A method for real-time decision-making of automatic driving vehicle lane changing considering multi-vehicle interaction

Abstract

The application relates to the technical field of automatic driving, and particularly discloses a real-time decision-making method for automatic driving vehicle lane changing considering multi-vehicle interaction, which comprises the following steps: assigning a leader or follower identity to a vehicle in an interaction scene, selecting a game mode according to vehicle position information, and assigning a multi-stage game order to the selected game mode; according to the multi-stage game order, sequentially executing two-vehicle interaction solving of leaders and followers in each stage; constructing a trajectory planning problem of the leader into a model predictive control optimization problem, constructing a longitudinal decision of the follower into a cooperative driver model, and adopting a parallel gradient calculation method to accelerate solving of the model predictive control optimization problem; and selecting a strategy with the minimum target function value as a final decision-making strategy of the ego vehicle. Through the design of the multi-stage game, the application ensures that the influence of vehicles in the same lane and adjacent lanes on the ego vehicle is fully considered when the ego vehicle changes lanes, and the decision-making safety in a multi-vehicle lane changing interaction scene is improved.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, specifically a real-time lane-changing decision-making method for autonomous vehicles that considers multi-vehicle interaction. Background Technology

[0002] Autonomous driving systems mainly consist of four modules: environmental perception, behavior decision-making, trajectory planning, and tracking control. Among these, the behavior decision-making system, as one of the key technologies of autonomous driving systems, is of great significance for research. The decision-making system is the core of the autonomous driving system, responsible for processing perceived information and coordinating subsequent control commands. Based on the road information and surrounding vehicle information obtained by the perception module, it provides the autonomous vehicle with the optimal strategy for safety and efficiency.

[0003] With the development of autonomous driving technology, it is foreseeable that autonomous vehicles will share roads with human-driven vehicles for a long time to come. However, if autonomous vehicles make erroneous decisions during interactions, it could cause traffic congestion or even serious accidents. Therefore, research on the interactive decision-making of autonomous vehicles is of great significance to driving efficiency and traffic safety.

[0004] Multi-vehicle lane-changing scenarios are a typical example of interactive decision-making research. These scenarios require vehicles to find safe routes and timing to complete lane changes amidst complex interactions among multiple vehicles. Current research on multi-vehicle lane-changing decision-making mainly focuses on the interaction between the vehicle and vehicles in adjacent lanes, as well as the timing of lane changes, while neglecting the indirect influence of vehicles within the vehicle's own lane on the vehicle itself. If the vehicles in front of or behind the vehicle in the vehicle's own lane change before the vehicle changes lanes, it will affect the decision-making behavior of vehicles traveling in the target lane, further influencing the vehicle's lane-changing decision. Therefore, it is necessary to consider the influence of vehicles in front of or behind the vehicle in the vehicle's own lane on the vehicle's lane-changing decision.

[0005] Furthermore, current research on multi-vehicle interactive decision-making largely relies on simple models, which can lead to inaccuracies and affect the precision of the decision-making system. However, improving model accuracy in multi-vehicle systems requires an exponential increase in the amount of data processed. Since the decision-making layer, as the core of autonomous driving, needs to rapidly process perceived information and make reasonable decisions, this increased data volume places a burden on the decision-making layer, ultimately reducing driving safety. Therefore, it is necessary to optimize the decision-making algorithm architecture and solution speed to ensure both driving efficiency and safety. Summary of the Invention

[0006] The purpose of this invention is to provide a real-time lane-changing decision-making method for autonomous vehicles that takes into account multi-vehicle interaction, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A real-time lane-changing decision-making method for autonomous vehicles considering multi-vehicle interaction, the method comprising:

[0009] Based on the input vehicle and environmental information, assign leader or follower roles to vehicles in the interactive scenario, select the game mode based on vehicle location information, and assign the order of multi-stage games to the selected game mode.

[0010] Following the order of the multi-stage game, the dual-vehicle interaction solution for each stage of the game is executed sequentially. Specifically, the leader's trajectory planning problem is constructed as a model predictive control optimization problem, and the follower's longitudinal decision-making is constructed as a collaborative driver model. Parallel gradient calculation method is used to accelerate the solution of the model predictive control optimization problem.

[0011] After all stages of the game have been solved, the strategy that minimizes the objective function value is selected as the final decision strategy for the vehicle.

[0012] As a further aspect of the present invention, the principle for assigning leader or follower status is as follows: vehicles with the intention to change lanes are designated as leaders, and vehicles in the target lane are designated as followers.

[0013] As a further embodiment of the present invention, the game mode is divided into four modes based on whether there are vehicles in front and behind with the intention to change lanes in the same lane as the self-vehicle, including: self-vehicle priority interaction mode with no vehicles in front and no vehicles behind, vehicle interference mode with only vehicles behind, vehicle guidance mode with only vehicles in front, and full-dimensional coupling interaction mode with both vehicles in front and behind.

[0014] As a further aspect of the present invention, the rules for allocating the multi-stage game order include: when there is a preceding vehicle or a following vehicle with the intention to change lanes in the same lane, the game order of the preceding vehicle or the following vehicle takes precedence over that of the following vehicle.

[0015] As a further aspect of the present invention, for the aforementioned rear vehicle interference mode, the multi-stage game sequence is as follows:

[0016] Solve the game between the vehicle following in the same lane and the vehicle following in the target lane;

[0017] After obtaining the updated strategies of vehicles behind in the target lane and the predicted trajectories of vehicles behind in the same lane, the vehicle then engages in a game with the vehicle in front in the target lane and the updated strategies of vehicles behind in the target lane.

[0018] As a further aspect of the present invention, for the aforementioned vehicle guidance mode, the multi-stage game sequence is as follows:

[0019] Solve the game between the vehicle in front in the same lane and the vehicle in front in the target lane;

[0020] Based on the strategy of the vehicle in front in the target lane after its interaction with the vehicle in front in the same lane, and combined with the state of the vehicle in front in the same lane after it changed lanes, the vehicle engages in a game with the vehicle in front in the target lane and the vehicle behind in the target lane.

[0021] As a further embodiment of the present invention, the leader's model predictive control optimization problem is constructed based on a three-degree-of-freedom coupled dynamics model of the vehicle, and a brush model is used to describe the interaction between the tire and the ground.

[0022] As a further aspect of the present invention, the parallel gradient calculation method is used to accelerate the solution, including: when calculating the gradient of the objective function with respect to the control quantity, the gradient of the system model and the gradient of the objective function are calculated in parallel.

[0023] As a further aspect of the present invention, performing a phase of leader-follower dual-vehicle interaction solution includes the following sub-steps:

[0024] Predict the future state of leaders and followers based on vehicle models;

[0025] The leader's control parameters and predicted state parameters are updated iteratively using gradient descent.

[0026] The follower updates its own control and predicted state quantities based on the updated leader's predicted state quantities, and then updates its own control and predicted state quantities through the collaborative driver model.

[0027] Based on its updated control variables and predicted state variables, the leader then uses the gradient descent algorithm to solve for the optimal control variable again.

[0028] As a further aspect of the present invention, in the final stage of the multi-stage game, the interactive game process between the self-vehicle and the vehicle in front of and behind the target lane is executed in parallel computing mode.

[0029] Compared with the prior art, the beneficial effects of the present invention are: by incorporating the lane-changing behavior of the vehicle in front and behind in the lane into the decision-making framework, designing a multi-stage game scheme, dividing the game mode into four types, selecting the game mode through actual scenarios and designing multi-stage games, ensuring that the influence of vehicles in the same lane and adjacent lanes is fully considered when the vehicle changes lanes, thus improving the decision-making safety in multi-vehicle lane-changing interaction scenarios.

[0030] The proposed dual-vehicle interactive solution algorithm improves accuracy by constructing a highly coupled three-degree-of-freedom vehicle model and constructing a nonlinear optimization problem for lane-changing vehicles to improve trajectory prediction accuracy. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention.

[0032] Figure 1 This is a flowchart illustrating a real-time lane-changing decision-making method for autonomous vehicles that considers multi-vehicle interaction, provided as an embodiment of the present invention.

[0033] Figure 2 This is an experimental scenario diagram provided for an embodiment of the present invention.

[0034] Figure 3 A comparison chart of trajectory curves provided for embodiments of the present invention.

[0035] Figure 4 A longitudinal displacement comparison diagram provided for an embodiment of the present invention.

[0036] Figure 5 A longitudinal velocity comparison diagram provided for an embodiment of the present invention. Detailed Implementation

[0037] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0038] Figure 1 Here is an architectural block diagram of a real-time lane-changing decision-making method for autonomous vehicles that considers multi-vehicle interaction, such as... Figure 1 As shown in the embodiment of the present invention, a real-time lane-changing decision-making method for autonomous vehicles considering multi-vehicle interaction is provided, the method comprising:

[0039] Based on the input vehicle and environmental information, assign leader or follower roles to vehicles in the interactive scenario, select the game mode based on vehicle location information, and assign the order of multi-stage games to the selected game mode.

[0040] Following the order of the multi-stage game, the dual-vehicle interaction solution for each stage of the game is executed sequentially. Specifically, the leader's trajectory planning problem is constructed as a model predictive control optimization problem, and the follower's longitudinal decision-making is constructed as a collaborative driver model. Parallel gradient calculation method is used to accelerate the solution of the model predictive control optimization problem.

[0041] After all stages of the game have been solved, the strategy that minimizes the objective function value is selected as the final decision strategy for the vehicle.

[0042] As a preferred embodiment of the present invention, the principle for assigning leader or follower status is as follows: vehicles with lane-changing intentions are designated as leaders, and vehicles in the target lane are designated as followers.

[0043] In this embodiment, a multi-stage game strategy is selected. Based on the input vehicle information and environmental information, different vehicles are assigned the roles of leader or follower. In the multi-vehicle lane-changing interaction scenario, we assume that the lane-changing vehicle has priority decision-making power, and the vehicles in the adjacent lane will adjust their own decision-making behavior according to the lane-changing vehicle's choice. Therefore, the lane-changing vehicle is set as the game leader.

[0044] As a preferred embodiment of the present invention, the game mode is divided into four modes based on whether there are any vehicles in front and behind the vehicle with the intention to change lanes in the same lane as the vehicle, including: the vehicle priority interaction mode with no vehicles in front and no vehicles behind, the vehicle interference mode with only vehicles behind, the vehicle guidance mode with only vehicles in front, and the full-dimensional coupling interaction mode with both vehicles in front and behind.

[0045] In this embodiment, the game mode is selected based on the position information of the leader and followers. The game mode is divided into four types based on the presence or absence of a preceding and following vehicle. After determining the game mode, a multi-stage game order is assigned. Different game modes have different game orders. This invention sets a priority assumption, that is, when a preceding or following vehicle exists in the same lane, its game order takes precedence over the vehicle itself. Through priority game theory, the preceding and following vehicles, by influencing the behavior of vehicles in the target lane, further affect the game outcome between the vehicle and vehicles in the target lane when changing lanes.

[0046] Common lane change interaction scenarios such as Figure 2 As shown, different game modes are determined by whether there is a preceding vehicle (AV) and a following vehicle (RV) in the same lane. Here, we denote the preceding and following vehicles in the target lane as LV and OV, respectively, and the vehicle itself as SV.

[0047] Based on whether there are interfering vehicles with lane-changing intentions in the same lane as the vehicle, the game mode is subdivided into four types. The first is the vehicle-priority interaction mode, which is applicable when there are no vehicles in front or behind the vehicle in the same lane as the vehicle. In this case, the decision-making process is simplified to a game between the vehicle and the vehicles in the target lane. The vehicle, as the leader, predicts the reactive behaviors of the vehicles in front and behind in the target lane. By calculating the objective function values ​​of different interaction results, the strategy with the smallest objective function value and that satisfies the safety constraints is selected as the final trajectory.

[0048] As a preferred embodiment of the present invention, the rule for allocating the multi-stage game order includes: when there is a preceding vehicle or a following vehicle with the intention to change lanes in the same lane, the game order of the preceding vehicle or the following vehicle takes precedence over that of the following vehicle.

[0049] In a preferred embodiment of the present invention, the multi-stage game sequence for the rear vehicle interference mode is as follows:

[0050] Solve the game between the vehicle following in the same lane and the vehicle following in the target lane;

[0051] After obtaining the updated strategies of vehicles behind in the target lane and the predicted trajectories of vehicles behind in the same lane, the vehicle then engages in a game with the vehicle in front in the target lane and the updated strategies of vehicles behind in the target lane.

[0052] In this embodiment, when the scenario transforms into a multi-stage game mode with interference from following vehicles, i.e., there are vehicles behind the vehicle that also need to change lanes, the decision-making logic is divided into two stages. The first stage first solves the game result between the following vehicle in the same lane and the following vehicle in the target lane, obtaining the updated strategy of the following vehicle in the target lane. In the second stage, based on the known new strategy of the following vehicle in the target lane and the predicted trajectory of the following vehicle in the same lane, the vehicle engages in a second game with the vehicle in front of the target lane and the updated strategy of the following vehicle in the target lane. During this process, the system considers the previous stage's game trajectory of the following vehicle in the same lane to ensure that the vehicle's lane-changing decision does not collide with the vehicle that changed lanes earlier.

[0053] In a preferred embodiment of the present invention, the multi-stage game sequence for the aforementioned vehicle guidance mode is as follows:

[0054] Solve the game between the vehicle in front in the same lane and the vehicle in front in the target lane;

[0055] Based on the strategy of the vehicle in front in the target lane after its interaction with the vehicle in front in the same lane, and combined with the state of the vehicle in front in the same lane after it changed lanes, the vehicle engages in a game with the vehicle in front in the target lane and the vehicle behind in the target lane.

[0056] In this embodiment, the multi-stage game theory mode guided by the preceding vehicle is similar to the logic described above. It mainly describes the situation where there is a vehicle ahead that has changed lanes. In this mode, the vehicle ahead in the same lane and the vehicle ahead in the target lane first engage in a game to determine whether the vehicle ahead in the target lane should yield or not. Subsequently, based on the feedback from the vehicle ahead in the target lane and the state of the vehicle ahead in the same lane after changing lanes, the vehicle adjusts the interaction weights between itself and the two vehicles in the target lane, thereby finding a safe entry point in complex traffic flow.

[0057] The last type is the full-dimensional coupled interaction mode, which requires considering the influence of both the vehicle in front and the vehicle behind simultaneously. Regardless of the mode, the final stage of the game involves the vehicle choosing the strategy with the minimum objective function value from the game results with both the vehicle in front and the vehicle behind in the target lane. Through this multi-stage game strategy, this invention ensures decision-making safety in multi-vehicle lane-changing scenarios.

[0058] As a preferred embodiment of the present invention, the leader's model predictive control optimization problem is constructed based on the three-degree-of-freedom coupled dynamics model of the vehicle, and a brush model is used to describe the interaction between the tire and the ground.

[0059] In this embodiment, after the multi-stage game strategy is formulated, the two-vehicle interactive solution needs to be performed sequentially. The two-vehicle interactive solution mainly includes the design of vehicle models, leader and follower driver models, and interactive solution algorithms.

[0060] Vehicle Model: Based on the vehicle input state, the vehicle in the lane change lane is defined as the leader, and the vehicle in the target lane is defined as the follower. Unlike single-cycling tasks, interactive tasks emphasize coordinated movement within an interactive scenario. Therefore, interactive scenarios require both accurate vehicle models and precise state predictions. This paper establishes a three-degree-of-freedom bicycle vehicle model based on vehicle dynamics and coupling characteristics to capture core longitudinal, lateral, and yaw dynamics. Furthermore, to improve the model's accuracy in interactive scenarios, this invention employs a brush model to describe the tire's state during vehicle motion.

[0061] As a preferred embodiment of the present invention, performing a phase of leader-follower dual-vehicle interaction solution includes the following sub-steps:

[0062] Predict the future state of leaders and followers based on vehicle models;

[0063] The leader's control parameters and predicted state parameters are updated iteratively using gradient descent.

[0064] The follower updates its own control and predicted state quantities based on the updated leader's predicted state quantities, and then updates its own control and predicted state quantities through the collaborative driver model.

[0065] Based on its updated control variables and predicted state variables, the leader then uses the gradient descent algorithm to solve for the optimal control variable again.

[0066] In this embodiment, the dual-vehicle interactive solution process is as follows: First, the future states of the leader and followers are predicted based on the vehicle model. Next, the control quantity and predicted state quantity of the leader vehicle are iteratively updated using gradient descent. Then, the followers update their control quantity and predicted state quantity using a cooperative driver model based on the updated leader's predicted state quantity. Finally, the leader uses the above information to solve for the optimal control quantity using the gradient descent algorithm. According to the multi-stage game process, the autonomous vehicle ultimately selects the outcome with the smaller objective function value as its strategy based on the game results with the vehicles in front and behind in the target lane.

[0067] The driver model is used to solve for vehicle control variables. Since the leader needs to plan road changes and requires precise lateral and longitudinal control, a model predictive control algorithm is used here to implement vehicle planning. The vehicle dynamics equations can be defined as follows:

[0068] ;

[0069] in, The derivative of a state variable express about and The function, For state variables, These are, respectively, longitudinal velocity, lateral velocity, angular velocity, angular velocity, longitudinal displacement, and lateral displacement. To control the quantity, , which are longitudinal acceleration and front wheel steering angle, respectively.

[0070] Define the prediction time domain as equal to the control time domain, denoted as The objective function is expressed as:

[0071] ;

[0072] in , , , For the corresponding variable weight coefficients, This indicates that the state variable follows the target. This indicates a penalty for changes in the control quantity. Indicates lane boundary constraints. The formulas for representing safety constraints are as follows:

[0073] ;

[0074] ;

[0075] ;

[0076] ;

[0077] in For reference values ​​of state variables, and These represent the upper and lower limits of the road boundary, respectively. and These represent the longitudinal and lateral displacements of the obstacle vehicle, respectively.

[0078] To ensure that the optimized control sequence meets the vehicle's physical limits and the safety requirements of the interaction scenario, the following constraints are set:

[0079] ;

[0080] in, and These represent the minimum and maximum values ​​of the control quantity, respectively.

[0081] The final leader optimization problem is constructed as follows:

[0082] ;

[0083] ;

[0084] ;

[0085] The optimal control quantity for the leader can be obtained by solving the above optimization problem.

[0086] Vehicles in the target lane must consider whether to accelerate to avoid or decelerate to allow the vehicle ahead to change lanes, and also consider whether their deceleration will result in them being too close to the vehicle behind them, potentially causing a hazard. Furthermore, since the target lane only requires longitudinal decision-making, a cooperative driver model is established for the following vehicles, with the longitudinal acceleration control equation being:

[0087] ;

[0088] in, Indicates the maximum acceleration. For longitudinal velocity, For longitudinal reference speed, For acceleration index, To maintain a safe distance from the vehicle in front, This indicates the desired spacing. Let be the function that determines the existence of the preceding vehicle, and the condition for determining it is:

[0089] ;

[0090] Expected speed Based on the change in distance from the vehicle behind, the formula for change is:

[0091] ;

[0092] in, Indicates the speed of collaboration. Indicates non-cooperative speed.

[0093] Expected Spacing The solution formula is:

[0094] ;

[0095] in, The stationary safe distance is the minimum net distance maintained between vehicles when one is completely stationary and the other. For safe headway, it indicates the desired time interval between the driver and the vehicle in front. It is the speed difference between this vehicle and the vehicle in front. For comfortable deceleration; through this follower-driver model, the follower decides on the appropriate acceleration to ensure safe driving during interaction.

[0096] As a preferred embodiment of the present invention, the parallel gradient calculation method is used to accelerate the solution, which includes: when calculating the gradient of the objective function with respect to the control quantity, the gradient of the system model and the gradient of the objective function are calculated in parallel.

[0097] In this embodiment, in order to reduce the amount of computation, gradient parallel computing is used to accelerate the leader planning problem, wherein the gradient of the system model and the gradient of the objective function are calculated in parallel when calculating the gradient.

[0098] In leader-follower games, followers need to consider the leader's decision-making behavior when making decisions. The leader predicts the followers' decision-making actions and optimizes its own decision-making behavior accordingly. Therefore, in the solution process, the future states of the leader and followers are first predicted by the model. The leader solves the problem for the first time based on the predicted information, and the followers then optimize their own trajectories based on the trajectory calculated by the leader. It is worth noting that when playing against SV games, followers OV in the rear vehicle interference mode, followers LV in the front vehicle guidance mode, and OV and LV in the full-dimensional coupled interaction mode all need to consider the strategy trajectories of the other two leaders RV and AV in the first stage of the game when optimizing their own trajectories to avoid collisions. Finally, the leader solves the problem again to complete a single two-vehicle interaction.

[0099] During the solution process, since the leader-finding algorithm needs to be executed twice and is relatively complex, consuming a significant amount of solution time, gradient parallel computation is adopted for the leader nonlinear optimization problem formula. Gradient computation has the advantage of fast iterative convergence and also has good parallelism. The gradient update formula is as follows:

[0100] ;

[0101] in, Indicates the control quantity. This indicates the updated control quantity. The iteration step size factor can improve the convergence efficiency of the algorithm by choosing an appropriate iteration step size. Let the gradient of the objective function with respect to the control variable be denoted by the sampling time. In calculation At this time, the gradient of the objective function with respect to the state variables at the next time step is needed. System model gradient and ,and and At the same time, they do not affect each other, and the gradient of the system model can be calculated in advance without affecting each other. Therefore, parallel computing is designed to improve the solution speed. In addition, at the algorithm level, the interaction game between the self vehicle and the two vehicles in the adjacent lane in the last stage is also independent. Since each game mode includes this final stage, parallel processing can significantly reduce the total solution time.

[0102] In a preferred embodiment of the present invention, in the last stage of the multi-stage game, the two-vehicle interactive game process between the self-vehicle and the vehicle in front of and behind the target lane is executed in parallel computing mode.

[0103] In this embodiment, to accelerate computation, a cooperative driver model is built for straight-going vehicles to save computational costs. The algorithm architecture is optimized by parallel computing the process of solving the optimal strategy of the vehicle in the last stage of the multi-stage game. Parallel computing is used to accelerate the solution process, and a fast-converging gradient descent algorithm is designed. When calculating the gradient, the gradient of the system model and the gradient of the objective function are calculated in parallel.

[0104] To further verify the advantages of multi-stage game-theoretic decision-making algorithms for autonomous vehicles that consider multi-vehicle interaction, a design was developed as follows: Figure 3 The scenario shown depicts a leader and followers using the same vehicle model with the following parameters: vehicle mass m = 1455 kg, center of gravity height hcg = 0.65 m, and moment of inertia about the vertical axis. The wheelbase lfr = 2.660m, of which the front wheelbase lf = 1.050m and the rear wheelbase lr = 1.610m. The front wheel lateral stiffness Caf = 92992.0 and the rear wheel lateral stiffness Car = 88783.0. With the lower left of the road as the zero coordinate, the vehicle coordinates are: RV(88,1.75), SV(125,1.75), AV(164,1.75), OV(100,5.25), LV(150,5.25). The initial speeds are RV = 14m / s, SV = 14m / s, AV = 14m / s, OV = 13.8m / s, and LV = 14m / s. In the simulation experiment, the leader uses a model predictive control algorithm, setting the prediction time domain equal to the control time domain at 40, and the maximum value of the control vector. minimum value The gradient update iteration step size factor d = 0.0001, and the maximum number of iterations is 5. In the follower collaborative decision-making model, the collaboration speed... Non-cooperative speed Maximum acceleration stationary safe distance Comfort deceleration Acceleration index Safe headway Simulation results in Figure 3 , Figure 4 and Figure 5 exhibit, Figure 3 This is a comparison chart of trajectory curves. Figure 4 This is a comparison diagram of longitudinal displacement. Figure 5 This is a comparison of longitudinal speeds. The experimental results show that the multi-stage game-theoretic decision-making algorithm for autonomous vehicles, which considers multi-vehicle interaction, can ensure the safety of the driver vehicle when the vehicle in front or behind changes lanes first in a forced lane change scenario.

[0105] For multi-vehicle lane-changing scenarios, current research focuses on the impact of adjacent lane vehicle decision-making behavior on the vehicle's own lane-changing decision, while neglecting the indirect changes caused by vehicles in the vehicle's own lane. If the vehicle in front of or behind the vehicle in the vehicle's own lane changes lanes before the vehicle in a multi-vehicle lane-changing scenario, it will affect the decision-making behavior of vehicles in the target lane, further influencing the vehicle's own decision. The multi-stage game decision-making method for autonomous vehicles that considers multi-vehicle interaction proposed in this invention integrates the lane-changing behavior of the vehicle in front of and behind the vehicle in the vehicle's own lane into the decision-making framework, designs a multi-stage game scheme, divides the game mode into four types, selects the game mode through actual scenarios, and designs multi-stage games to ensure that the influence of vehicles in the same lane and adjacent lanes is fully considered when the vehicle changes lanes, thereby improving the decision-making safety in multi-vehicle lane-changing scenarios.

[0106] Interactive decision-making algorithms require predicting vehicle trajectories during interaction. High-precision vehicle trajectories can improve decision-making efficiency and reduce unnecessary scenario simulations. However, the computational load also increases exponentially with model complexity and the number of interacting vehicles. This increased computational load affects the vehicle's reaction speed during interaction and reduces safety. To address these issues, this invention proposes a dual-vehicle interactive solution algorithm. To improve accuracy, it constructs a highly coupled three-degree-of-freedom vehicle model and performs a nonlinear optimization problem for lane-changing vehicles to enhance trajectory prediction accuracy. Through these methods, this invention improves decision-making efficiency and safety in multi-vehicle interactive scenarios.

[0107] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A real-time lane-changing decision-making method for autonomous vehicles considering multi-vehicle interaction, characterized in that, The method includes: Based on the input vehicle and environmental information, assign leader or follower roles to vehicles in the interactive scenario, select the game mode based on vehicle location information, and assign the order of multi-stage games to the selected game mode. Following the order of the multi-stage game, the dual-vehicle interaction solution for each stage of the game is executed sequentially. Specifically, the leader's trajectory planning problem is constructed as a model predictive control optimization problem, and the follower's longitudinal decision-making is constructed as a collaborative driver model. Parallel gradient calculation method is used to accelerate the solution of the model predictive control optimization problem. After solving the game at all stages, the strategy with the minimum objective function value is selected as the final decision strategy for the vehicle. The principle for assigning leader or follower roles is: vehicles with the intention to change lanes are designated as leaders, and vehicles in the target lane are designated as followers. The game mode is divided into four modes based on whether there are any vehicles in front and behind the vehicle with the intention to change lanes in the same lane as the vehicle: the vehicle priority interaction mode with no vehicles in front and no vehicles behind, the vehicle interference mode with only vehicles behind, the vehicle guidance mode with only vehicles in front, and the full-dimensional coupled interaction mode with both vehicles in front and behind. The rules for allocating the order of multi-stage games include: when there is a vehicle in front or behind that intends to change lanes in the same lane, the game order of the vehicle in front or behind takes precedence over that of the vehicle itself.

2. The real-time lane-changing decision-making method for autonomous vehicles considering multi-vehicle interaction according to claim 1, characterized in that, For the aforementioned rear vehicle interference mode, the multi-stage game sequence is as follows: Solve the game between the vehicle following in the same lane and the vehicle following in the target lane; After obtaining the updated strategies of vehicles behind in the target lane and the predicted trajectories of vehicles behind in the same lane, the vehicle then engages in a game with the vehicle in front in the target lane and the updated strategies of vehicles behind in the target lane.

3. The real-time lane-changing decision-making method for autonomous vehicles considering multi-vehicle interaction as described in claim 1, characterized in that, For the aforementioned lead vehicle guidance mode, the multi-stage game sequence is as follows: Solve the game between the vehicle in front in the same lane and the vehicle in front in the target lane; Based on the strategy of the vehicle in front in the target lane after its interaction with the vehicle in front in the same lane, and combined with the state of the vehicle in front in the same lane after it changed lanes, the vehicle engages in a game with the vehicle in front in the target lane and the vehicle behind in the target lane.

4. The real-time lane-changing decision-making method for autonomous vehicles considering multi-vehicle interaction as described in claim 1, characterized in that, The leader's model predictive control optimization problem is constructed based on a three-degree-of-freedom coupled dynamics model of the vehicle, and a brush model is used to describe the interaction between the tire and the ground.

5. The real-time lane-changing decision-making method for autonomous vehicles considering multi-vehicle interaction according to claim 1, characterized in that, Accelerating the solution by using parallel gradient calculation methods includes: when calculating the gradient of the objective function with respect to the control quantity, calculating the gradient of the system model and the gradient of the objective function in parallel.

6. The real-time lane-changing decision-making method for autonomous vehicles considering multi-vehicle interaction according to claim 1, characterized in that, The solution to the leader-follower two-vehicle interaction problem for a phase includes the following sub-steps: Predict the future state of leaders and followers based on vehicle models; The leader's control parameters and predicted state parameters are updated iteratively using gradient descent. The follower updates its own control and predicted state quantities based on the updated leader's predicted state quantities, and then updates its own control and predicted state quantities through the collaborative driver model. Based on its updated control variables and predicted state variables, the leader then uses the gradient descent algorithm to solve for the optimal control variable again.

7. The real-time lane-changing decision-making method for autonomous vehicles considering multi-vehicle interaction according to claim 1, characterized in that, In the final stage of the multi-stage game, the interactive game process between the voluntary vehicle and the vehicle in front of and behind the target lane is executed in parallel computing mode.

Images