Intelligent network connection vehicle real-time planning method based on shielding perception and game

Abstract

The invention relates to an intelligent network connection vehicle real-time planning method based on occlusion perception and game, and the method comprises the steps: collecting the surrounding environment data of a vehicle, and exchanging the state information of the vehicle with nearby vehicles and traffic infrastructures through the Internet of Vehicles communication; constructing a dynamic game model according to the vehicle surrounding environment data and other vehicle states obtained by vehicle networking communication; dividing a shielding open-loop game stage and a feedback game stage based on shielding objects and communication time delay conditions in the vehicle surrounding environment data; and based on the dynamic game model, adopting corresponding game methods for a shielding open-loop game stage and a feedback game stage, and carrying out vehicle track planning. By means of the method, the automatic driving vehicle can still conduct safe and efficient motion planning under the condition that partial perception is lacked, the strategy is adjusted in time after the view field is restored or key information is obtained, and therefore the robustness and reliability of an automatic driving system in the real traffic environment are improved.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving and intelligent vehicle control technology, and in particular to a real-time planning method for intelligent connected vehicles based on occlusion perception and game theory. Background Technology

[0002] Currently, the planning and decision-making processes of intelligent connected vehicles need to operate in complex and ever-changing traffic environments, but they face two major challenges: first, the incompleteness of perception caused by obstructions; and second, the uncertainty of human drivers' intentions. On actual roads, static obstacles such as buildings, large trucks, or road curves often block the vehicle's sensor line of sight, preventing autonomous vehicles from observing obstructed pedestrians or vehicles. When visibility is obstructed, traditional planning must consider potential hidden risks that may suddenly appear, otherwise, there is a risk of collision. On the other hand, human drivers, bicycles, and other traffic participants have their own driving strategies and goals, which are not entirely consistent with those of autonomous vehicles. Strategic intentions of human drivers, such as yielding or cutting in front, are difficult to predict in advance, directly exacerbating the uncertainty in the decision-making process. If planning does not consider the reactions of other drivers, it may lead to overly aggressive behavior and accidents; however, if it simply assumes that "no one will give way" as the worst-case scenario, the vehicle's driving strategy will become overly conservative, ultimately affecting traffic efficiency.

[0003] Current research status: To address the aforementioned problems, researchers from different fields have employed methods such as game theory and partially observable programming, but each has its limitations. Dynamic game theory was introduced into the field of robot planning as early as the 1990s to formally describe problems such as multi-robot pursuit and cooperation. Game models can treat multi-agent interactions as each party rationally optimizing its own goals, thus avoiding the uncertainty of a single vehicle predicting the open-loop behavior of other vehicles. However, in terms of information acquisition, classical game planning mainly has two extreme assumptions: the core assumption of open-loop games is that after the initial moment, each participant loses the ability to observe others—that is, they act "blindly." While this setting can effectively cover extreme occlusion situations, it can fall into the trap of over-conservatism due to the lack of dynamic feedback; feedback-based games assume that each vehicle can perfectly observe the state of all other vehicles at every moment, allowing mutual reactions, but this assumption does not hold when occlusion exists. Games with only these two information structures are difficult to cope with scenarios where the line of sight changes dynamically in reality. To address this, scholars began studying game theory models with hybrid information structures. Zhang et al. proposed a zero-sum game model for occlusion perception, alternating between open-loop and closed-loop information stages in car driving to simulate intermittent vision caused by static obstacles. However, zero-sum games assume complete adversarial relationships between opponents, making them only suitable for two-car competition scenarios such as collision avoidance, and unsuitable for general multi-car cooperative / game-based traffic. Gupta et al. subsequently extended hybrid information games to general draws, i.e., N-car non-zero-sum games, proposing the OGSolve algorithm. This algorithm solves a series of linear quadratic subgames to achieve multi-car planning in occlusion scenarios, achieving polynomial time complexity for the state dimension. However, OGSolve and related methods neglect the uncertainty of other vehicles' goals in the model. This method assumes that the driving intentions of all vehicles are specifically reflected in their cost functions, and that these intentions are known and fixed. This means that in highly interactive scenarios, there may still be situations where the planning is too conservative or the decision-making is incorrect: for example, if the actual behavior of other vehicles does not conform to the assumptions, the planning will fail.

[0004] Meanwhile, the fields of robotics and autonomous driving have also developed planning methods to address intent uncertainty. One approach is to incorporate multiple assumptions about driving intent into the planning considerations: Peters et al. proposed the "contingency game" framework, also known as contingency planning, which allows autonomous vehicles to plan a main trajectory and automatically switch to different branch strategies at some future moment based on the revelation of the true intent of other agents. This method is equivalent to preparing "Plan A" and "Plan B" in advance at key decision points, which can improve adaptability under uncertain interactions. However, existing contingency game models have a prerequisite: the system must have sufficient observation information about the environment before the strategy branch occurs. This setting is based on the premise of no perception of occlusion. This assumption is difficult to meet in many driving scenarios—if key intent information such as whether oncoming vehicles exist or whether to slow down and yield cannot be obtained in time due to occlusion, the contingency branch cannot be determined at the moment, and the planning may fall into a blind spot. In other words, previous intent contingency planning methods cannot directly handle situations where the line of sight is temporarily obstructed.

[0005] Besides game-theoretic planning, other methods address occlusion and intent uncertainty. For example, set-based reachability analysis and other reachability domain analysis methods conservatively plan by calculating the safe range where hidden vehicles might be located, ensuring safety but often sacrificing efficiency. To reduce occlusion uncertainty, proactive perception planning attempts to actively acquire a more comprehensive view by adjusting the vehicle's actions, such as slowing down and probing. However, it typically treats other traffic participants as passive obstacles, neglecting their strategic responses, thus limiting its effectiveness in highly interactive scenarios. Furthermore, traditional prediction techniques mostly handle intent prediction and trajectory planning separately, independently predicting the possible trajectories of other vehicles before planning, without being able to influence the behavior of other vehicles, resulting in a disconnect between prediction and decision-making.

[0006] In summary, a unified planning framework is currently lacking that can handle line-of-sight occlusion in dynamic environments, address the uncertainty of multi-person interactive decision-making, and meet real-time requirements. Existing technologies often focus only on one aspect, leading to either overly conservative waiting resulting in low efficiency or rash actions ignoring hidden dangers, resulting in poor safety. Therefore, it is necessary to provide a new planning method that considers both occlusion and human-vehicle interaction factors to overcome these shortcomings. Summary of the Invention

[0007] The purpose of this invention is to overcome the limitations of existing autonomous driving planning methods that operate in isolation, and to provide a real-time planning method for intelligent connected vehicles based on occlusion perception and game theory. This invention addresses the problems of overly conservative or recklessly aggressive approaches in existing technologies due to solely considering limited visibility or driver intent, proposing a unified solution that enables intelligent vehicles to plan safe and efficient driving strategies even in environments with incomplete perception and multi-vehicle interaction. Specifically, the technical problem this invention aims to solve is how to enable autonomous vehicles to anticipate various possible situations and react appropriately when two uncertainties exist simultaneously: temporary loss of visibility and unclear human driver strategies. This avoids both stalling due to unknown dangers and accidents caused by incorrect assumptions about other vehicles' behavior. Through the method of this invention, intelligent connected vehicles can adopt appropriate defensive driving strategies to ensure safety during periods of occlusion, and quickly switch to an aggressive strategy to complete the task after regaining visibility or clarifying the intentions of other vehicles, thus improving overall traffic efficiency and safety.

[0008] To achieve the above objectives, the present invention provides the following solution: A real-time planning method for intelligent connected vehicles based on occlusion perception and game theory includes: It is used to collect data on the environment around the vehicle and exchange its own status information with nearby vehicles and traffic infrastructure through vehicle-to-everything (V2X) communication. A dynamic game model is constructed based on the vehicle's surrounding environment data and the status of other vehicles obtained through vehicle-to-everything (V2X) communication. Based on the obstructions and communication latency in the vehicle's surrounding environment data, the obstruction open-loop game stage and the feedback game stage are divided. Based on the dynamic game model, corresponding game methods are used for the occlusion open-loop game stage and the feedback game stage to plan vehicle trajectories.

[0009] Optionally, constructing the dynamic game model includes: Each vehicle updates its state in discrete time, where the state variables include position, velocity, and heading; the state update equation for each vehicle is: in, Let represent the state of the i-th vehicle at time t. To control its input, For vehicle kinematics functions, The discrete time step; The dynamic game model is used to characterize the dynamic behavior of vehicles on a two-dimensional plane. Each vehicle makes decisions on control inputs by optimizing its own cost function, forming an N-player non-zero-sum dynamic game problem. Under the game framework, each vehicle tries to minimize its own cost, which includes deviation from the target trajectory, ride comfort, violation of traffic rules, and the risk of collision. The vehicle itself, as player 0 in the game, needs to interact and make decisions with other vehicle players 1...N.

[0010] Optionally, the division of the occlusion open-loop game stage and the feedback game stage includes: When the sensor line of sight of the vehicle is blocked, it enters the occlusion open-loop game stage; in the occlusion open-loop game stage, it can only make decisions based on its own state and the environmental information at the beginning, and cannot observe the real-time state changes of other vehicles. When there are no obstructions and the field of vision is unobstructed, or when the status of other vehicles is obtained through vehicle-to-everything (V2X) communication, the feedback game phase begins. During the feedback game phase, each vehicle is fully aware of the current status of all other vehicles and adjusts its strategy according to the real-time status to achieve interactive game. When the field of vision is limited due to static objects blocking it, the strategies of each vehicle automatically switch to the open-loop game stage of occlusion; when the occlusion disappears, the strategies switch to the feedback game stage.

[0011] Optionally, based on the dynamic game model, corresponding game theory methods are applied to the occlusion open-loop game stage and the feedback game stage to perform vehicle trajectory planning, including: Determine the driving intent of the other vehicle; wherein, the driving intent of the other vehicle includes: the intent to yield and the intent to force; Based on the driving intentions of the other vehicle, determine the timing of the strategy branch; Based on the branching time of the strategy, and combined with communication delay compensation, a vehicle trajectory including a main control sequence and a branch control sequence is planned. By integrating the strategy branch time, the planned vehicle trajectory, the occlusion open-loop game stage, and the feedback game stage into the dynamic game model, a hybrid information game model is obtained. An iterative linear quadratic approximation algorithm is used to solve the hybrid information game model and obtain the equilibrium strategy.

[0012] Optionally, determining the driving intention of the other vehicle includes: For each other participating vehicle, a discrete driving intention hypothesis library is pre-defined; Using Bayesian inference to determine the confidence level of a set of intentions This indicates information history Get off the vehicle Belongs to the The probability of each intention type Here is the confidence probability. In the history of information Under the given conditions, the event The probability of occurrence Let be the "intent parameter" for the j-th vehicle. Let k be the "intention hypothesis" for the j-th vehicle; The initial time is set based on prior knowledge. During the planning and execution process, whenever the actions of other vehicles are observed, the vehicle updates the posterior probability of its intention using the observation-update model: in, Assumption type Get off the vehicle The instantaneous cost function, The rationality coefficient, For vehicle j at time step The state vector, Let be the control input vector for the j-th vehicle at time step t. This is the global state vector at time step t; Based on the observation-update model described above, the confidence probability of each intent hypothesis is adjusted in real time through recursive Bayesian filtering. ; Using information entropy to measure intention uncertainty: ;like If the value is higher than a preset threshold, it indicates that the vehicle's intention is still unclear; conversely, if the value is lower than a preset threshold, it indicates that the vehicle's intention is still unclear. If the value is below a preset threshold, its type is determined.

[0013] Optionally, determining the policy branch time includes: The triggering time of the strategy branch is automatically determined by detecting the confidence entropy of the vehicle's intentions toward all other vehicles; in, For the planned branch point.

[0014] Optionally, the iterative linear quadratic approximation algorithm is used to solve the mixed information game model, including: a. Select an initial control sequence, simulate the system on the initial control sequence to obtain the initial state trajectory, perform Taylor expansion linearization on the dynamics of each vehicle at each time step at the current trajectory to obtain a linear approximation model, and perform a quadratic expansion approximation on the cost function of each player to obtain a quadratic cost approximation, thus forming a linear quadratic (LQ) game subproblem with linear dynamics and quadratic cost. b. Based on the preset piecewise Riccati equation backtracking algorithm, the feedback Nash solution of the LQ game subproblem is obtained, and the optimal linear feedback control law of each vehicle in each feedback stage is obtained. c. Using the optimal linear feedback control law, perform global forward simulation to obtain new state and belief trajectories. During the simulation, simultaneously track the evolution of intentional beliefs, determine new branch moments, and use the new trajectory for the next linearization and quadratic approximation. d. Repeat steps a, b, and c, continuously adjusting the strategy until the cost of each vehicle converges or the number of iterations reaches the preset upper limit, thus obtaining an approximate optimal strategy set for complex dynamic and uncertain environments.

[0015] Optionally, the piecewise Riccati equation backtracking algorithm includes: Full time domain The system is divided into several continuous intervals, each interval representing a visible feedback phase (where the policy depends on the current full state) or an occlusion open-loop phase (where the policy depends only on the state at the start of occlusion); from the endpoint... Starting with forward recursion, during the visible feedback phase, the optimal feedback strategy takes an affine form: ,in, , The optimal feedback gain matrix is; the corresponding quadratic form of the value function is ,in, for The quadratic form of the value function, also known as the optimal residual value, The coefficient matrix of the quadratic terms is The Hessian matrix (second derivative). Let be the vector of coefficients of the first-order terms. The gradient vector (first derivative). Given the next step value function parameters. Then, the optimal feedback gain and It is calculated using the Riccati recursive formula for classical discrete-time general and LQ games; In the open-loop occlusion stage, an improved open-loop optimization is applied to recursively solve for the corresponding value matrix. Let the parameters of the open-loop valued function be... ,definition: in, For open-loop coupling correction matrix, It is the identity matrix. The input matrix is ​​used to control the quadratic coefficients of the nonlinear cost function after a Taylor expansion. For the first jPenalty weight matrix of vehicle to its own control input Then, the value matrix obtained by backward recursion under the KKT conditions in the open-loop information structure is: in, Let be the state transition matrix of the system. for The weight matrix of the state terms in the cost function.

[0016] To ensure seamless connection of value functions at the boundaries of intervals, the connection rule is designed as follows: Let the time interval be... The subsequent time interval is In the visible feedback phase, the open-loop function parameters of the next segment are used. Alternative calculation During the occlusion open-loop phase, the open-loop value function parameters of the next segment are used. Alternative calculation This process of alternating calculations in segments continues until time 0.

[0017] In summary, this mixed-information Ricatti solution obtains the solution from... arrive During this period, the optimal linear feedback control law for each vehicle in each feedback game stage and the optimal open-loop control sequence in the occlusion open-loop game stage are determined.

[0018] Optionally, the vehicle trajectory planning also includes: The obtained near-optimal strategy set is sent to the vehicle's underlying execution to realize vehicle movement. At fixed time intervals, the current moment is used as a new starting point to repeatedly build the dynamic game model and re-plan the vehicle trajectory.

[0019] The beneficial effects of this invention are as follows: This invention integrates occlusion handling and intent inference into a single planning framework, directly bringing several beneficial effects: First, in terms of safety, the vehicle can effectively avoid accidents caused by hidden hazards. By adopting conservative strategies and branching plans during occlusion, the vehicle can react promptly once the hidden vehicle appears or its intent becomes clear, significantly reducing the risk of collision. Simulation data shows that compared to imperfect methods that only consider occlusion or only consider intent, the method of this invention reduces the collision accident rate by more than 60%, achieving a higher safety margin in complex scenarios. Second, in terms of efficiency and comfort, this invention avoids unnecessary waiting and sudden braking caused by excessive conservatism. This means that while ensuring safety, the vehicle can more actively complete overtaking, merging, and other actions, thereby shortening task time and reducing energy consumption. In the test scenario, the average driving time of the vehicle using the planning of this invention did not increase significantly, but fuel or electricity consumption was reduced by about 18% compared to traditional methods, and the driving process was smoother and more stable. Third, in terms of real-time performance and scalability, this invention utilizes the fusion and iterative optimization of analytical solutions from linear quadratic games to achieve a fast approximate solution to complex game problems. The computational load of the algorithm increases exponentially with the number of participating vehicles, and actual calculations show that it can meet the real-time replanning requirements of 20Hz or even higher frequencies in typical traffic scenarios. Therefore, the method of this invention is applicable to dynamic and multi-vehicle interactive real-world road environments and can be embedded into existing vehicle control systems for real-time operation. Finally, by uniformly considering visual and policy uncertainties, this invention enables autonomous vehicles to behave both cautiously and decisively: never acting rashly when caution is needed, and never delaying when it is permissible to proceed. This balanced intelligent decision-making greatly enhances the robustness of autonomous driving systems in various complex and unknown environments, and is expected to accelerate the large-scale safe deployment of intelligent connected vehicles in real-world traffic. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of a vehicle-road cooperative scenario according to an embodiment of the present invention; Figure 2 This is a block diagram of real-time planning for intelligent connected vehicles according to an embodiment of the present invention; Figure 3 A schematic diagram of the autonomous vehicle strategy branch trajectory in an embodiment of the present invention; Figure 4 This is a flowchart of the iterative linear quadratic approximation algorithm according to an embodiment of the present invention; Figure 5The driving trajectory diagram of the vehicle and other vehicles in the rural overtaking scenario of this invention embodiment. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] like Figure 1 and Figure 2 As shown, this embodiment proposes a real-time planning method for intelligent connected vehicles based on occlusion perception and game theory, including: It is used to collect data on the environment around the vehicle and exchange its own status information with nearby vehicles and traffic infrastructure through vehicle-to-everything (V2X) communication. A dynamic game model is constructed based on the vehicle's surrounding environment data and the status of other vehicles obtained through vehicle-to-everything (V2X) communication. Based on the obstructions and communication latency in the vehicle's surrounding environment data, the obstruction open-loop game stage and the feedback game stage are divided. Based on the dynamic game model, corresponding game methods are used for the occlusion open-loop game stage and the feedback game stage to plan vehicle trajectories.

[0025] Specifically, in this embodiment, step 1, constructing the dynamic game model, includes: A dynamic game model is constructed, involving an autonomous vehicle (hereinafter referred to as "autonomous vehicle") and N other vehicles participating in traffic. Each vehicle updates its state in discrete time, with state variables including position, speed, and heading. The vehicle kinematics are described using a bicycle kinematics model, and the state update equation for each vehicle is: in, Let represent the state of the i-th vehicle at time t. Determine the control inputs for it, including key parameters such as acceleration and rotation angle. For vehicle kinematics functions, The time step is discrete. This model can realistically describe the dynamic behavior of vehicles on a two-dimensional plane. Each player makes decisions about the control input by optimizing their own cost function, forming an N-player non-zero-sum dynamic game problem. Within the game framework, each vehicle attempts to minimize its own cost or maximize its own gain. Costs include deviation from the target trajectory, changes in ride comfort (i.e., the magnitude of the control input), violations of traffic rules (such as lane departure), and the risk of collision. The vehicle itself, as player 0 in the game, must interact with other vehicle players 1…N to make decisions.

[0026] Specifically, in this embodiment, step 2 defines a hybrid information structure and divides it into an open-loop occlusion game stage and a feedback game stage. Based on static obstructions in the road environment, such as buildings, parked vehicles, and large trucks, as well as vehicle-to-everything (V2X) communication, the information mode switching during the game process is determined. Specifically, the planning time axis T, for example, from 0 to T seconds, is discretized into several steps, divided into alternating open-loop intervals and feedback intervals; the open-loop intervals and feedback intervals represent the obstruction open-loop game stage and the feedback game stage, respectively. When a player's vehicle's sensor view is obstructed, such as by a large vehicle blocking the view ahead or by not being able to see oncoming vehicles when turning, the vehicle enters open-loop mode. Within the open-loop zone, each player can only make decisions based on their own status and the initial environmental information, and cannot observe the real-time status changes of other vehicles, which is equivalent to "blind driving".

[0027] Once there are no obstructions, a clear view, or the status of other vehicles is obtained via V2X communication, the system enters feedback mode. Within this feedback period, each player has full knowledge of the current status of all vehicles, allowing them to adjust their strategies based on real-time conditions and engage in interactive strategic gameplay.

[0028] remember The set of all moments in which the feedback pattern exists. The set of time points in the open-loop mode is given, and the union of these two sets represents the entire planning time domain T. The hybrid information structure dynamically matches the line-of-sight situation: when visibility is limited due to static obstruction, each vehicle's strategy automatically switches to open-loop prediction; when the obstruction disappears, such as when a vehicle exits a curve or visibility is restored, the strategy switches to feedback-responsive. This processing allows the model to more closely approximate real-world traffic conditions with intermittent visibility. It should be noted that in an intelligent connected vehicle environment, vehicles can share state via V2X communication. This invention considers the impact of communication latency: even if a vehicle obtains information about an obstructed vehicle via wireless communication, there will still be a delay. Seconds, for corresponding The delay is a discrete time step. Therefore, when partitioning mixed information, cases where the state is not geometrically visible but can be learned through communication with a delay are also modeled as an approximate open-loop stage, and are only incorporated into the feedback stage after delay compensation.

[0029] Furthermore, based on the aforementioned dynamic game model, corresponding game theory methods are applied to the occlusion open-loop game stage and the feedback game stage to perform vehicle trajectory planning, including: Determine the driving intent of the other vehicle; wherein, the driving intent of the other vehicle includes: the intent to yield and the intent to force; Based on the driving intentions of the other vehicle, determine the timing of the strategy branch; Based on the branching time of the strategy, and combined with communication delay compensation, a vehicle trajectory including a main control sequence and a branch control sequence is planned. By integrating the strategy branch time, the planned vehicle trajectory, the occlusion open-loop game stage, and the feedback game stage into the dynamic game model, a hybrid information game model is obtained. An iterative linear quadratic approximation algorithm is used to solve the hybrid information game model and obtain the equilibrium strategy.

[0030] Specifically, in this embodiment, step 3 determines the driving intention of the other vehicle: Intent Hypothesis Library and Bayesian Update: For each other participating vehicle, especially manned vehicles, a discrete driving intent hypothesis library is pre-defined. Each Representative vehicle One possible driving style or strategy tendency is the tendency to yield, for example, in yielding decisions, there might be two distinct types: "yielding" and "forced yielding." In game theory models, different intentions are reflected in the vehicle's cost function parameters, such as a preference for speed targets or a greater emphasis on safe distances. The vehicle itself is unaware of the true types of other vehicles. However, it can be estimated online during the planning process. Therefore, this invention employs a Bayesian inference framework to maintain a set of intent confidences. This indicates information history Get off the vehicle Belongs to the The probability of each intention type. Initial settings are based on prior knowledge; in the absence of prior information, a uniform distribution can be used for initialization. During planning and execution, whenever the actions of other vehicles are observed—for example, through visual observation of changes in acceleration during the feedback phase, or by receiving status information via communication—the vehicle updates its posterior probability of intention using the Boltzmann rationality model: in, Assumption type Get off the vehicle The instantaneous cost function, The rationality coefficient, For vehicle j at time step The state vector, Let be the control input vector for the j-th vehicle at time step t. Let be the global state vector at time step t. This model assumes that the probability of a human driver performing a certain control action is exponentially related to the degree of superiority or inferiority of that action in terms of its cost evaluation. Based on the above observation-update model, the confidence probabilities of each intent hypothesis can be adjusted in real time through recursive Bayesian filtering. .

[0031] In this embodiment, information entropy is used to measure the uncertainty of intent: .like H A larger value indicates that the vehicle's intention is still unclear; conversely, a smaller value indicates a smaller value. H A low value indicates that its type can be basically determined.

[0032] Specifically, in this embodiment, step 4 determines the policy branch time: Since the uncertainty of intent gradually decreases with increasing observations, an entropy threshold is set. This indicates that the intent is "basically clear." This embodiment automatically determines the trigger time of the strategy branch by detecting the confidence entropy of the vehicle's intent towards all other vehicles. In other words, after entering feedback mode, find the first global moment. At this point, the maximum uncertainty of all other players' intentions has been reduced to below the threshold, assuming that the behavior pattern of each other's vehicles is basically determined, plus communication delay compensation. Define this moment as the branch point of the plan. Before the strategy branch point, i.e. At any given moment, a driver faces unknown intentions from other vehicles and must consider multiple possibilities simultaneously; and once... The vehicle itself can assume it knows the true strategy types of other vehicles. Based on this, the present invention plans a trajectory for the vehicle that includes a "main trunk + branches": This is the main control sequence before the branch. It is divided into several sets to correspond to different intentions and situations. Assuming a predetermined branch control sequence, the vehicle will... Based on the identified types of other vehicles, the system selects and executes the appropriate trajectory branch. This design ensures that the vehicle adopts a conservative but not passive strategy when there is uncertainty in the early stages, and immediately executes a strategy optimized for the actual situation once the conditions become clear, thus improving the pertinence and flexibility of decision-making.

[0033] Specifically, in this embodiment, step 5 employs an iterative linear quadratic approximation algorithm to solve the mixed information game model, including: Game Theory and Trajectory Optimization: Directly finding the optimal solution for the dynamic game with mixed information structure and intention branches is very difficult, such as... Figure 3 and Figure 4 As shown, this embodiment proposes an Iterative Linear Quadratic (ILQ) approximation algorithm to achieve real-time computation of approximate Nash equilibrium solutions. The basic idea is to decompose the nonlinear, non-convex game problem into a series of locally linear quadratic subproblems. By solving for exact equilibria on these subproblems and iteratively improving the solution, the algorithm ultimately approximates the equilibrium strategy of the original game. The algorithm flow is as follows: a) Initial Trajectory and Linearization: First, determine the initial control sequence, which can be obtained through planning. Then, based on this, conduct simulation of the system to obtain the initial state trajectory. Then, on the current trajectory The dynamics of each vehicle at each time step are linearized using Taylor expansion to obtain a linear approximate model: And by performing a quadratic expansion approximation on the cost function of each player, a quadratic cost approximation is obtained. ,in , It is the parameter that imposes the strong convexity of the quadratic cost. This forms a linear dynamic, quadratic cost LQ game subproblem. This LQ game inherits the mixed information structure of the original problem: that is, in the open-loop phase, each car does not consider feedback when optimizing and can only use a predetermined strategy; in the feedback phase, each car's strategy can be a linear feedback function of the current state.

[0034] b) Mixed Information Riccati Recursive Solution: For the obtained LQ subgame, an analytical method is used to find its feedback Nash solution. This invention designs a piecewise Riccati equation backtracking algorithm: starting from the endpoint... We begin the forward recursion, and in the feedback game phase, we apply the Riccati recursive formula of classic LQ game to calculate the optimal feedback gain. and the substitution function An improved open-loop optimization recursive solution is applied to the corresponding value matrix within the open-loop interval. The value functions are seamlessly connected at the interval boundaries. Specifically, the end of the open-loop segment serves as the initial value for the feedback segment, and the end of the feedback segment is passed to the next open-loop segment. This segmented, alternating calculation continues until time 0. Through this hybrid information Ricatti solution, the value matrix can be obtained in one step from... arrive During this period, the optimal linear feedback control law for each player in each feedback phase. And the optimal open-loop control sequence in the open-loop phase. It is worth mentioning that when... When the intention is undecided, the vehicle's decision-making is actually optimizing the expected cost (weighted according to the current belief distribution); while when Upon entering a branch, for each defined scenario of another vehicle's intention, the vehicle itself has a corresponding Riccati solution and feedback strategy. To accommodate multiple scenarios, this embodiment maintains multiple sets of parameters during Riccati recursion, corresponding to the subsequent costs under each assumption, ensuring that... Different strategies will naturally branch out at any given moment.

[0035] The piecewise Riccati equation backtracking algorithm specifically includes: Full time domain The system is divided into several continuous intervals. Each interval is either a visible feedback phase, where the policy depends on the current full state, or an occlusion open-loop phase, where the policy depends only on the state at the start of occlusion. From the endpoint... Starting with forward recursion, during the visible feedback phase, the optimal feedback strategy takes an affine form: The quadratic form of the corresponding value function is , where. Given the next step value function parameters Then, the optimal feedback gain and It can be calculated using the Riccati recursive formula for classical discrete-time general and LQ games: In the open-loop occlusion stage, an improved open-loop optimization is applied to recursively solve for the corresponding value matrix. Let the parameters of the open-loop valued function be... ,definition: Then, the value matrix obtained by backward recursion under the KKT conditions in the open-loop information structure is: To ensure seamless connection of value functions at interval boundaries, the connection rule is designed as follows: Let the intervals be... Its successor interval is In the visible feedback phase, the open-loop function parameters of the next segment are used. Alternative calculation During the occlusion open-loop phase, the open-loop value function parameters of the next segment are used. Alternative calculation This process of alternating calculations in segments continues until time 0.

[0036] In summary, this mixed-information Ricatti solution obtains the solution from... arrive During this period, the optimal linear feedback control law for each vehicle in each feedback game stage and the optimal open-loop control sequence in the occlusion open-loop game stage are determined.

[0037] c) Forward Rollover Correction and Iteration: Using the control law obtained in step b, perform a global forward simulation of the system, including applying the control to the original nonlinear model to obtain new state and belief trajectories. During the simulation, the evolution of intentional beliefs is tracked simultaneously to determine new branching moments. If the branching point differs from the previous branching point, the bifurcation point of the game solution needs to be adjusted accordingly. The new trajectory is then used for the next linearization and quadratic approximation, repeating steps a, b, and c, continuously correcting the strategy until the costs of each player converge or the number of iterations reaches a preset upper limit. Since each iteration is based on a simulation of actual nonlinear dynamics, the algorithm can gradually approximate the equilibrium solution of the real game. Empirically, a satisfactory convergent solution can be obtained in about ten iterations for complex traffic scenarios. In the examples of this invention, 20 iterations per second can be achieved on a typical laptop CPU, meeting the requirements of real-time applications.

[0038] Through the above solution process, the vehicle obtains a set of approximately optimal strategies for complex, dynamic, and uncertain environments: including a path from... arrive The conservative main trajectory, and Then, several branch trajectories are optimized based on the intentions of different vehicles. Other vehicles, if autonomous, also receive their own strategies; if human-operated, their "strategies" can be considered to be reflected in their actual behavior. This set of strategies constitutes an approximate solution to a state-feedback Nash equilibrium strategy in a mixed-information game. Finally, when the autonomous vehicle executes its actions, it determines its game stage based on real-time information obtained from vehicle sensors and communication, considering occlusion or visibility, and takes corresponding feedback or open-loop control measures. Based on the determination of the other vehicle's intentions, switch to the corresponding trajectory branch at any time and continue to execute until the end of the planning or the next replanning.

[0039] Specifically, in this embodiment, step 6 is autonomous vehicle execution and loop planning: the control sequence obtained above is sent to the vehicle's underlying execution layer, i.e., steering, driving, and braking to realize vehicle movement. Since the actual traffic environment is constantly changing, this embodiment works in a rolling time-domain replanning manner: that is, every fixed time interval, such as 0.1 seconds or 0.05 seconds, the current moment is used as a new starting point to repeat steps (1)-(5), obtain a new optimal strategy, and update subsequent control commands. This continuous closed-loop operation can ensure that the autonomous vehicle responds in a timely manner to new occlusions appearing or disappearing, changes in the intentions of other vehicles, etc., in a dynamic environment, and realizes real-time safety decision-making and path planning.

[0040] This embodiment also provides a real-time planning system for intelligent connected vehicles to implement the planning method. The system includes the following main units: (1) an on-board environment perception module, used to collect environmental data around the vehicle, such as cameras and lidar to detect roads and obstacles, and identify the location and range of potential obstructions, thereby determining which vehicles or areas are currently in the sensor blind zone; (2) a vehicle-to-everything (V2X) communication module, used to exchange its own position, speed and other status information with nearby vehicles and traffic infrastructure through V2X, to a certain extent making up for the information loss caused by visual obstruction; (3) a planning decision control unit, whose built-in computing platform runs the above-mentioned planning algorithm based on obstruction perception game. This unit constructs a hybrid information game model based on the environmental status provided by the perception module and the status of other vehicles obtained through communication, and solves online to obtain the optimal control sequence of the vehicle, including the predicted main trajectory and necessary emergency branches. Then the decision result is converted into driving control commands and sent to the vehicle's steering system, power and braking system to execute the corresponding operations. The entire system works together to enable the vehicle to obtain robust motion planning capabilities in complex environments.

[0041] Through the above solutions, this invention achieves unified handling of occlusion and intent uncertainty, enabling autonomous vehicles to make rational decisions in highly dynamic and uncertain traffic environments. In summary, the game theory model constructed by this invention comprehensively considers: (i) temporary obstruction and recovery of vision—reflecting changes in sensor visibility through a hybrid information structure; (ii) gradual clarification of other vehicles' intentions—automatically bifurcating the vehicle's strategy as more information is acquired through Bayesian updates and a pre-defined branching mechanism; and (iii) rational decision-making by multiple agents—each vehicle is modeled as a rational agent, thus eliminating the need for overly conservative worst-case assumptions, but rather assuming that the other vehicle will react optimally based on its own interests and the actions of this embodiment. This method overcomes the limitations of previous approaches that only considered a single uncertainty, ensuring a balance between safety and efficiency. Especially in general multi-vehicle scenarios with N>2, the method of this invention remains applicable. To ensure real-time computation and scalability, this method employs iterative LQ solving, whose algorithm complexity increases cubically with the joint state dimension. In a typical 6-vehicle traffic scenario, a single planning solution can be completed within 50ms. Therefore, this invention can operate online on actual vehicle hardware, providing a practical and feasible technical approach for autonomous decision-making in intelligent vehicles under complex environments.

[0042] The following examples further illustrate the real-time planning method and system for intelligent connected vehicles based on occlusion perception and game theory according to the present invention. To verify the effectiveness of the method in real-world traffic scenarios, this embodiment uses a MATLAB-based traffic simulation platform and vehicle network for joint simulation experiments. The following example, a typical rural road following and overtaking scenario, illustrates the implementation process and effects of the method.

[0043] (I) Scenario Setting and Initial Conditions: This embodiment simulates a two-way single-lane rural road. The autonomous vehicle (i.e., a connected and automated vehicle) is traveling at a relatively high speed along this lane when it encounters a slow-moving truck (i.e., a static obstacle vehicle). The initial distance between the two vehicles is approximately 30 meters. The truck is large and almost completely obstructs the autonomous vehicle's forward view, preventing it from directly detecting any oncoming vehicles in the opposite lane using its onboard camera or radar. Simultaneously, there is a regular vehicle (marked as an orange car in the diagram) in the opposite lane, approaching the autonomous vehicle. However, initially, this oncoming vehicle is obscured by the truck and is invisible to the autonomous vehicle. It is assumed that both the autonomous vehicle and the oncoming vehicle are connected vehicles and can transmit status information via V2X communication. However, there is a delay of approximately 0.2 seconds, which corresponds to the next iteration step size of the distance step in this experiment. In this scenario, the autonomous vehicle's task is to overtake the slow-moving truck as quickly as possible, while ensuring that it does not collide with the oncoming vehicle. The autonomous vehicle's driving objective is to complete the overtaking maneuver within a limited road segment and return to its cruising speed before the oncoming vehicle appeared. Initial speed of the vehicle m / s, truck speed m / s, speed of oncoming vehicle m / s. All vehicles use the kinematic parameters of passenger cars. The time interval between the planning controller and the walk length is... The planned vision time is approximately 5 seconds, which means T = 25 steps.

[0044] (II) Construction of the Obstruction Game Model and Intent Setting: Based on the above scenario, the vehicle in From that moment on, the system enters open-loop information mode because the truck ahead severely obstructs the view, preventing the vehicle from directly observing the oncoming vehicle's status. Until the vehicle gradually moves out of the truck's rear and can see the oncoming vehicle again, the two vehicles are in an open-loop game phase. During this period, the vehicle can only make decisions based on the initial, rough information about the oncoming vehicle, such as its presence and approximate speed obtained through communication. The oncoming vehicle's human driver has not yet engaged in visual interaction with the vehicle. Approximately two seconds after the planning begins, when the vehicle attempts to change lanes to the left and extends a certain distance, a brief window appears in the forward field of vision, showing the position and speed of the oncoming vehicle. This indicates that the game has entered feedback information mode. At this point, because the vehicle has network connectivity, it can confirm the oncoming vehicle's driving status through both communication and visual perception. Once the vehicle and the oncoming vehicle become mutually visible, they begin real-time game interaction: the oncoming driver observes the vehicle's overtaking intentions and decides whether to slow down to let it pass or maintain speed; the vehicle also continuously adjusts its overtaking strategy based on the oncoming vehicle's immediate reactions. Next, consider the uncertainty of the driving intentions of oncoming vehicles: For simplicity, this embodiment assumes that the human driver intentions of oncoming vehicles are of two types: The "courteous" type refers to conservative drivers who slow down early to give way when someone tries to overtake them; the "aggressive" type refers to aggressive drivers who maintain their speed even when another vehicle is blocking their lane, sometimes even honking their horn, and do not slow down to give way. (The last sentence appears to be incomplete and possibly refers to a separate issue.) Since there is no prior preference for the type of oncoming vehicle, the initial belief can be set to equal. In the open-loop phase, because the oncoming vehicle is not visible, the vehicle cannot update this belief and can only assume that the other vehicle could be of any type and prepare for the worst-case scenario. Once the feedback phase begins and the behavior of oncoming vehicles is observed, the vehicle uses Bayesian inference to gradually update its probabilistic judgment of the oncoming vehicle's type. For example, if the oncoming vehicle begins to decelerate from a distance, the probability of the vehicle's "yielding" assumption will increase; conversely, if it does not decelerate or even accelerates, it tends to be of the "aggressive" type. As the distance between the vehicle and the oncoming vehicle decreases, this intention judgment will quickly become clear.

[0045] (III) Execution process of the planning algorithm: The vehicle is in The planning algorithm of this invention is invoked. The initial stage is an open-loop game of occlusion. The vehicle considers two possible intentions of oncoming vehicles, neither of which can be verified. Therefore, a conservative main trajectory is planned: this trajectory follows the truck for a period of time while gradually approaching the center line of the road and accelerating to the speed required for overtaking, aiming to complete the overtaking preparation as quickly as possible while maintaining a safe distance. During this process, the vehicle constantly maintains an emergency braking backup strategy: if an oncoming vehicle suddenly exits the truck's obstruction during the open-loop stage (i.e., in the worst-case scenario of a very close and high-speed approach), the vehicle's main strategy can still immediately apply emergency braking to avoid a collision. After approximately 2 seconds, the simulation reaches... As the vehicle approaches the rear of the truck and slightly changes lanes to the left, revealing its forward view, it sees an oncoming vehicle approaching from the front. At this moment, both the vehicle and the oncoming vehicle are visible to each other, entering a feedback game phase, which is also the key decision point in this overtaking scenario. According to the method of the present invention, the vehicle checks the belief entropy of the oncoming vehicle's intention at this time: since it has already obtained some information by observing the oncoming vehicle's deceleration, such as noticing that the oncoming vehicle has begun to decelerate slightly and the probability of yielding has increased, the uncertainty of the intention is reduced. If the entropy value is already below a preset threshold at this time... If so, it is determined that the intention is basically clear, and Set to the current time; otherwise, the autonomous vehicle will continue to observe for one or two more time steps. In this experiment, the oncoming vehicle did not significantly decelerate after seeing the autonomous vehicle approach, but maintained a high speed to continue approaching. The autonomous vehicle quickly judged it as a "forced approach," meaning the probability of its intention exceeded the threshold, and... At point 's', a strategy branch is triggered. Therefore, in response to the confirmed situation of "forced advance," the autonomous vehicle adopts the pre-planned "abandon overtaking" branch: specifically, it immediately stops changing lanes to the left, abandons the overtaking attempt, and decelerates back to the lane behind the truck; simultaneously, it communicates with the following vehicle via V2X or uses light signals to indicate that it has abandoned overtaking, thus coordinating to safely pass the oncoming vehicle. Conversely, if the oncoming vehicle is detected to be "yielding," meaning its speed has significantly decreased and it is moving to the right, the autonomous vehicle will... The alternative acceleration and overtaking branch was chosen: rapidly increasing speed to overtake the truck and merging into the lane ahead of the oncoming vehicle to complete the overtaking maneuver. In this simulation, because the oncoming vehicle was judged as an aggressive type that would not yield, the vehicle chose the safe branch. Through game-theoretic planning, the vehicle ensured that the distance to the oncoming vehicle was not less than the safe threshold while abandoning the overtaking maneuver, greatly reducing the risk of collision. Subsequently, the vehicle continued to follow the truck, waiting for the next safe overtaking opportunity or a change in the scenario. Throughout the process, the vehicle's planning and decision-making frequency was 20Hz, and the strategy was updated every 0.05 seconds. When the oncoming vehicle passed, the road ahead was clear again, and the visibility was unobstructed, the vehicle planned a new overtaking trajectory and successfully completed the overtaking maneuver.

[0046] (iv) Experimental Results and Effect Analysis: In the specific scenarios described above, the method of this invention achieves safe and reasonable decision-making, such as... Figure 5As shown. Even if the behavior of oncoming vehicles is not clear at the beginning, the autonomous vehicle effectively avoids the danger of head-on collisions caused by rashly overtaking at the beginning through the strategy of "probing first, then branching," and does not stagnate indefinitely. When an oncoming vehicle is identified as not yielding, the autonomous vehicle promptly converges its strategy to avoid it, allowing the oncoming vehicle to pass first; conversely, it also decisively completes the overtaking task when there is an opportunity. To verify the reliability of this method, this embodiment completed 100 Monte Carlo random tests for this typical interaction scenario. During the test, key variables such as the initial speed of the oncoming vehicle and the type of driver behavior were all subject to random perturbations. The results show that: using the planning of this invention, the autonomous vehicle passed the scenario with zero collisions in all tests; in contrast, if conventional modular planning that does not consider game interaction is used, about 12% of the tests result in collisions. This is mostly due to forced overtaking after misjudging the behavior of oncoming vehicles. Although the conservative planning that only considers occlusion has no collisions, nearly half of the tests failed to overtake successfully in time, which is inefficient. Furthermore, this embodiment statistically analyzed the average time for the vehicle to complete the overtaking task: the method achieved approximately 9.8 seconds, significantly better than the traditional conservative strategy's approximately 11 seconds and close to the ideal overtaking time under unobstructed conditions. Simultaneously, the vehicle's control output during the overtaking process was stable, and ride comfort indicators, such as acceleration changes and energy consumption, were improved by approximately 15%–20% compared to manually set safety strategies. Therefore, the method of this invention achieves higher efficiency and comfort while ensuring safety. It is worth mentioning that this embodiment also tested the invention in other typical scenarios, such as static obstruction at urban T-junctions and lane-merging scenarios on multi-lane highways, achieving significantly better results than the comparative methods: the invention comprehensively outperformed the methods in terms of collision avoidance rate, travel time, and fuel consumption. These examples verify the applicability and superiority of this invention in various complex traffic scenarios.

[0047] In summary, simulation demonstrations of specific embodiments demonstrate that the occlusion perception and game theory-based planning method proposed in this invention can effectively address the dual uncertainties of dynamic occlusion and human-vehicle interaction, ensuring that autonomous vehicles complete complex scenario tasks safely and efficiently. Those skilled in the art, after reading this specification, can adjust and modify the specific implementation as needed, such as introducing more types of intent assumptions or adjusting model parameters based on the vehicle's inertial characteristics; these changes are all within the scope of the methodological concept of this invention. The above are merely preferred embodiments of this invention and are not intended to limit the scope of protection of this invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

[0048] This invention aims to protect a real-time planning method and implementation system for intelligent connected vehicles based on occlusion perception and game theory. The method addresses the challenges of autonomous vehicles operating in environments with obstructed lines of sight and uncertainties in multi-vehicle interactions, enabling safe and efficient trajectory planning and decision-making control. Its core components include: First, addressing the issue of traditional planning methods lacking consideration for unknown hazards or the behavior of other vehicles, this invention introduces a hybrid information dynamic game model: during the planning process, it dynamically switches between open-loop and feedback information modes based on environmental geometric occlusion. This allows autonomous vehicles to adopt conservative predictive strategies when sensors temporarily cannot see other traffic participants, and to respond promptly to the actual state of other vehicles after they are seen again or after receiving information through communication. Compared to existing methods that only consider worst-case occlusion and are often overly conservative, this invention, through a game theory framework, ensures that the vehicle can proactively adjust its strategy when visibility is restored, without being bound by previous conservative assumptions, thus avoiding the drawback of continuous over-conservatism.

[0049] Secondly, to address the decision-making difficulties arising from the unknown intentions of other drivers, this invention integrates intention inference based on Bayesian updates and a strategy branching mechanism. Unlike existing decoupled prediction-planning methods, the autonomous vehicle simultaneously maintains probability estimates of multiple possible behavior patterns for each other vehicle during the planning process and embeds these estimates into the game decision. When accumulated information makes the intention of a certain other vehicle clearer, the autonomous vehicle's single trajectory automatically branches into targeted sub-trajectories, each corresponding to a different intention scenario. In this way, regardless of whether other drivers ultimately choose a cooperative or competitive strategy, the autonomous vehicle already has a pre-planned optimal or suboptimal response. Compared to traditional planning, which can only compromise or assume a fixed strategy when intentions are unclear, the method of this invention greatly enhances the robustness and flexibility of planning through a "one primary, multiple backup" contingency plan, avoiding the risks caused by blindly rushing in or making misjudgments.

[0050] Furthermore, the planning algorithm proposed in this invention possesses real-time solution capabilities and ease of integration. It employs an iterative linear quadratic approximation to solve complex games, achieving a balance between computational speed and solution quality. The algorithm primarily involves numerical calculations such as linearized dynamics modeling and Riccati equation recursion, with complexity on par with existing methods like iterative LQR, thus avoiding excessive burden on the onboard computing platform. This means that the method can be implemented on existing vehicle controllers as a software upgrade, embedded as an improvement to existing planning modules, without requiring special hardware acceleration. In practical applications, when environmental occlusion or abnormal behavior of other vehicles is detected, this method can seamlessly take over the planning task, generating safe, feasible, and efficient driving trajectories in real time. Even in general scenarios, this method works similarly, thereby improving the overall system's fault tolerance to unexpected situations. The method of this invention can run online in real time and also supports offline simulation, providing a powerful tool for autonomous driving decision-making in complex scenarios.

[0051] Furthermore, this invention also covers an intelligent connected vehicle planning system device, which includes modules for environmental perception, communication, and decision control, working together to achieve the functions of the aforementioned method. In this system, the onboard perception and communication units collect and share environmental information, and the decision control unit executes a game-theoretic planning algorithm and outputs vehicle control commands, enabling the vehicle to adjust its driving strategy based on real-time environmental changes and multi-agent interactions. Because the method itself is computationally rapid and reliable, the system can operate stably in scenarios such as autonomous vehicle overtaking, intersection passage, and highway convoy driving. Even when some vehicles are not connected to communication, or when pure visual perception is used, or when there is communication delay or packet loss, the method of this invention, through hybrid information game modeling, can still take these uncertainties into account, ensuring safe and effective planning, and preventing decision failure due to missing information.

[0052] The main problem addressed by this invention is that existing autonomous driving planning methods cannot simultaneously handle two key challenges: limited field of view and interaction uncertainty. Traditional occlusion handling solutions often assume worst-case scenarios, resulting in inefficiency, while interaction intent inference solutions frequently ignore the presence of hidden vehicles. Using these two approaches separately makes them difficult to handle real-world complex scenarios. By combining occlusion perception and game-theoretic intent inference, this invention provides a unified solution that allows autonomous vehicles to operate in complex environments without being overly conservative or reckless. Experiments and analysis show that this method significantly improves vehicle hazard avoidance capabilities and traffic efficiency, and enhances the system's robustness in unknown and complex environments, which is of great significance for achieving high-level autonomous driving. In summary, the planning method and system described in this invention maximize the performance of intelligent connected vehicles while ensuring safety, and have broad application prospects and promotional value.

[0053] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A real-time planning method for intelligent connected vehicles based on occlusion perception and game theory, characterized in that, include: It is used to collect data on the environment around the vehicle and exchange its own status information with nearby vehicles and traffic infrastructure through vehicle-to-everything (V2X) communication. A dynamic game model is constructed based on the vehicle's surrounding environment data and the status of other vehicles obtained through vehicle-to-everything (V2X) communication. Based on the obstructions and communication latency in the vehicle's surrounding environment data, the obstruction open-loop game stage and the feedback game stage are divided. Based on the dynamic game model, corresponding game methods are used for the occlusion open-loop game stage and the feedback game stage to plan vehicle trajectories.

2. The real-time planning method for intelligent connected vehicles based on occlusion perception and game theory as described in claim 1, characterized in that, Constructing the dynamic game model includes: Each vehicle updates its state in discrete time, where the state variables include position, velocity, and heading; the state update equation for each vehicle is: in, Let represent the state of the i-th vehicle at time t. To control its input, For vehicle kinematics functions, The discrete time step; The dynamic game model is used to characterize the dynamic behavior of vehicles on a two-dimensional plane. Each vehicle makes decisions on control inputs by optimizing its own cost function, forming an N-player non-zero-sum dynamic game problem. Under the game framework, each vehicle tries to minimize its own cost, which includes deviation from the target trajectory, ride comfort, violation of traffic rules, and the risk of collision. The vehicle itself, as player 0 in the game, needs to interact and make decisions with other vehicle players 1...N.

3. The real-time planning method for intelligent connected vehicles based on occlusion perception and game theory as described in claim 1, characterized in that, The division of the occlusion open-loop game stage and the feedback game stage includes: When the sensor line of sight of the vehicle is blocked, it enters the occlusion open-loop game stage; in the occlusion open-loop game stage, it can only make decisions based on its own state and the environmental information at the beginning, and cannot observe the real-time state changes of other vehicles. When there are no obstructions and the field of vision is unobstructed, or when the status of other vehicles is obtained through vehicle-to-everything (V2X) communication, the feedback game phase begins. During the feedback game phase, each vehicle is fully aware of the current status of all other vehicles and adjusts its strategy according to the real-time status to achieve interactive game. When the field of vision is limited due to static objects blocking it, the strategies of each vehicle automatically switch to the open-loop game stage of occlusion; when the occlusion disappears, the strategies switch to the feedback game stage.

4. The real-time planning method for intelligent connected vehicles based on occlusion perception and game theory as described in claim 1, characterized in that, Based on the dynamic game model, corresponding game methods are applied to the occlusion open-loop game stage and the feedback game stage to perform vehicle trajectory planning, including: Determine the driving intent of the other vehicle; wherein, the driving intent of the other vehicle includes: the intent to yield and the intent to force; Based on the driving intentions of the other vehicle, determine the timing of the strategy branch. Based on the branching time of the strategy, and combined with communication delay compensation, a vehicle trajectory including a main control sequence and a branch control sequence is planned. By integrating the strategy branch time, the planned vehicle trajectory, the occlusion open-loop game stage, and the feedback game stage into the dynamic game model, a hybrid information game model is obtained. An iterative linear quadratic approximation algorithm is used to solve the hybrid information game model and obtain the equilibrium strategy.

5. The real-time planning method for intelligent connected vehicles based on occlusion perception and game theory as described in claim 4, characterized in that, Determining the driving intent of the other vehicle includes: For each other participating vehicle, a discrete driving intention hypothesis library is pre-defined; Using Bayesian inference to determine the confidence level of a set of intentions This indicates information history Get off the vehicle Belongs to the The probability of each intention type Here is the confidence probability. In the history of information Under the given conditions, the event The probability of occurrence Let be the intent parameters of the j-th vehicle. Let k be the intention of the j-th vehicle. The initial time is set based on prior knowledge. During the planning and execution process, whenever the actions of other vehicles are observed, the vehicle updates the posterior probability of its intention using the observation-update model: in, Assumption type Get off the vehicle The instantaneous cost function, The rationality coefficient, For vehicle j at time step The state vector, Let be the control input vector for the j-th vehicle at time step t. This is the global state vector at time step t; Based on the observation-update model described above, the confidence probability of each intent hypothesis is adjusted in real time through recursive Bayesian filtering. ; Using information entropy to measure intention uncertainty: ;like If the value is higher than a preset threshold, it indicates that the vehicle's intention is still unclear; conversely, if the value is lower than a preset threshold, it indicates that the vehicle's intention is still unclear. If the value is below a preset threshold, its type is determined.

6. The real-time planning method for intelligent connected vehicles based on occlusion perception and game theory as described in claim 5, characterized in that, Determining the timing of policy branching includes: The triggering time of the strategy branch is automatically determined by detecting the confidence entropy of the vehicle's intentions toward all other vehicles; in, For the planned branch point, For time steps, Let j be the set of time steps in which all traffic participants can fully see each other, and let j be the j-th vehicle. This represents the equivalent time steps corresponding to the network communication delay.

7. The real-time planning method for intelligent connected vehicles based on occlusion perception and game theory as described in claim 4, characterized in that, The solution to the hybrid information game model is obtained by employing an iterative linear quadratic approximation algorithm, which includes: a. Select an initial control sequence, simulate the system on the initial control sequence to obtain the initial state trajectory, perform Taylor expansion linearization on the dynamics of each vehicle at each time step at the current trajectory to obtain a linear approximation model, and perform a quadratic expansion approximation on the cost function of each player to obtain a quadratic cost approximation, thus forming a linear quadratic game subproblem with linear dynamics and quadratic cost. b. Based on the pre-defined piecewise Riccati equation backtracking algorithm, the feedback Nash solution of the linear quadratic game subproblem is obtained, and the optimal linear feedback control law of each vehicle in each feedback stage is obtained. c. Using the optimal linear feedback control law, perform global forward simulation to obtain new state and belief trajectories. During the simulation, simultaneously track the evolution of intentional beliefs, determine new branch moments, and use the new trajectory for the next linearization and quadratic approximation. d. Repeat steps a, b, and c, continuously adjusting the strategy until the cost of each vehicle converges or the number of iterations reaches the preset upper limit, thus obtaining an approximate optimal strategy set for complex dynamic and uncertain environments.

8. The real-time planning method for intelligent connected vehicles based on occlusion perception and game theory as described in claim 7, characterized in that, The piecewise Riccati equation backtracking algorithm includes: Full time domain The system is divided into several continuous intervals, each interval representing a visible feedback phase (where the policy depends on the current full state) or an occlusion open-loop phase (where the policy depends only on the state at the start of occlusion); from the endpoint... Starting with forward recursion, during the visible feedback phase, the optimal feedback strategy takes an affine form: ,in, , For the optimal feedback gain matrix The quadratic form of the corresponding value function is ,in, for Quadratic form of value function The coefficient matrix of the quadratic terms is The Hessian matrix, Let be the vector of coefficients of the first-order terms. gradient vector, ; given the next step value function parameters Then, the optimal feedback gain and It is calculated using the Riccati recursive formula for classical discrete-time general and LQ games; In the open-loop occlusion stage, an improved open-loop optimization is applied to recursively solve for the corresponding value matrix. Let the parameters of the open-loop valued function be... ,definition: in, This is the open-loop coupling correction matrix. It is the identity matrix. To control the input matrix, For the first j Penalty weight matrix of vehicle to its own control input Then, the value matrix obtained by backward recursion under the KKT conditions in the open-loop information structure is: in, Let be the state transition matrix of the system. for The state term weight matrix in the cost function; To ensure seamless connection of value functions at the boundaries of intervals, the connection rule is designed as follows: Let the time interval be... The subsequent time interval is In the visible feedback phase, the open-loop function parameters of the next segment are used. Alternative calculation During the occlusion open-loop phase, the open-loop value function parameters of the next segment are used. Alternative calculation ; This process of alternating calculations continues until time 0; Through this mixed-information Ricatti solution, the information from... is obtained in one step. arrive During this period, the optimal linear feedback control law for each vehicle in each feedback game stage and the optimal open-loop control sequence in the occlusion open-loop game stage are determined.

9. The real-time planning method for intelligent connected vehicles based on occlusion perception and game theory as described in claim 7, characterized in that, The vehicle trajectory planning also includes: The obtained near-optimal strategy set is sent to the vehicle's underlying execution to realize vehicle movement. At fixed time intervals, the current moment is used as a new starting point to repeatedly build the dynamic game model and re-plan the vehicle trajectory.

Images