Automatic driving vehicle conflict avoidance method and device based on game theory, and vehicle

Abstract

The invention relates to the field of automatic driving, in particular to an automatic driving vehicle conflict avoiding method and device based on the game theory and a vehicle, and aims to solve the problem that the real-time dynamic conflict avoiding capability of an existing automatic driving vehicle is insufficient due to the fact that the existing automatic driving vehicle lacks a multi-agent collaborative decision-making mechanism in a complex traffic scene. The method comprises the following steps: constructing an environment data set of a vehicle; analyzing the environment data set by using a deep neural network to predict behavior tracks of other traffic participants around the vehicle, and identifying potential conflict points between the vehicle and the other traffic participants; constructing a game theory model taking the vehicle and at least one other traffic participant as game parties, and calculating an optimal action strategy for the vehicle by solving the game theory model based on the behavior track sum and the potential conflict points; according to the optimal action strategy, the vehicle is controlled to adjust at least one of the driving path, the speed or the direction, so that conflict avoidance is achieved.

Description

Technical Field

[0001] This application relates to the field of autonomous driving, specifically to a method, device, and vehicle for avoiding conflicts in autonomous vehicles based on game theory. Background Technology

[0002] Autonomous driving technology aims to improve traffic efficiency, safety, and the travel experience by enabling vehicles to drive autonomously through advanced sensors, communication systems, and intelligent algorithms. Current technologies primarily rely on onboard perception systems, such as radar, lidar, and cameras, to collect information about the surrounding environment and exchange data through vehicle-to-vehicle and vehicle-to-infrastructure communication technologies to assist vehicles in decision-making.

[0003] However, existing technologies still have significant limitations in real-time dynamic decision-making and multi-agent conflict avoidance. Specifically, most existing systems rely on preset rules and fixed-pattern decision-making logic, lacking flexibility and adaptability in complex and changing traffic scenarios. When multiple autonomous vehicles or traffic participants simultaneously enter the same interaction scenario, such as merging at unsignalized intersections or congested road sections, existing methods struggle to quickly and accurately understand the intentions of other vehicles and cannot effectively predict potential trajectory conflicts.

[0004] Furthermore, existing solutions often focus on single-vehicle decision optimization, lacking modeling and analysis of collaborative game mechanisms among multiple vehicles. This makes it difficult for the system to achieve dynamic balance and overall optimization among multiple objectives such as safety, traffic efficiency, and traffic rule compliance. Therefore, there is an urgent need in this field for a conflict avoidance method that can deeply integrate environmental perception, behavior prediction, and multi-agent collaborative decision-making to overcome the above-mentioned technical shortcomings and meet the application requirements of high-level autonomous driving. Summary of the Invention

[0005] This application provides a game theory-based method, device, and vehicle for avoiding conflicts in autonomous vehicles, which addresses the problem of insufficient real-time dynamic conflict avoidance capability of existing autonomous vehicles in complex traffic scenarios due to the lack of a multi-agent collaborative decision-making mechanism.

[0006] The technical solution of this invention is as follows:

[0007] Firstly, this application provides a game theory-based method for conflict avoidance in autonomous vehicles, comprising the following steps:

[0008] The vehicle acquires environmental data through its multimodal sensing system and receives environmental data sent by other vehicles or infrastructure through its communication system to construct the vehicle's environmental dataset.

[0009] The environmental dataset is analyzed using a deep neural network to predict the behavioral trajectories of other traffic participants around the vehicle and to identify potential conflict points between the vehicle and other traffic participants.

[0010] Construct a game theory model with the vehicle and at least one other traffic participant as the players. Based on the behavioral trajectory and the potential conflict points, calculate an optimal action strategy for the vehicle by solving the game theory model.

[0011] According to the optimal action strategy, the vehicle is controlled to adjust at least one of its driving path, speed, or direction to achieve conflict avoidance.

[0012] By integrating multimodal sensor data from the vehicle itself with external communication data to construct a complete environmental situation, the limitations of single-vehicle perception are addressed. By utilizing deep neural networks to predict the trajectories of other vehicles and identify potential conflict points, accurate understanding of complex traffic scenarios is achieved. A multi-agent collaborative decision-making mechanism is established by constructing a game theory model involving both the vehicle and other vehicles and solving for the optimal strategy. Finally, by controlling the vehicle's driving state in real time based on the game results, autonomous conflict avoidance in dynamic environments is achieved. The combination of these technical features effectively solves the core problem of insufficient real-time conflict avoidance capabilities in existing autonomous driving systems due to the lack of a collaborative decision-making mechanism.

[0013] In some possible embodiments, the environmental dataset is constructed as a time-series dataset: ;

[0014] in, This represents the environmental feature vector collected at time point t. Indicates the corresponding time tag;

[0015] The environmental feature vector Includes at least one of the following data: 3D surrounding terrain mapping data Dynamic weather conditions impact analysis data Road surface condition assessment data Other vehicle and pedestrian micro-movement data Other vehicle and pedestrian intention prediction data and the vehicle's real-time status data .

[0016] By constructing a time-series environmental dataset, multi-dimensional environmental features are organically integrated over time, forming a dynamically evolving environmental situation representation. This includes three-dimensional terrain mapping data. It provides a precise description of the spatial structure and weather data. and road condition data This reflects environmental constraints and minute motion data. and intention prediction data To achieve fine-grained perception of the behavior of traffic participants, combined with the vehicle's status data. Together, these elements form the complete information foundation for game theory decision-making. This time-series, multi-dimensional data organization significantly enhances the system's ability to capture the dynamic evolution of the traffic environment, providing high-quality data support for subsequent deep neural network analysis and game theory calculations, thereby improving the accuracy and foresight of conflict avoidance decisions.

[0017] In some possible embodiments, the steps of using a deep neural network to analyze the environmental dataset to predict the behavioral trajectories of other traffic participants around the vehicle and to identify potential conflict points between the vehicle and other traffic participants include:

[0018] The time series data in the environmental dataset are standardized and normalized preprocessed to obtain the preprocessed environmental feature vector;

[0019] The preprocessed environmental feature vector is input into a deep neural network, and the deep neural network outputs a behavioral pattern representation at the corresponding time point.

[0020] Based on current and historical behavioral patterns, predict the dynamic changes of other traffic participants in the next few seconds;

[0021] By combining the dynamic change prediction with the current environmental feature vector, potential conflict points between the vehicle and other traffic participants are identified.

[0022] By constructing a progressive analysis process encompassing data preprocessing, feature extraction, behavior prediction, and conflict identification, a deep understanding of the traffic environment is achieved. Standardization and normalization ensure the consistent quality of multi-source heterogeneous data, providing stable and reliable input for the deep neural network. The deep neural network, by uncovering hidden patterns in time-series data, transforms raw environmental features into high-level behavioral pattern representations, enabling deep extraction of traffic participant behavioral characteristics. Dynamic changes in historical behavioral patterns are then predicted. Finally, by fusing the prediction results with real-time environmental features, potential conflict points are accurately identified, providing crucial input for subsequent game theory decisions. This layered analytical architecture significantly enhances the system's depth of understanding and predictive accuracy in complex traffic scenarios.

[0023] In some possible embodiments, the step of inputting the preprocessed environmental feature vector into a deep neural network, and the deep neural network outputting a behavioral pattern representation corresponding to a given time point, includes:

[0024] The behavioral pattern representation is obtained through the following calculations. : ;

[0025] in, It represents the behavioral pattern at time t. It is the preprocessed environmental feature vector The i-th feature, Here, b is the weight corresponding to the i-th feature in the deep neural network, b is the bias term, and n is the preprocessed environmental feature vector. Dimensions It is a non-linear activation function;

[0026] Based on current and historical behavioral patterns, the steps to predict the dynamic changes of other traffic participants in the next few seconds include:

[0027] The following prediction function is used to predict the dynamic changes F(Bt) of other traffic participants in the next few seconds: ;

[0028] in, Indicates a prediction of future behavior. It is the behavioral pattern at the k-th time point in the past. These are weight parameters. This is the prediction error term.

[0029] This approach achieves an effective conversion from environmental characteristics to behavioral patterns through specific mathematical computational models. Deep neural networks, through weighted summation and nonlinear transformation, can extract representative behavioral pattern representations from complex, multi-dimensional environmental features. These representations effectively reflect the intrinsic behavioral characteristics of traffic participants. In the prediction phase, a prediction function based on a linear combination of historical behavioral patterns is employed. This considers the differences in the contribution of behavioral patterns at different time points while maintaining computational efficiency, enabling effective prediction of future dynamic changes under real-time requirements. This method, combining deep learning and time-series prediction, provides accurate and reliable input for subsequent conflict identification and game theory decision-making.

[0030] In some possible embodiments, the step of identifying potential conflict points between the vehicle and other traffic participants by combining the dynamic change prediction with the current environmental feature vector includes:

[0031] The following algorithm identifies potential conflict points between this vehicle and other road users. : ;

[0032] Among them, the function The future behavior prediction F(Bt) is transformed into the corresponding future position or state space, and the function Λ is based on the current environmental feature vector. Evaluate the current state space. It is a decision function;

[0033] when When the result exceeds a predetermined threshold, a potential conflict point is identified. .

[0034] Through function Map the predicted future behavior to a specific future state space, and simultaneously through a function This approach assesses the current state space based on current environmental characteristics, achieving a dual representation of the traffic situation across the time dimension. (Decision function) By calculating the difference between these two state spaces, the potential conflict risk between the vehicle and other traffic participants can be quantified. When this risk exceeds a predetermined threshold, the specific conflict point can be accurately identified. This conflict identification method, based on state-space difference measurement, can not only promptly detect potential collision risks, but also provide quantitative information on the degree of conflict for subsequent game-theoretic decisions, thereby supporting more refined decision-making.

[0035] In some possible embodiments, constructing a game theory model with the vehicle and at least one other traffic participant as the players, and calculating an optimal action strategy for the vehicle by solving the game theory model based on the behavioral trajectory and the potential conflict points, includes the following steps:

[0036] Based on vehicle dynamics constraints and traffic rules, a policy set containing multiple discretized behavior options is generated for this vehicle;

[0037] Define a game theory model, which includes a set of participants, a strategy space, and a set of utility functions;

[0038] Based on the behavioral trajectory and potential conflict points, the optimal action strategy for this vehicle is determined from the strategy set by calculating the Nash equilibrium of the game theory model.

[0039] By constructing a complete game-theoretic decision-making framework, optimal decision generation is achieved in scenarios involving multiple traffic participants. First, a policy set is generated based on vehicle dynamics and traffic rules, ensuring the feasibility and compliance of the decisions. A formal description of complex traffic interaction scenarios is established by defining a game-theoretic model that includes a set of participants, a policy space, and a set of utility functions. Finally, the optimal action strategy is determined by solving for the Nash equilibrium. This method ensures that, considering the optimal responses of other traffic participants, the strategy chosen by the vehicle achieves an optimal balance between multiple objectives such as safety and efficiency. This game-theoretic decision-making mechanism effectively solves the problem of multi-agent collaborative decision-making, providing a theoretical guarantee and implementation path for conflict avoidance in complex interaction scenarios for autonomous vehicles.

[0040] In some possible embodiments, the step of generating a policy set containing multiple discretized behavioral options for the vehicle based on vehicle dynamics constraints and traffic rules includes:

[0041] The discretized behavior options in the strategy set are defined by velocity adjustment and direction adjustment angle, and each behavior option si is represented as: ;

[0042] in, This is the speed adjustment amount, and its value range is... vd is the maximum deceleration limit, va is the maximum acceleration limit; θΔ is the direction adjustment angle, and its value range is... , θr is the maximum leftward deflection angle, and θr is the maximum rightward deflection angle;

[0043] The game theory model is represented as G=(N,S,U), where:

[0044] N is the set of traffic participants involved in the decision-making process, including the vehicle itself and at least one other traffic participant; S is the policy space, consisting of the policy sets of all participants; U is the set of utility functions, containing the utility functions defined for each participant.

[0045] Utility function defined for each participant Represented as: ;

[0046] Where si represents the strategy of this vehicle, , , These represent evaluation functions for safety, efficiency, and regulatory compliance, respectively, with α, β, and γ being weighting coefficients used to adjust the importance of different objectives.

[0047] By establishing a utility function framework for multi-objective optimization, the conflict avoidance problem is transformed into a quantifiable optimization problem. The discretization of speed and direction adjustments ensures the feasibility and computational efficiency of the strategy, while the comprehensive consideration of safety, efficiency, and regulatory compliance within the utility function ensures that the final decision effectively avoids conflict while also balancing traffic efficiency and traffic rules. The introduction of weighting coefficients α, β, and γ provides a flexible decision preference adjustment mechanism, enabling adjustments to the decision focus based on different scenario requirements. This game-theoretic model based on a multi-objective utility function provides a systematic solution for decision-making in complex traffic environments for autonomous vehicles, achieving collaborative optimization among multiple participants while ensuring safety.

[0048] In some possible embodiments, the step of determining the optimal action strategy of the vehicle from the strategy set by calculating the Nash equilibrium of the game theory model based on the behavioral trajectory and potential conflict points includes:

[0049] Initialize and set the initial strategy for this vehicle. Initial strategy groups with other traffic participants ;

[0050] To update the policy, perform the following calculations in sequence:

[0051] Fix the current policies of other traffic participants and update the current policy of this vehicle: ;

[0052] Fix the current policy of this vehicle and update the current policies of other traffic participants: ;

[0053] When the policy change during continuous iteration Less than the preset threshold Stop iteration when the time is up, and output the optimal action strategy for this vehicle. .

[0054] This paper proposes a computationally feasible decision-making scheme for solving the Nash equilibrium of a game theory model through iterative optimization. An alternating optimization strategy is employed: in each iteration, the strategies of other participants are fixed to optimize the vehicle's own strategy, and then the strategies of other participants are optimized based on the updated vehicle strategy. This alternating update mechanism effectively simulates the policy interaction process among traffic participants. A convergence threshold is set to terminate the iteration, ensuring both computational efficiency and the stability of the final strategy. This iterative optimization-based Nash equilibrium solution method can effectively compute the optimal action strategy for a vehicle in complex multi-agent interaction scenarios, providing a reliable technical guarantee for achieving safe and efficient conflict avoidance.

[0055] Secondly, this application also provides a game theory-based collision avoidance device for autonomous vehicles, comprising:

[0056] The acquisition module is used to acquire environmental data through the vehicle's multimodal sensing system and receive environmental data sent by other vehicles or infrastructure through the vehicle's communication system to construct the vehicle's environmental dataset.

[0057] The behavior trajectory and potential conflict point prediction module is used to analyze the environmental dataset using a deep neural network to predict the behavior trajectories of other traffic participants around the vehicle and identify potential conflict points between the vehicle and other traffic participants.

[0058] The optimal action strategy prediction module is used to construct a game theory model with the vehicle and at least one other traffic participant as the players. Based on the behavior trajectory and the potential conflict points, it calculates an optimal action strategy for the vehicle by solving the game theory model.

[0059] The control module is used to control the vehicle to adjust at least one of its driving path, speed, or direction according to the optimal action strategy in order to achieve conflict avoidance.

[0060] Thirdly, this application also provides a vehicle including the aforementioned game theory-based autonomous vehicle conflict avoidance device. Attached Figure Description

[0061] Figure 1 This is a flowchart illustrating the game theory-based conflict avoidance method for autonomous vehicles in an embodiment of this application.

[0062] Figure 2 This is a flowchart illustrating the game theory-based conflict avoidance method for autonomous vehicles in this application embodiment. Detailed Implementation

[0063] refer to Figure 1 and Figure 2 This application provides a game theory-based conflict avoidance method for autonomous vehicles, comprising the following steps:

[0064] S1. Acquire environmental data through the vehicle's multimodal sensing system and receive environmental data sent by other vehicles or infrastructure through the vehicle's communication system to construct the vehicle's environmental dataset.

[0065] S2, Analyze the environmental dataset using a deep neural network to predict the behavioral trajectories of other traffic participants around the vehicle and identify potential conflict points between the vehicle and other traffic participants.

[0066] S3. Construct a game theory model with the vehicle and at least one other traffic participant as the players. Based on the behavioral trajectory and the potential conflict points, calculate an optimal action strategy for the vehicle by solving the game theory model.

[0067] S4, according to the optimal action strategy, control the vehicle to adjust at least one of the following: driving path, speed, or direction, in order to achieve conflict avoidance.

[0068] In this embodiment, the constructed environmental dataset is a time-series dataset: ;

[0069] in, This represents the environmental feature vector collected at time point t. Indicates the corresponding time tag;

[0070] Environmental feature vector Includes 3D surrounding terrain mapping data Capture the physical spatial structure around the vehicle and analyze the impact of dynamic weather conditions. It includes weather parameters such as temperature, humidity, rainfall, and visibility, as well as road surface condition assessment data. It can identify whether the road is wet, icy, or otherwise slippery, and it also collects data on the subtle movements of other vehicles and pedestrians. It captures subtle changes in motion through high-precision sensors, and predicts the intentions of other vehicles and pedestrians. Based on minute movements and behavioral history, it predicts short-term intentions using the vehicle's real-time status data. This includes position, velocity, and acceleration;

[0071] Each of these data items All data are multidimensional, with each data point collected at time point t, and combined to form an environmental feature vector. This is used for subsequent in-depth analysis and strategy calculation.

[0072] In this embodiment, three-dimensional surrounding terrain mapping data The collection includes:

[0073] LiDAR emits laser pulses and receives the reflected pulses. A stereo vision system uses two or more cameras to capture the scene from different angles and generate point cloud data. Corresponding points in the image Point cloud data Represented as a series of points: ;

[0074] Where i is the index of the point. These are the coordinates of a point in space;

[0075] Combine the point cloud data from the LiDAR with the image data generated by the stereo vision system: ;

[0076] in, This represents the position of the i-th point after merging. It is the coordinate value of the i-th point in the k-th dimension in the point cloud data. It is the coordinate value of the i-th point in the k-th dimension in the image data. and These are the weights of the point cloud data and the image data in the k-th dimension, respectively. k is the number of dimensions of the data. For 3D point cloud data and image data, k is 3.

[0077] Use the merged data Based on a point cloud and image data fusion algorithm, a 3D terrain mapping around the vehicle is generated through a 3D reconstruction algorithm. ;

[0078] in, It is a scalar function representing the position of point P in 3D space relative to the fused point cloud dataset. The fitting error is denoted by N, where N is the total number of points in the fused point cloud dataset. is the weight of the i-th point in relation to the reconstruction quality, determined based on the point's reliability or other indicators. P is a point in 3D space, representing a predicted location during the 3D reconstruction process. It is P and The Euclidean distance between them.

[0079] In this embodiment, road surface condition assessment data Based on comprehensive analysis of data from multiple sensors, including visual image analysis and environmental perception sensor data, road surface conditions are identified and assessed by combining image processing techniques and machine learning methods. ;

[0080] in, This represents the road surface condition assessed at time point t. Represents road surface image data captured at the same point in time. This represents the data collected at time point t by the temperature and humidity sensor, the light sensor, and the environmental sensing sensor. Used to extract image data and sensor data Features are extracted and road surface conditions are classified or scored, including slippery, icy, and dry. These are weighting coefficients, reflecting the weighting of each feature. The importance of N in road surface condition assessment, where N is the total number of characteristic functions, represents the comprehensive assessment of road surface conditions based on multiple characteristics.

[0081] In this embodiment, data on the minute movements of other vehicles and pedestrians The motion of each pixel in a video sequence is estimated by using optical flow to analyze minute motion changes of objects in the sequence. ;

[0082] in, Let represent the motion vector of the pixel located at coordinates (x, y) at time t, where and These are the velocity components in the horizontal and vertical directions, respectively;

[0083] The pixel intensity in the scene is constant across consecutive time frames. For times t and t+dt, the pixel intensity at positions (x,y) and (x+dx, y+dy) is... The optical flow equation is obtained as follows: ;

[0084] Under small motions, we can expand using Taylor series and ignore higher-order terms: ;

[0085] in, and These are the spatial derivatives of the image in the x and y directions, respectively. It is the time derivative, and u and v are the velocity components in the horizontal and vertical directions, respectively.

[0086] In step S1, the vehicle-to-vehicle communication system and the vehicle-to-infrastructure communication system are used to exchange data in the environmental dataset in real time, and to predict and share possible action plans in the next few seconds.

[0087] In this embodiment, step S2 specifically includes:

[0088] S21. For time series datasets After standardization and normalization, the preprocessed data is represented as follows: The environmental feature vector at each time point t After being converted by the preprocessing function P ;

[0089] S22, Utilizing deep neural network models Input preprocessed feature vector Output the behavior pattern representation at the corresponding time point. The model identifies the behavioral characteristics of different entities by learning the relationship between environmental features and behavioral patterns. The model expression is: ;

[0090] in, It is the preprocessed feature vector The i-th feature, Here, b is the weight corresponding to the i-th feature in the deep neural network, b is the bias term, and n is the feature vector. Dimensions It is an activation function used for nonlinear transformations;

[0091] S23, Behavioral Pattern Representation The deep neural network further predicts dynamic changes in the environment over the next few seconds, including the movement trajectories of other vehicles and pedestrians. The prediction function is expressed as: ;

[0092] in, This represents a prediction of future behavior based on behavioral patterns over the past K time points. Perform linear combination prediction. These are weight parameters, representing the importance of time point tk to the current prediction. It is the prediction error term, which includes random variations that cannot be explained by the model;

[0093] S24. Combining dynamic change prediction with current environmental characteristics, through algorithms... Analyze and identify the locations and times where conflicts may occur; potential conflict points are represented as... This includes the types of conflicts between vehicles and between vehicles and pedestrians, as well as the predicted timing and location of such conflicts. ;

[0094] in, It is a function that transforms future behavior predictions into corresponding future positions or state spaces. It is a function that, based on the current environmental feature vector Evaluate the current state space. It is a decision function used to calculate the difference between the future state and the current state to identify potential conflict points. When the result exceeds a certain threshold, it indicates a high risk of conflict.

[0095] In this embodiment, step S3 specifically includes:

[0096] S31, Based on vehicle dynamics constraints and traffic rules, generate a policy set for the vehicle containing multiple discretized behavior options;

[0097] S32, Define a game theory model, which includes a set of participants, a strategy space, and a set of utility functions;

[0098] S33, Based on the behavioral trajectory and potential conflict points, the optimal action strategy for this vehicle is determined from the strategy set by calculating the Nash equilibrium of the game theory model.

[0099] Specifically, a set of policies is built for each participating vehicle: ;

[0100] in, The i-th strategy that the vehicle may adopt is represented by the strategy set, which contains all possible behavioral responses and their corresponding consequences. The selection of the strategy is used to achieve conflict avoidance, efficiency maximization and safe distance maintenance.

[0101] Define the game theory model:

[0102] G = (N, S, U);

[0103] Where N is the set of vehicles participating in the decision-making process, S is the set of strategies, and U is the set of utility functions: ;

[0104] in, , , These represent evaluation functions for safety, efficiency, and regulatory compliance, respectively. , , These are weighting coefficients used to adjust the importance of different objectives;

[0105] Using the concept of Nash equilibrium (NE), we find the equilibrium point in a game theory model, specifically the Nash equilibrium point. When all participants choose the optimal strategy, no participant can gain higher utility by changing their strategy. ; ;

[0106] in, Let represent the set of strategy choices made by all vehicles except vehicle i. Indicates any other possible strategies considered by vehicle i;

[0107] Calculate the utility value of each vehicle under each possible strategy, including the success rate of conflict avoidance, minimization of required time, ability to maintain a safe distance, and adaptability to environmental changes: ;

[0108] in, Let j represent the cost or utility function corresponding to the j-th criterion, including conflict avoidance or time minimization. These are the corresponding weights, indicating the importance of the criterion;

[0109] Based on the concept of Nash equilibrium, each vehicle adjusts its behavior according to the equilibrium strategy in the game theory model to ensure optimal conflict avoidance and efficiency maximization in a dynamically changing road environment, while maintaining a safe distance. ; ;

[0110] For each vehicle i, select one that maximizes its utility. Maximization strategy As the optimal strategy, the optimal strategies of other vehicles are taken into account. .

[0111] In this embodiment, step S4 specifically includes:

[0112] S41. The optimal action strategy determined in step S3 Assign specific behavioral operations to each participating vehicle, including adjustments to the driving path P, speed V, and direction D, to achieve conflict avoidance;

[0113] S42. Define the driving path adjustment function According to the optimal action strategy Automatically calculate and adjust parameters:

[0114] Driving route adjustment takes into account environmental factors E and current location and target location : ;

[0115] in, It represents the gradient of the environment's influence on the path under a given optimal strategy, integrating factors such as traffic flow and obstacles;

[0116] Speed ​​adjustment combined with current speed Target speed and surrounding traffic speed Impact: ;

[0117] in, It is an adjustment factor. It is a balancing factor that adjusts the current speed towards the target speed or traffic flow speed preference;

[0118] Direction adjustment is based on the current direction Target direction and road conditions : ;

[0119] in, It is an adjustment factor. This represents the influence function of road conditions and optimal strategy on direction adjustment.

[0120] Continuously monitor vehicle status and changes in the surrounding environment, update data in real time, and repeat steps S1 to S3 as needed until the driving ends.

[0121] In this embodiment, by monitoring vehicle status and changes in the surrounding environment in real time, if new potential conflicts or changes in road conditions are encountered, the system will re-execute steps S1 to S4 to calculate a new optimal action strategy: ; ; ; ;

[0122] in, This represents the utility function of strategy s under new environments and states. The new adjustment function takes into account the latest environmental information and vehicle status, enabling real-time dynamic adjustment.

[0123] The aforementioned method integrates environmental data collected by a multimodal sensing system and utilizes deep neural networks for data analysis, enabling precise understanding of the surrounding environment and behavioral patterns. This improves the ability of autonomous vehicles to make real-time dynamic decisions in complex traffic environments. Furthermore, by employing game theory models to calculate optimal action strategies, potential conflicts are effectively prevented, thereby significantly enhancing the safety of autonomous vehicles.

[0124] Through advanced algorithms and intelligent software architecture, the system can quickly process large amounts of environmental data and vehicle status information, ensuring high efficiency in information processing. Based on this, it can automatically implement optimal driving strategies, which not only improves the accuracy of decision-making and response speed, but also enhances the adaptability and stability of the entire autonomous driving system.

[0125] This study applies game theory to the conflict avoidance decision-making process of autonomous vehicles, analyzing and resolving potential conflict problems in multi-vehicle shared road environments. By calculating game theory strategies such as Nash equilibrium, the decision-making process is optimized to minimize accident risks, establishing an effective cooperation or competition mechanism among autonomous vehicles.

[0126] Example 1:

[0127] Applying this invention to peak hours, this embodiment will demonstrate how to apply a game theory-based conflict avoidance method for autonomous vehicles to optimize the behavioral decisions of autonomous vehicles, thereby improving road use efficiency and reducing potential conflict risks.

[0128] On a busy city thoroughfare, two autonomous vehicles (AV1 and AV2) are about to simultaneously approach a four-way intersection without traffic lights. Pedestrians are crossing the zebra crossing. AV1 attempts to proceed straight through the intersection, while AV2 plans to travel straight from the opposite direction and then turn left. Without effective traffic management, this scenario could easily lead to reduced traffic efficiency and even accidents.

[0129] Guided by the method of this invention, AV1 and AV2 collect real-time data of the surrounding environment through their multimodal sensing systems, including vehicle dynamics at intersections, pedestrian positions, and road conditions. Simultaneously, the vehicles exchange information in real time using onboard communication systems (V2V and V2I), sharing their intentions and planned actions.

[0130] By analyzing the collected data through deep neural networks, the system can predict pedestrian movement trajectories and the intentions of other vehicles, identifying potential conflict points. Subsequently, game theory models are applied to calculate optimal action strategies for AV1 and AV2 respectively, including adjusting vehicle speed, changing driving routes, or altering the order of passing through intersections, in order to avoid potential conflicts and optimize traffic efficiency.

[0131] Table 1 Comparison of traffic conditions at intersections ,

[0132] As shown in Table 1 above, this embodiment verifies the method through simulation experiments. Without applying the method of this invention, the average travel time of AV1 and AV2 at the intersection is 45 seconds, and there are instances of minor sudden braking, increasing the risk of accidents. After applying the method of this invention, the average travel time of the two vehicles is reduced to 30 seconds, and there are no instances of sudden braking, effectively improving traffic efficiency and driving safety.

[0133] The specific experiment was conducted at a simulated intersection in a city during the morning rush hour (Monday morning, 8:00-9:00). By simulating 100 similar traffic scenarios, the traffic conflict rate was 20.36% without the application of the method of this invention; however, after applying the method of this invention, the conflict rate dropped to 2.84%, clearly demonstrating the effectiveness of this invention in conflict avoidance.

[0134] Furthermore, it was observed that after applying the method of the present invention, the average speed of AV1 and AV2 when passing through intersections increased by 15.7%, and the driving path was smoother, reducing unnecessary acceleration and deceleration, further improving fuel economy and ride comfort.

[0135] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A conflict avoidance method for autonomous vehicles based on game theory, characterized in that, Includes the following steps: The vehicle acquires environmental data through its multimodal sensing system and receives environmental data sent by other vehicles or infrastructure through its communication system to construct the vehicle's environmental dataset. The environmental dataset is analyzed using a deep neural network to predict the behavioral trajectories of other traffic participants around the vehicle and to identify potential conflict points between the vehicle and other traffic participants. Construct a game theory model with the vehicle and at least one other traffic participant as the players. Based on the behavioral trajectory and the potential conflict points, calculate an optimal action strategy for the vehicle by solving the game theory model. According to the optimal action strategy, the vehicle is controlled to adjust at least one of its driving path, speed, or direction to achieve conflict avoidance.

2. The method according to claim 1, characterized in that, The environmental dataset was constructed as a time-series dataset: ； in, This represents the environmental feature vector collected at time point t. Indicates the corresponding time tag; The environmental feature vector Includes at least one of the following data: 3D surrounding terrain mapping data Dynamic weather conditions impact analysis data Road surface condition assessment data Other vehicle and pedestrian micro-movement data Other vehicle and pedestrian intention prediction data and the vehicle's real-time status data .

3. The method according to claim 1, characterized in that, The steps of analyzing the environmental dataset using a deep neural network to predict the behavioral trajectories of other traffic participants around the vehicle and identifying potential conflict points between the vehicle and other traffic participants include: The time series data in the environmental dataset are standardized and normalized preprocessed to obtain the preprocessed environmental feature vector; The preprocessed environmental feature vector is input into a deep neural network, and the deep neural network outputs a behavioral pattern representation at the corresponding time point. Based on current and historical behavioral patterns, predict the dynamic changes of other traffic participants in the next few seconds; By combining the dynamic change prediction with the current environmental feature vector, potential conflict points between the vehicle and other traffic participants are identified.

4. The method according to claim 3, characterized in that, The step of inputting the preprocessed environmental feature vector into a deep neural network, and having the deep neural network output a behavioral pattern representation at the corresponding time point, includes: The behavioral pattern representation is obtained through the following calculations. : ； in, It represents the behavioral pattern at time t. It is the preprocessed environmental feature vector The i-th feature, Here, b is the weight corresponding to the i-th feature in the deep neural network, b is the bias term, and n is the preprocessed environmental feature vector. Dimensions It is a non-linear activation function; Based on current and historical behavioral patterns, the steps to predict the dynamic changes of other traffic participants in the next few seconds include: The following prediction function is used to predict the dynamic changes F(Bt) of other traffic participants in the next few seconds: ； in, Indicates a prediction of future behavior. It is the behavioral pattern at the k-th time point in the past. These are weight parameters. This is the prediction error term.

5. The method according to claim 3, characterized in that, The steps for identifying potential conflict points between the vehicle and other traffic participants, by combining the dynamic change prediction with the current environmental feature vector, include: The following algorithm identifies potential conflict points between this vehicle and other road users. : Among them, the function The future behavior prediction F(Bt) is transformed into the corresponding future position or state space, and the function Λ is based on the current environmental feature vector. Evaluate the current state space. It is a decision function; when When the result exceeds a predetermined threshold, a potential conflict point is identified. .

6. The method according to claim 1, characterized in that, The steps of constructing a game theory model with the vehicle and at least one other traffic participant as the players, and calculating an optimal action strategy for the vehicle based on the behavioral trajectory and the potential conflict points by solving the game theory model, include: Based on vehicle dynamics constraints and traffic rules, a policy set containing multiple discretized behavior options is generated for this vehicle; Define a game theory model, which includes a set of participants, a strategy space, and a set of utility functions; Based on the behavioral trajectory and potential conflict points, the optimal action strategy for this vehicle is determined from the strategy set by calculating the Nash equilibrium of the game theory model.

7. The method according to claim 6, characterized in that, The steps for generating a policy set containing multiple discretized behavioral options for the vehicle, based on vehicle dynamics constraints and traffic rules, include: The discretized behavior options in the strategy set are defined by velocity adjustment and direction adjustment angle, and each behavior option si is represented as: in, This is the speed adjustment amount, and its value range is... vd is the maximum deceleration limit, and va is the maximum acceleration limit; This is the direction adjustment angle, and its range is... , For the maximum leftward deviation angle, This is the maximum rightward deflection angle; The game theory model is represented as G=(N,S,U), where: N is the set of traffic participants involved in the decision-making process, including the vehicle itself and at least one other traffic participant; S is the policy space, consisting of the policy sets of all participants; U is the set of utility functions, containing the utility functions defined for each participant. Utility function defined for each participant Represented as: ； in, This indicates the strategy of this vehicle. , , These represent evaluation functions for safety, efficiency, and regulatory compliance, respectively, with α, β, and γ being weighting coefficients used to adjust the importance of different objectives.

8. The method according to claim 6, characterized in that, Based on the behavioral trajectory and potential conflict points, the steps for determining the optimal action strategy of this vehicle from the strategy set by calculating the Nash equilibrium of the game theory model include: Initialize and set the initial strategy for this vehicle. Initial strategy groups with other traffic participants ; To update the policy, perform the following calculations in sequence: Fix the current policies of other traffic participants and update the current policy of this vehicle: ； Fix the current policy of this vehicle and update the current policies of other traffic participants: ； When the policy change during continuous iteration The iteration stops when the value is less than a preset threshold ϵ, and the optimal action strategy for this vehicle is output. .

9. A conflict avoidance device for autonomous vehicles based on game theory, characterized in that, include: The acquisition module is used to acquire environmental data through the vehicle's multimodal sensing system and receive environmental data sent by other vehicles or infrastructure through the vehicle's communication system to construct the vehicle's environmental dataset. The behavior trajectory and potential conflict point prediction module is used to analyze the environmental dataset using a deep neural network to predict the behavior trajectories of other traffic participants around the vehicle and identify potential conflict points between the vehicle and other traffic participants. The optimal action strategy prediction module is used to construct a game theory model with the vehicle and at least one other traffic participant as the players. Based on the behavior trajectory and the potential conflict points, it calculates an optimal action strategy for the vehicle by solving the game theory model. The control module is used to control the vehicle to adjust at least one of its driving path, speed, or direction according to the optimal action strategy in order to achieve conflict avoidance.

10. A vehicle, characterized in that, Includes the game theory-based autonomous vehicle conflict avoidance device as described in claim 9.

Images