Multi-intention interactive game decision method and system for automatic driving path planning

Abstract

The application provides a multi-intention interactive game decision method and system for automatic driving path planning, which fuses current environment information and current state information by acquiring the current state information and the current environment information, extracts state features and intention features of a moving target from current scene features, fuses the intention features of the state features of the moving target, constructs a game relationship graph, calculates attention weights between an automatic driving vehicle and the moving target, weights and aggregates a value vector by a prediction probability and the attention weights of different preset intention features executed by the automatic driving vehicle, performs feature splicing, nonlinear conversion and decoding on the aggregated features, evaluates each game candidate trajectory, and selects a game candidate trajectory with the maximum game value score as the automatic driving path; the application reduces uncertainty in the interactive process, avoids the risk of sudden braking or collision caused by misjudgment, and has stronger environmental adaptability and generalization ability.

Description

Technical Field

[0001] This application relates to the field of autonomous driving technology, and in particular to a multi-intent interactive game decision-making method and system for autonomous driving path planning. Background Technology

[0002] In complex urban traffic flows, autonomous vehicles need to interact frequently and closely with human drivers, cyclists, and others, such as when making unprotected left turns, merging lanes, or navigating narrow roads. Traditional methods (e.g., physics-driven methods, behavior-based methods) typically treat other traffic participants as dynamic obstacles and perform simple reactive planning. However, this approach performs poorly in scenarios requiring game theory, often leading to vehicles being overly conservative (being "bullied") or making incorrect decisions.

[0003] In some cutting-edge research, such as FlowPlanner (an autonomous driving trajectory planning algorithm based on a flow matching model), attention has been paid to modeling interactive behavior. However, it mainly enhances the model's understanding of the scene implicitly, without explicitly performing intent inference and game-theoretic decision-making. While methods like GameFormer (a hierarchical interaction modeling framework that combines game theory with the Transformer architecture) introduce game theory ideas, their models are complex and highly dependent on the later refinement of rules, resulting in limited generalization ability.

[0004] Therefore, there is an urgent need for a learnable planning method that can explicitly understand the intentions of others and make interactive decisions. Summary of the Invention

[0005] In view of the shortcomings of the prior art described above, this application provides a multi-intent interactive game decision-making method and system for autonomous driving path planning to solve the above-mentioned technical problems.

[0006] According to one aspect of the embodiments of this application, a multi-intent interactive game decision-making method for autonomous driving path planning is provided. The multi-intent interactive game decision-making method is implemented based on a multi-intent interactive game decision-making and planning model. The multi-intent interactive game decision-making method includes: acquiring the current state information of an autonomous vehicle and the current environmental information surrounding the autonomous vehicle; the current environmental information includes: multi-view images, radar point clouds, and map vectors; fusing the current environmental information and the current state information with multi-modal features to obtain current scene features; extracting state features and intent features of moving targets from the current scene features, and fusing the state features and intent features of the moving targets to obtain interactive target features with intent information; the moving targets include: pedestrians and vehicles; constructing a game relationship graph with the autonomous vehicle and the moving targets as nodes, the lines connecting the nodes as edges, and configured with edge weights; the edge weights are determined by the state features and intent features of the nodes; the state features of the autonomous vehicle are determined by the... The current state information is determined, and the intent features of the autonomous vehicle are determined by navigation planning information. Using the map vector as a mapping condition and based on different preset intent features, the current state features of the autonomous vehicle are mapped into a set of query vectors, and the interactive target features are mapped into key vectors and value vectors. Based on the set of query vectors, the key vectors, and the edge weights, the attention weight between the autonomous vehicle and the moving target is calculated. The value vectors are weighted and aggregated using the predicted probability of the autonomous vehicle executing different preset intent features and the attention weights to obtain aggregated features. The aggregated features are then subjected to feature concatenation, nonlinear transformation, and decoding to obtain multiple game candidate trajectories. The predicted probability of the autonomous vehicle executing different preset intent features is predicted based on the current scene features. The safety, comfort, and game success rate of each game candidate trajectory are evaluated to obtain a game value score for each game candidate trajectory. The game candidate trajectory with the highest game value score among all game value scores is selected as the autonomous driving path.

[0007] According to another aspect of the embodiments of this application, a multi-intent interactive game decision-making system for autonomous driving path planning is also provided, comprising: an information acquisition module for acquiring current state information of an autonomous vehicle and current environmental information surrounding the autonomous vehicle; the current environmental information includes: multi-view images, radar point clouds, and map vectors; a feature fusion module for fusing the current environmental information and current state information into multi-modal features to obtain current scene features; an intent prediction and encoding module for extracting state features and intent features of moving targets from the current scene features, and fusing the state features and intent features of moving targets to obtain interactive target features with intent information; the moving targets include: pedestrians and vehicles; and a game relationship construction module for constructing a game relationship graph with the autonomous vehicle and moving targets as nodes, the lines connecting the nodes as edges, and edge weights configured; the edge weights are determined by the state features and intent features of the nodes; the state features of the autonomous vehicle are determined by the current state information. The intent features of the autonomous vehicle are determined by navigation planning information. The multimodal planning module uses map vectors as mapping conditions and different preset intent features to map the current state features of the autonomous vehicle into a set of query vectors and the interactive target features into key vectors and value vectors. Based on a set of query vectors, key vectors, and edge weights, it calculates the attention weight between the autonomous vehicle and the moving target. The value vectors are weighted and aggregated using the predicted probabilities of the autonomous vehicle executing different preset intent features and the attention weights to obtain aggregated features. These aggregated features are then concatenated, nonlinearly transformed, and decoded to obtain multiple game candidate trajectories. The predicted probabilities of the autonomous vehicle executing different preset intent features are predicted based on the current scene features. The game decision module evaluates the safety, comfort, and success rate of each game candidate trajectory, obtaining a game value score for each trajectory. The game candidate trajectory with the highest game value score among all game value scores is selected as the autonomous driving path.

[0008] The beneficial effects of this application are as follows: This application obtains the current state information of the autonomous vehicle and the current environmental information around the autonomous vehicle, performs multimodal feature fusion on the current environmental information and the current state information to obtain the current scene features, extracts the state features and intent features of the moving target from the current scene features, and fuses the state features and intent features of the moving target to obtain interactive target features with intent information. A game relationship graph is constructed with the autonomous vehicle and the moving target as nodes, the lines between nodes as edges, and edge weights configured. Map vectors are used as mapping conditions, and based on different preset intent features, the current state features of the autonomous vehicle are mapped to a set of query vectors, and the interactive target features are mapped to key vectors and value vectors. Based on a set of query vectors, key vectors, and edge weights, the attention weight between the autonomous vehicle and the moving target is calculated. The value vectors are weighted and aggregated by the predicted probability of the autonomous vehicle executing different preset intent features and the attention weights to obtain aggregated features. The aggregated features are then subjected to feature concatenation, nonlinear transformation, and decoding to obtain multiple game candidate trajectories. The safety, comfort, and game success rate of each game candidate trajectory are evaluated to obtain each... The game value score of candidate trajectories is used to select the candidate trajectory with the highest game value score as the autonomous driving path. This process extracts the state and intent features of moving targets from the current scene features and generates interactive target features with intent information. This explicit intent understanding mechanism significantly reduces uncertainty in the interaction process and avoids the risk of sudden braking or collisions due to misjudgment. Based on the game relationship graph and dynamic attention mechanism, the influence of different moving targets on the autonomous vehicle's decision-making can be dynamically measured. By generating multiple candidate trajectories based on different preset intent features and comprehensively evaluating them from three dimensions—safety, comfort, and game success rate—it effectively overcomes the "being bullied" or deadlock problems caused by traditional reactive planning. By integrating multimodal feature fusion, intent inference, game decision-making, and trajectory decoding into a unified framework, complex inter-module coupling and rule dependencies are avoided, enabling the model to automatically learn the optimal interaction strategy through data-driven methods. Compared to methods that heavily rely on manually designed rules, this solution has stronger environmental adaptability and generalization ability, and can cope with diverse interaction scenarios in complex urban traffic flows.

[0009] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0010] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings:

[0011] Figure 1 This is a schematic diagram illustrating an exemplary system architecture as shown in an exemplary embodiment of this application;

[0012] Figure 2 This is a flowchart illustrating a multi-intent interactive game decision-making method for autonomous driving path planning, as shown in an exemplary embodiment of this application.

[0013] Figure 3 This is an exemplary embodiment of the multi-intent interactive game decision-making and planning model illustrated in this application;

[0014] Figure 4 This is an exemplary embodiment of the present application illustrating the internal structure of an interactive game planner;

[0015] Figure 5 This is a schematic diagram illustrating an intent classification and interactive game scenario, as shown in an exemplary embodiment of this application;

[0016] Figure 6 This is a block diagram illustrating a multi-intent interactive game decision-making system for autonomous driving path planning, as shown in an exemplary embodiment of this application. Detailed Implementation

[0017] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0018] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0019] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the present application. However, it will be apparent to those skilled in the art that embodiments of the present application may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the present application.

[0020] Figure 1 This is a schematic diagram illustrating an exemplary system architecture as shown in an exemplary embodiment of this application.

[0021] Reference Figure 1 As shown, the system architecture may include a data acquisition device 101 and a processing device 102. The processing device 102 may be at least one of a desktop graphics processing unit (GPU) computer, a GPU computing cluster, or a neural network computer. Those skilled in the art can use this processing device 102 to acquire the current state information of the autonomous vehicle and the current environmental information surrounding the vehicle. They can then perform multimodal feature fusion of the current environmental information and the current state information to obtain current scene features. From these current scene features, they can extract the state features and intent features of the moving target. Furthermore, they can fuse the state features and intent features of the moving target to obtain interactive target features with intent information. A game theory graph is constructed with the autonomous vehicle and the moving target as nodes, the lines connecting the nodes as edges, and edge weights configured. Map vectors are used as mapping conditions, and based on different preset intent features, the current state features of the autonomous vehicle are processed. The feature mapping is a set of query vectors, mapping the interactive target features to key vectors and value vectors. Based on a set of query vectors, key vectors, and edge weights, the attention weight between the autonomous vehicle and the moving target is calculated. The value vectors are weighted and aggregated by the predicted probability of the autonomous vehicle executing different preset intention features and the attention weights to obtain aggregated features. The aggregated features are then concatenated, nonlinearly transformed, and decoded to obtain multiple game candidate trajectories. The safety, comfort, and game success rate of each game candidate trajectory are evaluated to obtain a game value score for each game candidate trajectory. The game candidate trajectory with the highest game value score among all game value scores is used as the autonomous driving path. The acquisition device 101 is used to collect the current state information of the autonomous vehicle and the current environmental information around the autonomous vehicle, and provides the current state information and current environmental information to the processing device 102 for processing.

[0022] The implementation details of the technical solutions in the embodiments of this application are described in detail below:

[0023] Figure 2This is a flowchart illustrating a multi-intent interactive game decision-making method for autonomous driving path planning, as shown in an exemplary embodiment of this application. (Refer to...) Figure 2 As shown, the multi-intent interactive game decision-making method for autonomous driving path planning includes at least steps S210 to S260, which are described in detail below:

[0024] In step S210, the current state information of the autonomous vehicle and the current environmental information surrounding the autonomous vehicle are acquired. In one embodiment of this application, the current environmental information includes: multi-view images, radar point clouds, and map vectors. The multi-view images are acquired by onboard cameras at different locations on the vehicle, and can reflect the environmental information surrounding the vehicle from multiple angles. The radar point clouds are acquired by onboard LiDAR or millimeter-wave radar, and can accurately provide the distance, relative speed, and obstacle contour information between surrounding moving targets and the autonomous vehicle. The map vectors contain static road information such as the current road topology, lane line positions, traffic rules, speed limit information, and traffic light status, which are used to provide road constraints for subsequent path planning. The current state information includes the autonomous vehicle's current driving speed, acceleration, steering wheel angle, vehicle attitude, and current positioning coordinates in the global map. This information can accurately reflect the current driving basic state of the autonomous vehicle and is an important constraint basis for subsequent trajectory generation.

[0025] In step S220, the current environment information and the current state information are fused using multimodal features to obtain the current scene features. In one embodiment of this application, the current scene features fuse the current environment information and the current state information, ensuring that information from different modalities can be complementary and fused. This preserves both the richness of visual semantics and the accuracy of point cloud ranging and the constraints of map information, enabling a comprehensive and structured expression of the core information of the current traffic scene and helping subsequent modules to accurately understand the scene interaction relationships.

[0026] In step S230, the state features and intent features of the moving target are extracted from the current scene features, and the state features and intent features of the moving target are fused to obtain interactive target features with intent information. In one embodiment of this application, the moving target includes: pedestrians and vehicles. The process of extracting the state features of the moving target from the current scene features includes: extracting the state features of the moving target from the current scene features, determining the intent features of the moving target based on the state features of the moving target, and using the state features and intent features of the moving target as output. The process of extracting the state features of the moving target from the current scene features is implemented through an intent classifier.

[0027] In step S240, a game theory graph is constructed with autonomous vehicles and moving targets as nodes, lines connecting the nodes as edges, and edge weights configured. In one embodiment of this application, the edge weights are determined by the state features and intent features of the nodes; the state features of the autonomous vehicle are determined by the current state information, and the intent features of the autonomous vehicle are determined by the navigation planning information. The process of determining the state features of the autonomous vehicle from the current state information includes: preprocessing and normalizing the current state information; inputting the preprocessed and normalized current state information into a state encoder (e.g., a multilayer perceptron) to obtain a state feature vector; and performing temporal fusion on the state feature vector to obtain the state features. The process of determining the intent features of the autonomous vehicle from the navigation planning information includes: determining the driving instructions of the autonomous vehicle at the current location point from the navigation planning information, such as left turn, right turn, straight ahead, etc.; encoding the current location point, mapping the driving instructions to one-hot encoding, and concatenating the location encoding and the driving instruction encoding to form the intent features of the autonomous vehicle. The formula for calculating the edge weights is as follows:

[0028] Equation (1)

[0029] in, Represents a node With nodes Edge weights between them This represents the learnable adaptive equilibrium parameters. Represents a node With nodes The normalized Euclidean distance between them Represents the distance normalization coefficient. Represents a node With nodes The relative speed between them Represents the velocity normalization coefficient. The hyperparameter representing the intent weights controls the degree to which intent features influence edge weights. Characteristics representing the intent of autonomous vehicles. Indicating the intentional characteristics of a moving target, This represents the probability of intent conflict. The probability of intent conflict is calculated based on the Euclidean distance or cosine distance between the intent features of the autonomous vehicle and the intent features of the moving target. It can also be obtained by inputting the intent features of the autonomous vehicle and the intent features of the moving target into a conflict discriminator (or an interaction risk predictor).

[0030] In step S250, using map vectors as mapping conditions and based on different preset intent features, the current state features of the autonomous vehicle are mapped to a set of query vectors, and the interactive target features are mapped to key vectors and value vectors. Based on the set of query vectors, key vectors, and edge weights, the attention weight between the autonomous vehicle and the moving target is calculated. The value vectors are weighted and aggregated using the predicted probability of the autonomous vehicle executing different preset intent features and the attention weights to obtain aggregated features. The aggregated features are then subjected to feature concatenation, nonlinear transformation, and decoding to obtain multiple game-theoretic candidate trajectories. In one embodiment of this application, the predicted probability of the autonomous vehicle executing different preset intent features is obtained based on the current scene features. Specifically, the current scene features are input into a classifier based on a fully connected layer to obtain scores for the autonomous vehicle executing different preset intent features. The scores for the autonomous vehicle executing different preset intent features are normalized to obtain the predicted probability of the autonomous vehicle executing different preset intent features. The formula for calculating the attention weight is as follows:

[0031] Equation (2)

[0032] in, Indicates at node With nodes Between the first Attention weights calculated from pre-defined intent features Indicates the first A query vector with predefined intent features. Represents the key vector. The dimension of the key vector. Represents a node With nodes The edge weights between them are scalar biases.

[0033] The expression for the aggregated feature is as follows:

[0034] Equation (3)

[0035] in, Indicates aggregation features, Indicates that the autonomous vehicle is performing the first The predicted probability of a pre-defined intention feature. Indicates at node With nodes Between the first Attention weights calculated from pre-defined intent features Indicates the total number of moving targets. Represents a value vector;

[0036] No. The expression for the query vector of a preset intent feature is as follows:

[0037] Equation (4)

[0038] in, Indicates the first A query vector with predefined intent features. Indicates querying the projection matrix. Indicates the current state characteristics. Indicates the first A preset intention feature, A local map vector representing the location of an autonomous vehicle. This indicates a splicing operation. This represents a multilayer perceptron used for query vectors;

[0039] The expression for the key vector is as follows:

[0040] Equation (5)

[0041] in, Represents the key vector. Represents the key projection matrix. This represents a multilayer perceptron used for key vectors. Indicates the characteristics of the interaction target. A local map vector representing the location of a moving target. Indicates a splicing operation;

[0042] The expression for the value vector is as follows:

[0043] Equation (6)

[0044] in, Represents a value vector. Represents the value projection matrix, This represents a multilayer perceptron used for value vectors. Indicates the characteristics of the interaction target. A local map vector representing the location of a moving target. This indicates a splicing operation.

[0045] In one embodiment of this application, the process of performing feature concatenation, nonlinear transformation, and decoding on aggregated features to obtain multiple game candidate trajectories includes: concatenating aggregated features with features of different intentions to obtain concatenated vectors with different intentions; performing nonlinear transformation on the concatenated vectors with different intentions (e.g., through a multilayer perceptron and activation function) to obtain multiple nonlinear transformation vectors; and decoding the multiple nonlinear transformation vectors (e.g., through a Transformer decoder) to obtain multiple game candidate trajectories.

[0046] In step S260, the safety, comfort, and success rate of each candidate trajectory are evaluated to obtain a game value score for each candidate trajectory. The candidate trajectory with the highest game value score among all the candidate trajectories is selected as the autonomous driving path. In one embodiment of this application, by extracting the state features and intent features of moving targets from the current scene features and generating interactive target features with intent information, this explicit intent understanding mechanism greatly reduces the uncertainty in the interaction process and avoids the risk of sudden braking or collision due to misjudgment. By using a game relationship graph and dynamic attention mechanism, the influence of different moving targets on the decision-making of autonomous vehicles can be dynamically measured. By generating multiple game candidate trajectories based on different preset intent features and comprehensively evaluating them from three dimensions of safety, comfort, and game success rate, the "being bullied" or deadlock problem caused by traditional reactive planning is effectively overcome. By integrating multimodal feature fusion, intent inference, game decision-making, and trajectory decoding into a unified framework, complex inter-module coupling and rule dependence are avoided, enabling the model to automatically learn the optimal interaction strategy through data-driven methods. Compared with methods that heavily rely on manually designed rules, this solution has stronger environmental adaptability and generalization ability and can cope with diverse interaction scenarios in complex urban traffic flows.

[0047] In one embodiment of this application, the process of fusing current environmental information and current state information using multimodal features to obtain current scene features includes:

[0048] The current state information is encoded to obtain a current state feature vector and modulation parameters. In one embodiment of this application, the process of encoding the current state information to obtain a current state feature vector and modulation parameters includes: inputting the current state information into a linear projection layer to obtain an initial embedding vector; inputting the initial embedding vector into a shared encoder (composed of a multilayer perceptron or Transformer module) to obtain deep features; inputting the deep features into a first independent prediction head to obtain a current state feature vector; and inputting the deep features into a second independent prediction head to obtain modulation parameters (e.g., scaling factor, bias factor).

[0049] Feature extraction and perspective transformation are performed on multi-view images to obtain bird's-eye view features. In one embodiment of this application, the process of extracting features and transforming perspectives on multi-view images to obtain bird's-eye view features includes: extracting image features from multi-view images to obtain image features under different perspectives; transforming the image features under different perspectives into bird's-eye view space based on the LSS (Lift-Splat-Shoot) method to obtain bird's-eye view space features; and encoding the bird's-eye view space features through a BEV encoder to obtain the bird's-eye view features.

[0050] Voxelization and sparse convolution are performed on the radar point cloud to obtain radar bird's-eye view features. In one embodiment of this application, the process of voxelization and sparse convolution of the radar point cloud includes: uniformly dividing the radar point cloud into different three-dimensional voxel grids according to a preset bird's-eye view resolution; for each non-empty voxel grid, extracting the original features such as coordinates, reflection intensity, and radial velocity of all points in the grid, and calculating the mean, variance, and extreme values ​​of the original features as the overall features of the current voxel; using three-dimensional sparse convolution to downsample and aggregate the overall features of the current voxel, thereby compressing the three-dimensional voxel features into a two-dimensional bird's-eye view form to obtain radar bird's-eye view features.

[0051] The map vectors are vectorized and rasterized to obtain the map bird's-eye view features. In one embodiment of this application, the process of vectorizing and rasterizing the map vectors includes: first, classifying different types of map elements, including lane center lines, lane boundaries, stop lines, pedestrian crossings, intersection directional arrows, curbs, traffic signs, etc.; for each type of map element, it is converted into a vector sequence composed of multiple ordered polyline segments, and the nodes (i.e., coordinate points) on each ordered polyline segment are encoded in both coordinates and categories to obtain encoded features; the encoded features are input into a Transformer encoder to extract the context features of the vector sequence to obtain the vectorized features of each map element; then, all vectorized features are projected onto a two-dimensional grid in the current autonomous vehicle coordinate system according to a preset bird's-eye view resolution to complete the rasterization process. If multiple map element features exist in the same grid, average pooling or max pooling operations are used to obtain the final rasterized features of the grid; finally, the overall features after rasterization are further extracted and fused through multi-layer two-dimensional convolution to obtain the final map bird's-eye view features.

[0052] The bird's-eye view features from the image, radar, and map are stitched together, and the stitched features are modulated using the current state feature vector and modulation parameters to obtain the current scene features. In one embodiment of this application, the calculation formula for the current scene features includes:

[0053] Equation (7)

[0054] in, Indicates the characteristics of the current scene. Indicates the scaling factor. Represents the feature vector of the current state. Represents the bird's-eye view features of the image. Indicates features in a radar bird's-eye view. Indicates the features of a bird's-eye view of a map. This represents the bias coefficient. The modulation coefficients include the scaling coefficient and the bias coefficient.

[0055] In one embodiment of this application, the process of fusing the state features and intention features of a moving target to obtain interactive target features with intention information includes:

[0056] The state features of the moving target are mapped to a first query vector, and the intention features of the moving target are mapped to a first key vector and a first value vector. A cross-attention operation is then performed between the first query vector, the first key vector, and the first value vector to obtain intermediate fusion features. In one embodiment of this application, the cross-attention operation enables the state features of the moving target to actively focus on the effective information of the intention features of the moving target, improving the accuracy of subsequent trajectory prediction and decision-making. The intermediate fusion features are obtained by taking the dot product of the first query vector and the first key vector and performing normalization to obtain an attention score, and then using the attention score to weightedly sum the first value vector.

[0057] The probability distribution corresponding to the intention features of the moving target is used to weight and aggregate the intermediate fusion features to obtain the final fusion feature. In one embodiment of this application, the probability distribution corresponding to the intention features of the moving target is predicted based on the current environmental features and the historical trajectory of the moving target. The prediction model can be implemented using a model based on a recurrent neural network. Weighting and aggregating the intermediate fusion features using the probability distribution corresponding to the intention features of the moving target is beneficial to the final fusion feature.

[0058] To better adapt to interaction scenarios with different probabilities of intent occurrence, avoid low-probability intents occupying too much feature weight, improve the ability to capture the true interaction tendencies of moving targets, and reduce decision-making errors caused by intent uncertainty.

[0059] The final fused features and state features are residually connected and layer-normalized to obtain the interactive target features. In one embodiment of this application, the process of residually connecting and layer-normalizing the final fused features and state features includes: residually connecting the final fused features and state features to obtain a residual fusion result; and performing layer-normalization on the residual fusion result to obtain the interactive target features.

[0060] In one embodiment of this application, the process of evaluating the safety, comfort, and success rate of each candidate game trajectory to obtain a game value score for each candidate trajectory includes:

[0061] The acceleration and longitudinal travel distance of each candidate trajectory are obtained, along with collision prediction results for each candidate trajectory. In one embodiment of this application, the collision prediction results include: collision occurred, safe, and collision hazard. These are determined based on a comparison of the collision time of the candidate trajectory with a minimum hazard time threshold and a maximum safe time threshold. The collision time of the candidate trajectory is detected or estimated using methods such as geometric detection of spatiotemporal occupancy grids, analytical calculation based on kinematic models, and temporal difference learning and probability estimation. The longitudinal travel distance characterizes the projected length of the candidate trajectory on a reference line, which is a smooth and drivable baseline path generated based on the lane centerline. The acceleration of each candidate trajectory is obtained by taking the third derivative of the corresponding mathematical expression for that candidate trajectory.

[0062] Based on the collision prediction results of each candidate trajectory, a safety score is determined for each candidate trajectory. In one embodiment of this application, the candidate trajectory... The formula for calculating the security score is as follows:

[0063] Equation (8)

[0064] in, Representing the candidate trajectories of the game Safety rating Representing the candidate trajectories of the game Collision time, This represents the minimum dangerous time threshold. This represents the maximum safe time threshold; other scenarios with collision risks are considered as candidate trajectories in the game theory. The collision time is greater than or equal to the minimum danger time threshold and less than or equal to the maximum safe time threshold.

[0065] A comfort score is determined for each candidate game trajectory based on its accelerometer. In one embodiment of this application, the candidate game trajectory... The formula for calculating the comfort score is as follows:

[0066] Equation (9)

[0067] in, Representing the candidate trajectories of the game Comfort rating, Indicates jerk. This indicates the tolerance threshold.

[0068] The success rate of each candidate game trajectory is determined based on its longitudinal travel distance. In one embodiment of this application, the success rate is positively correlated with the longitudinal travel distance, and the candidate game trajectories... The formula for calculating the success rate of the game is as follows:

[0069] Equation (10)

[0070] in, Representing the candidate trajectories of the game The success rate of the game, Indicates the longitudinal driving distance. This indicates the theoretical maximum driving distance.

[0071] The game value score for each candidate trajectory is obtained by weighted summing of its safety score, comfort score, and game success rate. In one embodiment of this application, the formula for calculating the game value score is as follows:

[0072] Equation (11)

[0073] in, Representing the candidate trajectories of the game The game value score, Representing the candidate trajectories of the game Safety rating Indicates the weighting of comfort scores. Representing the candidate trajectories of the game Comfort rating, Indicates the success rate weight of the game. Representing the candidate trajectories of the game The success rate of the game is such that the sum of the comfort score weight and the success rate weight equals 1.

[0074] In one embodiment of this application, before obtaining the multi-intention interactive game decision and planning model, the multi-intention interactive game decision method further includes:

[0075] This invention acquires historical environmental information and historical state information of autonomous vehicles, as well as the historical trajectory of autonomous vehicles under the given historical environmental and state information. In one embodiment of this application, the historical environmental information includes: historical multi-view images, historical radar point clouds, and historical map vectors; both the historical environmental information and the historical state information are time-series data.

[0076] Historical scene features are obtained by fusing historical environmental information and historical state information using multimodal features. In one embodiment of this application, the process of fusing historical environmental information and historical state information using multimodal features is the same as the process of fusing current environmental information and current state information using multimodal features.

[0077] Based on historical scene features, the state features of the moving target are extracted as historical estimated state features of the moving target; and based on historical scene features, the intention features of the moving target are extracted as historical estimated intention features of the moving target; the historical estimated state features and historical estimated intention features are fused to obtain interactive target features with predicted intention information, denoted as historical interactive target features. In one embodiment of this application, the process of extracting the state features of the moving target based on historical scene features is the same as the process of extracting the state features of the moving target from the current scene features; the process of extracting the intention features of the moving target based on historical scene features is the same as the process of extracting the intention features of the moving target from the current scene features; and the process of fusing the historical estimated state features and historical estimated intention features to obtain interactive target features with predicted intention information is the same as the process of fusing the state features and intention features of the moving target to obtain interactive target features with intention information.

[0078] A historical game relationship graph is constructed, with autonomous vehicles and moving targets as historical nodes, and the lines connecting historical nodes as historical edges, and historical edge weights configured. In one embodiment of this application, the historical edge weights are determined by the state characteristics and intention characteristics of the historical nodes; the calculation process of the historical edge weights is the same as the calculation process of the edge weights. The intention characteristics of the historical nodes include: the historical intention characteristics of the autonomous vehicles and the historical estimated intention characteristics of the moving targets; the historical intention characteristics of the autonomous vehicles are determined by the historical trajectories of the autonomous vehicles. Specifically, they are obtained by feature encoding of the historical trajectories of the autonomous vehicles based on deep neural networks, or by cluster analysis of the historical trajectories of the autonomous vehicles based on driving style perception, or by interactive analysis or situation quantification of the historical trajectories of the autonomous vehicles based on physical potential fields and rules; the state characteristics of the historical nodes include: the historical state characteristics of the autonomous vehicles and the historical estimated state characteristics of the moving targets. The historical state characteristics of the autonomous vehicles are determined by historical state information; the process of determining the historical state characteristics of the autonomous vehicles from historical state information is the same as the process of determining the state characteristics of the autonomous vehicles from current state information.

[0079] Using historical map vectors as mapping conditions and based on different preset intent features, the historical state features of autonomous vehicles are mapped to a set of historical query vectors, and the historical interaction target features are mapped to historical key vectors and historical value vectors. Based on a set of historical query vectors, historical key vectors, and historical edge weights, the predicted attention weight between the autonomous vehicle and the moving target is calculated. The historical value vectors are weighted and aggregated using the historical predicted probabilities of the autonomous vehicle executing different preset intent features and the predicted attention weights to obtain predicted aggregate features. These predicted aggregate features are then subjected to feature concatenation, nonlinear transformation, and decoding to obtain multiple predicted game candidate trajectories. In one embodiment of this application, the historical predicted probabilities of the autonomous vehicle executing different preset intent features are predicted based on historical scene features. The process of predicting the historical predicted probabilities of the autonomous vehicle executing different preset intent features based on historical scene features is the same as the process of predicting the predicted probabilities of the autonomous vehicle executing different preset intent features based on current scene features. The calculation method for the predicted attention weight is the same as the calculation method for the attention weight. The calculation method for the predicted aggregate features is the same as the calculation method for the aggregate features. The process of performing feature concatenation, nonlinear transformation, and decoding on the predicted aggregate features is the same as the process of performing feature concatenation, nonlinear transformation, and decoding on the aggregate features.

[0080] The safety, comfort, and success rate of each predicted game candidate trajectory are evaluated to obtain a predicted game value score for each candidate trajectory. The predicted game candidate trajectory with the highest predicted game value score is then selected as the output trajectory. In one embodiment of this application, the method for calculating the predicted game value score of each predicted game candidate trajectory is the same as the process for calculating the game value score of each candidate trajectory.

[0081] Based on the predicted game value score of the predicted candidate trajectory, the actual game value score of the predicted candidate trajectory, the coordinate position of the output trajectory, the coordinate position of the historical trajectory, the historical estimated intent features of the moving target, and the actual intent features of the moving target, a loss function is constructed. With minimizing the loss function as the objective, the parameters of a preset multi-intent interactive game decision-making and planning model are adjusted to obtain the multi-intent interactive game decision-making and planning model. In one embodiment of this application, the actual intent features of the moving target are determined based on the historical trajectory of the moving target. The method of determining the actual intent features of the moving target based on its historical trajectory is the same as the method of determining the historical intent features of the autonomous vehicle based on its historical trajectory. The expression of the loss function is as follows:

[0082] Equation (12)

[0083] in, Represents the loss function. Indicates the weight of the loss term in trajectory regression. Indicates the trajectory regression loss. Indicates the weight of the loss item in the valuation. Indicates a loss in value assessment. Indicates meaning Figure 1 Weight of consistency loss term Indicates meaning Figure 1 Consistent loss; weights of loss terms in trajectory regression, weights of loss terms in value assessment, and intentional loss. Figure 1 The sum of the weights of the consistency loss terms equals 1. This represents the scaling factor of the trajectory regression loss (e.g., the average value of the trajectory regression loss). This represents a scaling factor for valuation loss (e.g., the average value of valuation loss). Indicates meaning Figure 1 Scaling factor for consistency loss (e.g., intention) Figure 1 (Average value of loss of homogeneity).

[0084] The formula for calculating trajectory regression loss is as follows:

[0085] Equation (13)

[0086] in, Indicates the trajectory regression loss. Indicates the total number of time steps. Indicates the output trajectory at the th The coordinates of each time step. Indicates the historical trajectory at the 1st The coordinates of each time step;

[0087] The formula for calculating the loss in value assessment is as follows:

[0088] Equation (14)

[0089] in, Indicates a loss in value assessment. This represents the total number of candidate trajectories for predicting the game. Indicates the first The predictive game value score of the candidate trajectories in the predictive game. Indicates the first The actual game value score of each predicted candidate trajectory;

[0090] meaning Figure 1 The formula for calculating loss of toxicity is as follows:

[0091] Equation (15)

[0092] in, Indicates meaning Figure 1 Sexual damage, Indicates the first The true label of class intent is represented in the form of one-hot encoding. Indicates the first The predicted probability of class intent.

[0093] Figure 3 This is an exemplary embodiment of the present application illustrating the architecture of a multi-intent interactive game decision-making and planning model. Figure 3In this multi-intent interactive game decision and planning model, the following components are included: (1) Multimodal perception input (located in the perception layer): perceiving the current state information of the autonomous vehicle and the current environmental information around the autonomous vehicle; (2) Scene encoder (located in the representation layer): used to obtain the current state information of the autonomous vehicle and the current environmental information around the autonomous vehicle, and to fuse the current environmental information and the current state information with multimodal features to obtain the current scene features; (3) Target intent prediction and encoding module (located in the intent layer): including an intent classifier and an intent encoding module, wherein the intent classifier is used to extract the state features of the moving target from the current scene features, and Based on the state characteristics of the moving target, the intention characteristics of the moving target are determined, and the state characteristics and intention characteristics of the moving target are used as outputs; the intention encoding module is used to fuse the state characteristics and intention characteristics of the moving target to obtain interactive target characteristics with intention information; (4) The interactive game planner (located in the game decision layer) includes: game relationship construction module, game strategy optimization module and multimodal planning module, wherein the game relationship construction module is used to construct a game relationship graph with autonomous vehicles and moving targets as nodes, the lines between nodes as edges, and edge weights configured; the game strategy optimization module is used to use map vectors as mapping conditions and based on different preset intentions The graph features map the current state features of the autonomous vehicle to a set of query vectors and the interactive target features to key vectors and value vectors; based on a set of query vectors, key vectors and edge weights, the attention weights between the autonomous vehicle and the moving target are calculated; the value vectors are weighted and aggregated by the predicted probability and attention weights of the autonomous vehicle executing different preset intention features to obtain aggregated features; the multimodal planning module is used to perform feature splicing, nonlinear transformation and decoding on the aggregated features to obtain multiple game candidate trajectories; (5) Trajectory evaluation and selector (located in the evaluation layer): evaluate the safety, comfort and game success rate of each game candidate trajectory to obtain each game candidate trajectory. The game value score of the candidate trajectory is used as the autonomous driving path (final planned trajectory) of the current frame. (6) The autonomous vehicle is controlled by the controller to drive according to the autonomous driving path of the current frame. At the same time, the multimodal perception input automatically collects the multi-view image of the next frame, the radar point cloud of the next moment and the map vector of the next moment, as well as the state information of the autonomous vehicle at the next moment. Based on the multi-view image of the next frame, the radar point cloud of the next moment and the map vector of the next moment, as well as the state information of the next moment, a multi-intent interactive game decision-making process is carried out to obtain the autonomous driving path of the next frame.

[0094] Figure 4 This is an exemplary embodiment of the present application illustrating the internal structure of an interactive game planner. Figure 4In the interactive game planner, there are: (1) an input layer: used for current scene information and interactive target features with intent information; (2) a game relationship construction module: used to construct a game relationship graph with autonomous vehicles and moving targets as nodes, the connection between nodes as edges, and edge weights configured; (3) a game strategy optimization module including: a first game layer, a second game layer and a third game layer, wherein the first game layer uses map vectors as mapping conditions and, based on different preset intent features, maps the current state features of autonomous vehicles to a set of query vectors and maps the interactive target features to key vectors and value vectors; based on a set of query vectors, key vectors and edge weights, calculates the first attention weight between autonomous vehicles and moving targets; weights and aggregates the value vectors by the prediction probability and attention weights of autonomous vehicles executing different preset intent features to obtain the first aggregated feature; the second game layer uses map vectors as mapping conditions and, based on different preset intent features, maps the current state features of autonomous vehicles to a set of query vectors and maps the first aggregated feature to a set of query vectors and edge weights. The updated key vector and the updated value vector are used to calculate the second attention weight between the autonomous vehicle and the moving target based on a set of query vectors, the updated key vector and the first attention weight; the updated value vector is weighted and aggregated by the predicted probability of the autonomous vehicle executing different preset intention features and the second attention weight to obtain the second aggregated feature; the third game layer uses the map vector as the mapping condition and maps the current state feature of the autonomous vehicle to a set of query vectors based on different preset intention features, and maps the second aggregated feature to the updated key vector and the updated value vector; the third attention weight between the autonomous vehicle and the moving target is calculated based on a set of query vectors, the updated key vector and the second attention weight; the updated value vector is weighted and aggregated by the predicted probability of the autonomous vehicle executing different preset intention features and the third attention weight to obtain the third aggregated feature; (4) The multimodal planning module is used to perform feature splicing, nonlinear transformation and decoding on the third aggregated feature to obtain multiple game candidate trajectories.

[0095] In one embodiment of this application, the game process calculates the attention weight between the autonomous vehicle and the moving target based on the self-attention mechanism, and adjusts the attention weight according to the predicted probability of the autonomous vehicle executing different preset intention features, thereby adjusting the interaction intensity between the autonomous vehicle and the moving target. Through multiple rounds of iterative optimization through the first game layer, the second game layer and the third game layer, multiple game candidate trajectories are output, thereby enabling more accurate capture of the dynamic game relationship in the multi-vehicle interaction process, solving the problem that traditional decision-making methods only perform single-round intention prediction and cannot cope with the dynamic changes of target intention.

[0096] Figure 5This is a schematic diagram illustrating an intent classification and interactive game scenario, as shown in an exemplary embodiment of this application. Figure 5 As shown, in an unprotected left-turn scenario, the blue vehicle represents an autonomous vehicle, and the green vehicles represent other vehicles, including vehicle A and vehicle B. Vehicle A's intention is to maintain its position, and vehicle B's intention is to maintain its position. The purple vehicle is a merging vehicle, vehicle C, whose intention is to merge. The yellow vehicle represents a pedestrian whose intention is to cross the road. The game-theoretic decision-making process is as follows: The autonomous vehicle (i.e., the self-driving car) has a left-turn intention, so it needs to determine whether it can pass vehicle A if it accelerates. If vehicle B accelerates faster, it should wait. The pedestrian may cross the road, so a safe distance needs to be maintained. The final game-theoretic outcome is: first, slow down and observe, and then quickly complete the left turn after vehicle A has passed.

[0097] The advantages of this application compared to existing technologies are shown in the table below:

[0098] Table 1

[0099]

[0100] As summarized in Table 1, this application significantly improves the pass rate in complex interactive scenarios such as unprotected left turns and lane changes; the design of the multi-intent interactive game decision-making system enables it to understand and predict the intentions of others like a human driver, resulting in more natural and smooth driving behavior and a significant improvement in the human-likeness of driving behavior; through active game theory, unnecessary waiting and deceleration are reduced, improving traffic efficiency; and through intent modeling and game theory mechanisms, the multi-intent interactive game decision-making system can anticipate potential conflicts in advance and take preventive measures.

[0101] Unlike implicit interaction processing, this application uses an explicit intent classifier to enable the multi-intent interactive game decision-making system to "understand" the intentions of other road users, just like humans, providing clearer input for subsequent game decisions. Utilizing a learnable game mechanism, the complex interactive game process is transformed into a learnable module based on an attention mechanism. The multi-intent interactive game decision-making system can automatically learn the optimal response strategy under different intent combinations from massive amounts of driving data, avoiding tedious manual rule formulation. By simulating different interaction hypotheses, the game planner can generate multiple strategy trajectories, reflecting the diversity and uncertainty of driving decisions. This "preparedness is better than cure" approach enhances the safety redundancy of multi-intent interactive game decision-making systems in complex interactive scenarios. The entire process, from perceptual features to the final trajectory, is end-to-end, reducing information loss and allowing for joint optimization, enabling various modules to work collaboratively. Autonomous vehicles are no longer simply passively avoiding obstacles, but can actively participate in the game based on their understanding of others' intentions, striving for right-of-way or choosing appropriate opportunities, thereby achieving smoother and more efficient driving, especially in scenarios such as congested merging and intersection passage, thus improving the initiative and efficiency of the game.

[0102] The following describes an apparatus embodiment of this application, which can be used to execute the multi-intent interactive game decision-making system for autonomous driving path planning described in the above embodiments of this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the multi-intent interactive game decision-making method for autonomous driving path planning described above in this application.

[0103] Figure 6 This is a block diagram illustrating a multi-intent interactive game decision-making system for autonomous driving path planning, as shown in an exemplary embodiment of this application.

[0104] like Figure 6 As shown, the exemplary multi-intent interactive game decision-making system 600 for autonomous driving path planning includes:

[0105] The information acquisition module 601 is used to acquire the current status information of the autonomous vehicle and the current environmental information around the autonomous vehicle.

[0106] The feature fusion module 602 is used to perform multimodal feature fusion of current environmental information and current state information to obtain current scene features.

[0107] The intent prediction and encoding module 603 is used to extract the state features and intent features of the moving target from the current scene features, and to fuse the state features and intent features of the moving target to obtain interactive target features with intent information.

[0108] The game relationship construction module 604 is used to construct a game relationship graph with autonomous vehicles and moving targets as nodes, the lines connecting the nodes as edges, and edge weights configured.

[0109] The multimodal planning module 605 is used to map the current state features of the autonomous vehicle into a set of query vectors and the interactive target features into key vectors and value vectors, based on map vectors as mapping conditions and different preset intent features. Based on a set of query vectors, key vectors and edge weights, it calculates the attention weights between the autonomous vehicle and the moving target. It then performs weighted aggregation on the value vectors by the predicted probabilities of the autonomous vehicle executing different preset intent features and the attention weights to obtain aggregated features. Finally, it performs feature concatenation, nonlinear transformation and decoding on the aggregated features to obtain multiple game candidate trajectories.

[0110] The game decision module 606 is used to evaluate the safety, comfort and success rate of each game candidate trajectory, obtain the game value score of each game candidate trajectory, and take the game candidate trajectory with the highest game value score as the autonomous driving path.

[0111] In one embodiment of this application, the current environmental information includes: multi-view images, radar point clouds, and map vectors. The multi-view images are acquired by onboard cameras at different locations on the vehicle, and can reflect the environmental information around the vehicle from multiple angles. The radar point clouds are acquired by onboard LiDAR or millimeter-wave radar, and can accurately provide information on the distance, relative speed, and obstacle contours between surrounding moving targets and the autonomous vehicle. The map vectors contain static road information such as the current road topology, lane positions, traffic rules, speed limits, and traffic light status, which are used to provide road constraints for subsequent path planning. The current state information includes the autonomous vehicle's current speed, acceleration, steering wheel angle, vehicle attitude, and current location coordinates in the global map. This information can accurately reflect the current driving status of the autonomous vehicle and is an important constraint basis for subsequent trajectory generation.

[0112] In one embodiment of this application, the current scene features integrate current environmental information and current state information, ensuring that information from different modalities can complement and integrate, thus preserving the richness of visual semantics, the accuracy of point cloud ranging, and the constraints of map information. This allows for a comprehensive and structured expression of the core information of the current traffic scene, helping subsequent modules to accurately understand the scene interaction relationships.

[0113] In one embodiment of this application, the moving target includes: pedestrians and vehicles. The process of extracting the state features of the moving target from the current scene features includes: extracting the state features of the moving target from the current scene features, determining the intention features of the moving target based on the state features of the moving target, and using the state features and intention features of the moving target as output. The process of extracting the state features of the moving target from the current scene features is implemented by an intention classifier.

[0114] In one embodiment of this application, the edge weights are determined by the state features and intent features of the nodes; the state features of the autonomous vehicle are determined by the current state information, and the intent features of the autonomous vehicle are determined by the navigation planning information. The process of determining the state features of the autonomous vehicle from the current state information includes: preprocessing and normalizing the current state information; inputting the preprocessed and normalized current state information into a state encoder (e.g., a multilayer perceptron) to obtain a state feature vector; and performing temporal fusion on the state feature vector to obtain the state features. The process of determining the intent features of the autonomous vehicle from the navigation planning information includes: determining the driving instructions of the autonomous vehicle at the current location point from the navigation planning information, such as left turn, right turn, straight ahead, etc.; encoding the current location point, mapping the driving instructions to one-hot encoding, and concatenating the location encoding and the driving instruction encoding to form the intent features of the autonomous vehicle. The calculation formula for the edge weights is shown in formula (1).

[0115] In one embodiment of this application, the predicted probability of an autonomous vehicle executing different preset intent features is obtained based on the current scene features. Specifically, the current scene features are input into a classifier based on a fully connected layer to obtain the score of the autonomous vehicle executing different preset intent features; the score of the autonomous vehicle executing different preset intent features is normalized to obtain the predicted probability of the autonomous vehicle executing different preset intent features. The calculation formula for attention weight is shown in formula (2), and the calculation formula for aggregated features is shown in formula (3).

[0116] In one embodiment of this application, the process of performing feature concatenation, nonlinear transformation, and decoding on aggregated features to obtain multiple game candidate trajectories includes: concatenating aggregated features with features of different intentions to obtain concatenated vectors with different intentions; performing nonlinear transformation on the concatenated vectors with different intentions (e.g., through a multilayer perceptron and activation function) to obtain multiple nonlinear transformation vectors; and decoding the multiple nonlinear transformation vectors (e.g., through a Transformer decoder) to obtain multiple game candidate trajectories.

[0117] In one embodiment of this application, by extracting the state features and intent features of moving targets from the current scene features and generating interactive target features with intent information, this explicit intent understanding mechanism greatly reduces the uncertainty in the interaction process and avoids the risk of sudden braking or collision due to misjudgment. By using a game relationship graph and dynamic attention mechanism, the influence of different moving targets on the decision-making of autonomous vehicles can be dynamically measured. By generating multiple game candidate trajectories based on different preset intent features and comprehensively evaluating them from three dimensions of safety, comfort, and game success rate, the "being bullied" or deadlock problem caused by traditional reactive planning is effectively overcome. By integrating multimodal feature fusion, intent inference, game decision-making, and trajectory decoding into a unified framework, complex inter-module coupling and rule dependence are avoided, enabling the model to automatically learn the optimal interaction strategy through data-driven methods. Compared with methods that heavily rely on manually designed rules, this solution has stronger environmental adaptability and generalization ability and can cope with diverse interaction scenarios in complex urban traffic flows.

[0118] It should be noted that the multi-intent interactive game decision-making system for autonomous driving path planning provided in the above embodiments and the multi-intent interactive game decision-making method for autonomous driving path planning provided in the above embodiments belong to the same concept. The specific methods of execution of each module and unit have been described in detail in the method embodiments and will not be repeated here. In practical applications, the multi-intent interactive game decision-making system for autonomous driving path planning provided in the above embodiments can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. This is not a limitation here.

[0119] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A multi-intent interactive game decision-making method for autonomous driving path planning, characterized in that, The multi-intention interactive game decision-making method is implemented based on a multi-intention interactive game decision and planning model, and includes: The system acquires the current state information of the autonomous vehicle and the current environmental information surrounding the autonomous vehicle; the current environmental information includes: multi-view images, radar point clouds, and map vectors. The current environment information and the current state information are fused using multimodal features to obtain the current scene features; The state features and intent features of moving targets are extracted from the current scene features, and the state features and intent features of moving targets are fused to obtain interactive target features with intent information; the moving targets include: pedestrians and vehicles; A game theory graph is constructed with the autonomous vehicle and the moving target as nodes, the lines connecting the nodes as edges, and edge weights configured. The edge weights are determined by the state features and intent features of the nodes. The state features of the autonomous vehicle are determined by the current state information, and the intent features of the autonomous vehicle are determined by the navigation planning information. Using the map vectors as mapping conditions and based on different preset intent features, the current state features of the autonomous vehicle are mapped into a set of query vectors, and the interactive target features are mapped into key vectors and value vectors. Based on the set of query vectors, the key vectors, and the edge weights, the attention weight between the autonomous vehicle and the moving target is calculated. The value vectors are weighted and aggregated using the predicted probabilities of the autonomous vehicle executing different preset intent features and the attention weights to obtain aggregated features. These aggregated features are then subjected to feature concatenation, nonlinear transformation, and decoding to obtain multiple game-theoretic candidate trajectories. The predicted probabilities of the autonomous vehicle executing different preset intent features are predicted based on the current scene features. The safety, comfort, and success rate of each candidate trajectory are evaluated to obtain the game value score of each candidate trajectory. The candidate trajectory with the highest game value score among all the candidate trajectories is taken as the autonomous driving path. The process of evaluating the safety, comfort, and success rate of each candidate game trajectory to obtain a game value score for each candidate trajectory includes: The acceleration and longitudinal travel distance of each candidate trajectory in the game are obtained, as well as the collision prediction results for each candidate trajectory. The collision prediction results include: collision occurred, safe, and collision risk. The longitudinal travel distance is used to characterize the projection length of the candidate trajectory on the reference line. Based on the collision prediction results of each candidate trajectory, a safety score is determined for each candidate trajectory. Based on the acceleration of each candidate game trajectory, determine the comfort score of each candidate game trajectory. The success rate of each candidate game trajectory is determined based on the longitudinal travel distance of each candidate game trajectory; the success rate of the game is positively correlated with the longitudinal travel distance. The game value score of each candidate trajectory is obtained by weighting and summing the safety score, comfort score and game success rate of each candidate trajectory. The formula for calculating the game value score includes: ， in, Representing the candidate trajectories of the game The game value score, Representing the candidate trajectories of the game Safety rating Indicates the weight of comfort rating. Representing the candidate trajectories of the game Comfort rating Indicates the success rate weight of the game. Representing the candidate trajectories of the game The success rate of the game is such that the sum of the comfort score weight and the success rate weight equals 1. The game candidate trajectory The formula for calculating the safety score includes: ， in, Representing the candidate trajectories of the game The collision time, This represents the minimum dangerous time threshold. This represents the maximum safe time threshold; other scenarios with collision risks are considered as candidate trajectories in the game theory. The collision time is greater than or equal to the minimum dangerous time threshold, but less than or equal to the maximum safe time threshold. The game candidate trajectory The formula for calculating the comfort score includes: ， in, Indicates jerk. Indicates the tolerance threshold; The game candidate trajectory The formulas for calculating the success rate of a game include: ， in, Indicates the longitudinal driving distance. This indicates the theoretical maximum driving distance.

2. The multi-intent interactive game decision-making method for autonomous driving path planning according to claim 1, characterized in that, The process of fusing the current environment information and the current state information using multimodal features to obtain the current scene features includes: The current state information is state encoded to obtain the current state feature vector and modulation parameters; Feature extraction and perspective transformation are performed on the multi-view images to obtain bird's-eye view features; Voxelization and sparse convolution are performed on the radar point cloud to obtain radar bird's-eye view features; The map vectors are vectorized and rasterized to obtain the bird's-eye view features of the map; The image bird's-eye view features, the radar bird's-eye view features, and the map bird's-eye view features are stitched together, and the stitched features are modulated using the current state feature vector and the modulation parameters to obtain the current scene features.

3. The multi-intent interactive game decision-making method for autonomous driving path planning according to claim 2, characterized in that, The calculation formula for the current scene features includes: ， in, Indicates the characteristics of the current scene. Indicates the scaling factor. Represents the feature vector of the current state. Represents the bird's-eye view features of the image. Indicates features in a radar bird's-eye view. Indicates the features of a bird's-eye view of a map. This represents the bias coefficient. The modulation coefficients include the scaling coefficient and the bias coefficient.

4. The multi-intent interactive game decision-making method for autonomous driving path planning according to claim 1 or 2, characterized in that, The process of fusing the state features and intent features of the moving target to obtain interactive target features with intent information includes: The state features of the moving target are mapped to a first query vector, and the intention features of the moving target are mapped to a first key vector and a first value vector. The first query vector is then subjected to cross-attention operation with the first key vector and the first value vector to obtain intermediate fusion features. The probability distribution corresponding to the intention features of the moving target is used to weight and aggregate the intermediate fusion features to obtain the final fusion features; The final fused features and the state features are residually connected and layer normalized to obtain the interactive target features.

5. The multi-intent interactive game decision-making method for autonomous driving path planning according to claim 1 or 2, characterized in that, Before obtaining the multi-intention interactive game decision and planning model, the multi-intention interactive game decision method also includes: The system acquires historical environmental information and historical state information of the autonomous vehicle, as well as the historical trajectory of the autonomous vehicle under the conditions of the historical environmental information and the historical state information; the historical environmental information includes: historical multi-view images, historical radar point clouds, and historical map vectors; The historical environment information and the historical state information are fused using multimodal features to obtain historical scene features; Based on the historical scene features, the state features of the moving target are extracted as the historical estimated state features of the moving target; and based on the historical scene features, the intention features of the moving target are extracted as the historical estimated intention features of the moving target; the historical estimated state features and the historical estimated intention features are fused to obtain interactive target features with predicted intention information, which are denoted as historical interactive target features. A historical game relationship graph is constructed, with the autonomous vehicle and the moving target as historical nodes, the lines connecting the historical nodes as historical edges, and weights assigned to the historical edges. The weights of the historical edges are determined by the state features and intent features of the historical nodes. The intent features of the historical nodes include the historical intent features of the autonomous vehicle and the historical estimated intent features of the moving target. The historical intent features of the autonomous vehicle are determined by the historical trajectory of the autonomous vehicle. The state features of the historical nodes include the historical state features of the autonomous vehicle and the historical estimated state features of the moving target. The historical state features of the autonomous vehicle are determined by the historical state information. Using the historical map vectors as mapping conditions and based on different preset intent features, the historical state features of the autonomous vehicle are mapped into a set of historical query vectors, and the historical interaction target features are mapped into historical key vectors and historical value vectors. Based on the set of historical query vectors, the historical key vectors, and the historical edge weights, the predicted attention weights between the autonomous vehicle and the moving target are calculated. The historical value vectors are weighted and aggregated using the historical predicted probabilities of the autonomous vehicle executing different preset intent features and the predicted attention weights to obtain predicted aggregate features. The predicted aggregate features are then subjected to feature concatenation, nonlinear transformation, and decoding to obtain multiple predicted game candidate trajectories. The historical predicted probabilities of the autonomous vehicle executing different preset intent features are predicted based on the historical scene features. The safety, comfort, and success rate of each predicted game candidate trajectory are evaluated to obtain the predicted game value score of each predicted game candidate trajectory. The predicted game candidate trajectory with the highest predicted game value score is taken as the output trajectory. Based on the predicted game value score of the predicted candidate trajectory, the actual game value score of the predicted candidate trajectory, the coordinate position of the output trajectory, the coordinate position of the historical trajectory, the historical estimated intent features of the moving target, and the actual intent features of the moving target, a loss function is constructed. With minimizing the loss function as the objective, the parameters of the preset multi-intent interactive game decision and planning model are adjusted to obtain the multi-intent interactive game decision and planning model. The actual intent features of the moving target are determined based on the historical trajectory of the moving target.

6. The multi-intent interactive game decision-making method for autonomous driving path planning according to claim 5, characterized in that, The expression for the loss function includes: ， in, Represents the loss function. Indicates the weight of the loss term in trajectory regression. Indicates the trajectory regression loss. Indicates the weight of the loss item in the valuation. Indicates a loss in value assessment. This indicates the weight of the intention consistency loss term. Indicates loss of consistency of intent. The scaling factor representing the trajectory regression loss. The scaling factor representing the loss in valuation. The scaling factor representing the loss of intent consistency; The formula for calculating the trajectory regression loss includes: ， in, Indicates the total number of time steps. Indicates the output trajectory at the th The coordinates of each time step. Indicates the historical trajectory at the 1st The coordinates of each time step; The formula for calculating the loss in the valuation includes: ， in, This represents the total number of candidate trajectories for predicting the game. Indicates the first The predictive game value score of the candidate trajectories in the predictive game. Indicates the first The actual game value score of each predicted candidate trajectory; The formula for calculating the intent consistency loss includes: ， in, Indicates the first The true label of class intent is represented in the form of one-hot encoding. Indicates the first The predicted probability of class intent.

7. The multi-intent interactive game decision-making method for autonomous driving path planning according to claim 1 or 2, characterized in that, The formula for calculating the edge weight includes: ， in, Represents a node With nodes Edge weights between them This represents the learnable adaptive equilibrium parameters. Represents a node With nodes The normalized Euclidean distance between them Represents the distance normalization coefficient. Represents a node With nodes The relative speed between them Represents the velocity normalization coefficient. The hyperparameter representing the intent weights controls the degree to which intent features influence edge weights. This indicates the intent characteristics of the autonomous vehicle. This indicates the intentional features of the moving target. Indicates the probability of conflict of intent; The formula for calculating the attention weight includes: ， in, Indicates at node With nodes Between the first Attention weights calculated from pre-defined intent features Indicates the first A query vector with predefined intent features. Represents the key vector. This represents the dimension of the key vector. Represents a node With nodes The edge weights between them are scalar biases; The calculation formula for the aggregation feature includes: ， in, Indicates aggregation features, Indicates that the autonomous vehicle is performing the first The predicted probability of a pre-defined intention feature. Indicates the total number of moving targets. Represents a value vector; The first The expression for a query vector of preset intent features includes: ， in, Indicates querying the projection matrix. Indicates the current state characteristics. Indicates the first A preset intention feature, A local map vector representing the location of an autonomous vehicle. This indicates a splicing operation. This represents a multilayer perceptron used for query vectors; The expression for the key vector includes: ， in, Represents the key projection matrix. This represents a multilayer perceptron used for key vectors. Indicates the characteristics of the interaction target. A local map vector representing the location of a moving target. Indicates a splicing operation; The expression for the value vector includes: ， in, Represents the value projection matrix, This represents a multilayer perceptron used for value vectors.

8. A multi-intent interactive game decision-making system for autonomous driving path planning, characterized in that, The system is used to implement the multi-intent interactive game decision-making method for autonomous driving path planning as described in claim 1, and the system includes: The information acquisition module is used to acquire the current status information of the autonomous vehicle and the current environmental information around the autonomous vehicle; the current environmental information includes: multi-view images, radar point clouds and map vectors; The feature fusion module is used to perform multimodal feature fusion of the current environment information and the current state information to obtain the current scene features; The intent prediction and encoding module is used to extract the state features and intent features of moving targets from the current scene features, and to fuse the state features and intent features of moving targets to obtain interactive target features with intent information; the moving targets include: pedestrians and vehicles; The game relationship construction module is used to construct a game relationship graph with the autonomous vehicle and the moving target as nodes, the lines connecting the nodes as edges, and edge weights configured; the edge weights are determined by the state characteristics and intention characteristics of the nodes; the state characteristics of the autonomous vehicle are determined by the current state information, and the intention characteristics of the autonomous vehicle are determined by the navigation planning information; A multimodal planning module is used to map the current state features of the autonomous vehicle into a set of query vectors and the interactive target features into key vectors and value vectors, based on the map vectors as mapping conditions and different preset intent features. Based on the set of query vectors, the key vectors, and the edge weights, the module calculates the attention weights between the autonomous vehicle and the moving target. It then performs weighted aggregation on the value vectors using the predicted probabilities of the autonomous vehicle executing different preset intent features and the attention weights to obtain aggregated features. Finally, it performs feature concatenation, nonlinear transformation, and decoding on the aggregated features to obtain multiple game-theoretic candidate trajectories. The predicted probabilities of the autonomous vehicle executing different preset intent features are predicted based on the current scene features. The game decision-making module is used to evaluate the safety, comfort, and success rate of each candidate trajectory, obtain the game value score of each candidate trajectory, and select the candidate trajectory with the highest game value score as the autonomous driving path.

Images