacent lane sub-region ( Figure 2 1 in Figure 2 ), the rear sub-region ( Figure 2 2 in Figure 2 ), the front sub-region (

[0061] Step 2: Definition of key scenario parameter extraction for vehicles within the region.

[0062] Extract key scenario parameters within the defined area after the observation area is divided. These parameters include road information, the positions of background vehicles and the ego vehicle, lane ID, speed, and acceleration. Subsequently, these parameters are used to calculate the TTC value and evaluate the threat level of each sub-region, which is the basis for constructing scenario nodes.

[0063] Step 3: Construction of scenario nodes based on the regional hazard level (TTC).

[0064] After dividing the observation area and extracting the scenario parameters, calculate the TTC value and its corresponding level for each sub-region, thus facilitating the construction of the corresponding scenario nodes.

[0065] Step 3.1: Determine the TTC values of the four sub-regions

[0066] Step 3.1.1: Calculate the TTC value between the leading vehicle and the following vehicle

[0067] First, the TTC value between the leading vehicle and the following vehicle is determined by the following factors:

[0068]

[0069] where the subscripts ev and bv represent the leading vehicle and the following vehicle respectively. v ev and v bv are their speeds, and d is the distance between them. δ is a small positive constant to prevent division by zero.

[0070] Step 3.1.2: Calculate the sub-region TTC value of a single background vehicle

[0071] Taking the scenario in Figure 2 as an example. For a sub-region that contains only one background vehicle, such as sub-region 1, the sub-region TTC value is equal to the TTC value between the ego vehicle and the background vehicle, that is

[0072] S 1 = TTC ev,bv (2)

[0073] where ev and bv represent the ego vehicle and the background vehicle respectively, and TTC ev,bv represents the TTC value between these two vehicles.

[0074] Step 3.1.3: Determine the TTC value of the sub-region without vehicle interaction

[0075] Since there are no background vehicles, sub-regions 2 and 3 are different. To distinguish them from other sub-regions that contain vehicles, the TTC value is set to -1, indicating no vehicle interaction. Therefore, the TTC values of these sub-regions are defined as

[0076] S 2 = S 3= -1 (3)

[0077] Step 3.1.4: Determine the TTC value of the sub-region containing multiple background vehicles

[0078] For a sub-region containing multiple background vehicles, such as sub-region 4, the sub-region TTC value is set to the minimum TTC value between the host vehicle and the background vehicles, i.e.,

[0079]

[0080] where i is the index of the background vehicle in this sub-region.

[0081] At this stage, the TTC values of four different sub-regions have been determined.

[0082] Step 3.2: Discretize the TTC value into classification levels

[0083] Directly representing the scenario with these sub-region TTC values will lead to a combinatorial explosion in the number of scenario nodes. To mitigate this complexity, the TTC value is discretized into classification levels. This discretization significantly reduces the complexity of scenario representation. The sub-region TTC level L j is calculated according to formula (5) as follows

[0084]

[0085] where j ∈ {1, 2, 3, 4} represents the index of the sub-region, and S j represents the TTC value of the sub-region. When there are no background vehicles in the sub-region, the sub-region TTC level is -1. The parameter T threshold is set to a relatively high TTC threshold of 4 seconds. It is assumed that the threat of background vehicles to the ego vehicle is minimal beyond this threshold. M is the maximum TTC level, which is set to 4 in the present invention. Therefore, there are 6 sub-region TTC levels: -1, 0, 1, 2, 3, 4. The higher the sub-region TTC level L j is, the greater the threat of this sub-region to the host vehicle.

[0086] Step 3.3: Comprehensive risk assessment of scenario nodes

[0087] In addition to focusing on the TTC levels of different sub-regions, the present invention's attention to the scenario also includes a comprehensive risk assessment.

[0088] Step 3.3.1: Set the weights of four sub-regions

[0089] According to experience, set the weights w related to the sub-regions j .

[0090] Step 3.3.2: Define the overall risk value of the scenario node

[0091] Recognizing that each sub-region poses different levels of threat to the host vehicle, the present invention defines the overall risk value of the scenario node N k as the weighted average of the TTC levels of the sub-regions, that is

[0092]

[0093] where k is the index of the scenario node.

[0094] Step 3.4: Scenario node construction.

[0095] Combining the TTC level of the sub-region and the scenario risk to represent the scenario node, as follows

[0096]

[0097] Considering the number of scenario nodes and the diversity of scenarios, the present invention divides the observation area into 4 different sub-regions. The TTC of each sub-region is divided into 6 levels. Although there are alternative methods for dividing the scenario sub-regions and classifying the TTC levels, the proposed scenario node construction method is still applicable to these alternative methods.

[0098] In another embodiment of the present invention, it further includes: Step 4: Constructing a knowledge graph.

[0099] Step 4.1: Data collection and standardization.

[0100] Extract all the characteristic parameters (such as L 1 , L 2 , L 3 , L 4 , ) defined in the scenario node, and standardize them into the scenario node attributes of the graph.

[0101] Step 4.2: Generation of nodes and relationships.

[0102] Generate corresponding entity nodes (such as "front vehicle area node") according to different scenario areas, and connect the sub-region nodes to the scenario nodes according to the weight relationship defined in formula (6).

[0103] Step 4.3: Graph storage and optimization.

[0104] Use a graph database (Neo4j) to store the knowledge graph, and utilize node aggregation and indexing technologies to improve query efficiency.

[0105] In summary, the present invention realizes the balance between the richness of scenario information and the number of scenarios by combining region division and TTC level discretization. Region division enables the host vehicle to focus on the vehicles with the greatest impact on safety, while TTC level discretization simplifies the scenario representation and reduces the complexity of scenario nodes.

[0106] Specifically, the present invention realizes self-vehicle attention within a specific range based on the division of the observation area. In actual driving scenarios, the self-vehicle usually only needs to pay attention to background vehicles within a specific range that may pose a safety threat. Based on this discovery, the present invention defines an observation rectangular area and only pays attention to the vehicles within this area. As Figure 2 shown, the observation area is divided into 4 different sub-areas: the left adjacent lane sub-area ( Figure 2 1 in), the rear sub-area ( Figure 2 2 in), the front sub-area ( Figure 2 3 in) and the right adjacent lane sub-area ( Figure 2 4 in). Restricting and subdividing this observation area brings two significant benefits: 1) Reducing background information: By excluding vehicles that are far from the self-vehicle and have little impact on the self-vehicle, the number of background vehicles that need to be paid attention to is reduced. 2) Avoiding SEKG node explosion: By dividing the area and selecting the TTC value of the vehicle with the greatest threat in each sub-area as the characteristic element, the number of scenario nodes in SEKG can be significantly reduced, avoiding the explosive growth of the number of scenario nodes. After the observation area is divided, key scenario parameters within the defined area are extracted. These parameters include road information, the positions of background vehicles and the self-vehicle, lane ID, speed, and acceleration. Subsequently, these parameters are used to calculate the TTC value and evaluate the threat level of each sub-area, which is the basis for constructing scenario nodes.

[0107] The present invention realizes discrete classification based on the regional danger level (TTC). After completing the division of the observation area and the extraction of scenario parameters, the TTC value of each sub-area is calculated and corresponding levels are assigned to construct scenario nodes. For a sub-area with a single background vehicle, the TTC value is the TTC value between the self-vehicle and this vehicle; for a sub-area without background vehicles, the TTC value is marked as -1; for a sub-area with multiple background vehicles, the minimum TTC value is taken. After determining the TTC values of the four sub-areas, directly using them to represent the scenario will greatly increase the number of scenario nodes and complexity. Therefore, the TTC value is discretized into different classification levels, greatly reducing the complexity of scenario representation. When there are no background vehicles in the sub-area, its TTC level is set to -1. If the TTC value exceeds a set threshold, such as 4 seconds, we consider that the background vehicle poses the least threat to the self-vehicle, and at this time the TTC level is 0. For other situations, the TTC level is determined according to the ratio of the TTC value to the threshold. The higher the level, the greater the threat of the sub-area to the self-vehicle. The present invention also conducts a comprehensive risk assessment, defining the overall risk value of the scenario node as the weighted average of the TTC levels of the sub-areas, and calculating the overall risk value by assigning weights according to the threat degree of the sub-areas to the self-vehicle. Finally, the TTC level of the sub-area and the scenario risk are combined to represent the scenario node. This method provides a flexible and effective way to construct scenario nodes, balancing the richness of scenario information and the number of scenarios.

[0108] Other embodiments of the present disclosure will be readily apparent to those skilled in the art upon consideration of the specification and practice of the present disclosure. The present disclosure is intended to cover any variations, uses, or adaptations of the present disclosure, which follow the general principles of the present disclosure and include known common general knowledge or conventional technical means in the technical field not disclosed in the present disclosure. The specification and examples are only to be considered as exemplary, and the present disclosure is not limited to the exact structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for constructing scene nodes in an autonomous driving knowledge graph, characterized in that: The method comprises: Define the observation area of ​​the vehicle in the current scene and divide the observation area into several sub-areas; Extracting key scene parameters within the observation area; Based on the key scene parameters, the regional danger level of each sub-area and the overall risk of the scene node are calculated, and the regional danger level and the overall risk value are used as attributes of the scene node in the knowledge graph.

2. The method for constructing scene nodes in an autonomous driving knowledge graph according to claim 1, characterized in that: The sub-areas include: a left adjacent lane sub-area, a rear sub-area, a front sub-area and a right adjacent lane sub-area.

3. The method for constructing scene nodes in an autonomous driving knowledge graph according to claim 1, characterized in that: The key scene parameters include: road information and the positions, lane IDs, speeds, and accelerations of background vehicles and the vehicle itself.

4. The method for constructing scene nodes in an autonomous driving knowledge graph according to claim 3, characterized in that: The calculating the regional danger level of each sub-area based on the key scene parameters includes: According to the positions of the background vehicles and the ego vehicle, the distance between each background vehicle and the ego vehicle is obtained; Based on the distance between each background vehicle and the vehicle, the speed of the background vehicle and the speed of the vehicle, obtaining a TTC value between the background vehicle and the vehicle; Based on the road information and the lane IDs of the background vehicles, the distribution of the background vehicles in each sub-area is obtained; When there is a background vehicle in a sub-area, the TTC value of the sub-area is the TTC value between the background vehicle and the ego vehicle; When there is no background vehicle in a sub-area, the TTC value of the sub-area is a set value; When there are multiple background vehicles in a sub-area, the TTC value of the sub-area is the minimum TTC value between the ego vehicle and the background vehicles; Based on the set value and a set threshold, the TTC value of each sub-region is discretized to obtain the regional danger level of each sub-region.

5. The method for constructing scene nodes in an autonomous driving knowledge graph according to claim 1, characterized in that: The process of calculating the overall risk value of the scene node includes: Get the weight of each sub-region; The regional danger level of each sub-region is weighted based on the weight of the sub-region to obtain the overall risk value of the scene node.

6. A method for constructing an autonomous driving knowledge graph, characterized in that: The method comprises: Define the observation area of ​​the vehicle in each scenario and divide the observation area into several sub-areas; Constructing a scene node, wherein the attributes of the scene node are generated based on the method for constructing a scene node in an autonomous driving knowledge graph according to any one of claims 1 to 5; Generate entity nodes based on sub-areas in different scenarios; After connecting the entity node with the corresponding scene node, the autonomous driving knowledge graph is obtained.

7. The method for constructing an autonomous driving knowledge graph according to claim 6, characterized in that: A graph database is used to store the autonomous driving knowledge graph.

8. An autonomous driving knowledge graph scene node construction system, characterized in that: The system comprises: The area definition module is used to define the observation area of ​​the vehicle in the current scene and divide the observation area into several sub-areas; A parameter extraction module, used to extract key scene parameters in the observation area; A node construction module is used to calculate the regional danger level of each sub-area and the overall risk value of the scene node based on the key scene parameters, and use the regional danger level and the overall risk value as attributes of the scene node in the knowledge graph.

9. An autonomous driving knowledge graph construction system, characterized in that: The system comprises: The area definition module defines the observation area of ​​the vehicle in each scenario and divides the observation area into several sub-areas; A scene node construction module, used to construct a scene node, wherein the attributes of the scene node are generated based on the method for constructing a scene node in an autonomous driving knowledge graph according to any one of claims 1 to 5; The entity node construction module is used to generate entity nodes based on sub-areas in different scenarios; The knowledge graph generation module is used to connect the entity node with the corresponding scene node to obtain the autonomous driving knowledge graph.

10. An electronic device, characterized in that: The electronic device includes: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, it implements the autonomous driving knowledge graph scene node construction method as described in any one of claims 1-5 or the autonomous driving knowledge graph construction system construction method as described in any one of claims 6-7.

Images