Driving scene evaluation method and device, computer equipment and storage medium

Abstract

The invention relates to a driving scene evaluation method and device, computer equipment and a storage medium. The method comprises the following steps of: acquiring driving scene information through a plurality of sensors of an automatic driving vehicle by using the driving scene evaluation method and device, the computer equipment and the storage medium; generating a knowledge graph of the driving scene according to the driving scene information and a preset knowledge graph structure; the node types of the preset knowledge graph structure comprise concept layer nodes, entity layer nodes and attribute layer nodes; according to the knowledge graph and the evaluation model of the driving scene, obtaining an evaluation result of the driving scene; the evaluation model comprises a global spatial feature extraction sub-model, a node feature extraction sub-model and a self-attention prediction sub-model. And the accuracy of evaluating the driving scene is improved.

Description

Technical Field

[0001] This application relates to the field of autonomous driving technology, and in particular to a method, apparatus, computer device, and storage medium for evaluating driving scenarios. Background Technology

[0002] In recent years, autonomous driving technology, as a core area of ​​intelligentization, has achieved remarkable breakthroughs. However, the complex and ever-changing real-world road environment, as well as the inherent limitations and uncertainties of the system itself, mean that autonomous vehicles constantly face potential safety risks during operation. If these risks are not identified and effectively addressed in a timely manner, they can lead to a degraded driving experience or even serious traffic accidents. Therefore, state assessment of autonomous driving scenarios is of paramount importance.

[0003] However, traditional methods for evaluating driving scenarios suffer from low accuracy. Summary of the Invention

[0004] Therefore, it is necessary to provide an assessment method, device, computer equipment, and storage medium for driving scenarios that can improve the accuracy of assessment, addressing the aforementioned technical problems.

[0005] Firstly, this application provides a method for evaluating driving scenarios, the method comprising:

[0006] Information about the driving scenario is collected through multiple sensors in autonomous vehicles;

[0007] A knowledge graph of the driving scenario is generated based on the driving scenario information and a preset knowledge graph structure; the node types of the preset knowledge graph structure include concept layer nodes, entity layer nodes, and attribute layer nodes.

[0008] The evaluation result of the driving scenario is obtained based on the knowledge graph and evaluation model of the driving scenario; the evaluation model includes a global spatial feature extraction sub-model, a node feature extraction sub-model, and a self-attention prediction sub-model.

[0009] In one embodiment, obtaining the evaluation result of the driving scenario based on the knowledge graph and evaluation model of the driving scenario includes:

[0010] The global spatial feature extraction sub-model processes and compresses the node features corresponding to the knowledge graph of the driving scenario to obtain global spatial structure features, which are then broadcast to the node feature extraction sub-model.

[0011] Based on the node feature extraction sub-model, the processed node features and the global spatial structure features are fused to obtain the processed node features;

[0012] The processed node features are input into the self-attention prediction sub-model for prediction to obtain the evaluation result of the driving scenario.

[0013] In one embodiment, the step of processing and compressing the node features corresponding to the knowledge graph of the driving scenario based on the global spatial feature extraction sub-model to obtain global spatial structure features includes:

[0014] For each node in the knowledge graph of the driving scenario, an aggregation function is used to aggregate information from the features of neighboring nodes and update the node's own features to obtain the updated node features.

[0015] The updated node characteristics are evaluated based on a third-party open-source library to determine which nodes to retain.

[0016] The features of the retained nodes and the updated nodes are processed and compressed to obtain the global spatial structure features.

[0017] In one embodiment, the process of processing and compressing the retained nodes and the updated node features to obtain the global spatial structure features includes:

[0018] The features of the retained nodes and the updated nodes are added together to obtain multiple summed node features;

[0019] The multiple summed node features are fused and compressed to obtain the global spatial structure features.

[0020] In one embodiment, the self-attention prediction sub-model includes a first attention mechanism layer and a second attention mechanism layer. The step of inputting the processed node features into the self-attention prediction sub-model for prediction to obtain the evaluation result of the driving scenario includes:

[0021] The processed node features are input into the first attention mechanism layer to obtain the first processing result;

[0022] The first processing result and the processed node features are input into the second attention mechanism layer, and the evaluation result of the driving scenario is obtained after linear layer processing.

[0023] In one embodiment, the method further includes:

[0024] Construct a relational graph convolutional network model; the relational graph convolutional network model includes an initial global spatial feature extraction sub-model, an initial node feature extraction sub-model, and an initial self-attention prediction sub-model;

[0025] Obtain simulation scene information of the target scene; the information type of the simulation scene information is related to the preset knowledge graph structure;

[0026] The relationship graph convolutional network model is trained based on the simulation scenario information to obtain the evaluation model.

[0027] Secondly, this application also provides a driving scenario evaluation device, including:

[0028] The data acquisition module is used to collect driving scene information through multiple sensors in autonomous vehicles;

[0029] The generation module is used to generate a knowledge graph of the driving scenario based on the driving scenario information and a preset knowledge graph structure; the nodes in the knowledge graph include concept layer nodes, entity layer nodes, and attribute layer nodes.

[0030] The evaluation module is used to obtain the evaluation result of the driving scenario based on the knowledge graph and evaluation model of the driving scenario; the evaluation model includes a global spatial feature extraction sub-model, a node feature extraction sub-model, and a self-attention prediction sub-model.

[0031] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0032] Information about the driving scenario is collected through multiple sensors in autonomous vehicles;

[0033] A knowledge graph of the driving scenario is generated based on the driving scenario information and a preset knowledge graph structure; the node types of the preset knowledge graph structure include concept layer nodes, entity layer nodes, and attribute layer nodes.

[0034] The evaluation result of the driving scenario is obtained based on the knowledge graph and evaluation model of the driving scenario; the evaluation model includes a global spatial feature extraction sub-model, a node feature extraction sub-model, and a self-attention prediction sub-model.

[0035] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0036] Information about the driving scenario is collected through multiple sensors in autonomous vehicles;

[0037] A knowledge graph of the driving scenario is generated based on the driving scenario information and a preset knowledge graph structure; the node types of the preset knowledge graph structure include concept layer nodes, entity layer nodes, and attribute layer nodes.

[0038] The evaluation result of the driving scenario is obtained based on the knowledge graph and evaluation model of the driving scenario; the evaluation model includes a global spatial feature extraction sub-model, a node feature extraction sub-model, and a self-attention prediction sub-model.

[0039] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0040] Information about the driving scenario is collected through multiple sensors in autonomous vehicles;

[0041] A knowledge graph of the driving scenario is generated based on the driving scenario information and a preset knowledge graph structure; the node types of the preset knowledge graph structure include concept layer nodes, entity layer nodes, and attribute layer nodes.

[0042] The evaluation result of the driving scenario is obtained based on the knowledge graph and evaluation model of the driving scenario; the evaluation model includes a global spatial feature extraction sub-model, a node feature extraction sub-model, and a self-attention prediction sub-model.

[0043] The aforementioned driving scenario evaluation method, device, computer equipment, and storage medium collect driving scenario information through multiple sensors of the autonomous vehicle; generate a knowledge graph of the driving scenario based on the driving scenario information and a pre-defined knowledge graph structure; the node types of the pre-defined knowledge graph structure include concept layer nodes, entity layer nodes, and attribute layer nodes; and obtain the evaluation result of the driving scenario based on the knowledge graph and the evaluation model; the evaluation model includes a global spatial feature extraction sub-model, a node feature extraction sub-model, and a self-attention prediction sub-model. By constructing a more comprehensive node type for the knowledge graph, the comprehensiveness and reliability of the knowledge graph are improved, thereby improving the accuracy of driving scenario evaluation. Furthermore, through the global spatial feature extraction sub-model, node feature extraction sub-model, and self-attention prediction sub-model, the three-tiered model analyzes the knowledge graph, providing a more accurate evaluation of the driving scenario, further improving the accuracy of driving scenario evaluation. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is an application environment diagram of the evaluation method for driving scenarios in one embodiment;

[0046] Figure 2 This is a flowchart illustrating a method for evaluating a driving scenario in one embodiment;

[0047] Figure 3 This is a flowchart of an evaluation method for a driving scenario in one embodiment;

[0048] Figure 4 This is a flowchart illustrating the evaluation method for a driving scenario in another embodiment;

[0049] Figure 5 This is a network structure diagram of the evaluation method for driving scenarios in another embodiment;

[0050] Figure 6 This is a flowchart illustrating the evaluation method for a driving scenario in another embodiment;

[0051] Figure 7 This is a flowchart illustrating the evaluation method for a driving scenario in another embodiment;

[0052] Figure 8 This is a flowchart illustrating the evaluation method for a driving scenario in another embodiment;

[0053] Figure 9 This is a flowchart illustrating the evaluation method for a driving scenario in another embodiment;

[0054] Figure 10 This is one of the schematic diagrams of the scene data acquisition environment in one embodiment;

[0055] Figure 11 This is the second schematic diagram of the scene data acquisition environment in one embodiment;

[0056] Figure 12 This is a structural block diagram of a driving scenario evaluation device in one embodiment;

[0057] Figure 13 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0059] The driving scenario evaluation method provided in this application embodiment can be applied to, for example, driving scenarios. Figure 1In the application environment shown, sensor 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104, or it can be located in the cloud or on another network server. Sensor 102 sends the collected driving scenario information to server 104, and server 104 evaluates the driving scenario based on the driving scenario information to obtain the evaluation result. Server 104 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server providing cloud computing services.

[0060] In one embodiment, such as Figure 2 As shown, a method for evaluating driving scenarios is provided, which can be applied to... Figure 1 Taking the server in the example of this, the explanation includes:

[0061] S201 collects driving scenario information through multiple sensors in autonomous vehicles.

[0062] These sensors may include an onboard camera array, LiDAR, millimeter-wave radar, and a GPS / IMU integrated navigation system. Optionally, map information corresponding to the autonomous vehicle can also be acquired, which may include road structure type, road structure parameters, lane type, speed limit information, etc.

[0063] In this embodiment, the parameters of multiple sensors installed on the autonomous vehicle are synchronously calibrated to ensure spatiotemporal consistency, and the multiple sensors send the driving scene information they collect to the server.

[0064] Optionally, driving scenario information may include visual data, point cloud data, radar data, and positioning data. Visual data can be RGB images and video streams captured by cameras for target detection and semantic segmentation (such as lane lines and traffic signs); point cloud data can generate 3D environmental point clouds for LiDAR, providing information on obstacle location, distance, and shape; radar data can output relative velocity, acceleration, and angle information of targets for millimeter-wave radar; and positioning data can provide global coordinates for GPS, with IMU supplementing vehicle attitude (pitch / yaw angle).

[0065] Optionally, after acquiring driving scenario information, multi-source heterogeneous data can be spatiotemporally synchronized and unified to the vehicle coordinate system; noise reduction (such as point cloud filtering), normalization, and format conversion can also be performed. The driving scenario can then be evaluated based on the processed data.

[0066] S202, Generate a knowledge graph of the driving scenario based on driving scenario information and a preset knowledge graph structure; the node types of the preset knowledge graph structure include concept layer nodes, entity layer nodes, and attribute layer nodes.

[0067] In this embodiment, the concept layer nodes can be initially divided into eight major categories according to the completely different categories designed for the scenario: "road structure," "road facility," "road change," "traffic participants," "weather environment," "digital information," "ego information," and "frame sequence." In the knowledge graph, these can be represented sequentially as "road structure," "road facility," "road change," "traffic participants," "weather environment," "digital information," "ego information," and "frame," where:

[0068] (1) The parent node “road structure” is further subdivided into seven sub-nodes: “straight road”, “curve”, “merger ramp”, “exit ramp”, “intersection”, “roundabout”, and “tunnel”. In the knowledge graph, these can be represented as “straight”, “curve”, “mergeramp”, “exit ramp”, “intersection”, “roundabout”, and “tunnel”.

[0069] (2) The parent node “road facility” is further subdivided into five sub-nodes: “lane”, “road marking”, “traffic sign”, “traffic light”, and “guardrail”. In the knowledge graph, these can be represented as “lane”, “marker”, “sign”, “light”, and “barrier” respectively.

[0070] (3) The parent node “road change” is further subdivided into four sub-nodes: “area change”, “road damage”, “debris accumulation”, and “temporary event”. In the knowledge graph, these can be represented as “area”, “road defect”, “clutter”, and “temporal event”, respectively.

[0071] (4) The parent node “traffic participants” is further subdivided into thirteen sub-nodes: “car”, “van”, “bus”, “pickup”, “semi-trailer”, “trailer”, “tricycle”, “bicycle”, “police car”, “ambulance”, “fire truck”, “pedestrian”, and “animal”. In the knowledge graph, these can be represented as “car”, “van”, “bus”, “pickup”, “semi-trailer”, “trailer”, “tricycle”, “bicycle”, “police car”, “ambulance”, “fire truck”, “pedestrians”, and “animal”, respectively.

[0072] (5) The parent node “weather environment” is further subdivided into six child nodes: “wind”, “snow”, “rain”, “fog”, “haze” and “dust”. In the knowledge graph, these can be represented as “wind”, “snow”, “rain”, “fog”, “haze” and “dust” respectively.

[0073] (6) The parent node “digital information” is further subdivided into a child node “communication quality”, which can be represented as “communication quality” in the knowledge graph.

[0074] (7) The parent node “ego information” is further divided into three child nodes: “automotive entity”, “driver” and “passenger”, which can be represented as “ego”, “driver” and “passenger” in the knowledge graph, respectively.

[0075] (8) The parent node “frame” is subdivided into ten child nodes: “frame1”, “frame2”, “frame3”, “frame4”, “frame5”, “frame6”, “frame7”, “frame8”, “frame9”, and “frame10”. In the knowledge graph, these can be represented as “frame1”, “frame2”, “frame3”, “frame4”, “frame5”, “frame6”, “frame7”, “frame8”, “frame9”, and “frame10”.

[0076] In this embodiment, the entity layer nodes consist of target detection algorithms from the autonomous driving system and detection data from other sensors, and dynamically change with the current driving scenario; the attribute layer nodes can be further subdivided into seven categories of parent nodes: "width," "height," "speed," "length," "state," "angle," and "distance," which can be represented in the knowledge graph as "width," "height," "speed," "length," "state," "angle," and "distance," respectively.

[0077] (1) The parent node “width” is discretely divided into seven values ​​according to the actual value detected by the sensor: “0-1”, “1-2”, “2-3”, “3-4”, “4-6”, “6-10”, and “over10”, with the unit being meters.

[0078] (2) The parent node “height” is discrete into five values: “0-1”, “1-2”, “2-3”, “3-4” and “over4”, according to the actual value detected by the sensor. The unit is meters.

[0079] (3) The parent node “speed” is also discretely divided into ten values ​​according to the actual value detected by the sensor: “0-10”, “10-20”, “20-30”, “30-40”, “40-50”, “50-60”, “60-80”, “80-100”, “100-120”, and “over120”, with the unit being kilometers per hour.

[0080] (4) The parent node “length” is discretely divided into eight values ​​according to the actual value detected by the sensor: “0-1”, “1-2”, “2-3”, “3-4”, “4-5”, “5-8”, “8-10”, and “over10”, with the unit being meters.

[0081] (5) The parent node “state” is subdivided into five values: “sudden braking and deceleration”, “deceleration”, “uniform speed”, “acceleration”, and “rapid acceleration”. In the knowledge graph, these values ​​can be represented as “hash slow”, “speed down”, “uniform”, “speed up”, and “hashup”, respectively.

[0082] (6) The parent node “angle” is subdivided into nine values: “0-15”, “15-25”, “25-35”, “35-45”, “45-55”, “55-65”, “65-75”, “75-90”, and “90-180”, with the unit being °.

[0083] (7) The parent node “distance” is subdivided into eight values: “0-1 meter”, “1-2 meter”, “2-4 meter”, “4-8 meter”, “8-16 meter”, “16-32 meter”, “32-64 meter”, and “64 meter and above”. In the knowledge graph, these values ​​can be represented as “dis0-1”, “dis1-2”, “dis2-4”, “dis4-8”, “dis8-16”, “dis16-32”, “dis32-64”, and “dis64-”, respectively, with the unit being meters.

[0084] In this embodiment, the relationships between conceptual layer nodes, between conceptual layer and entity layer nodes, and between entity layer nodes are described using fifteen types: "has," "subclass of," "has frame," "has next frame," "left of," "right of," "has effect," "is a," "next to," "is on," "left side," "right side," "left front," "right front," and "front." These correspond to the following types: "has," "subclass of," "has frame," "has next frame," "left of," "right of," "has effect," "is a," "next to," "is on," "left side," "right side," "left front," "right front," and "front." The relationships between entity layer nodes and attribute layer nodes are described using seven categories: "width," "height," "speed," "length," "state," "angle," and "distance," corresponding to all values ​​and using colors such as "red," "yellow," and "green."

[0085] In the embodiments of this application, such as Figure 3 As shown, after acquiring driving scene information, a knowledge graph is constructed for every ten frames of sensor data. Each knowledge graph starts from the "scene" node and connects to eight nodes in the concept layer sequentially, selecting the "has" relationship. Then, based on the results of the autonomous driving object detection algorithm, corresponding entity layer nodes are established and connected to the corresponding concept layer nodes. At the same time, corresponding attribute layer nodes are created based on each sensor data, and the correct edge type is selected.

[0086] S203, based on the knowledge graph and evaluation model of the driving scenario, obtains the evaluation result of the driving scenario; the evaluation model includes a global spatial feature extraction sub-model, a node feature extraction sub-model, and a self-attention prediction sub-model.

[0087] In the embodiments of this application, such as Figure 3 As shown, the knowledge graph of the current frame and historical sequence data can be input into the evaluation model. The model then undergoes global feature extraction, node feature enhancement, and self-attention weighted fusion to obtain the evaluation result of the driving scenario. Optionally, the evaluation result can be the scenario safety level, such as safe, warning, or dangerous, as well as a list of targets requiring emergency obstacle avoidance and suggested trajectories.

[0088] In the aforementioned driving scenario evaluation method, driving scenario information is collected through multiple sensors of the autonomous vehicle; a knowledge graph of the driving scenario is generated based on the driving scenario information and a pre-defined knowledge graph structure; the node types of the pre-defined knowledge graph structure include concept layer nodes, entity layer nodes, and attribute layer nodes; and the evaluation result of the driving scenario is obtained based on the knowledge graph and the evaluation model; the evaluation model includes a global spatial feature extraction sub-model, a node feature extraction sub-model, and a self-attention prediction sub-model. By constructing a more comprehensive node type for the knowledge graph, the comprehensiveness and reliability of the knowledge graph are improved, thereby improving the accuracy of driving scenario evaluation. Furthermore, the three models—the global spatial feature extraction sub-model, the node feature extraction sub-model, and the self-attention prediction sub-model—analyze the knowledge graph to more accurately evaluate the driving scenario, further improving the accuracy of driving scenario evaluation.

[0089] In one embodiment, one implementation of S203 above is provided, such as... Figure 4 As shown, the above-mentioned "obtaining the evaluation results of the driving scenario based on the knowledge graph and evaluation model of the driving scenario" includes:

[0090] S301: Based on the global spatial feature extraction sub-model, the node features corresponding to the knowledge graph of the driving scenario are processed and compressed to obtain global spatial structure features, which are then broadcast to the node feature extraction sub-model.

[0091] In the embodiments of this application, such as Figure 5 As shown, the evaluation model receives node features (such as entity type and attribute value) and topological relationships (connection edges between nodes) from the knowledge graph. It performs multi-hop neighborhood aggregation on the node features in the knowledge graph and gradually captures the spatial dependency relationship between the node and its neighboring nodes through multiple processing layers, generating global spatial structure features containing global context information.

[0092] For example, global spatial structure features may include the connectivity of road structures, the distribution patterns of traffic participants, etc.

[0093] In this embodiment, the global spatial structure features can be expanded according to the number of nodes in the graph using the Broadcast mechanism function and broadcast to the node feature extraction sub-model, so that the global spatial structure features can be fused with the node features in the node feature extraction sub-model.

[0094] S302, based on the node feature extraction sub-model, the processed node features and global spatial structure features are fused to obtain the processed node features.

[0095] In the embodiments of this application, such as Figure 5As shown, the node feature extraction sub-model is composed of stacked Relational Graph Convolutional Networks (RGCNs). During the stacking process, RGCNs also use a shortcut method to connect and fuse the processed node features. This has a structural advantage during training, making it less prone to gradient vanishing during backpropagation training. At the same time, it improves the ability of subsequent layers to have features calculated by the upper and uppermost layers, increasing the amount of information contained in the features and further enhancing the feature extraction capability of subsequent layers.

[0096] S303 inputs the processed node features into the self-attention prediction sub-model for prediction to obtain the evaluation results of the driving scenario.

[0097] In this embodiment of the application, the processed node features are input into the self-attention prediction sub-model to determine the risk level of the driving scenario and the priority ranking of key entities.

[0098] In the above-mentioned application embodiments, the evaluation process achieves a closed loop from raw sensor data to interpretable risk assessment through a feature iteration mechanism from global to local and back to global, which significantly improves the decision reliability and safety of the autonomous driving system in complex scenarios.

[0099] In one embodiment, one implementation of S301 above is provided, such as... Figure 6 As shown, the above-mentioned "processing and compressing the node features corresponding to the knowledge graph of the driving scene based on the global spatial feature extraction sub-model to obtain global spatial structure features" includes:

[0100] S401: For each node in the knowledge graph of the driving scenario, use the aggregation function to aggregate information from the features of neighboring nodes and update the node's own features to obtain the updated node features.

[0101] In this embodiment, the global spatial feature extraction sub-model may include multiple processing models, each including the RGCN processing method. RGCN mainly uses an aggregation function to aggregate information from the features of neighboring nodes and update its own node features. During the aggregation process, the types of edges in different knowledge graphs are considered, so the aggregation function artificially increases the weights according to the edge type. The aggregation function is shown in Equation 1:

[0102] (Equation 1)

[0103] Among them, It is the first Adjacency matrix of class relationships. It is the first Weight matrix of class relationships. It is the first Bias terms in class relationships. It is the first Layer node feature matrix. The weight matrix is ​​determined by humans based on the importance of training experience, as shown in Table 1. This represents the activation function.

[0104] Table 1

[0105]

[0106]

[0107]

[0108] S402 evaluates the updated node characteristics based on a third-party open-source library to determine which nodes to retain.

[0109] In this embodiment, each processing module also includes a TopKPool processing method. This method calculates the score of each node in the output result of the RGCN network layer step using FastRGCN provided by the third-party open-source library torch_geometric. Then, it retains the K nodes with the highest scores, matches these K nodes with the input nodes, and adds their features. The specific formula for calculating the scores of the K nodes is shown in Equation 2:

[0110] (Equation 2)

[0111] S403 processes and compresses the features of the retained nodes and the updated nodes to obtain the global spatial structure features.

[0112] In this embodiment of the application, after determining the retained nodes based on the scores of each node, the nodes are then fused and compressed according to the node features of each retained node and the updated node features to obtain the global spatial structure features.

[0113] In the above-mentioned embodiments, the iterative optimization of node features is achieved by dynamically aggregating the feature information of adjacent nodes in the knowledge graph, thereby improving the completeness and relevance of the driving scene representation. By using a third-party open-source library to conduct quantitative evaluation of the updated nodes, key nodes with semantic value are effectively screened out, and redundant or noisy information is eliminated, making the knowledge graph structure more refined and reliable. Finally, the global spatial structure features constructed based on the retained nodes and their optimized features not only preserve the spatial relationships and attribute details of scene elements, but also enhance the feature expression capability through cross-node information fusion.

[0114] In one embodiment, one implementation of S403 described above is provided, such as... Figure 7As shown, the above-mentioned "processing and compressing the features of the retained nodes and the updated nodes to obtain global spatial structure features" includes:

[0115] S501, the features of the retained nodes and the updated nodes are added together to obtain multiple summed node features.

[0116] S502 performs feature fusion and compression on the summed node features to obtain global spatial structure features.

[0117] In this embodiment of the application, the feature addition of the retained node and the updated node features can be performed as shown in Equation 3:

[0118] (Equation 3)

[0119] Furthermore, the processing model can also include a Global Average Pooling (GAP) function, which can compress all nodes and their features into a single feature representation. Optionally, GAP can be implemented by calling the third-party open-source library torch_geometric.nn to compress all nodes and their features into a single feature representation.

[0120] For example, the GAP processing procedure is shown in Equation 4:

[0121] (Equation 4)

[0122] in, It is a feature of the entire graph, graph G=(V,E), where V is the set of nodes and E is the set of edges. These are the characteristics of each node.

[0123] In the above-mentioned application embodiments, by adding the retained nodes and their updated node features one by one, the complementarity and enhancement of multi-dimensional information are achieved. This not only preserves the basic attributes of the nodes themselves but also incorporates the contextual information of the neighborhood relationships. The hierarchical and progressive feature processing mechanism enables the final global spatial structure features to have both local detail sensitivity and global pattern capture capabilities, which significantly improves the understanding accuracy and generalization ability of complex driving scenarios.

[0124] In one embodiment, an implementation of S303 above is provided, wherein the self-attention prediction sub-model includes a first attention mechanism layer and a second attention mechanism layer, such as... Figure 8 As shown, the above-mentioned "inputting the processed node features into the self-attention prediction sub-model for prediction to obtain the evaluation results of the driving scenario" includes:

[0125] S601, the processed node features are input into the first attention mechanism layer to obtain the first processing result.

[0126] S602, the first processing result and the processed node features are input into the second attention mechanism layer, and the evaluation result of the driving scenario is obtained after linear layer processing.

[0127] In this embodiment, the self-attention prediction sub-model comprises three components: Multihead Attention, Linear, and GAP. Multihead Attention and Linear are implemented using the third-party open-source library torch.nn. Multihead Attention consists of two layers: a first attention mechanism layer and a second attention mechanism layer, with inputs of Q, K, and V. The first attention mechanism layer receives the output of the node feature extraction sub-model, while the second attention mechanism layer receives Q (the input of the node feature extraction sub-model) and K and V (the calculation results of the first attention mechanism layer). This ensures the second attention mechanism layer also has original data, preventing overfitting. Then, Linear, GAP, and Linear calculations are performed sequentially. The first Linear layer is a basic linear neural network layer that processes the results of the second attention mechanism layer. The final Linear layer compresses the number of channels to 2, representing the risk level of the prediction.

[0128] In the above-mentioned application embodiments, key association information of node features is dynamically captured by first-level attention, and local features with semantic value are automatically selected. Then, the refined features are deeply integrated with the original node attributes and global spatial structure features of the scene. Cross-level information interaction and global reasoning are completed with the help of the second-level attention mechanism. This progressive feature processing architecture enables the model to make risk judgment by combining local anomalies and global context, and finally outputs evaluation results with both high timeliness and strong robustness.

[0129] In one embodiment, such as Figure 9 As shown, the evaluation method for the above driving scenarios includes:

[0130] S701, Construct a relational graph convolutional network model; the relational graph convolutional network model includes an initial global spatial feature extraction sub-model, an initial node feature extraction sub-model, and an initial self-attention prediction sub-model.

[0131] In this embodiment, a core model is constructed based on the Relational Graph Convolutional Network (RGCN) framework. This model consists of three parts: an initial global spatial feature extraction sub-model uses two-dimensional convolutional layers to reduce the dimensionality of input scene images and other information, extracting global contextual features; an initial node feature extraction sub-model aggregates feature information from adjacent nodes in the knowledge graph through graph convolutional layers, enhancing the semantic representation of the nodes themselves; and an initial self-attention prediction sub-model introduces a multi-head self-attention mechanism to dynamically calculate the importance weights between nodes, focusing on key entities and their relationships. These three parts form a cascaded structure through residual connections, achieving progressive feature learning from local node features to global scene understanding.

[0132] It should be noted that both node features and global features are aggregated using graph convolution, but global features are extracted more deeply and have a wider receptive field.

[0133] S702, Obtain simulation scene information of the target scene; the information type of the simulation scene information is related to the preset knowledge graph structure.

[0134] In this embodiment, typical potential risk scenarios are constructed and data is collected on a simulation platform. Based on a preset knowledge graph structure, potential risk scenarios are built in the Carla simulation platform. These scenarios are divided into two main categories: "ghost peek" incidents caused by emergency braking of adjacent vehicles and emergency lane-changing incidents caused by adjacent vehicles. Each category requires the design of corresponding safety scenarios. Each scenario involves three road structures: four-lane, three-lane, and two-lane. Left-side and right-side events are designed for each road structure. During each scenario reproduction, information such as time, weather, number and position of surrounding vehicles, and initial speed are controlled by random parameters, as shown in Table 2 below. The safety scenario corresponding to the "ghost peek" event is that the pedestrian motion controller is not triggered, and the safety scenario corresponding to the emergency lane-changing incident is that the adjacent vehicle slowly and normally changes lanes. The top-down view of the constructed potential dangerous driving scenarios is shown below. Figure 10 and Figure 11 As shown, each time a random hyperparameter is applied, corresponding safe and insecure data are collected, thereby enabling rapid batch collection of scenario data.

[0135] Table 2

[0136]

[0137] S703 trains the relational graph convolutional network model based on the simulation scene information to obtain the evaluation model.

[0138] In this embodiment, the simulation scene information of the target scene is converted into a data format that can be processed by the relational graph convolutional network model using a one-hot encoding method. Optionally, the data format may include four parts: New node feature, New edge index, New edge attr, and Predict. Each row of the New node feature represents a node, and each node is processed using one-hot encoding. Each row of the New edge index represents the connection index between nodes. Each row of the New edge attr represents the connection type. Predict represents the predicted label, with 0.0 indicating safety and 1.0 indicating danger.

[0139] In this embodiment of the application, the relationship graph convolutional network model is trained using the simulated scene information of the processed target scene, and the model outputs the probability of potential risks in the current knowledge graph.

[0140] In the above-mentioned embodiments, a relational graph convolutional network model including initial global spatial feature extraction, node feature extraction, and self-attention prediction sub-model is constructed to achieve joint modeling of multi-level features of driving scenarios. By training with simulation scenario information that is highly consistent with the preset knowledge graph structure, the model can fully learn the implicit rules of spatial distribution, node attributes, and relationships of scenario elements. This not only improves the model's ability to represent complex scenarios, but also enables the evaluation model to have both global context awareness and local detail capture capabilities through end-to-end optimization. The final evaluation model exhibits stronger robustness and generalization ability in dynamic driving scenarios.

[0141] In one embodiment, a complete method for evaluating driving scenarios is provided, including:

[0142] S1, Construct a relational graph convolutional network model; the relational graph convolutional network model includes an initial global spatial feature extraction sub-model, an initial node feature extraction sub-model, and an initial self-attention prediction sub-model.

[0143] S2, obtain simulation scene information of the target scene; the information type of the simulation scene information is related to the preset knowledge graph structure.

[0144] S3: Train the relational graph convolutional network model based on the simulation scenario information to obtain the evaluation model.

[0145] S4 collects driving scenario information through multiple sensors in autonomous vehicles.

[0146] S5 generates a knowledge graph of the driving scenario based on driving scenario information and a preset knowledge graph structure; the node types of the preset knowledge graph structure include concept layer nodes, entity layer nodes, and attribute layer nodes.

[0147] S6 uses an aggregation function to aggregate information from the features of neighboring nodes and update its own node features for each node in the knowledge graph of the driving scenario, thus obtaining the updated node features.

[0148] S7 evaluates the updated node characteristics based on third-party open-source libraries to determine which nodes to retain.

[0149] S8 adds the features of the retained nodes and the updated nodes together to obtain multiple summed node features.

[0150] S9 performs feature fusion and compression on multiple summed node features to obtain global spatial structure features, and broadcasts them to the node feature extraction sub-model.

[0151] S10: Based on the node feature extraction sub-model, feature fusion processing is performed on node features and global spatial structure features to obtain processed node features.

[0152] S11, input the processed node features into the first attention mechanism layer to obtain the first processing result.

[0153] S12, the first processing result and the processed node features are input into the second attention mechanism layer, and the evaluation result of the driving scenario is obtained after linear layer processing.

[0154] In the aforementioned driving scenario evaluation method, driving scenario information is collected through multiple sensors of the autonomous vehicle; a knowledge graph of the driving scenario is generated based on the driving scenario information and a pre-defined knowledge graph structure; the node types of the pre-defined knowledge graph structure include concept layer nodes, entity layer nodes, and attribute layer nodes; and the evaluation result of the driving scenario is obtained based on the knowledge graph and the evaluation model; the evaluation model includes a global spatial feature extraction sub-model, a node feature extraction sub-model, and a self-attention prediction sub-model. By constructing a more comprehensive node type for the knowledge graph, the comprehensiveness and reliability of the knowledge graph are improved, thereby improving the accuracy of driving scenario evaluation. Furthermore, the three models—the global spatial feature extraction sub-model, the node feature extraction sub-model, and the self-attention prediction sub-model—analyze the knowledge graph to more accurately evaluate the driving scenario, further improving the accuracy of driving scenario evaluation.

[0155] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0156] Based on the same inventive concept, this application also provides a driving scenario evaluation device for implementing the driving scenario evaluation method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations of one or more driving scenario evaluation device embodiments provided below can be found in the limitations of the driving scenario evaluation method described above, and will not be repeated here.

[0157] In one embodiment, such as Figure 12 As shown, a driving scenario evaluation device is provided, including: a data acquisition module 10, a generation module 11, and an evaluation module 12, wherein:

[0158] The data acquisition module 10 is used to collect driving scene information through multiple sensors of the autonomous vehicle.

[0159] The generation module 11 is used to generate a knowledge graph of the driving scenario based on the driving scenario information and the preset knowledge graph structure; the nodes in the knowledge graph include concept layer nodes, entity layer nodes, and attribute layer nodes.

[0160] Evaluation module 12 is used to obtain the evaluation results of the driving scenario based on the knowledge graph and evaluation model of the driving scenario; the evaluation model includes a global spatial feature extraction sub-model, a node feature extraction sub-model and a self-attention prediction sub-model.

[0161] In one embodiment, the evaluation module 12 includes: a first processing unit, a second processing unit, and an evaluation unit, wherein:

[0162] The first processing unit is used to process and compress the node features corresponding to the knowledge graph of the driving scenario based on the global spatial feature extraction sub-model, obtain the global spatial structure features, and broadcast them to the node feature extraction sub-model.

[0163] The second processing unit is used to perform feature fusion processing on node features and global spatial structure features based on the node feature extraction sub-model to obtain the processed node features.

[0164] The evaluation unit is used to input the processed node features into the self-attention prediction sub-model for prediction, and obtain the evaluation results of the driving scenario.

[0165] In one embodiment, the first processing unit specifically targets each node in the knowledge graph of the driving scenario, using an aggregation function to aggregate information from the features of neighboring nodes and update its own node features to obtain updated node features; evaluates the updated node features based on a third-party open-source library to determine the nodes to be retained; and processes and compresses the retained nodes and the updated node features to obtain global spatial structure features.

[0166] In one embodiment, the first processing unit is specifically used to add the features of the retained nodes and the updated node features to obtain multiple added node features; and to fuse and compress the multiple added node features to obtain global spatial structure features.

[0167] In one embodiment, the evaluation unit described above includes a self-attention prediction sub-model comprising a first attention mechanism layer and a second attention mechanism layer, specifically used to input the processed node features into the first attention mechanism layer to obtain a first processing result;

[0168] The first processing result and the processed node features are input into the second attention mechanism layer, and after processing by the linear layer, the evaluation result of the driving scenario is obtained.

[0169] In one embodiment, the above-mentioned driving scenario evaluation device further includes: a construction module, an acquisition module, and a training module, wherein:

[0170] The building module is used to construct the relational graph convolutional network model; the relational graph convolutional network model includes an initial global spatial feature extraction sub-model, an initial node feature extraction sub-model, and an initial self-attention prediction sub-model.

[0171] The acquisition module is used to acquire simulation scene information of the target scene; the information type of the simulation scene information is related to the preset knowledge graph structure.

[0172] The training module is used to train the relational graph convolutional network model based on the simulation scenario information to obtain the evaluation model.

[0173] Each module in the aforementioned driving scenario evaluation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0174] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 13 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores evaluation data for driving scenarios. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for evaluating driving scenarios.

[0175] Those skilled in the art will understand that Figure 13 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0176] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0177] Information about the driving scenario is collected through multiple sensors in autonomous vehicles;

[0178] A knowledge graph of the driving scenario is generated based on driving scenario information and a preset knowledge graph structure; the node types of the preset knowledge graph structure include concept layer nodes, entity layer nodes, and attribute layer nodes.

[0179] The evaluation results of the driving scenario are obtained based on the knowledge graph and evaluation model of the driving scenario. The evaluation model includes a global spatial feature extraction sub-model, a node feature extraction sub-model, and a self-attention prediction sub-model.

[0180] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0181] The global spatial feature extraction sub-model processes and compresses the node features corresponding to the knowledge graph of the driving scenario to obtain global spatial structure features, which are then broadcast to the node feature extraction sub-model.

[0182] Based on the node feature extraction sub-model, feature fusion processing is performed on node features and global spatial structure features to obtain processed node features;

[0183] The processed node features are input into the self-attention prediction sub-model for prediction to obtain the evaluation results of the driving scenario.

[0184] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0185] For each node in the knowledge graph of the driving scenario, an aggregation function is used to aggregate information from the features of neighboring nodes and update the node's own features to obtain the updated node features.

[0186] The updated node characteristics are evaluated based on third-party open-source libraries to determine which nodes to retain.

[0187] The features of the retained nodes and the updated nodes are processed and compressed to obtain the global spatial structure features.

[0188] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0189] The features of the retained nodes and the updated nodes are summed to obtain multiple summed node features;

[0190] The features of multiple summed nodes are fused and compressed to obtain global spatial structure features.

[0191] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0192] The processed node features are input into the first attention mechanism layer to obtain the first processing result;

[0193] The first processing result and the processed node features are input into the second attention mechanism layer, and after processing by the linear layer, the evaluation result of the driving scenario is obtained.

[0194] In one embodiment, the processor, when executing a computer program, also performs the following steps:

[0195] Construct a relational graph convolutional network model; the relational graph convolutional network model includes an initial global spatial feature extraction sub-model, an initial node feature extraction sub-model, and an initial self-attention prediction sub-model;

[0196] Obtain simulation scene information of the target scenario; the information type of the simulation scene information is related to the preset knowledge graph structure;

[0197] The relational graph convolutional network model is trained based on the simulation scenario information to obtain the evaluation model.

[0198] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0199] Information about the driving scenario is collected through multiple sensors in autonomous vehicles;

[0200] A knowledge graph of the driving scenario is generated based on driving scenario information and a preset knowledge graph structure; the node types of the preset knowledge graph structure include concept layer nodes, entity layer nodes, and attribute layer nodes.

[0201] The evaluation results of the driving scenario are obtained based on the knowledge graph and evaluation model of the driving scenario. The evaluation model includes a global spatial feature extraction sub-model, a node feature extraction sub-model, and a self-attention prediction sub-model.

[0202] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0203] The global spatial feature extraction sub-model processes and compresses the node features corresponding to the knowledge graph of the driving scenario to obtain global spatial structure features, which are then broadcast to the node feature extraction sub-model.

[0204] Based on the node feature extraction sub-model, feature fusion processing is performed on node features and global spatial structure features to obtain processed node features;

[0205] The processed node features are input into the self-attention prediction sub-model for prediction to obtain the evaluation results of the driving scenario.

[0206] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0207] For each node in the knowledge graph of the driving scenario, an aggregation function is used to aggregate information from the features of neighboring nodes and update the node's own features to obtain the updated node features.

[0208] The updated node characteristics are evaluated based on third-party open-source libraries to determine which nodes to retain.

[0209] The features of the retained nodes and the updated nodes are processed and compressed to obtain the global spatial structure features.

[0210] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0211] The features of the retained nodes and the updated nodes are summed to obtain multiple summed node features;

[0212] The features of multiple summed nodes are fused and compressed to obtain global spatial structure features.

[0213] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0214] The processed node features are input into the first attention mechanism layer to obtain the first processing result;

[0215] The first processing result and the processed node features are input into the second attention mechanism layer, and after processing by the linear layer, the evaluation result of the driving scenario is obtained.

[0216] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0217] Construct a relational graph convolutional network model; the relational graph convolutional network model includes an initial global spatial feature extraction sub-model, an initial node feature extraction sub-model, and an initial self-attention prediction sub-model;

[0218] Obtain simulation scene information of the target scenario; the information type of the simulation scene information is related to the preset knowledge graph structure;

[0219] The relational graph convolutional network model is trained based on the simulation scenario information to obtain the evaluation model.

[0220] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, performs the following steps:

[0221] Information about the driving scenario is collected through multiple sensors in autonomous vehicles;

[0222] A knowledge graph of the driving scenario is generated based on driving scenario information and a preset knowledge graph structure; the node types of the preset knowledge graph structure include concept layer nodes, entity layer nodes, and attribute layer nodes.

[0223] The evaluation results of the driving scenario are obtained based on the knowledge graph and evaluation model of the driving scenario. The evaluation model includes a global spatial feature extraction sub-model, a node feature extraction sub-model, and a self-attention prediction sub-model.

[0224] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0225] The global spatial feature extraction sub-model processes and compresses the node features corresponding to the knowledge graph of the driving scenario to obtain global spatial structure features, which are then broadcast to the node feature extraction sub-model.

[0226] Based on the node feature extraction sub-model, feature fusion processing is performed on node features and global spatial structure features to obtain processed node features;

[0227] The processed node features are input into the self-attention prediction sub-model for prediction to obtain the evaluation results of the driving scenario.

[0228] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0229] For each node in the knowledge graph of the driving scenario, an aggregation function is used to aggregate information from the features of neighboring nodes and update the node's own features to obtain the updated node features.

[0230] The updated node characteristics are evaluated based on third-party open-source libraries to determine which nodes to retain.

[0231] The features of the retained nodes and the updated nodes are processed and compressed to obtain the global spatial structure features.

[0232] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0233] The features of the retained nodes and the updated nodes are summed to obtain multiple summed node features;

[0234] The features of multiple summed nodes are fused and compressed to obtain global spatial structure features.

[0235] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0236] The processed node features are input into the first attention mechanism layer to obtain the first processing result;

[0237] The first processing result and the processed node features are input into the second attention mechanism layer, and after processing by the linear layer, the evaluation result of the driving scenario is obtained.

[0238] In one embodiment, when the computer program is executed by a processor, it also performs the following steps:

[0239] Construct a relational graph convolutional network model; the relational graph convolutional network model includes an initial global spatial feature extraction sub-model, an initial node feature extraction sub-model, and an initial self-attention prediction sub-model;

[0240] Obtain simulation scene information of the target scenario; the information type of the simulation scene information is related to the preset knowledge graph structure;

[0241] The relational graph convolutional network model is trained based on the simulation scenario information to obtain the evaluation model.

[0242] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0243] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0244] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for evaluating driving scenarios, characterized in that, The method includes: Information about the driving scenario is collected through multiple sensors in autonomous vehicles; A knowledge graph of the driving scenario is generated based on the driving scenario information and a preset knowledge graph structure; the node types of the preset knowledge graph structure include concept layer nodes, entity layer nodes, and attribute layer nodes. The evaluation result of the driving scenario is obtained based on the knowledge graph and evaluation model of the driving scenario; the evaluation model includes a global spatial feature extraction sub-model, a node feature extraction sub-model, and a self-attention prediction sub-model.

2. The method according to claim 1, characterized in that, The step of obtaining the evaluation result of the driving scenario based on the knowledge graph and evaluation model of the driving scenario includes: The global spatial feature extraction sub-model processes and compresses the node features corresponding to the knowledge graph of the driving scenario to obtain global spatial structure features, which are then broadcast to the node feature extraction sub-model. Based on the node feature extraction sub-model, the node features and the global spatial structure features are fused to obtain the processed node features; The processed node features are input into the self-attention prediction sub-model for prediction to obtain the evaluation result of the driving scenario.

3. The method according to claim 2, characterized in that, The step of processing and compressing the node features corresponding to the knowledge graph of the driving scenario based on the global spatial feature extraction sub-model to obtain global spatial structure features includes: For each node in the knowledge graph of the driving scenario, an aggregation function is used to aggregate information from the features of neighboring nodes and update the node's own features to obtain the updated node features. The updated node characteristics are evaluated based on a third-party open-source library to determine which nodes to retain. The features of the retained nodes and the updated nodes are processed and compressed to obtain the global spatial structure features.

4. The method according to claim 3, characterized in that, The process of processing and compressing the retained nodes and the updated node features to obtain the global spatial structure features includes: The features of the retained nodes and the updated nodes are added together to obtain multiple summed node features; The multiple summed node features are fused and compressed to obtain the global spatial structure features.

5. The method according to claim 2, characterized in that, The self-attention prediction sub-model includes a first attention mechanism layer and a second attention mechanism layer. The processed node features are input into the self-attention prediction sub-model for prediction to obtain the evaluation result of the driving scenario, including: The processed node features are input into the first attention mechanism layer to obtain the first processing result; The first processing result and the processed node features are input into the second attention mechanism layer, and the evaluation result of the driving scenario is obtained after linear layer processing.

6. The method according to any one of claims 1-5, characterized in that, The method further includes: Construct a relational graph convolutional network model; the relational graph convolutional network model includes an initial global spatial feature extraction sub-model, an initial node feature extraction sub-model, and an initial self-attention prediction sub-model; Obtain simulation scene information of the target scene; the information type of the simulation scene information is related to the preset knowledge graph structure; The relationship graph convolutional network model is trained based on the simulation scenario information to obtain the evaluation model.

7. A driving scenario evaluation device, characterized in that, The device includes: The data acquisition module is used to collect driving scene information through multiple sensors in autonomous vehicles; The generation module is used to generate a knowledge graph of the driving scenario based on the driving scenario information and a preset knowledge graph structure; the nodes in the knowledge graph include concept layer nodes, entity layer nodes, and attribute layer nodes. The evaluation module is used to obtain the evaluation result of the driving scenario based on the knowledge graph and evaluation model of the driving scenario; the evaluation model includes a global spatial feature extraction sub-model, a node feature extraction sub-model, and a self-attention prediction sub-model.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images