Vehicle dynamic obstacle avoidance real-time decision-making method, device and medium

Abstract

The invention relates to a vehicle dynamic obstacle avoidance real-time decision-making method and device and a medium, and belongs to the technical field of path planning, and the vehicle dynamic obstacle avoidance real-time decision-making method comprises the steps: building a physical information neural network based on a three-degree-of-freedom model of a vehicle, and obtaining a vehicle state and a control signal at a current moment; inputting the vehicle state and the control signal into a physical information neural network to obtain the vehicle state at the next moment; based on the road reference path, the real-time position of the dynamic obstacle and the vehicle state at the current moment, a grid planner is adopted to carry out local planning on the vehicle track, and a reference track is obtained; based on the reference trajectory, the vehicle state at the current moment and the vehicle dynamics relation described by the physical information neural network, a target function of model prediction control is built, the target function is solved, and a control signal at the next moment is obtained; the vehicle is controlled based on the control signal at the next moment and the vehicle state, dynamic obstacle avoidance of the vehicle is achieved, and the stability of the vehicle is improved.

Description

Technical Field

[0001] This invention relates to the field of path planning technology, and in particular to a method, device and medium for real-time decision-making on vehicle dynamic obstacle avoidance. Background Technology

[0002] In off-road scenarios, such as unstructured terrains like mountains, jungles, and deserts, autonomous vehicles face complex and varied environmental challenges, including uneven roads, loose soil, sudden obstacles (such as fallen trees and rolling stones), and dynamically changing obstacles (such as moving wild animals and other off-road vehicles).

[0003] Traditional model predictive control (MPC) methods based on accurate vehicle dynamics models suffer from significantly increased prediction errors in complex off-road environments due to simplified model assumptions (such as ignoring the impact of terrain on tire adhesion and neglecting the dynamic characteristics of the vehicle in roll states). This leads to a decline in the control performance and driving safety of autonomous vehicles. Furthermore, in off-road environments, dynamic obstacles frequently appear and their behavior is difficult to predict, causing MPC to fail to perfectly track the planned trajectory path, resulting in a severe disconnect between planning and control, which in turn affects the stability of autonomous vehicles. Summary of the Invention

[0004] In view of this, it is necessary to provide a method, device and medium for real-time decision-making on vehicle dynamic obstacle avoidance, in order to solve the technical problems of low state prediction accuracy and insufficient stability of autonomous vehicles in complex off-road environments.

[0005] To address the aforementioned problems, in a first aspect, the present invention provides a real-time decision-making method for vehicle dynamic obstacle avoidance, comprising: A physical information neural network is constructed based on a three-degree-of-freedom model of the vehicle to obtain the vehicle state and control signals at the current moment. The vehicle state and control signals are then input into the physical information neural network to obtain the vehicle state at the next moment. Based on the road reference path, the real-time position of dynamic obstacles, and the current vehicle state, a grid planner is used to locally plan the vehicle trajectory to obtain a reference trajectory. Based on the reference trajectory, the current vehicle state, and the vehicle dynamics relationship described by the physical information neural network, a model predictive control objective function is constructed. The objective function is solved to obtain the control signal for the next moment. The vehicle is controlled based on the control signal and vehicle state at the next moment, and the vehicle state and control signal at the next moment are used as the new vehicle state and control signal at the current moment for iterative updates to achieve dynamic obstacle avoidance of the vehicle.

[0006] In one possible implementation, the vehicle state includes position, heading angle, and speed; the control signals include steering angle signals and throttle / brake signals; the three-degree-of-freedom model is: , , , , , in, For time Changing vehicle position, For time Changing vehicle heading angle For time Changing vehicle speed, For dynamic model, This is the vehicle's steering angle signal. This is the accelerator / brake signal. For model parameters, , , , Initial value; The physical information neural network is: , in, This is a state-space physical information neural network. For the hyperparameters of a physical information neural network, This represents the current vehicle status. The control signal at the current moment. This is a prediction of the vehicle's state at the next moment.

[0007] In one possible implementation, before inputting the vehicle state and control signals into the physical information neural network to obtain the vehicle state at the next moment, the process includes: Random observation data and discrete physical data generated based on a three-degree-of-freedom real physical model are acquired, and the physical information neural network is trained based on the random observation data and the discrete data.

[0008] In one possible implementation, the three-degree-of-freedom real physical model is: , , , , in, This is the steering angle signal at the current moment. This represents the accelerator / brake signal at the current moment. Let be the time step of the discrete space. This represents the vehicle's current position.

[0009] In one possible implementation, the loss function of the physical information neural network is: , , in, This is the weighted loss function for the physical information neural network. For the loss of random observation data, For the loss of discrete physical data, for A continuous observation interval, for A discrete physical data point, The prediction results of the physical information neural network. The residuals of physical laws in a physical information neural network. These are physical parameters.

[0010] In one possible implementation, the process of using a grid planner to locally plan the vehicle trajectory based on the road reference path, the real-time position of dynamic obstacles, and the current vehicle state to obtain the reference trajectory includes: Based on the road reference path, the real-time position of dynamic obstacles, and the current vehicle status, a grid planner is used to plan the vehicle trajectory and obtain multiple sets of horizontal and vertical candidate trajectories. The multiple sets of horizontal and vertical candidate trajectories are transformed into Cartesian coordinates to obtain multiple sampled trajectories; Collision detection is performed on the multiple sampled trajectories, and the total cost of the multiple sampled trajectories after collision detection is calculated to obtain the reference trajectory with the minimum total cost.

[0011] In one possible implementation, the lateral candidate trajectory is: , , in, The target time for trajectory prediction. This represents the vehicle's current lateral position. The vehicle's current lateral speed, The vehicle's current lateral acceleration, The lateral target position is predicted by the trajectory. , , , , , These are the polynomial coefficients; The longitudinal candidate trajectory is: , , in, For the current longitudinal position of the vehicle, The vehicle's current longitudinal speed, The vehicle's current longitudinal acceleration, , , , , Let be the polynomial coefficients; The sampling trajectory is as follows: , in, Let be the coordinates in the vertical position. The tangent direction angle; The total cost is: , in, For the total cost, As a cost-weighted coefficient, As a cost of smoothness, For offset cost, For the price of speed, For the cost of time, As a result of the collision, This is the cost of dynamic constraints.

[0012] In one possible implementation, the objective function is: , in, The process state of the vehicle. The reference state provided for the reference trajectory, The optimal control signal. For Mayer's term, For Lagrange terms.

[0013] In a second aspect, the present invention also provides a vehicle control device, comprising: a processor and a memory; The memory stores a computer-readable program that can be executed by the processor; When the processor executes the computer-readable program, it implements the steps in the real-time decision-making method for vehicle dynamic obstacle avoidance as described above.

[0014] Thirdly, the present invention also provides a computer-readable storage medium for storing a computer-readable program or instruction, wherein the program or instruction, when executed by a processor, is capable of implementing the steps in the vehicle dynamic obstacle avoidance real-time decision-making method described in any of the above-mentioned method items.

[0015] The beneficial effects of this invention are as follows: The current vehicle state and control signal are acquired, and the vehicle state and control signal are input into a physical information neural network to obtain the vehicle state at the next moment. The prediction accuracy of the vehicle state is improved through the prediction of the vehicle state by the physical information neural network. Based on the road reference path, the real-time position of dynamic obstacles, and the current vehicle state, a grid planner is used to locally plan the vehicle trajectory to obtain a reference trajectory. Based on the reference trajectory, the current vehicle state, and the vehicle dynamics relationship described by the physical information neural network, a model predictive control objective function is constructed. The objective function is solved to obtain the control signal at the next moment. The vehicle is controlled based on the control signal at the next moment and the vehicle state. The grid planner plans the vehicle trajectory, generating an optimized trajectory that takes into account both global path tracking and dynamic obstacle avoidance. This simplifies the nonlinear constraints of the MPC, accelerates the convergence speed of the objective function, achieves efficient optimal control solutions, and improves the stability of autonomous vehicles. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0017] Figure 1 A flowchart of an embodiment of the vehicle dynamic obstacle avoidance real-time decision-making method provided by the present invention; Figure 2 A schematic diagram illustrating the prediction effect of the physical information neural network in the real-time decision-making method for vehicle dynamic obstacle avoidance provided by the present invention. Figure 3 A Frenet coordinate system diagram of the real-time decision-making method for vehicle dynamic obstacle avoidance provided by the present invention; Figure 4a A schematic diagram illustrating the vehicle control effect of the real-time decision-making method for dynamic obstacle avoidance provided by the present invention. Figure 4b A schematic diagram illustrating the vehicle control effect of the real-time decision-making method for dynamic obstacle avoidance provided by the present invention. Figure 4c A schematic diagram illustrating the vehicle control effect of the real-time decision-making method for dynamic obstacle avoidance provided by the present invention. Figure 5 This is a schematic diagram of an embodiment of the vehicle control device provided by the present invention. Detailed Implementation

[0018] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0019] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0020] This invention discloses a method, device, and medium for real-time decision-making regarding vehicle dynamic obstacle avoidance, which can be used in a computer. The method, device, or computer-readable storage medium involved in this invention can be integrated with the aforementioned device or exist relatively independently.

[0021] One specific embodiment of the present invention discloses a real-time decision-making method for vehicle dynamic obstacle avoidance, which can be executed by a computer, specifically by one or more processors of the computer. For example... Figure 1 As shown, the real-time decision-making method for vehicle dynamic obstacle avoidance includes: S101. Construct a physical information neural network based on the three-degree-of-freedom model of the vehicle to obtain the vehicle state and control signal at the current moment, and input the vehicle state and control signal into the physical information neural network to obtain the vehicle state at the next moment. It should be noted that the Physical Information Neural Network (PINN) uses vehicle dynamics equations as training constraints, which allows it to fully utilize existing incomplete physical knowledge and directly learn the state transition laws of nonlinear systems, thus providing a more accurate prediction model for MPC.

[0022] S102. Based on the road reference path, the real-time position of dynamic obstacles and the current vehicle state, a grid planner is used to perform local planning of the vehicle trajectory to obtain the reference trajectory. Based on the reference trajectory, the current vehicle state and the vehicle dynamics relationship described by the physical information neural network, a model predictive control objective function is constructed. The objective function is solved to obtain the control signal for the next moment. It should be noted that by generating a reference trajectory that takes into account both global path tracking and dynamic obstacle avoidance through the grid planner, the nonlinear constraints of MPC are simplified and the convergence speed of the objective function is accelerated.

[0023] S103. Control the vehicle based on the control signal and vehicle state at the next moment, and use the vehicle state and control signal at the next moment as the new vehicle state and control signal at the current moment for iterative updates to achieve dynamic obstacle avoidance of the vehicle.

[0024] In some embodiments, in step S101, a physical information neural network is constructed based on the vehicle's three-degree-of-freedom model. Regarding the vehicle model, a three-degree-of-freedom (3-DOF) model is used to represent its dynamic characteristics, which include position, velocity, acceleration, steering angle, and yaw angle. Its three-degree-of-freedom model is as follows: , , , , , in, For time Changing vehicle position, For time Changing vehicle heading angle For time Changing vehicle speed, For dynamic model, This is the vehicle's steering angle signal, and it is the first control signal. This is the accelerator / brake signal, the second control signal. These are the model parameters, used to accurately map the input to the output. , , , Initial value; The physical information neural network can be represented in state-space form as follows: , in, This is a state-space physical information neural network. For the hyperparameters of a physical information neural network, This represents the current vehicle status. The control signal at the current moment. The predicted vehicle state for the next moment; The Physical Information Neural Network (PINN) incorporates dynamic constraints into the neural network training process. It can utilize a feedforward neural network. The training of the PINN involves acquiring random observation data and discrete physical data generated based on a three-degree-of-freedom (DOF) real physical model. The PINN is trained using both random observation data and discrete data to obtain a trained PINN. Gaussian random fields are used to generate the required data samples. Initial values ​​for the vehicle's state are first randomly generated. Continuous observation segments are generated using the vehicle's initial state and a continuous control signal generated by the random field, processed through real physical equations. Discrete physical data points are generated using discrete control signals. The PINN's loss function converges to its minimum during training, resulting in the final neural network model. Its three-degree-of-freedom real physical model is as follows: , , , , in, This is the steering angle signal at the current moment. This represents the accelerator / brake signal at the current moment. Let be the time step of the discrete space. This represents the vehicle's current position. The loss function of its physical information neural network is: , , in, This is the weighted loss function for the physical information neural network. For the loss of random observation data, For the loss of discrete physical data, for A continuous observation interval, for A discrete physical data point, The prediction results of the physical information neural network. The residuals of physical laws in a physical information neural network. These are physical parameters; for a schematic diagram illustrating the prediction performance of the physical information neural network, please refer to [link / reference]. Figure 2 ,like Figure 2 As shown, Figure 2 (a) and Figure 2 In (b), the solid lines represent the vehicle's actual position on the X and Y axes at different time points, while the dashed lines represent the vehicle's predicted position on the X and Y axes by the physical information neural network at different time points. Figure 2In (c), the solid line represents the vehicle's actual steering angle at different time points, and the dashed line represents the steering angle predicted by the physical information neural network at different time points. Figure 2 In (d), the solid line represents the actual speed of the vehicle at different time points, and the dashed line represents the speed predicted by the physical information neural network at different time points. The physical information neural network is used as the vehicle dynamics model in MPC to predict the vehicle state at the next moment in real time. As the prediction model of MPC, the physical information neural network can more accurately simulate the behavior of the vehicle in the off-road environment and provide more reliable prediction of the future state. The system obtains the current vehicle state and control signals, inputs them into a physical information neural network, and then predicts the vehicle state at the next moment. In other words, by inputting the current vehicle state and control signals into a trained physical information neural network, the system can predict the vehicle state at the next moment.

[0025] In some embodiments, in step S102, based on the road reference path, the real-time position of dynamic obstacles, and the current vehicle state, a grid planner is used to perform local planning of the vehicle trajectory to obtain a reference trajectory. Based on the road reference path, the real-time position of dynamic obstacles, and the current vehicle state, a grid planner is used to plan the vehicle trajectory to obtain multiple sets of horizontal and vertical candidate trajectories. The horizontal candidate trajectory is as follows: , , in, The target time for trajectory prediction. This represents the vehicle's current lateral position. The vehicle's current lateral speed, The vehicle's current lateral acceleration, The lateral target position is predicted by the trajectory. , , , , , These are the polynomial coefficients; Its longitudinal candidate trajectory is: , , in, For the current longitudinal position of the vehicle, The vehicle's current longitudinal speed, The vehicle's current longitudinal acceleration, , , , , Let be the polynomial coefficients; A Cartesian coordinate transformation was performed on multiple sets of candidate horizontal and vertical trajectories to obtain multiple sampled trajectories, which are as follows: , in, Let be the coordinates in the vertical position. The tangent direction angle; Collision detection is performed on multiple sampled trajectories, i.e., to determine whether any trajectories exceed the runway range or encounter obstacles. After collision detection, infeasible trajectories are removed from all sampled trajectories. The total cost of the multiple sampled trajectories after collision detection is calculated to obtain the reference trajectory with the minimum total cost. The total cost of the remaining trajectories is then calculated, and the remaining trajectories are sorted according to their total cost. The trajectory with the minimum total cost is selected as the reference trajectory and provided as the output path to MPC. Its total cost is: , in, For the total cost, As a cost-weighted coefficient, The cost of smoothness is the trajectory acceleration and the magnitude of the acceleration. The offset cost is the degree of deviation from the reference path. The cost is the speed difference between the trajectory terminal speed and the target speed. The less time required, the lower the cost. The collision cost is whether the trajectory collides with an obstacle. This comes at the cost of dynamic constraints, namely exceeding limits such as maximum acceleration and maximum curvature; the constraints of its grille planner include constraints on vehicle state and constraints on control signals, and these constraints are as follows: , , , , , , in, , For vehicles in The minimum and maximum positions of the axis. , The vehicles are respectively in The minimum and maximum positions of the axis. , These are the minimum and maximum heading angles of the vehicle, respectively. , These are the vehicle's minimum and maximum speeds, respectively. , These are the minimum and maximum values ​​of the accelerator / brake signals, respectively. , These are the minimum and maximum values ​​of the steering angle signal, respectively. The grid planner generates multiple sets of horizontal and vertical candidate trajectories in the Frenet coordinate system based on the road reference path, current vehicle state, and real-time positions of dynamic obstacles. For a schematic diagram of the Frenet coordinate system, please refer to [link / reference needed]. Figure 3 ,like Figure 3 As shown, S is the road reference path, and L is the lateral distance perpendicular to the reference path. For the coordinates of points on the reference path, middle Using the lateral offset of the reference path, cost evaluation and collision detection are performed on multiple sets of lateral and longitudinal candidate trajectories, and the optimal feasible trajectory is selected as the local path planning output.

[0026] Based on the reference trajectory, the current vehicle state, and the vehicle dynamics relationship described by the physical information neural network, a model predictive control (MPC) objective function is constructed. Solving the objective function yields the control signal for the next moment. After the grid planner provides the reference trajectory and constraints in real time, an MPC framework for collaborative optimization between the PINN model and the grid planner can be formulated, i.e., the MPC objective function, which is: , in, The process state of the vehicle. Reference state provided for reference trajectory , , , The optimal control signal. , For the Mayer term, the cost of the terminal state is given. Let be the Lagrange term, and be the cost of each stage. The objective function is solved using the IPOPT interior-point optimizer, which is a nonlinear optimization problem based on CasADi.

[0027] In some embodiments, in step S103, the vehicle is controlled based on the control signal and vehicle state at the next moment, and the vehicle state and control signal at the next moment are used as the new vehicle state and control signal at the current moment for iterative updates to achieve dynamic obstacle avoidance of the vehicle. After obtaining the control signal and vehicle state at the next moment, the MPC controls the vehicle according to the control signal and vehicle state at the next moment; the control signal and vehicle state at the current moment are iteratively updated using the control signal and vehicle state at the next moment; the vehicle state at the next moment is predicted using a physical information neural network at different time steps; a grid planner is used to generate a reference trajectory at different time steps; and then the vehicle state at the next moment is determined by the MPC. Within the prediction range, the optimal control signal solved for each time instance is... However, only the first control signal is applied to the vehicle, propagating the vehicle state from the current time step to the next time step; dynamic obstacle avoidance is achieved by controlling the vehicle to travel along the reference trajectory; please refer to the diagram illustrating the vehicle control effect. Figure 4a , Figure 4b , Figure 4c ,like Figure 4a , 4b As shown in 4c, the lines are the road reference path, the line segments are the reference trajectory planned by the grid planner, and the dots are obstacles that suddenly appear in the environment. In the off-road environment, under the road reference path, the MPC controls the vehicle to move along the reference trajectory. When an obstacle appears, the vehicle is controlled to avoid the suddenly appearing obstacle according to the planned reference trajectory, so as to realize the vehicle's dynamic obstacle avoidance. Finally, the vehicle moves to the end of the road reference path.

[0028] A vehicle control method that integrates PINN-modeled MPC with grid planner co-optimization achieves high-precision tracking and real-time obstacle avoidance in dynamic obstacle environments through an intelligent decision-making framework driven by both physical knowledge and data. It replaces the analytical dynamics model in traditional MPC with a physical information neural network, constructing a predictive model that fuses physical constraints and data-driven approaches by using vehicle kinematic equations as training penalties for the neural network. This design significantly reduces the dependence of pure data-driven methods on large amounts of training data, while avoiding modeling errors in complex scenarios (such as complex off-road environments) using traditional physical models, thus improving the accuracy and generalization ability of state prediction. Based on a global reference path and real-time obstacle information (position, velocity), and utilizing the longitudinal-lateral decoupling characteristics of the Frenet coordinate system, the originally complex two-dimensional motion planning problem is transformed into two parallel one-dimensional optimization problems, significantly improving trajectory generation efficiency. Multiple smooth candidate trajectories are generated through polynomial curve fitting, and trajectory feasibility is evaluated in real-time during MPC rolling optimization, ensuring that the path both conforms to the global route and enables real-time dynamic obstacle avoidance. Simultaneously, the grid planning method effectively simplifies the nonlinear constraints of MPC, significantly accelerating the convergence speed of the objective function and further improving the real-time performance of the system.

[0029] In summary, the real-time decision-making method for vehicle dynamic obstacle avoidance provided by this invention constructs a physical information neural network based on a three-degree-of-freedom model of the vehicle to obtain the vehicle state and control signals at the current moment. The vehicle state and control signals are then input into the physical information neural network to obtain the vehicle state at the next moment. Based on the road reference path, the real-time position of dynamic obstacles, and the vehicle state at the current moment, a grid planner is used to perform local planning of the vehicle trajectory to obtain a reference trajectory. Based on the reference trajectory, the vehicle state at the current moment, and the vehicle dynamics relationship described by the physical information neural network, a model predictive control objective function is constructed. The objective function is solved to obtain the control signal at the next moment. The vehicle is controlled based on the control signal at the next moment and the vehicle state, and the vehicle state and control signals at the next moment are used as the new vehicle state and control signals at the current moment for iterative updates to achieve dynamic obstacle avoidance and improve vehicle stability.

[0030] like Figure 5 As shown, the present invention also provides a vehicle control device 500, which can be a mobile terminal, desktop computer, laptop, handheld computer, server, or other computing device. The vehicle control device 500 includes a processor 501, a memory 502, and a display 503. Figure 5 Some components of the vehicle control device 500 are shown, but it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0031] In some embodiments, memory 502 may be an internal storage unit of the vehicle control device 500, such as a hard disk or memory of the vehicle control device 500. In other embodiments, memory 502 may be an external storage device of the vehicle control device 500, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the vehicle control device 500. Further, memory 502 may include both internal and external storage units of the vehicle control device 500. Memory 502 is used to store application software and various types of data installed on the vehicle control device 500, such as program code installed on the vehicle control device 500. Memory 502 may also be used to temporarily store data that has been output or will be output. In one embodiment, memory 502 stores a real-time vehicle dynamic obstacle avoidance decision-making program, which can be executed by processor 501 to implement the real-time vehicle dynamic obstacle avoidance decision-making method of various embodiments of the present invention.

[0032] In some embodiments, processor 501 may be a central processing unit (CPU), microprocessor or other data processing chip, used to run program code stored in memory 502 or process data, such as a real-time decision-making method for vehicle dynamic obstacle avoidance.

[0033] In some embodiments, display 503 may be an LED display, a liquid crystal display, a touch-screen liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. Display 503 is used to display identification information from the vehicle's dynamic obstacle avoidance real-time decision-making process and to display a visual user interface. Components 501-503 of the vehicle control device 500 communicate with each other via a system bus.

[0034] In some embodiments, when the processor 501 executes the vehicle dynamic obstacle avoidance real-time decision program in the memory 502, it implements each step of the vehicle dynamic obstacle avoidance real-time decision method as described in the above embodiments. Since the vehicle dynamic obstacle avoidance real-time decision method has been described in detail above, it will not be repeated here.

[0035] Accordingly, the present invention also provides a computer-readable storage medium for storing a computer-readable program or instruction, which, when executed by a processor, can implement the steps or functions of the vehicle dynamic obstacle avoidance real-time decision-making method provided in the above-described method embodiments.

[0036] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0037] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A vehicle dynamic obstacle avoidance real-time decision method, characterized in that, The method comprises the following steps: a physical information neural network is constructed based on a three-degree-of-freedom model of a vehicle, a current vehicle state and a control signal are input into the physical information neural network, and a next vehicle state is obtained; a reference trajectory is obtained by locally planning a vehicle trajectory based on a road reference path, real-time positions of dynamic obstacles and the current vehicle state, a target function of model predictive control is constructed based on the reference trajectory, the current vehicle state and a relationship between vehicle dynamics described by the physical information neural network, the target function is solved, and a next control signal is obtained; the vehicle is controlled based on the next control signal and the vehicle state, and the next vehicle state and the control signal are taken as new current vehicle state and control signal for iterative updating, so as to realize dynamic obstacle avoidance of the vehicle.

2. The vehicle dynamic obstacle avoidance real-time decision method of claim 1, wherein, The vehicle state comprises position, heading angle and speed, the control signal comprises steering angle signal and throttle / brake signal, and the three-degree-of-freedom model is as follows: ， ， ， ， ， wherein is a vehicle position that varies over time, is a vehicle heading angle that varies over time, is a vehicle speed that varies over time, is a vehicle speed that varies over time, is a vehicle speed that varies over time, is a vehicle speed that varies over time, is a dynamics model, is a steering angle signal of the vehicle, is a throttle / brake signal, is a model parameter, , , , is an initial value; The physical information neural network is as follows: ， wherein, is a state space physics informed neural network, is a hyper parameter of the physics informed neural network, is a current time vehicle state, is a current time control signal, is a predicted next time vehicle state.

3. The vehicle dynamic obstacle avoidance real-time decision method of claim 2, wherein, Before the vehicle state and the control signal are input into the physical information neural network to obtain the next vehicle state, the following steps are included: random observation data and discrete physical data generated based on a three-degree-of-freedom real physical model are obtained, and the physical information neural network is trained based on the random observation data and the discrete data.

4. The vehicle dynamic obstacle avoidance real-time decision method of claim 3, wherein, The three-degree-of-freedom real physical model is as follows: ， ， ， ， wherein, is a steering angle signal at the current time instant, is an accelerator / brake signal at the current time instant, is a time step of the discrete space, is a position of the vehicle at the current time instant.

5. The vehicle dynamic obstacle avoidance real-time decision method of claim 3, wherein, The loss function of the physical information neural network is as follows: ， ， wherein, is a weighted loss function of the physical information neural network, is a loss of random observation data, is a loss of discrete physical data, is a segment continuous observation interval, is a discrete physical data point, is a prediction result of the physical information neural network, is a residual of a physical law in the physical information neural network, is a physical parameter.

6. The vehicle dynamic obstacle avoidance real-time decision method of claim 1, wherein, The reference trajectory is obtained by locally planning a vehicle trajectory based on a road reference path, real-time positions of dynamic obstacles and the current vehicle state, and the target function of model predictive control is constructed based on the reference trajectory, the current vehicle state and a relationship between vehicle dynamics described by the physical information neural network, the target function is solved, and a next control signal is obtained, and the method comprises the following steps: a plurality of groups of lateral and longitudinal candidate trajectories are obtained by planning a vehicle trajectory based on a road reference path, real-time positions of dynamic obstacles and the current vehicle state; a plurality of sampling trajectories are obtained by converting the plurality of groups of lateral and longitudinal candidate trajectories in a Cartesian coordinate system; a total cost of the plurality of sampling trajectories after collision detection is calculated to obtain a reference trajectory with the minimum total cost.

7. The vehicle dynamic obstacle avoidance real-time decision method of claim 6, wherein, The lateral candidate trajectory is as follows: ， ， wherein is a target time for the trajectory prediction, is a current lateral position of the vehicle, is a current lateral velocity of the vehicle, is a current lateral acceleration of the vehicle, is a target lateral position for the trajectory prediction, , , , , , is a polynomial coefficient; The longitudinal candidate trajectory is as follows: ， ， wherein is the current longitudinal position of the vehicle, is the current longitudinal speed of the vehicle, is the current longitudinal acceleration of the vehicle, , , , , is a polynomial coefficient; The sampling trajectory is as follows: ， wherein is the coordinate in the longitudinal position, is the tangent direction angle; The total cost is as follows: ， wherein, is a total cost, is a cost weighting coefficient, is a smoothness cost, is an offset cost, is a speed cost, is a time cost, is a collision cost, is a dynamics constraint cost.

8. The vehicle dynamic obstacle avoidance real-time decision method of claim 6, wherein, The target function is as follows: ， wherein a process state of the vehicle, a reference state provided for the reference trajectory, an optimal control signal, a Mayer term, a Lagrange term.

9. A vehicle control apparatus characterized by comprising: The method comprises a memory and a processor; The memory stores a computer readable program which can be executed by the processor; The processor executes the computer readable program to realize the steps in the vehicle dynamic obstacle avoidance real-time decision method according to any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The program or instructions stored in the memory can be executed by the processor to realize the steps in the vehicle dynamic obstacle avoidance real-time decision method according to any one of claims 1-8.

Images