An intelligent driving decision-making method and device based on a multi-level risk perception graph neural network

Abstract

The application discloses a kind of intelligent driving decision-making method and device based on multistage risk perception graph neural network, to solve the technical problem that the safety of driving decision is poor that existing intelligent driving decision-making method will lose the complex topological relationship between traffic participants by compressing the features of vehicle, pedestrian, signal light and other elements in traffic scene into one-dimensional vector.The method comprises obtaining observation data of the vehicle and a plurality of neighbor vehicles, and inputting a preset multistage risk perception graph neural network (including a hierarchical attention mechanism module, a three-level risk assessment module and a risk perception feature fusion module).The hierarchical attention mechanism module extracts scene perception features, and the three-level risk assessment module generates a risk score graph neural network feature.The two are fused through the risk perception feature fusion module to output an initial driving strategy, and the target driving strategy is output through preset driving constraint optimization.

Description

Technical Field

[0001] This invention relates to the field of neural network technology, and in particular to an intelligent driving decision-making method and device based on a multi-level risk perception graph neural network. Background Technology

[0002] The dynamic and complex nature of traffic environments, along with the interactivity among traffic participants, poses significant challenges to the decision-making and control systems of intelligent vehicles. Among these challenges, the uncertainty of traffic participant movement trends (e.g., possible future trajectories or intentions) has a substantial impact on the decision-making and control outcomes of intelligent vehicles. Therefore, accurately predicting the movement trends of surrounding traffic participants is crucial for the decision-making and control of intelligent vehicles in dynamic traffic environments.

[0003] In the development of intelligent driving decision-making technology, knowledge-driven, data-driven, and hybrid approaches have formed three major technological threads. Early knowledge-driven methods, centered on expert systems and fuzzy logic, addressed structured scenarios through pre-set rules, but their effectiveness gradually diminished in complex urban traffic due to insufficient rule coverage.

[0004] Most existing intelligent driving decision-making methods are based on reinforcement learning. However, these methods compress the features of elements such as vehicles, pedestrians, and traffic lights in traffic scenarios into one-dimensional vectors, which will lose the complex topological relationships between traffic participants. For example, at intersections, the structured information such as the spatial position association and driving direction coupling of straight-going vehicles, left-turning vehicles, and pedestrians crossing the road is simplified, resulting in poor safety of driving decisions. Summary of the Invention

[0005] This invention provides an intelligent driving decision-making method and device based on a multi-level risk perception graph neural network, which solves the technical problem that existing intelligent driving decision-making methods compress the features of elements such as vehicles, pedestrians, and traffic lights in traffic scenarios into one-dimensional vectors, resulting in the loss of complex topological relationships between traffic participants and thus poor driving decision safety.

[0006] The first aspect of this invention provides an intelligent driving decision-making method based on a multi-level risk perception graph neural network, comprising:

[0007] Obtain observation data of its own vehicle and observation data of multiple neighboring vehicles corresponding to its own vehicle;

[0008] The observation data of the vehicle itself and the observation data of each of the neighboring vehicles are input into a pre-set multi-level risk perception graph neural network, which includes a hierarchical attention mechanism module, a three-level risk assessment module, and a risk perception feature fusion module.

[0009] The hierarchical attention mechanism module extracts scene features based on the observation data of the vehicle itself and the observation data of each of the neighboring vehicles, and outputs scene perception features.

[0010] The three-level risk assessment module performs risk assessment based on the observation data of its own vehicle and the observation data of each of its neighboring vehicles, and generates risk score graph neural network features.

[0011] The risk score map neural network features and the scene perception features are input into the risk perception feature fusion module for feature fusion, and an initial driving strategy is output.

[0012] The initial driving decision is optimized using preset driving constraints, and a target driving strategy is output.

[0013] Optionally, the observation data includes historical trajectory data and map path data; the hierarchical attention mechanism module extracts scene features based on the observation data of its own vehicle and the observation data of each of its neighboring vehicles, and outputs scene perception features, including:

[0014] The historical trajectory data of the vehicle itself and the historical trajectory data of each of the neighboring vehicles are time-series encoded by a recurrent neural network, and the vehicle trajectory encoding features of the vehicle itself and the vehicle trajectory encoding features of each of the neighboring vehicles are output.

[0015] The vehicle trajectory encoding features of the vehicle itself and the vehicle trajectory encoding features of each of the neighboring vehicles are modeled using a temporal attention mechanism to perform long-short-term dependency modeling, and the vehicle motion trend features of the vehicle itself and the vehicle motion trend features of each of the neighboring vehicles are output.

[0016] A map encoder is used to encode the map path data of the vehicle itself and the map path data of each of the neighboring vehicles to generate map encoding features of the vehicle itself and map encoding features of each of the neighboring vehicles.

[0017] The map coding features and vehicle motion trend features of the vehicle itself are fused to generate map-constrained vehicle features of the vehicle itself.

[0018] The map coding features and vehicle motion trend features of each neighboring vehicle are fused to generate map-constrained vehicle features for each neighboring vehicle.

[0019] Spatial attention processing is performed on the map-constrained vehicle features of the vehicle itself and the map-constrained vehicle features of each of the neighboring vehicles to generate the vehicle interaction features of the vehicle itself and the vehicle interaction features of each of the neighboring vehicles.

[0020] The vehicle's own map path data and vehicle interaction features, as well as the vehicle interaction features of each of the neighboring vehicles, are fused together to output vehicle map fusion features.

[0021] The vehicle map fusion features are decoded to generate scene-aware features.

[0022] Optionally, the observation data further includes vehicle state vectors; the risk assessment using the three-level risk assessment module based on the observation data of the vehicle itself and the observation data of each of the neighboring vehicles, generating risk score graph neural network features, includes:

[0023] A vehicle dynamic graph is constructed based on the vehicle state vector of the vehicle itself and the vehicle state vectors of each of the neighboring vehicles using a pre-set adaptive adjustment strategy.

[0024] A six-dimensional feature vector is constructed from the edges in the vehicle dynamic graph, and the six-dimensional edge feature vector between the vehicle itself and each of the neighboring vehicles is output.

[0025] Based on the position coordinates in the vehicle state vector of the vehicle itself and the position coordinates in the vehicle state vectors of each of the neighboring vehicles, an instant risk assessment is performed, and the instant risk assessment value between the vehicle itself and each of the neighboring vehicles is output.

[0026] Based on the speed in the vehicle state vector of the vehicle itself, the speed in the vehicle state vectors of each of the neighboring vehicles, and the distance and relative speed modulus in each of the six-dimensional edge feature vectors, future risk prediction is performed to generate a future risk assessment value between the vehicle itself and each of the neighboring vehicles.

[0027] A comprehensive risk assessment is performed based on the position coordinates and speed in the vehicle state vector of the vehicle itself, the position coordinates and speed in the vehicle state vectors of each of the neighboring vehicles, and the six-dimensional edge feature vectors of each vehicle, and the comprehensive risk assessment value between the vehicle itself and each of the neighboring vehicles is output.

[0028] The future risk assessment values ​​between the vehicle itself and each of the neighboring vehicles are weighted, and the weighted future risk assessment value between the vehicle itself and each of the neighboring vehicles is output.

[0029] Among the immediate risk assessment value, weighted future risk assessment value, and comprehensive risk assessment value between the vehicle itself and each of the neighboring vehicles, the largest risk assessment value is selected as the target risk assessment value between the vehicle itself and the neighboring vehicles.

[0030] Based on the target risk assessment value between the vehicle itself and each of the neighboring vehicles, the distance and relative speed modulus in each of the six-dimensional edge feature vectors, and the vehicle state vector of the vehicle itself and the vehicle state vector of each of the neighboring vehicles, a normalized attention weight between the vehicle itself and each of the neighboring vehicles is generated.

[0031] The normalized attention weights and target risk assessment values ​​between the vehicle itself and each of the neighboring vehicles are weighted and aggregated to output a graph representation of the relationship between the vehicle itself and each of the neighboring vehicles.

[0032] The graph representations of the vehicle itself and each of its neighboring vehicles are input into a graph neural network, and the risk score graph neural network features are output.

[0033] Optionally, the step of constructing a vehicle dynamic graph based on the vehicle state vector of the vehicle itself and the vehicle state vectors of each of the neighboring vehicles using a preset adaptive adjustment strategy includes:

[0034] Calculate the average speed of the vehicle itself and the speeds of each of the neighboring vehicles using the speed of the vehicle itself and the speeds of each of the neighboring vehicles;

[0035] The preset adaptive adjustment strategy is used to calculate the dynamic distance threshold based on the average speed;

[0036] The vehicle and its neighboring vehicles are used as graph nodes. Based on the comparison between the distance between the vehicle and its neighboring vehicles and the dynamic distance threshold, edge connections are established between the vehicle and its neighboring vehicles to construct a vehicle dynamic graph.

[0037] Optionally, generating normalized attention weights between the vehicle and each of the neighboring vehicles based on the target risk assessment value between the vehicle and each of the neighboring vehicles, the distance and relative speed modulus in each of the six-dimensional edge feature vectors, and the vehicle state vector of the vehicle and the vehicle state vectors of the neighboring vehicles includes:

[0038] Based on the distance, relative velocity modulus, vehicle state vector of the vehicle itself, and vehicle state vector of each of the neighboring vehicles in the six-dimensional edge feature vectors, basic attention is calculated, and the basic attention score between the vehicle itself and each of the neighboring vehicles is output.

[0039] Based on the basic attention weights and target risk assessment values ​​between the vehicle itself and each of the neighboring vehicles, a risk modulation attention score is calculated between the vehicle itself and each of the neighboring vehicles.

[0040] Based on the basic attention score and target risk assessment value between the vehicle itself and each of the neighboring vehicles, output the attention adjustment factor between the vehicle itself and each of the neighboring vehicles;

[0041] The attention adjustment factor and risk-modulated attention score between the vehicle itself and each of the neighboring vehicles are multiplied respectively to output the target attention weight between the vehicle itself and each of the neighboring vehicles.

[0042] The target attention weights between the vehicle itself and each of the neighboring vehicles are normalized, and the normalized attention weights between the vehicle itself and each of the neighboring vehicles are output.

[0043] Optionally, the step of inputting the risk score map neural network features and the scene perception features into the risk perception feature fusion module for feature fusion and outputting an initial driving strategy includes:

[0044] Multi-head attention calculation is performed on the scene perception features and the risk score map neural network features to output fused multimodal features;

[0045] Adaptive gating fusion is performed on the scene perception features and the risk score graph neural network features to generate gating fusion features;

[0046] Based on the scene perception features, the fused multimodal features, and the gated fusion features, the final fused features are determined;

[0047] An initial driving strategy is generated based on the final fused features and the scene perception features.

[0048] Optionally, the model training process of the pre-set multi-level risk perception graph neural network is as follows:

[0049] Obtain observation data for model training of its own vehicle and observation data for model training of multiple neighboring vehicles corresponding to its own vehicle.

[0050] The observation data of the vehicle itself used for model training and the observation data of each of the neighboring vehicles used for model training are input into the initial multi-level risk perception graph neural network, and the initial driving strategy used for model training is output.

[0051] Substitute the initial driving strategy used for model training into the preset loss function and take the derivative to output the model gradient;

[0052] The model parameters of the initial multi-level risk perception graph neural network are updated using the model gradient to determine the intermediate multi-level risk perception graph neural network, and the number of updates is counted in real time.

[0053] Determine whether the number of updates has reached the preset number of training iterations;

[0054] If this is achieved, the intermediate multi-level risk perception graph neural network will be used as the pre-trained multi-level risk perception graph neural network.

[0055] A second aspect of the present invention provides an intelligent driving decision-making device based on a multi-level risk perception graph neural network, comprising:

[0056] A computer device provided in a third aspect of the present invention includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the intelligent driving decision-making method based on a multi-level risk perception graph neural network as described in any of the preceding claims.

[0057] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the steps of the intelligent driving decision-making method based on a multi-level risk perception graph neural network as described in any of the preceding claims.

[0058] As can be seen from the above technical solutions, the present invention has the following advantages:

[0059] The above-mentioned technical solution of the present invention provides an intelligent driving decision-making method based on a multi-level risk perception graph neural network. This method acquires observation data of the vehicle itself and observation data of multiple neighboring vehicles. The observation data of the vehicle itself and the observation data of each neighboring vehicle are input into a pre-set multi-level risk perception graph neural network. This pre-set multi-level risk perception graph neural network includes a hierarchical attention mechanism module, a three-level risk assessment module, and a risk perception feature fusion module. The hierarchical attention mechanism module extracts scene features based on the observation data of the vehicle itself and the observation data of each neighboring vehicle, outputting scene perception features. The three-level risk assessment module extracts scene features based on the observation data of the vehicle itself and the observation data of each neighboring vehicle. The system performs risk assessment on vehicle observation data, generating a risk score map neural network feature. This feature, along with scene perception features, is then input into a risk perception feature fusion module for feature fusion, outputting an initial driving strategy. Pre-set driving constraints are used to optimize the initial driving decision, outputting a target driving strategy. Based on this approach, the invention accurately captures the complex interaction relationships between itself and neighboring vehicles through a hierarchical attention mechanism, avoiding the loss of topological information caused by feature vectorization. A three-level risk assessment module refines multi-dimensional risks, combining feature fusion and driving constraint optimization, effectively solving the problem of insufficient understanding of dynamic interactions among traffic participants in existing methods, and significantly improving the safety of driving decisions. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a flowchart illustrating the steps of an intelligent driving decision-making method based on a multi-level risk perception graph neural network, as provided in Embodiment 1 of the present invention.

[0062] Figure 2 This is an overall framework diagram of an intelligent driving decision-making method based on a multi-level risk perception graph neural network provided in Embodiment 1 of the present invention;

[0063] Figure 3 This is an overall framework diagram of the three-level risk assessment module provided in Embodiment 1 of the present invention;

[0064] Figure 4 This is a schematic diagram of the six-dimensional edge feature vector provided in Embodiment 1 of the present invention;

[0065] Figure 5 This is a flowchart illustrating the steps of training a pre-set multi-level risk perception graph neural network model according to Embodiment 2 of the present invention.

[0066] Figure 6 This is a schematic diagram of the vehicle's trajectory provided in Embodiment 2 of the present invention;

[0067] Figure 7 This is a structural block diagram of an intelligent driving decision-making device based on a multi-level risk perception graph neural network, provided in Embodiment 3 of the present invention. Detailed Implementation

[0068] This invention provides an intelligent driving decision-making method and device based on a multi-level risk perception graph neural network, which solves the technical problem that existing intelligent driving decision-making methods compress the features of elements such as vehicles, pedestrians, and traffic lights in traffic scenarios into one-dimensional vectors, resulting in the loss of complex topological relationships between traffic participants and thus poor driving decision-making safety.

[0069] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0070] Please see Figure 1 , Figure 1 This is a flowchart illustrating the steps of an intelligent driving decision-making method based on a multi-level risk perception graph neural network, as provided in Embodiment 1 of the present invention.

[0071] This invention provides an intelligent driving decision-making method based on a multi-level risk perception graph neural network, comprising:

[0072] Step 101: Obtain the observation data of your own vehicle and the observation data of multiple neighboring vehicles corresponding to your own vehicle.

[0073] It should be noted that you should refer to [link / reference]. Figure 2 The overall framework of the intelligent driving decision-making method based on a multi-level risk perception graph neural network proposed in this invention adopts an end-to-end deep reinforcement learning architecture, such as... Figure 2 As shown, the entire system consists of seven core modules, forming a complete perception-decision-execution closed loop. The system receives raw traffic environment observations, processes them through multimodal feature processing and risk perception mechanisms, and ultimately outputs safe vehicle control commands.

[0074] Furthermore, this invention constructs a TrafficEnvironment module based on the CARLA simulation platform, which is responsible for generating complex and ever-changing traffic scenarios. The system is implemented using the InterSection class, supporting the dynamic generation and management of various traffic scenarios. Specifically, during system initialization, 10 observation vehicles are generated at predefined locations, with uniform vehicle types to maintain consistency. When a vehicle enters a specific trigger area, the system dynamically generates additional vehicles (i.e., multiple neighboring vehicles corresponding to the vehicle) to increase scenario complexity. The trigger condition is that the vehicle's position satisfies: and To avoid collisions, the system implements a location occupancy detection mechanism to ensure that newly generated vehicles maintain a safe distance of at least 5 meters from existing vehicles. Data output structure: The system collects data in real time through the traffic environment module and outputs standardized observation data for each frame, mainly including: 1) `neighbor_trajs`: a trajectory matrix of dimension , recording the historical state of the vehicle and its 5 nearest neighbors for 10 frames; 2) `ego_current_state`: a current state vector of dimension (i.e., vehicle state vector); 3) `neighbor_waypoints`: path planning information of dimension (map path data); 4) `gnn_features`: feature data required for graph neural network processing; 5) `temporal_data`: historical data for time-series modeling. Among these, the feature data includes: node features: vehicle position (x, y), heading angle (...). ), velocity (v) x ,v yEdge features: distance between vehicles, relative speed, heading difference, time to collision (TTC), collision risk score, and relative heading angle; Graph construction data: real-time vehicle position, speed, and acceleration used to calculate dynamic connectivity thresholds and edge weights; Historical data includes: vehicle state sequences over multiple past time steps (temporal_steps frames), each frame containing the [x, y, yaw, v] values ​​of all vehicles. x ,v y Historical trajectory (i.e., historical trajectory data).

[0075] It is worth mentioning that for dynamic vehicle management, the system calculates vehicle priority based on distance and selects the five traffic participants closest to its own vehicle as observation objects (i.e., neighboring vehicles). For dynamically generated vehicles, the system sets their speed difference to -50% to -20% and the following distance to 8-12 meters.

[0076] Step 102: Input the observation data of your own vehicle and the observation data of each neighboring vehicle into the pre-set multi-level risk perception graph neural network. The pre-set multi-level risk perception graph neural network includes a hierarchical attention mechanism module, a three-level risk assessment module, and a risk perception feature fusion module.

[0077] It should be noted that the observation data of vehicles (own vehicles and neighboring vehicles) collected in real time by the traffic environment module, namely the 6×10×5 dimension vehicle trajectory information (historical trajectory data), the current vehicle state vector (including position coordinates, speed, and heading angle), and the 18×50×2 path data (map path data), are input into a pre-built multi-level risk perception graph neural network. The system maintains an update frequency of 10Hz to ensure real-time decision-making.

[0078] Step 103: Using a hierarchical attention mechanism module, scene features are extracted based on the observation data of the vehicle itself and the observation data of each neighboring vehicle, and scene perception features are output.

[0079] It should be noted that after obtaining the observation data, the parallel feature processing stage begins. For this stage, this invention designs two parallel processing paths: one is the traditional RL (Reinforcement Learning) encoding path (i.e., using a hierarchical attention mechanism module to extract scene features based on the observation data of the vehicle itself and the observation data of each neighboring vehicle). The raw observation data is processed by a hierarchical Transformer, and 10 frames of historical data (i.e., historical trajectory data) are processed through a temporal attention mechanism to model the spatial relationships between vehicles. Finally, map information is fused to generate scene perception features. The other is the GNN (Graph Neural Network) processing path (i.e., using a three-level risk assessment module to perform risk assessment based on the observation data of the vehicle itself and the observation data of each neighboring vehicle). The neighboring trajectory data is first used to construct a dynamic graph, with the connection threshold set between 20-40 meters. Then, through 6-dimensional edge feature engineering and three-level risk assessment, the final output is a GNN feature containing a risk score (i.e., risk score graph neural network feature).

[0080] Specifically, step 103 may include the following sub-steps S31-S38:

[0081] Step S31: Use a recurrent neural network to perform time-series encoding on the historical trajectory data of the vehicle itself and the historical trajectory data of each neighboring vehicle, and output the vehicle trajectory encoding features of the vehicle itself and the vehicle trajectory encoding features of each neighboring vehicle.

[0082] Step S32: Through the temporal attention mechanism, perform long-short-term dependency modeling on the vehicle trajectory encoding features of the vehicle itself and the vehicle trajectory encoding features of each neighboring vehicle, and output the vehicle motion trend features of the vehicle itself and the vehicle motion trend features of each neighboring vehicle.

[0083] Step S33: Use a map encoder to encode the map path data of your own vehicle and the map path data of each neighboring vehicle to generate the map encoding features of your own vehicle and the map encoding features of each neighboring vehicle.

[0084] Step S34: Perform feature fusion between the map coding features and vehicle motion trend features of the vehicle itself to generate map-constrained vehicle features of the vehicle itself.

[0085] Step S35: Perform feature fusion on the map coding features and vehicle motion trend features of each neighboring vehicle to generate map-constrained vehicle features for each neighboring vehicle.

[0086] Step S36: Perform spatial attention processing on the map-constrained vehicle features of the vehicle itself and the map-constrained vehicle features of each neighboring vehicle to generate the vehicle interaction features of the vehicle itself and the vehicle interaction features of each neighboring vehicle.

[0087] Step S37: Perform feature fusion on the map path data and vehicle interaction features of the vehicle itself, as well as the vehicle interaction features of each neighboring vehicle, and output the vehicle map fusion features.

[0088] Step S38: Decode the vehicle map fusion features to generate scene-aware features.

[0089] Historical trajectory data is a vehicle state vector organized in a time series, containing 10 historical time steps. Each time step includes position coordinates (x, y), heading angle, and velocity information (v). x ,v y These 5 characteristics.

[0090] Map path data is drivable path information organized in a spatial sequence, including path selections for all vehicles (the vehicle itself and all neighboring vehicles). Each vehicle has 3 selectable paths (left lane path, current lane path, right lane path). Each path contains 50 sequentially arranged path points, and each path point contains two features: location coordinates (x, y).

[0091] It should be noted that, firstly, a recurrent neural network is used to perform temporal encoding on the historical trajectory data of the vehicle itself and the historical trajectory data of each neighboring vehicle, outputting the vehicle trajectory encoding features of the vehicle itself and the vehicle trajectory encoding features of each neighboring vehicle. Next, this invention utilizes a hierarchical attention mechanism module (Hierarchical Transformer module) to handle complex spatiotemporal relationships, achieving the transformation from raw observation to structured scene features. Specifically, firstly, the vehicle trajectory encoding features of each vehicle (i.e., the vehicle itself and neighboring vehicles) are processed through a temporal attention mechanism (temporal attention layer). Specifically, firstly, long-short-term dependency modeling is performed based on the input features, constructing a causal attention mask matrix to ensure that the model focuses only on historical information at the current moment. Then, a multi-head self-attention mechanism is used to calculate the dependencies and attention weights between each moment in the trajectory sequence. After processing with the ReLU activation function, global max pooling is used to extract the most significant feature representation from the temporal dimension, finally outputting a fixed-dimensional vehicle motion trend feature vector.

[0092] This process involves constructing a causal attention mask (lower triangular matrix) to ensure that the current moment only focuses on historical information, avoiding information leakage. The attention mask calculation process includes: 1) constructing a basic mask matrix based on the effective time step length; 2) applying the lower triangular causal mask to block future information; and 3) multiplying the two masks to obtain the final attention mask. Through a multi-head self-attention mechanism, the system can capture important temporal patterns in the trajectory sequence. Finally, global max pooling is used to extract the temporal feature representation of each vehicle, thereby obtaining the vehicle motion trend features of its own vehicle and the vehicle motion trend features of each neighboring vehicle.

[0093] Next, a map encoder (RL feature encoder) is used to receive 18×50×2 map path data, and a structured scene representation is generated through multi-layer processing. The encoder adopts a hierarchical design. First, it processes the temporal information, then models the spatial relationships to obtain the map coding features of each vehicle. After that, the map coding features of each vehicle and the vehicle motion trend features are fused to obtain the map-constrained vehicle features of each vehicle.

[0094] Furthermore, modeling the spatial relationships between vehicles is crucial for understanding traffic scenarios. The system first separates the features of the user vehicle and neighboring vehicles (map-constrained vehicle features of the user vehicle and neighboring vehicles). Then, it models the interaction relationships between vehicles through a relational attention layer, obtaining the vehicle interaction features of the user vehicle and neighboring vehicles. Specifically, the relational attention mechanism uses the user vehicle's features as the query and all vehicle features (including the user vehicle's) as key / value pairs. Attention weights are dynamically allocated based on the relative importance of vehicles, with vehicles that are close, fast, or in conflicting directions receiving higher attention.

[0095] Finally, based on the path planning information (map path data) processed by the map encoder, a feature representation of the static environment (map encoded features) is generated. Through a cross-attention mechanism, the static path information is fused with the dynamic vehicle state; that is, the map path data and interaction features of the vehicle itself, as well as the interaction features of neighboring vehicles, are fused together to obtain vehicle-map fusion features. After decoding, scene-aware features are obtained. In this cross-attention mechanism, vehicle features serve as queries, and map features serve as key-value pairs. This design allows the system to selectively focus on relevant path information based on the current traffic conditions, achieving scene-aware feature representation (scene-aware features).

[0096] Step 104: Using a three-level risk assessment module, risk assessment is performed based on the observation data of the vehicle itself and the observation data of each neighboring vehicle, generating a risk score graph neural network feature.

[0097] It should be noted that you should refer to [link / reference]. Figure 3The three-level risk assessment module proposed in this invention (Enhanced GNN with Integrated Risk Assessment) models the relationships between traffic participants through graph neural networks and integrates a multi-level risk assessment mechanism.

[0098] In this embodiment, a risk-aware fusion strategy is adopted in the feature fusion stage. A four-head attention mechanism is used to process the two feature streams in parallel, and adaptive gating is used to intelligently adjust the fusion ratio according to the current risk level, while retaining residual connections to ensure the integrity of the information.

[0099] Specifically, step 104 may include the following sub-steps S41-S410:

[0100] Step S41: Using a preset adaptive adjustment strategy, construct a vehicle dynamic graph based on the vehicle state vector of the vehicle itself and the vehicle state vectors of each neighboring vehicle.

[0101] Further, step S41 may include the following sub-steps S411-S413:

[0102] Step S411: Using the speed of your own vehicle and the speeds of all neighboring vehicles, calculate the average speed of your own vehicle and all neighboring vehicles.

[0103] Step S412: Calculate the dynamic distance threshold based on the average speed using a preset adaptive adjustment strategy;

[0104] Step S413: Use the vehicle itself and each neighboring vehicle as graph nodes, and based on the comparison between the distance between the vehicle itself and each neighboring vehicle and the dynamic distance threshold, establish edge connections between the vehicle itself and each neighboring vehicle to construct a vehicle dynamic graph.

[0105] It should be noted that this invention dynamically establishes connection relationships based on the spatial distance and relative speed between vehicles. The connection threshold adopts a preset adaptive adjustment strategy, specifically as follows:

[0106] ;

[0107] in, The average speed of all vehicles (including the vehicle itself and its neighbors' vehicles). , Let be the speed of the i-th vehicle; the base threshold is set to 20.0 meters. When the average speed of the vehicle group reaches 10 m / s, the connection threshold is increased to 40 meters to ensure safe modeling during high-speed travel.

[0108] In this embodiment, when constructing the vehicle dynamic graph, the vehicle itself and its neighbors are first defined as nodes in the graph. Next, the average speed modulus of all vehicles (including the vehicle itself and its neighbors) is calculated: the speeds of each vehicle are summed and then divided by the total number of vehicles. Using 20 meters as the base distance threshold, the average speed is divided by 10, and the result is limited to between 0 and 1 using the clip function. This result is then incremented by 1 and multiplied by the base threshold to obtain the dynamic distance threshold. If the actual distance between a vehicle and its neighbors is less than this dynamic threshold, an edge connection is established between them; otherwise, it is not. When the average speed of the vehicle group reaches 10 m / s, the threshold increases to 40 meters, allowing the vehicle connection relationships to dynamically adjust with the group speed, ensuring the safety of high-speed scene modeling.

[0109] Step S42: Construct a six-dimensional feature vector for the edges in the vehicle dynamic graph, and output the six-dimensional edge feature vector between the vehicle itself and each of its neighboring vehicles.

[0110] It should be noted that you should refer to [link / reference]. Figure 4 This invention constructs a 6-dimensional feature vector (six-dimensional edge feature vector) for each graph edge to comprehensively describe the relationship between vehicles. The dimensions of the edge features can be adjusted according to the requirements of the task. In simple tasks, only the first two dimensions can be selected. Figure 4 Red indicates that the two vehicles are too close, which will trigger a warning; yellow indicates that the distance between the two vehicles needs to be carefully monitored; and blue indicates that the distance between the two vehicles is safe.

[0111] Furthermore, the six-dimensional edge feature vector includes: 1) Distance: the straight-line distance between vehicles, reflecting spatial proximity; 2) Relative velocity magnitude: the magnitude of the velocity vector difference, measuring relative motion intensity; 3) Directional difference: the difference in vehicle heading angles, characterizing the consistency of motion direction; 4) Time-to-Collision (TTC): the predicted collision time based on the current motion state, a key safety indicator; 5) Collision risk score: a comprehensive risk quantification indicator combining distance, speed, and direction; 6) Relative heading angle: the angle between vehicle motion directions, supplementing motion relationship information. TTC calculation considers the relative position and relative speed of the vehicles. When a possible collision trajectory is detected, the system calculates the time to reach the collision point. The risk score comprehensively considers the reciprocal of the distance and relative speed; the closer the distance and the greater the relative speed, the higher the risk score.

[0112] Step S43: Perform real-time risk assessment based on the position coordinates in the vehicle state vector of your own vehicle and the position coordinates in the vehicle state vectors of each neighboring vehicle, and output the real-time risk assessment value between your own vehicle and each neighboring vehicle.

[0113] It should be noted that this invention constructs a three-level assessment system covering different time scales and risk types, using complete edge feature information: Real-time risk assessor: Employs a 6→16→1 fully connected network architecture, with input features as 6-dimensional vectors: ;in, For input to the real-time risk assessor; Let x be the x-coordinate of the position coordinate of the i-th vehicle itself; Let be the ordinate of the position coordinates of the i-th vehicle itself; Let x be the x-coordinate of the position coordinates of the j-th neighboring vehicle; Let y be the ordinate of the position coordinates of the j-th neighboring vehicle.

[0114] Furthermore, the network structure of the instant risk assessor can be represented as follows:

[0115] ;

[0116] in, The first layer weight matrix is ​​[16×6], with 6 dimensions input. Mapped to a 16-dimensional hidden layer; The first layer bias vector is [16×1]; The second layer weight matrix [1×16] maps the 16-dimensional hidden layer to the 1-dimensional output. This is the second-level bias scalar; The Sigmoid activation function has an output range of [0,1], representing the risk probability. It is a linear rectified activation function. (x) =MAX(0,x); This is a real-time risk assessment value.

[0117] In this embodiment, the present invention constructs a corresponding 6-dimensional vector based on the position coordinates in the vehicle state vector of its own vehicle and the position coordinates in the vehicle state vectors of each neighboring vehicle. The data is then input into the real-time risk assessor (which uses a 6→16→1 fully connected network architecture) to obtain the real-time risk assessment values ​​between the vehicle itself and its neighboring vehicles.

[0118] Step S44: Based on the speed in the vehicle state vector of the vehicle itself, the speed in the vehicle state vectors of each neighboring vehicle, and the distance and relative speed modulus in each six-dimensional edge feature vector, predict the future risk and generate the future risk assessment value between the vehicle itself and each neighboring vehicle.

[0119] It should be noted that this invention constructs a corresponding 6-dimensional vector based on the velocity in the vehicle's own vehicle state vector, the velocities in the vehicle state vectors of each neighboring vehicle, and the distance and relative velocity modulus in each 6-dimensional edge feature vector. This data is then input into a future risk predictor (using a 6→16→1 fully connected network architecture) to obtain future risk assessment values ​​for the vehicle itself and its neighboring vehicles. ,in, As input for the future risk predictor; Let be the velocity component of the i-th vehicle in the x-direction; Let be the velocity component of the i-th vehicle in the y-direction; Let be the velocity component of the j-th neighboring vehicle in the x-direction; Let be the velocity component of the j-th neighboring vehicle in the y-direction; Let be the distance in the six-dimensional edge feature vector between the i-th vehicle and the j-th neighbor vehicle; Let be the relative velocity modulus in the six-dimensional edge feature vector between the i-th vehicle and the j-th neighbor vehicle.

[0120] Step S45: Perform a comprehensive risk assessment based on the position coordinates and velocity in the vehicle state vector of the vehicle itself, the position coordinates and velocity in the vehicle state vectors of each neighboring vehicle, and the six-dimensional edge feature vectors, and output the comprehensive risk assessment value between the vehicle itself and each neighboring vehicle.

[0121] It should be noted that this invention constructs a corresponding 6-dimensional vector based on the position coordinates and velocity in the vehicle state vector of the vehicle itself, the position coordinates and velocities in the vehicle state vectors of each neighboring vehicle, and the 6-dimensional edge feature vectors. This data is then input into a comprehensive risk network (using a 32→16→8→1 deep architecture) to obtain a comprehensive risk assessment value between the vehicle itself and its neighboring vehicles. , As input to the integrated risk network; Let be the position vector corresponding to the position coordinates of the i-th vehicle itself; Let be the position vector corresponding to the position coordinates of the j-th neighboring vehicle; This is the velocity vector corresponding to the velocity of the i-th vehicle itself. This is the velocity vector corresponding to the speed of the j-th neighboring vehicle; Let be the relative position vector between the i-th vehicle itself and the j-th neighboring vehicle; Let be the six-dimensional edge feature vector between the i-th vehicle itself and the j-th neighboring vehicle. , Let be the distance in the six-dimensional edge feature vector between the i-th vehicle and the j-th neighboring vehicle. Let be the relative velocity magnitude in the six-dimensional edge feature vector between the i-th vehicle and its j-th neighbor vehicle. Let be the directional difference in the six-dimensional edge feature vector between the i-th vehicle and its j-th neighbor vehicle. The time-to-collision parameter in the six-dimensional edge feature vector between the i-th vehicle and its j-th neighbor vehicle is given. Let be the collision risk score in the six-dimensional edge feature vector between the i-th vehicle and its j-th neighbor vehicle. Let T be the relative heading angle in the six-dimensional edge feature vector between the i-th vehicle and the j-th neighbor vehicle, and T be the transpose.

[0122] Step S46: Weight the future risk assessment values ​​between your own vehicle and each of your neighboring vehicles, and output the weighted future risk assessment value between your own vehicle and each of your neighboring vehicles.

[0123] Step S47: Select the highest risk assessment value among the instant risk assessment value, weighted future risk assessment value, and comprehensive risk assessment value between your own vehicle and each neighbor's vehicle as the target risk assessment value between your own vehicle and each neighbor's vehicle.

[0124] It should be noted that the immediate risk assessment value, future risk assessment value, and comprehensive risk assessment value between the vehicle itself and the neighboring vehicles are processed by taking the maximum value in a tiered manner. Specifically, after applying a weighting coefficient of 0.8 to the future risk assessment value, a tiered comparison strategy is adopted. First, the immediate risk assessment value is compared with the weighted future risk assessment value and the larger one is taken. Then, this result is compared with the comprehensive risk assessment value and the maximum value is taken. This ensures that any type of high-risk situation can be effectively captured, and finally, the target risk assessment value between the vehicle itself and the neighboring vehicles is generated.

[0125] For example, if the maximum risk assessment value between your vehicle and neighbor vehicle A is the immediate risk assessment value, then this value is used as the target risk assessment value between your vehicle and neighbor vehicle A. If the maximum risk assessment value between your vehicle and neighbor vehicle B is the weighted future risk assessment value, then this value is used as the target risk assessment value between your vehicle and neighbor vehicle B.

[0126] Step S48: Generate normalized attention weights between the vehicle and its neighbors based on the target risk assessment values ​​between the vehicle and each neighboring vehicle, the distance and relative speed modulus in each six-dimensional edge feature vector, and the vehicle state vector of the vehicle and the vehicle state vectors of each neighboring vehicle.

[0127] Further, step S48 may include the following sub-steps S481-S485:

[0128] Step S481: Calculate basic attention based on the distance, relative velocity modulus, vehicle state vector of the vehicle itself, and vehicle state vector of each neighboring vehicle in each six-dimensional edge feature vector, and output the basic attention score between the vehicle itself and each neighboring vehicle.

[0129] Step S482: Based on the basic attention weights between the vehicle itself and each neighboring vehicle and the target risk assessment value, calculate the risk modulation attention score between the vehicle itself and each neighboring vehicle.

[0130] Step S483: Based on the basic attention scores and target risk assessment values ​​between your own vehicle and each neighboring vehicle, output the attention adjustment factor between your own vehicle and each neighboring vehicle.

[0131] Step S484: Multiply the attention adjustment factor and risk-modulated attention score between the vehicle itself and each neighboring vehicle respectively, and output the target attention weight between the vehicle itself and each neighboring vehicle.

[0132] Step S485: Normalize the target attention weights between the vehicle itself and each of its neighboring vehicles, and output the normalized attention weights between the vehicle itself and each of its neighboring vehicles.

[0133] Step S49: Weight and aggregate the normalized attention weights and target risk assessment values ​​between the vehicle itself and each of its neighboring vehicles, and output the graph representation between the vehicle itself and each of its neighboring vehicles.

[0134] Step S410: Input the graph representations of the vehicle itself and each of its neighboring vehicles into the graph neural network, and output the risk score graph neural network features.

[0135] It should be noted that the Level 3 risk assessment module employs a risk-modulated graph attention mechanism to output risk score graph neural network features based on the target risk assessment values ​​between the vehicle and its neighbors, the distance and relative speed modulus in each of the six-dimensional edge feature vectors, and the vehicle state vectors of both the vehicle and its neighbors. Specifically, this process involves integrating the risk score into the graph attention calculation to achieve feature learning for risk perception. The basic attention calculation is as follows:

[0136] ;

[0137] in, Let be the base attention score between the i-th vehicle itself and the j-th neighboring vehicle; Use the LeakyReLU activation function; This is the transpose of the learnable attention weight vector of the vehicle itself. Let be the vehicle state vector of the i-th vehicle itself; This is the transpose of the learnable attention weight vector of neighboring vehicles. Let be the vehicle state vector of the j-th neighboring vehicle; This is the transpose of the learnable attention weight vector for edge features; The 2D edge features used in graph attention. That is, the distance and relative velocity modulus in the six-dimensional edge eigenvector.

[0138] Furthermore, risk-modulated attention is calculated as follows:

[0139] ;

[0140] in, The risk modulation attention score between the i-th vehicle and the j-th neighboring vehicle is defined, ranging from [0,1]. The target risk assessment value between the i-th vehicle and the j-th neighbor vehicle is 1.0, which represents the basic weight; 2.0 represents the risk amplification factor.

[0141] Furthermore, this invention optimizes through an adaptive attention modulator:

[0142] ;

[0143] in, For adaptive attention adjustment parameters; This represents the instantaneous risk assessment value between the i-th vehicle and its j-th neighboring vehicle.

[0144] Furthermore, based on the adaptive attention adjustment parameters, the attention adjustment factor is determined:

[0145] ;

[0146] in, Let be the attention adjustment factor (adjustment factor) between the i-th own vehicle and the j-th neighbor vehicle, ranging from [0,1]. , , This represents the weight matrix of a three-layer network; , , This represents the bias vector of a three-layer network; This is the activation function.

[0147] Furthermore, the calculation process for the target attention weights is as follows: ,in, Let be the target attention weight between the i-th vehicle and the j-th neighboring vehicle.

[0148] Furthermore, segmented softmax is used to normalize the target attention weights between the vehicle itself and each of its neighboring vehicles. Softmax calculation is performed by grouping by source node, and the normalized attention weights between the vehicle itself and each of its neighboring vehicles are output. This process is specifically as follows:

[0149] ;

[0150] in, The normalized attention weights between the i-th vehicle itself and the j-th neighboring vehicle; This is the maximum value among all target attention weights (for numerical stability); Let be the target attention weight between the i-th vehicle itself and the l-th neighboring vehicle; Let be the target attention weight between the i-th vehicle itself and the k-th neighboring vehicle; It is the set of neighbors of a node, that is, the set of neighboring vehicles corresponding to its own vehicle; This is a numerical stability parameter, with a value of [value missing]. .

[0151] Furthermore, the normalized attention weights and target risk assessment values ​​between the vehicle itself and each of its neighbors are weighted and aggregated to output a graph representation between the vehicle itself and each of its neighbors; the graph representation between the vehicle itself and each of its neighbors is then input into a graph neural network (GNN) to output a risk score graph neural network feature.

[0152] Step 105: Input the risk score map neural network features and scene perception features into the risk perception feature fusion module for feature fusion, and output the initial driving strategy.

[0153] It should be noted that the Risk-Aware FeatureFusion module proposed in this invention is used to intelligently fuse GNN features (risk score graph neural network features) and RL encoded features (scene-aware features), employing a combination of multi-head attention and gating mechanisms. Multi-head attention fusion: Four attention heads are used for parallel processing, with each head having a key-value dimension of 16.

[0154] Specifically, step 105 may include the following sub-steps S51-S54:

[0155] Step S51: Perform multi-head attention calculation on scene perception features and risk score map neural network features, and output fused multimodal features;

[0156] Step S52: Perform adaptive gating fusion on scene perception features and risk score map neural network features to generate gating fusion features;

[0157] Step S53: Determine the final fused features based on scene-aware features, multimodal features, and gated fusion features;

[0158] Step S54: Generate an initial driving strategy based on the final fused features and scene perception features.

[0159] It should be noted that the calculation process for fusing multimodal features is as follows:

[0160] ;

[0161] in, To fuse multimodal features; Features for scene perception. , The dimension corresponding to the scene-aware features; For risk score graph neural network features, , The dimension corresponding to the features of the risk score graph neural network; This is a multi-head attention mechanism.

[0162] Furthermore, this invention utilizes a gating mechanism to dynamically adjust the feature fusion ratio based on the risk level, that is, to perform adaptive gating fusion of scene-aware features and risk score graph neural network features:

[0163] ;

[0164] in, This is a gating fusion feature; For layer normalization; The risk perception features (i.e., 32-dimensional risk perception features) are obtained through processing by the risk perception layer: , This is the weight vector of the risk processing layer. For layer normalization, This is the risk score vector for all surrounding vehicles (i.e., the vector composed of the target risk assessment values ​​of all neighboring vehicles). This is the bias vector for the risk processing layer.

[0165] Furthermore, based on the gating fusion characteristics, the gating value is determined:

[0166] ;

[0167] in, This is the gate value, ranging from [0,1]. For the bias scalar of the gated network; represents the weight vector of the gating network.

[0168] Furthermore, based on the gating value, the fusion of multimodal features, and scene-aware features, the final fused features are determined. :

[0169] ;

[0170] Furthermore, residual connections are implemented through a projection layer to ensure the integrity of information transmission and output the initial driving strategy. :

[0171] ;

[0172] in, It is a fully connected processing layer; As a projection layer, ensure dimension matching.

[0173] Step 106: Optimize the initial driving decision using preset driving constraints and output the target driving strategy.

[0174] It should be noted that this invention establishes a complete safety assurance system to ensure the safety of autonomous driving. A risk assessment mechanism is the foundation of this safety assurance. The system sets a general risk threshold of 0.8 and an emergency risk threshold of 0.9, enabling real-time monitoring of the safety of current decisions. More importantly, it considers not only collision risks but also comprehensively assesses various potential threats such as deviation from the path. A tiered constraint strategy employs different responses based on the risk level. When a general risk is detected, the system implements progressive speed limits and steering constraints; in emergency situations, it activates forced braking and emergency stop protection. Simultaneously, the system also sets a minimum distance protection based on following distance, which serves as the final safety baseline. The motion control execution section is responsible for translating the safety strategy into specific control actions. Steering control ensures the precision of the steering wheel angle, acceleration control ensures the smoothness of throttle operation, braking control provides tiered braking force, and lane change decisions ensure the safety of lane changes. The entire control system works in coordination, ensuring both driving smoothness and driving safety.

[0175] Specifically, the pre-set driving constraints are expressed as follows: ,in, This is a smoothing function; This is the distance constraint function; For risk constraint functions; The preset minimum safe distance (such as following distance or side collision avoidance distance) is the "hard threshold" for physical safety; The system provides the initial driving decision after constraints. The risk constraint function performs real-time risk assessment based on the risk score output by a graph neural network. Using preset safety thresholds of 0.8 and emergency thresholds of 0.9, it progressively limits vehicle speed and steering actions when high risk is detected. The distance constraint function calculates the distance between vehicles in real-time based on the position information in the vehicle's state vector. Combined with a preset minimum safe distance threshold, it dynamically adjusts the forward speed when the following distance is too close. The smoothing function uses a weighted average method with a smoothing factor of 0.7 to smooth the constrained decision actions, transforming abrupt control commands into a continuous and smooth sequence of operations, preventing vehicles from engaging in sudden stops or other undesirable driving behaviors. These three constraint functions work collaboratively through software algorithms to form a complete safety decision optimization process.

[0176] Furthermore, the constrained initial driving decision (the constrained decision action) and the initial driving decision are smoothed to obtain the target driving strategy. :

[0177] ;

[0178] in, This is a smoothing factor with a value of 0.7.

[0179] Finally, the execution control phase translates the target driving strategy into specific control commands. The motion control module is responsible for mapping and transforming the action space, integrating path planning for dynamic waypoint selection, and ultimately achieving control through the vehicle's execution system.

[0180] For comparison of technical effectiveness, existing technologies can be used as a reference. Current risk assessment technologies use a combination of deterministic computation and probabilistic estimation to quantify the degree of danger. They make preliminary risk judgments through environmental perception and dynamically update the risk assessment results by combining real-time sensor observation data, comprehensively considering multiple factors such as environmental conditions, traffic participant behavior, and road infrastructure. In terms of decision optimization technology, deep reinforcement learning algorithms such as Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC) have been widely used in the field of autonomous driving. Multi-agent systems can handle complex multi-vehicle interaction scenarios, and attention mechanism technology, by focusing on key areas and targets, has shown good technical performance in core tasks such as vehicle detection and trajectory prediction.

[0181] Currently, several technologies are relatively close to the current level: one category is reinforcement learning decision-making methods, which address the problem of scarce reward signals during the learning process by identifying key safety states and establishing connections; another category combines deep learning and predictive control to enable the system to continuously learn and improve. Another category is safety constraint techniques, such as using control obstacle functions to define safety boundaries for vehicles, or directly embedding safety conditions into the learning algorithm. A third category is attention mechanism modeling methods, which can simultaneously process spatial, temporal, and frequency information, or use multi-head attention to analyze the complex relationships in traffic scenarios. While these technologies each have their advantages, they still have significant shortcomings in establishing multi-layered risk perception systems, deeply integrating graph neural networks with risk assessment, and achieving truly real-time safety intervention.

[0182] Based on the above, the shortcomings of existing driving decision-making systems can be mainly divided into four parts:

[0183] 1) Disadvantages of traditional decision-making methods based on reinforcement learning: relying solely on vectorized state representations makes it impossible to effectively model the complex topological relationships and dynamic interactions among traffic participants; the lack of explicit modeling of traffic scene graph structure information leads to insufficient understanding of spatial relationships between vehicles; and the limited feature representation capabilities make it difficult to capture fine-grained interaction patterns in multi-vehicle cooperation and competition.

[0184] 2) Disadvantages of traditional attention mechanisms: Attention weight allocation is mainly based on feature similarity and does not incorporate risk factors for modulation; it cannot prioritize dangerous areas and key traffic participants in high-risk scenarios; it lacks risk perception guidance and the matching degree between attention distribution and safety needs is not high.

[0185] 3) Disadvantages of existing security constraint technologies: They adopt a hard constraint method with fixed thresholds, lacking the ability to adaptively adjust based on real-time risk assessment; they cannot take differentiated intervention measures with corresponding strengths according to different risk levels; and they lack a hierarchical risk response mechanism, making it impossible to handle multiple security threats in a refined manner.

[0186] 4) Disadvantages of traditional environmental modeling methods: They mainly rely on Euclidean distance for neighbor selection, without considering the logical relationships and interaction strength between vehicles; they lack in-depth modeling of time-space coupling characteristics; and they cannot effectively handle the uncertainties and complex dependencies in dynamic traffic scenarios.

[0187] To address the aforementioned problems, this invention proposes an intelligent driving decision-making method based on a multi-level risk perception graph neural network. It utilizes an enhanced graph neural network to construct a space-time graph structure among traffic participants, integrating immediate risk, future risk, and comprehensive risk assessment mechanisms. Through a risk-modulated attention mechanism and an adaptive reinforcement learning strategy, it achieves intelligent driving control and safety assurance in complex traffic scenarios. Specifically, 1) it introduces graph neural network modeling: modeling the traffic scenario as a dynamic graph structure to effectively capture the topological relationships and interaction patterns between vehicles; 2) it constructs a multi-level risk perception system: designing a three-level assessment mechanism for immediate risk, future risk, and comprehensive risk; 3) it designs a risk-modulated attention mechanism: integrating risk scores into attention calculation to achieve feature learning for risk perception; 4) it establishes a hierarchical safety constraint system: providing differentiated safety intervention strategies based on risk levels; and 5) it enhances edge features: designing 6-dimensional enhanced edge features to comprehensively describe the complex relationships between vehicles.

[0188] In this embodiment of the invention, an intelligent driving decision-making method based on a multi-level risk perception graph neural network is provided. The method acquires observation data of the vehicle itself and observation data of multiple neighboring vehicles. The observation data of the vehicle itself and the observation data of each neighboring vehicle are input into a pre-set multi-level risk perception graph neural network, which includes a hierarchical attention mechanism module, a three-level risk assessment module, and a risk perception feature fusion module. The hierarchical attention mechanism module extracts scene features based on the observation data of the vehicle itself and the observation data of each neighboring vehicle, outputting scene perception features. The three-level risk assessment module extracts scene features based on the observation data of the vehicle itself and the observation data of each neighboring vehicle. The system performs risk assessment on vehicle observation data, generating a risk score map neural network feature. This feature, along with scene perception features, is input into a risk perception feature fusion module for feature fusion, outputting an initial driving strategy. Pre-set driving constraints are then used to optimize the initial driving decision, outputting a target driving strategy. Based on this approach, the invention accurately captures the complex interaction relationships between itself and neighboring vehicles through a hierarchical attention mechanism, avoiding the loss of topological information caused by feature vectorization. A three-level risk assessment module refines multi-dimensional risks, combining feature fusion and driving constraint optimization, effectively solving the problem of insufficient understanding of dynamic interactions among traffic participants in existing methods, and significantly improving the safety of driving decisions.

[0189] For better explanation, refer to Figure 5 The diagram illustrates the steps of training a pre-set multi-level risk perception graph neural network model according to Embodiment 2 of the present invention. This process may include the following steps:

[0190] Step 501: Obtain observation data for model training of your own vehicle and observation data for model training of multiple neighboring vehicles corresponding to your own vehicle.

[0191] Step 502: Input the observation data of your own vehicle used for model training and the observation data of each neighboring vehicle used for model training into the initial multi-level risk perception graph neural network, and output the initial driving strategy used for model training.

[0192] Step 503: Substitute the initial driving strategy used for model training into the preset loss function and calculate the derivative to output the model gradient.

[0193] Step 504: Update the model parameters of the initial multi-level risk perception graph neural network using model gradients, determine the intermediate multi-level risk perception graph neural network, and count the number of updates in real time.

[0194] Step 505: Determine whether the number of updates has reached the preset number of training iterations.

[0195] Step 506: If the condition is met, use the intermediate multi-level risk perception graph neural network as the pre-trained multi-level risk perception graph neural network.

[0196] It should be noted that for the model training phase, this invention proposes the SAC (Soft Actor-Critic) decision system, which employs a Soft Actor-Critic reinforcement learning framework. The actor network is primarily responsible for generating specific driving decisions. Its policy network specifically handles the continuous action space, enabling it to output precise control commands. Notably, the system introduces an entropy regularization mechanism, which effectively balances exploration and exploitation, making the policy more robust. Furthermore, this network can simultaneously process traditional features and GNN-enhanced features, demonstrating comprehensive input processing capabilities. The critic network employs a dual-Q network structure, using two independent value networks to evaluate the quality of actions; this design significantly improves the stability of the learning process. The system also includes a target network, using soft updates to progressively adjust parameters, ensuring stable convergence during training. Additionally, an experience replay mechanism is integrated, enabling more efficient utilization of historical samples.

[0197] Regarding the training components, this invention implements priority experience replay, allowing important samples to be learned first, significantly improving training efficiency. The learning rate is automatically adjusted according to the training progress, and batch normalization ensures the stability of the entire network training process.

[0198] Furthermore, for the pre-defined loss function (SAC loss function), a collaborative optimization strategy is adopted, including three loss functions that are logically related to the algorithm: Q-network loss, policy loss, and temperature parameter loss. Different optimizers are used to update the parameters and output the model gradient. Specifically, the Q-network loss trains the value evaluation network, the policy loss guides policy optimization based on the Q-network output and the temperature parameter α, and the temperature parameter loss adaptively adjusts the temperature coefficient according to the policy entropy to balance exploration and exploitation. The three loss functions form an algorithmic closed loop through shared state-action information and policy output. In implementation, different optimizers are used to update the parameters to ensure training stability.

[0199] The Q-network loss is specifically as follows:

[0200] ;

[0201] in, This is the vehicle state vector used for model training of the current Q-network at time t. and the initial driving strategy used for model training at the current moment. Value estimation; For instant rewards; Discount factor; The target Q-network uses the vehicle state vector for model training at the next time step (t+1). and the initial driving strategy used for model training in the next moment. Value estimation; This is the entropy regularization coefficient; The action probability output by the policy network; For the sampled state in the experience playback buffer D ,action ,award The state at the next moment Calculate the expectation.

[0202] Furthermore, regarding the strategy loss, specifically:

[0203] ;

[0204] in, The current state of sampling the experience playback buffer D. and strategies Output action Calculate the mathematical expectation; The action probability output by the policy network; This is the vehicle state vector used for model training of the current Q-network at time t. and the initial driving strategy used for model training at the current moment. Value estimate.

[0205] Furthermore, regarding the loss of temperature parameters, specifically:

[0206] ;

[0207] in, as a strategy Output action Calculate the mathematical expectation; The target entropy value; This refers to the temperature parameter.

[0208] Furthermore, if the number of model updates has not reached the preset number of training iterations, the intermediate multi-level risk perception graph neural network is used as a new initial multi-level risk perception graph neural network. The soft actor-critic reinforcement learning algorithm is then used to generate new observation data for the vehicle itself and new observation data for multiple neighboring vehicles based on the output initial driving strategy used for model training. Step 502 is then executed until the number of model updates reaches the preset number of training iterations. The intermediate multi-level risk perception graph neural network determined when the number of model updates reaches the preset number of training iterations is then used as the trained preset multi-level risk perception graph neural network.

[0209] It is worth mentioning that, during the decision generation phase, the SAC system generates driving actions based on these enhanced features. The actor network outputs continuous speed control [-1,1] and steering control [-1,1], the critic network provides value assessments to guide policy optimization, and entropy regularization maintains a moderate level of exploratory capability.

[0210] In the experiment, the method of this invention was compared with the traditional method on the CARLA (Car Learning to Act) simulation platform. Each method was trained for 1M steps, followed by testing. Each test consisted of 100 rounds, and the tests were conducted three times. The average performance indicators are shown in Table 1 below:

[0211] Table 1 Average Indicators

[0212]

[0213] The experimental results are shown in Table 1, which verify the effectiveness of the present invention in risk prediction and safe driving in complex environments.

[0214] For example, please refer to Figure 6Assuming the scenario is a city intersection, the scenario description is as follows: a vehicle needs to complete a left turn at a three-way intersection and a straight-ahead maneuver at a crossroads in the Carla Town10HD_Opt map, with multiple vehicles and pedestrians around. The specific implementation process is as follows: Sensor data acquisition: acquiring vehicle state (x, y, θ, vx, vy) and surrounding environment information; State encoding: using Ego_Neighbours_Encoder (i.e., a pre-built multi-level risk perception graph neural network) to process the relationship between the vehicle and 5 surrounding vehicles; Hierarchical processing: Temporal layer: analyzing the motion trajectory of 10 historical frames; Relationship layer: modeling the interaction relationships between vehicles; Decision layer: generating a left turn decision; Action output: outputting speed and steering control commands; Safety check: verifying the safety of the action and making necessary corrections; Experimental results: in 100 tests, the success rate reached 89%, and the collision rate was only 6%. Figure 6 The solid green line (wp1) represents the vehicle's first path (path 1), the dashed green line (wp2) represents the vehicle's second path (path 2), the blue dots represent road points, and the red dots (Ego) represent the vehicle's initial position.

[0215] In this embodiment of the invention, by simultaneously acquiring observation data from its own vehicle and multiple neighboring vehicles, the dynamic interaction information of multiple subjects in the traffic scene can be comprehensively captured, avoiding the one-sidedness of scene understanding caused by a single perspective. Relying on the initial multi-level risk perception graph neural network to process multi-vehicle data, the model's ability to model the topological relationship between vehicles can be fully utilized, enabling the model to deeply learn the complex relationships between traffic participants from the training stage. By using the derivative of the loss function and gradient updates to achieve iterative optimization of parameters, an end-to-end training closed loop is formed, allowing the model to continuously correct policy deviations during continuous learning. At the same time, using a preset number of training iterations as the termination condition ensures the sufficiency of model training to uncover potential patterns in the data, while controlling the iteration scale to avoid overfitting. The final model can more accurately adapt to the dynamic characteristics of actual traffic scenes, improving the reliability and safety of driving strategies.

[0216] Please see Figure 7 , Figure 7 This is a structural block diagram of an intelligent driving decision-making device based on a multi-level risk perception graph neural network, provided in Embodiment 3 of the present invention.

[0217] This invention provides an intelligent driving decision-making device based on a multi-level risk perception graph neural network, comprising:

[0218] The acquisition module 701 is used to acquire the observation data of its own vehicle and the observation data of multiple neighboring vehicles corresponding to its own vehicle.

[0219] The input module 702 is used to input the observation data of its own vehicle and the observation data of each neighboring vehicle into the pre-set multi-level risk perception graph neural network. The pre-set multi-level risk perception graph neural network includes a hierarchical attention mechanism module, a three-level risk assessment module, and a risk perception feature fusion module.

[0220] The extraction module 703 is used to extract scene features based on the observation data of its own vehicle and the observation data of each neighboring vehicle using a hierarchical attention mechanism module, and output scene perception features.

[0221] The assessment module 704 is used to perform risk assessment based on the observation data of its own vehicle and the observation data of each neighboring vehicle using the three-level risk assessment module, and generate risk score graph neural network features.

[0222] The fusion module 705 is used to input the risk score map neural network features and scene perception features into the risk perception feature fusion module for feature fusion and output the initial driving strategy.

[0223] The optimization module 706 is used to optimize the initial driving decision using preset driving constraints and output the target driving strategy.

[0224] Furthermore, the observation data includes historical trajectory data and map path data; the extraction module 703 is specifically used for:

[0225] The historical trajectory data of the vehicle itself and the historical trajectory data of each neighboring vehicle are time-series encoded by a recurrent neural network, and the vehicle trajectory encoding features of the vehicle itself and each neighboring vehicle are output.

[0226] By using a temporal attention mechanism, long-term and short-term dependency modeling is performed on the vehicle trajectory encoding features of the vehicle itself and the vehicle trajectory encoding features of each neighboring vehicle, and the vehicle motion trend features of the vehicle itself and the vehicle motion trend features of each neighboring vehicle are output.

[0227] The map encoder is used to encode the map path data of the vehicle itself and the map path data of each neighboring vehicle to generate the map encoding features of the vehicle itself and the map encoding features of each neighboring vehicle.

[0228] The map coding features and vehicle motion trend features of the vehicle itself are fused to generate map-constrained vehicle features of the vehicle itself.

[0229] The map coding features and vehicle motion trend features of each neighboring vehicle are fused to generate map-constrained vehicle features for each neighboring vehicle.

[0230] Spatial attention processing is performed on the map-constrained vehicle features of the vehicle itself and the map-constrained vehicle features of each neighboring vehicle to generate the vehicle interaction features of the vehicle itself and the vehicle interaction features of each neighboring vehicle.

[0231] The vehicle map path data and vehicle interaction features of its own vehicle, as well as the vehicle interaction features of each neighboring vehicle, are fused together to output vehicle map fusion features.

[0232] The vehicle map fusion features are decoded to generate scene-aware features.

[0233] Furthermore, the observation data also includes vehicle state vectors; the evaluation module 704 includes:

[0234] The first submodule is used to construct a vehicle dynamic graph based on the vehicle state vector of its own vehicle and the vehicle state vectors of each neighboring vehicle using a preset adaptive adjustment strategy.

[0235] The second submodule is used to construct a six-dimensional feature vector for the edges in the vehicle dynamic graph and output the six-dimensional edge feature vector between the vehicle itself and each of its neighboring vehicles.

[0236] The third submodule is used to perform real-time risk assessment based on the position coordinates in the vehicle state vector of its own vehicle and the position coordinates in the vehicle state vectors of each neighboring vehicle, and output the real-time risk assessment value between its own vehicle and each neighboring vehicle.

[0237] The fourth submodule is used to predict future risks based on the speed in the vehicle state vector of its own vehicle, the speed in the vehicle state vectors of each neighboring vehicle, and the distance and relative speed modulus in each six-dimensional edge feature vector, and to generate future risk assessment values ​​between its own vehicle and each neighboring vehicle.

[0238] The fifth submodule is used to perform a comprehensive risk assessment based on the position coordinates and speed in the vehicle state vector of its own vehicle, the position coordinates and speed in the vehicle state vectors of each neighboring vehicle, and the feature vectors of each six-dimensional edge, and outputs the comprehensive risk assessment value between its own vehicle and each neighboring vehicle.

[0239] The sixth submodule is used to weight the future risk assessment values ​​between the vehicle itself and each of its neighbors' vehicles, and output the weighted future risk assessment value between the vehicle itself and each of its neighbors' vehicles.

[0240] The seventh submodule is used to select the highest risk assessment value among the instant risk assessment value, weighted future risk assessment value, and comprehensive risk assessment value between the vehicle itself and all neighboring vehicles as the target risk assessment value between the vehicle itself and all neighboring vehicles.

[0241] The eighth submodule is used to generate normalized attention weights between the vehicle itself and each neighboring vehicle based on the target risk assessment value between the vehicle itself and each neighboring vehicle, the distance and relative speed modulus in each six-dimensional edge feature vector, as well as the vehicle state vector of the vehicle itself and the vehicle state vectors of each neighboring vehicle.

[0242] The ninth submodule is used to weight and aggregate the normalized attention weights and target risk assessment values ​​between the vehicle itself and each of its neighbors, and output the graph representation between the vehicle itself and each of its neighbors.

[0243] The tenth submodule is used to input the graph representations of the vehicle itself and its neighbors into the graph neural network, and output the risk score graph neural network features.

[0244] Furthermore, the first submodule is specifically used for:

[0245] Calculate the average speed of your own vehicle and the speeds of all neighboring vehicles using your own vehicle's speed and the speeds of all neighboring vehicles.

[0246] A pre-set adaptive adjustment strategy is adopted to calculate the dynamic distance threshold based on the average speed. The vehicle itself and each neighboring vehicle are used as graph nodes. Based on the comparison between the distance between the vehicle itself and each neighboring vehicle and the dynamic distance threshold, edge connections are established between the vehicle itself and each neighboring vehicle to construct a vehicle dynamic graph.

[0247] Furthermore, the seventh submodule is specifically used for:

[0248] Basic attention is calculated based on the distance, relative velocity magnitude, vehicle state vector of the vehicle itself, and vehicle state vector of each neighboring vehicle in each six-dimensional edge feature vector, and the basic attention score between the vehicle itself and each neighboring vehicle is output.

[0249] Based on the basic attention weights and target risk assessment values ​​between the vehicle itself and each of its neighboring vehicles, calculate the risk modulation attention score between the vehicle itself and each of its neighboring vehicles.

[0250] Based on the basic attention score and target risk assessment value between its own vehicle and each of its neighboring vehicles, the system outputs the attention adjustment factor between its own vehicle and each of its neighboring vehicles.

[0251] Multiply the attention modulation factor and risk modulated attention score between the vehicle itself and each of its neighbors, respectively, and output the target attention weight between the vehicle itself and each of its neighbors.

[0252] Normalize the target attention weights between the vehicle itself and each of its neighbors, and output the normalized attention weights between the vehicle itself and each of its neighbors.

[0253] Furthermore, the fusion module 705 is specifically used for:

[0254] Multi-head attention computation is performed on scene perception features and risk score map neural network features to output fused multimodal features;

[0255] Adaptive gating fusion of scene perception features and risk score graph neural network features is performed to generate gating fusion features.

[0256] The final fused features are determined based on scene-aware features, multimodal fusion features, and gated fusion features.

[0257] An initial driving strategy is generated based on the final fused features and scene-aware features.

[0258] In one optional device embodiment, it further includes:

[0259] The first module is used to acquire observation data for model training of its own vehicle and observation data for model training of multiple neighboring vehicles corresponding to its own vehicle.

[0260] The second module is used to input the observation data of its own vehicle used for model training and the observation data of each neighboring vehicle used for model training into the initial multi-level risk perception graph neural network, and output the initial driving strategy used for model training.

[0261] The third module is used to substitute the initial driving strategy used for model training into the preset loss function and calculate the derivative to output the model gradient;

[0262] The fourth module is used to update the model parameters of the initial multi-level risk perception graph neural network using model gradients, determine the intermediate multi-level risk perception graph neural network, and count the number of updates in real time.

[0263] The fifth module is used to determine whether the number of updates has reached the preset number of training iterations;

[0264] The sixth module is used to, if necessary, use the intermediate multi-level risk perception graph neural network as a pre-trained multi-level risk perception graph neural network.

[0265] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, modules, and sub-modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0266] This invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program; when the computer program is executed by the processor, the processor performs the steps of the intelligent driving decision-making method based on a multi-level risk perception graph neural network as described in any of the above embodiments.

[0267] This invention also provides a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implement the steps of the intelligent driving decision-making method based on a multi-level risk perception graph neural network as described in any of the above embodiments.

[0268] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0269] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0270] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent driving decision-making method based on a multi-level risk perception graph neural network, characterized in that, include: Obtain observation data of its own vehicle and observation data of multiple neighboring vehicles corresponding to its own vehicle; The observation data of the vehicle itself and the observation data of each of the neighboring vehicles are input into a pre-set multi-level risk perception graph neural network. The pre-set multi-level risk perception graph neural network includes a hierarchical attention mechanism module, a three-level risk assessment module, and a risk perception feature fusion module. The observation data includes historical trajectory data and map path data. The hierarchical attention mechanism module extracts scene features based on the observation data of the vehicle itself and the observation data of each neighboring vehicle, and outputs scene perception features, including: The historical trajectory data of the vehicle itself and the historical trajectory data of each of the neighboring vehicles are time-series encoded by a recurrent neural network, and the vehicle trajectory encoding features of the vehicle itself and the vehicle trajectory encoding features of each of the neighboring vehicles are output. The vehicle trajectory encoding features of the vehicle itself and the vehicle trajectory encoding features of each of the neighboring vehicles are modeled using a temporal attention mechanism to perform long-short-term dependency modeling, and the vehicle motion trend features of the vehicle itself and the vehicle motion trend features of each of the neighboring vehicles are output. A map encoder is used to encode the map path data of the vehicle itself and the map path data of each of the neighboring vehicles to generate map encoding features of the vehicle itself and map encoding features of each of the neighboring vehicles. The map coding features and vehicle motion trend features of the vehicle itself are fused to generate map-constrained vehicle features of the vehicle itself. The map coding features and vehicle motion trend features of each neighboring vehicle are fused to generate map-constrained vehicle features for each neighboring vehicle. Spatial attention processing is performed on the map-constrained vehicle features of the vehicle itself and the map-constrained vehicle features of each of the neighboring vehicles to generate the vehicle interaction features of the vehicle itself and the vehicle interaction features of each of the neighboring vehicles. The vehicle's own map path data and vehicle interaction features, as well as the vehicle interaction features of each of the neighboring vehicles, are fused together to output vehicle map fusion features. The vehicle map fusion features are decoded to generate scene-aware features; The three-level risk assessment module performs risk assessment based on the observation data of its own vehicle and the observation data of each of its neighboring vehicles, and generates risk score graph neural network features. The risk score map neural network features and the scene perception features are input into the risk perception feature fusion module for feature fusion, and an initial driving strategy is output. The initial driving strategy is optimized using preset driving constraints to output the target driving strategy.

2. The intelligent driving decision-making method based on a multi-level risk perception graph neural network according to claim 1, characterized in that, The observation data also includes vehicle state vectors; the three-level risk assessment module performs risk assessment based on the observation data of its own vehicle and the observation data of each of its neighboring vehicles, generating risk score graph neural network features, including: A vehicle dynamic graph is constructed based on the vehicle state vector of the vehicle itself and the vehicle state vectors of each of the neighboring vehicles using a pre-set adaptive adjustment strategy. A six-dimensional feature vector is constructed from the edges in the vehicle dynamic graph, and the six-dimensional edge feature vector between the vehicle itself and each of the neighboring vehicles is output. Based on the position coordinates in the vehicle state vector of the vehicle itself and the position coordinates in the vehicle state vectors of each of the neighboring vehicles, an instant risk assessment is performed, and the instant risk assessment value between the vehicle itself and each of the neighboring vehicles is output. Based on the speed in the vehicle state vector of the vehicle itself, the speed in the vehicle state vectors of each of the neighboring vehicles, and the distance and relative speed modulus in each of the six-dimensional edge feature vectors, future risk prediction is performed to generate a future risk assessment value between the vehicle itself and each of the neighboring vehicles. A comprehensive risk assessment is performed based on the position coordinates and speed in the vehicle state vector of the vehicle itself, the position coordinates and speed in the vehicle state vectors of each of the neighboring vehicles, and the six-dimensional edge feature vectors of each vehicle, and the comprehensive risk assessment value between the vehicle itself and each of the neighboring vehicles is output. The future risk assessment values ​​between the vehicle itself and each of the neighboring vehicles are weighted, and the weighted future risk assessment value between the vehicle itself and each of the neighboring vehicles is output. Among the immediate risk assessment value, weighted future risk assessment value, and comprehensive risk assessment value between the vehicle itself and each of the neighboring vehicles, the largest risk assessment value is selected as the target risk assessment value between the vehicle itself and the neighboring vehicles. Based on the target risk assessment value between the vehicle itself and each of the neighboring vehicles, the distance and relative speed modulus in each of the six-dimensional edge feature vectors, and the vehicle state vector of the vehicle itself and the vehicle state vector of each of the neighboring vehicles, a normalized attention weight between the vehicle itself and each of the neighboring vehicles is generated. The normalized attention weights and target risk assessment values ​​between the vehicle itself and each of the neighboring vehicles are weighted and aggregated to output a graph representation of the relationship between the vehicle itself and each of the neighboring vehicles. The graph representations of the vehicle itself and each of its neighboring vehicles are input into a graph neural network, and the risk score graph neural network features are output.

3. The intelligent driving decision-making method based on a multi-level risk perception graph neural network according to claim 2, characterized in that, The step of constructing a vehicle dynamic graph using a pre-set adaptive adjustment strategy based on the vehicle state vector of the vehicle itself and the vehicle state vectors of each of its neighboring vehicles includes: Calculate the average speed of the vehicle itself and the speeds of each of the neighboring vehicles using the speed of the vehicle itself and the speeds of each of the neighboring vehicles; The preset adaptive adjustment strategy is used to calculate the dynamic distance threshold based on the average speed; The vehicle and its neighboring vehicles are used as graph nodes. Based on the comparison between the distance between the vehicle and its neighboring vehicles and the dynamic distance threshold, edge connections are established between the vehicle and its neighboring vehicles to construct a vehicle dynamic graph.

4. The intelligent driving decision-making method based on a multi-level risk perception graph neural network according to claim 2, characterized in that, The step of generating normalized attention weights between the vehicle and each of its neighboring vehicles based on the target risk assessment value between the vehicle and each of its neighboring vehicles, the distance and relative speed modulus in each of the six-dimensional edge feature vectors, and the vehicle state vectors of the vehicle and each of its neighboring vehicles, includes: Based on the distance, relative velocity modulus, vehicle state vector of the vehicle itself, and vehicle state vector of each of the neighboring vehicles in the six-dimensional edge feature vectors, basic attention is calculated, and the basic attention score between the vehicle itself and each of the neighboring vehicles is output. Based on the basic attention weights and target risk assessment values ​​between the vehicle itself and each of the neighboring vehicles, a risk modulation attention score is calculated between the vehicle itself and each of the neighboring vehicles. Based on the basic attention score and target risk assessment value between the vehicle itself and each of the neighboring vehicles, output the attention adjustment factor between the vehicle itself and each of the neighboring vehicles; The attention adjustment factor and risk-modulated attention score between the vehicle itself and each of the neighboring vehicles are multiplied respectively to output the target attention weight between the vehicle itself and each of the neighboring vehicles. The target attention weights between the vehicle itself and each of the neighboring vehicles are normalized, and the normalized attention weights between the vehicle itself and each of the neighboring vehicles are output.

5. The intelligent driving decision-making method based on a multi-level risk perception graph neural network according to claim 2, characterized in that, The step of inputting the risk score map neural network features and the scene perception features into the risk perception feature fusion module for feature fusion and outputting an initial driving strategy includes: Multi-head attention calculation is performed on the scene perception features and the risk score map neural network features to output fused multimodal features; Adaptive gating fusion is performed on the scene perception features and the risk score graph neural network features to generate gating fusion features; Based on the scene perception features, the fused multimodal features, and the gated fusion features, the final fused features are determined; An initial driving strategy is generated based on the final fused features and the scene perception features.

6. The intelligent driving decision-making method based on a multi-level risk perception graph neural network according to claim 1, characterized in that, The model training process of the pre-set multi-level risk perception graph neural network is as follows: Obtain observation data for model training of its own vehicle and observation data for model training of multiple neighboring vehicles corresponding to its own vehicle. The observation data of the vehicle itself used for model training and the observation data of each of the neighboring vehicles used for model training are input into the initial multi-level risk perception graph neural network, and the initial driving strategy used for model training is output. Substitute the initial driving strategy used for model training into the preset loss function and take the derivative to output the model gradient; The model parameters of the initial multi-level risk perception graph neural network are updated using the model gradient to determine the intermediate multi-level risk perception graph neural network, and the number of updates is counted in real time. Determine whether the number of updates has reached the preset number of training iterations; If this is achieved, the intermediate multi-level risk perception graph neural network will be used as the pre-trained multi-level risk perception graph neural network.

7. An intelligent driving decision-making device based on a multi-level risk perception graph neural network, applied to the intelligent driving decision-making method based on a multi-level risk perception graph neural network as described in claim 1, characterized in that, include: The acquisition module is used to acquire the observation data of its own vehicle and the observation data of multiple neighboring vehicles corresponding to its own vehicle; The input module is used to input the observation data of the vehicle itself and the observation data of each of the neighboring vehicles into a pre-set multi-level risk perception graph neural network. The pre-set multi-level risk perception graph neural network includes a hierarchical attention mechanism module, a three-level risk assessment module, and a risk perception feature fusion module. The extraction module is used to extract scene features based on the observation data of its own vehicle and the observation data of each of the neighboring vehicles using the hierarchical attention mechanism module, and output scene perception features. The assessment module is used to perform risk assessment based on the observation data of its own vehicle and the observation data of each of the neighboring vehicles using the three-level risk assessment module, and generate risk score graph neural network features. The fusion module is used to input the risk score map neural network features and the scene perception features into the risk perception feature fusion module for feature fusion and output an initial driving strategy; The optimization module is used to optimize the initial driving strategy using preset driving constraints and output the target driving strategy.

8. A computer device, characterized in that, The system includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the intelligent driving decision-making method based on a multi-level risk perception graph neural network as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the intelligent driving decision-making method based on a multi-level risk perception graph neural network as described in any one of claims 1-6.

Images