Parking control method based on artificial intelligence

Abstract

The invention relates to the technical field of automatic driving, and discloses an artificial intelligence-based parking control method, which comprises the following steps that: step 1, a vehicle dynamic model is adopted to carry out modeling on the motion of a vehicle, the motion model of the vehicle describes the relationship among the position, the speed, the direction angle and the acceleration of the vehicle, and the vehicle dynamic model is a nonlinear differential equation; step 2, after the dynamic model of the vehicle is determined, constructing a target function of path planning; a weighting coefficient in a path planning objective function is dynamically adjusted into an environment-related adaptive coefficient, and the priority of path planning and the smoothness of control input are adjusted in real time by combining the obstacle density and the vehicle motion state in the parking lot, so that the problem of dynamic change of obstacles in a complex parking lot environment is solved, and the path planning efficiency is improved. The stable parking performance in a high-density dynamic environment is realized, and the efficient and flexible path planning capability is obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous driving, and in particular to a parking control method based on artificial intelligence. Background Art

[0002] With the rapid development of intelligent transportation systems, automatic parking technology, as an important part of intelligent driving, has been applied to modern vehicles. The automatic parking system is mainly used to help the driver park the vehicle at a designated location without human intervention. The implementation of the technology depends on the vehicle's sensors and positioning system, and requires accurate path planning, control algorithms, and real-time adaptability to dynamic environments. The existing automatic parking technology has made some progress, but there are still the following technical problems:

[0003] Most current automatic parking systems rely on static obstacle detection and path planning. As the complexity of parking environments increases, the dynamic changes of obstacles become a major challenge. Existing methods ignore dynamic changes, resulting in poor adaptability of the system in high-density parking lots, low path planning efficiency, and even parking failures.

[0004] When traditional reinforcement learning algorithms perform path planning in complex environments, the overestimation of Q values ​​will lead to unstable strategies. Instability will affect the selection of parking paths, resulting in the inability of intelligent systems to make the best decisions in real time, especially in dynamic environments, and unable to quickly respond to changes in obstacles and challenges in dynamic environments.

[0005] The automatic parking task needs to consider the accuracy of the parking position and multiple sets of goals, such as path smoothness, obstacle avoidance, and speed control. Existing systems usually focus on a single task and cannot effectively balance the relationship between multiple sets of goals, resulting in the inability to simultaneously optimize path planning, control input smoothness, and safety. Therefore, how to solve multiple sets of tasks at the same time and coordinate them in a complex environment has become an important issue in improving the performance of automatic parking systems.

[0006] Therefore, those skilled in the art provide a parking control method based on artificial intelligence to solve the above-mentioned problems. Summary of the invention

[0007] In view of the deficiencies in the prior art, the present invention provides a parking control method based on artificial intelligence to solve the problems raised in the above background technology.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: a parking control method based on artificial intelligence, comprising:

[0009] Step 1: Modeling the motion of a vehicle using a vehicle dynamics model, wherein the vehicle motion model describes the relationship between the position, velocity, direction angle and acceleration of the vehicle, and the vehicle dynamics model is a nonlinear differential equation;

[0010] Step 2: After determining the dynamic model of the vehicle, construct an objective function for path planning, which includes the distance between the current position of the vehicle and the target parking space and the smoothness of the control input. The objective function aims to minimize the path length from the starting position to the target parking space and the cost of the control input to achieve the optimal motion trajectory.

[0011] Step 3: Combined with the Lagrange multiplier method, the optimal path solution problem is established by means of constraint optimization, wherein the constraints include dynamic constraints, obstacle constraints, boundary constraints, starting point and end point constraints, control input constraints, smoothness constraints and dynamic environment constraints;

[0012] Step 4: Based on the objective function and constraints obtained in the previous steps, the deep Q learning algorithm in reinforcement learning is used to achieve real-time optimization of the path. When the vehicle plans the path in a dynamic environment, reinforcement learning updates the strategy by interacting with the environment. In each interaction, the reward is calculated according to the current state and action, and the strategy is adjusted;

[0013] Step 5: Generate real-time control instructions based on the path planning results obtained by the deep Q learning algorithm. The control instructions include the acceleration and steering angle of the vehicle. The control instructions will be implemented through the vehicle control system to adjust the vehicle's motion state in real time.

[0014] Preferably, the control input weighted term β in the path planning objective function is adjusted to a dynamic adjustment term to perform adaptive adjustment according to different environmental conditions, obstacle density and vehicle speed, and the path planning objective function is adjusted to:

[0015]

[0016] Where x(t) is the position vector of the vehicle at time t, x goal is the position vector of the target parking space, α is the weight coefficient of the position error in the control path planning, u(t) is the control input vector of the vehicle at time t, β(t) is the dynamic weighted term of the vehicle control input, t0 and t1 are the start time and end time of the path planning, where β(t) is calculated as:

[0017]

[0018] Among them, β(t) is the weighting coefficient of the control input, γ env is the environmental response coefficient, o env (t) represents the influence of environmental obstacles.

[0019] Preferably, the dynamic adjustment term β(t) is affected by the vehicle speed. If the vehicle speed is higher than a certain threshold, β(t) increases, which reduces the smoothness of the path and the cost of the control input. The calculation formula of β(t) is:

[0020]

[0021] Among them, β(t) is the weighting coefficient of the control input, γ env is the environmental response coefficient, o env (t) represents the influence of environmental obstacles, γ v is the adjustment coefficient related to the vehicle speed, and v(t0 is the speed of the vehicle.

[0022] Preferably, the control input constraint takes into account the dynamic response time limit of the vehicle, the acceleration of the control input and steering angle The rate of change of is subject to the following constraints:

[0023]

[0024] Among them, κ a is the maximum rate of change of acceleration, κ φ is the maximum rate of change of the steering angle;

[0025] The change rate is adaptively adjusted according to the change of the vehicle speed v(t) to optimize the dynamic response time of the vehicle. The adaptive adjustment rule is:

[0026]

[0027] Among them, γ v is the adjustment coefficient related to vehicle speed, is the acceleration, is the steering angle, κ a is the maximum rate of change of acceleration, κ φ is the maximum rate of change of the steering angle.

[0028] Preferably, the constraint condition includes a time window limit based on a prediction model of the vehicle, the prediction model is a real-time prediction of the vehicle path, and the future state prediction of the vehicle is represented by the following equation:

[0029]

[0030] Among them, f(x(t), u(t)) is the dynamic equation of the vehicle, Δt is the prediction time window, x(t) is the state vector of the vehicle at time t, and u(t) is the control input of the vehicle at time t.

[0031] x(t+Δt) is the state vector of the vehicle at time t+Δt.

[0032] Preferably, the reinforcement learning algorithm adopts a dual Q learning structure to reduce the bias of overestimating the Q value, and updates the Q value in the following form:

[0033]

[0034] Where Q1(s, a) and Q2(s, a) are the values ​​of the Q network, η is the learning rate, γ is the discount factor, s is the current state, a is the current action, s′ is the next state, a′ is the next action, R is the immediate reward, and represents the Q-value of the optimal action in the next state s′.

[0035] Preferably, in the Q-value updating process, an experience replay mechanism is added, by storing a 4-tuple (s, a, R, s′) of historical state, action, reward, and next state, and sampling small batch data for training.

[0036] Preferably, the path planning method is combined with the state transition probability function in the Markov decision process to dynamically adjust the path planning strategy, and the transition probability function is:

[0037]

[0038] Among them, s is the current state, s′ is the next state, and o env (t) is the position of the obstacle, α is the adjustment coefficient, x(t) is the state vector of the vehicle at time t,

[0039] P(s′|s, a) represents the probability of transitioning to the next state s′ given the current state s and action a.

[0040] Preferably, the path planning optimization of the vehicle is a multi-stage decision-making process, and the path strategy is updated in each stage according to the real-time status of the vehicle, and the optimal action of the current stage is calculated by the following formula:

[0041]

[0042] where a * (t) is the optimal control action determined at time t,

[0043] Q(s(t), a) is the Q value of executing action a in state s(t),

[0044] Action a that maximizes Q(s(t), a) is the optimal control strategy under the current state.

[0045] Preferably, the control instruction generation of the system is combined with a multi-task learning algorithm to share learned features among multiple groups of tasks, so that the system can quickly adapt and optimize the path in a complex environment. Specifically, multiple groups of tasks are jointly trained in the following form:

[0046]

[0047] Among them, L i (x, u) represents the loss function of the i-th task, N is the total number of tasks, L(x, u) is the total loss function, x is the state of the vehicle, and u is the control input of the vehicle.

[0048] The present invention provides a parking control method based on artificial intelligence, which has the following beneficial effects:

[0049] 1. The present invention dynamically adjusts the weighted coefficient in the path planning objective function to an environment-related adaptive coefficient, combines the obstacle density and vehicle motion state in the parking lot, adjusts the priority of path planning and the smoothness of control input in real time, solves the problem of dynamic changes of obstacles in complex parking lots, achieves stable parking performance in high-density dynamic environments, and obtains efficient and flexible path planning capabilities.

[0050] 2. The present invention solves the problem of reduced strategy stability caused by Q value deviation in complex parking environments by introducing a dual Q learning update mechanism, achieves fast and stable strategy learning, and obtains reliable parking path planning in a dynamic parking environment.

[0051] 3. The present invention designs a multi-task loss function so that the system can take into account multiple sets of objectives at the same time, solves the problem that traditional methods cannot dynamically balance the performance of multiple sets of objectives, achieves a balance between different objectives in a complex parking environment, and obtains a comprehensive and efficient parking control method. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0053] In order to make the technical personnel in the technical field understand the scheme of the present invention, the technical scheme in the embodiment of the present invention will be clearly and completely described below in combination with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a partial embodiment of the present invention, not a complete embodiment. Based on the embodiment of the present invention, other embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present invention.

[0054] The present invention is described in detail below in conjunction with the accompanying drawings:

[0055] Example:

[0056] Please refer to the attached Figure 1 , an embodiment of the present invention provides a parking control method based on artificial intelligence, comprising:

[0057] Step 1: Use a vehicle dynamics model to model the motion of the vehicle. The vehicle motion model describes the relationship between the position, speed, direction angle and acceleration of the vehicle. The vehicle dynamics model is a nonlinear differential equation.

[0058] Step 2: After determining the vehicle's dynamic model, construct the objective function of path planning. The objective function includes the distance between the vehicle's current position and the target parking space and the smoothness of the control input. The objective function aims to minimize the path length from the vehicle's starting position to the target parking space and the cost of the control input to achieve the optimal motion trajectory.

[0059] Step 3: Combined with the Lagrange multiplier method, the optimal path solution problem is established through constraint optimization. The constraints include dynamic constraints, obstacle constraints, boundary constraints, starting and end point constraints, control input constraints, smoothness constraints and dynamic environment constraints.

[0060] Step 4: Based on the objective function and constraints obtained in the previous steps, the deep Q learning algorithm in reinforcement learning is used to achieve real-time optimization of the path. When the vehicle plans the path in a dynamic environment, reinforcement learning updates the strategy by interacting with the environment. In each interaction, the reward is calculated according to the current state and action, and the strategy is adjusted;

[0061] Step 5: Generate real-time control instructions based on the path planning results obtained by the deep Q learning algorithm. The control instructions include the acceleration and steering angle of the vehicle. The control instructions will be implemented through the vehicle control system to adjust the vehicle's motion state in real time.

[0062] Benefits of step 1: By using a nonlinear differential equation model to describe the dynamic characteristics of the vehicle in different states, including position, velocity, direction angle and acceleration, the system can understand and predict the vehicle's motion behavior, providing a theoretical basis for subsequent path planning and control;

[0063] Benefits of step 2: By setting the objective function, the vehicle's path planning considers the shortest path, including the smoothness of the control input, avoiding violent movements caused by rapid acceleration and steering angles, ensuring the efficiency and smoothness of the path, and improving the comfort and safety during parking;

[0064] Benefits of step 3: Combining dynamic constraints, obstacle constraints, boundary constraints, start and end point constraints, control input constraints, smoothness constraints and dynamic environment constraints ensures that the path planning is physically feasible, avoiding collisions and vehicles leaving the parking lot boundary. Through optimization methods, the optimal path can be obtained under the constraints, improving the reliability of path planning;

[0065] Benefits of step 4: Reinforcement learning, especially deep Q-learning algorithms, enables the system to adjust path planning strategies in real time in dynamic environments. Through continuous interaction with the environment, the vehicle can autonomously learn the best control strategy to cope with complex and changing environments;

[0066] Benefits of step 5: Through the path planning results obtained based on the deep Q learning algorithm, the system can generate control instructions in real time, such as acceleration and steering angle, to ensure that the vehicle can perform parking tasks flexibly and accurately in the actual environment, and the vehicle can respond quickly to environmental changes to ensure the smooth completion of the parking process.

[0067] Summary: The artificial intelligence-based parking control method provided by the present invention solves the path planning and control problems in a dynamic environment through the organic combination of multiple steps. The accuracy and safety of the path are ensured by using the vehicle dynamics model and the constraint optimization method; the path is optimized in real time through the deep Q learning algorithm, so that the system has the ability to cope with dynamically changing environments; finally, through the generation and execution of control instructions, the vehicle can complete the parking task efficiently, smoothly and safely.

[0068] The control input weighted term β in the path planning objective function is adjusted to a dynamic adjustment term to make adaptive adjustments according to different environmental conditions, obstacle density and vehicle speed. The path planning objective function is adjusted to:

[0069]

[0070] Where x(t) is the position vector of the vehicle at time t, x goal is the position vector of the target parking space, α is the weight coefficient of the position error in the control path planning, u(t) is the control input vector of the vehicle at time t, β(t) is the dynamic weighted term of the vehicle control input, t0 and t1 are the start time and end time of the path planning, where β(t) is calculated as:

[0071]

[0072] Among them, β(t) is the weighting coefficient of the control input, γ env is the environmental response coefficient, o env (t) represents the influence of environmental obstacles.

[0073] The dynamic adjustment term β(t) is affected by the vehicle speed. If the vehicle speed is higher than a certain threshold, β(t) increases, which reduces the smoothness of the path and the cost of control input. The calculation formula of β(t) is:

[0074]

[0075] Among them, β(t) is the weighting coefficient of the control input, γ env is the environmental response coefficient, o env (t) represents the influence of environmental obstacles, γ v is the adjustment coefficient related to the vehicle speed, and v(t) is the speed of the vehicle.

[0076] By adjusting the control input weighted term β(t) in the path planning objective function to a dynamic adjustment term, adaptive adjustments are made according to different environmental conditions, obstacle density, and vehicle speed. The dynamic adjustment term ensures that in different parking environments, path planning can flexibly respond to changes in obstacles and different vehicle motion states. This feature is particularly suitable for complex dynamic environments and can optimize path selection in real time to ensure that the vehicle can efficiently avoid obstacles and park stably.

[0077] If the vehicle speed is higher than a certain threshold, the weighted term β(t) of the control input is increased to promote the smoothness of the path and reduce the cost of the control input. This solves the problem of sharp turns and acceleration changes caused by path planning at high speeds, thereby improving the smoothness of the driving experience. At the same time, it reduces vibration and excessive control in path planning, allowing the vehicle to maintain stability when driving at high speeds.

[0078] The weighted term β(t) of the control input is related to the environmental response coefficient and the vehicle speed. The system can automatically adjust the control input strategy according to different driving environments. If the vehicle speed is high, the system will enhance the smoothness of the path, reduce excessive acceleration and sharp steering, and improve the accuracy and safety of vehicle control.

[0079] The method enables path planning to make accurate decisions in a dynamic environment, taking into account the vehicle's motion performance and the impact of obstacles in the environment. By dynamically adjusting the weights of control inputs, path planning can find a suitable path in a short time, reducing parking time and improving parking efficiency.

[0080] The control input constraint considers the dynamic response time limit of the vehicle and the acceleration of the control input. and steering angle The rate of change of is subject to the following constraints:

[0081]

[0082] Among them, κ a is the maximum rate of change of acceleration, κ φ is the maximum rate of change of the steering angle;

[0083] The rate of change is adaptively adjusted according to the change of vehicle speed v(t) to optimize the dynamic response time of the vehicle. The adaptive adjustment rule is:

[0084]

[0085] Among them, γ v is the adjustment coefficient related to vehicle speed, is the acceleration, is the steering angle, κ a is the maximum rate of change of acceleration, κ φ is the maximum rate of change of the steering angle.

[0086] By introducing adaptive regulation of vehicle speed, the rate of change of acceleration and steering angle can be flexibly adjusted according to the current speed of the vehicle, effectively optimizing the dynamic response time of the vehicle, making the vehicle perform smoothly and responsively under different driving conditions and avoiding unnecessary overreaction.

[0087] Limiting the rate of change of acceleration and steering angle plays an important role in avoiding the dangers caused by sudden control inputs. Especially at high speeds, reducing the sudden changes in control inputs effectively reduces the risk of accidents caused by excessive control and improves the safety of the system. By adaptively adjusting the rate of change, the system can optimize in real time according to the vehicle's motion state, reduce unnecessary sharp reactions, and ensure a more stable and safe operation process.

[0088] By limiting the rate of change of vehicle acceleration and steering angle, the discomfort of passengers in the vehicle caused by excessive control is avoided, and the driving experience is improved. The vehicle can smoothly transition during parking, avoiding rapid turns and braking, reducing vibration and discomfort of passengers, and improving the comfort of parking operations.

[0089] The change of vehicle speed affects the rate of change of acceleration and steering angle. By setting the adjustment coefficient related to vehicle speed, the system can dynamically optimize the response according to the driving requirements at different speeds. The adaptive adjustment mechanism enables the system to respond quickly at high speeds, and avoid unnecessary sharp operations at low speeds and parking, improving the vehicle's reaction efficiency in different situations, ensuring fast response and short parking time.

[0090] By limiting the rate of change of vehicle control inputs and adaptively adjusting according to speed, the system can accurately control the movement of the vehicle, ensuring precise parking maneuvers in complex and narrow parking scenarios. Especially when parking spaces are narrow, the optimization method ensures that the vehicle completes the parking task in a delicate and smooth manner, reducing errors and deviations during the operation.

[0091] The constraints include the time window limit of the vehicle-based prediction model. The prediction model is a real-time prediction of the vehicle path. The future state prediction of the vehicle is expressed by the following equation:

[0092]

[0093] Among them, f(x(t), u(t)) is the dynamic equation of the vehicle, Δt is the prediction time window, x(t) is the state vector of the vehicle at time t, and u(t) is the control input of the vehicle at time t.

[0094] x(t+Δt) is the state vector of the vehicle at time t+Δt.

[0095] By predicting the future state based on the vehicle's prediction model, the system can calculate the vehicle's trajectory in real time at the next moment. The vehicle's motion is modeled using dynamic equations to ensure that path planning can foresee the dynamic changes of the vehicle in the future. The prediction model can help the system identify potential risks and obstacles in advance, adjust the path planning strategy in time, avoid collisions and path deviations, and improve decision-making efficiency during parking.

[0096] The prediction of the vehicle's future state relies on real-time feedback of the vehicle's current state vector and control input, allowing the system to accurately calculate the trajectory of future steps. By using a dynamic prediction model, the system can optimize the path according to the vehicle's immediate state and improve the accuracy of path planning, especially in complex or narrow parking scenarios, and can accurately control the vehicle's driving path.

[0097] Based on the time window limit of the prediction model, the system can consider the future movement changes of the vehicle within a certain time range and make adjustments, so that the path planning can be adjusted according to the movement state of the vehicle to cope with different environments and obstacle configurations. For complex dynamic parking environments, the system can adapt to rapidly changing parking situations and effectively avoid unstable behaviors caused by environmental changes.

[0098] By predicting the future state of the vehicle, potential risks leading to collisions can be discovered early, and the path can be optimized to avoid accidents. In addition, the application of the prediction model can reduce the delay in calculation and decision-making, making path planning efficient. In the process of automatic parking, unnecessary repeated adjustments and unnecessary path corrections can be avoided to ensure the safety of the parking process.

[0099] Based on the prediction of the vehicle's future state, the system can reasonably evaluate the impact of control inputs. Within the prediction time window, the vehicle's control inputs can be dynamically adjusted according to changes in future states, thereby improving control accuracy and response speed. By considering the impact of future states on current decisions, the system optimizes the decision-making process of control inputs and improves the stability and robustness of the system.

[0100] The reinforcement learning algorithm uses a double Q learning structure to reduce the bias of overestimating the Q value and updates the Q value in the following form:

[0101]

[0102] Where Q1(s, a) and Q2(s, a) are the values ​​of the Q network, η is the learning rate, γ is the discount factor, s is the current state, a is the current action, s′ is the next state, a′ is the next action, R is the immediate reward, and represents the Q-value of the optimal action in the next state s′.

[0103] During the Q-value update process, an experience replay mechanism is added, which stores the four-tuple (s, a, R, s′) of historical state, action, reward, and next state, and samples small batches of data for training.

[0104] By adopting the dual Q learning structure, the system uses the Q network to estimate and update the Q value, effectively avoiding the common problem of overestimating the Q value in the Q learning algorithm. In traditional Q learning, the overestimated Q value will lead to unstable strategies and slow learning convergence. Dual Q learning reduces deviations by introducing the Q network and alternating updates, making the system's learning process stable and accurate in complex environments.

[0105] The introduction of the experience replay mechanism allows the intelligent agent to store the quadruple of historical state, action, reward, and next state and sample small batches of data for training, helping the intelligent agent to break the correlation between samples, enhance the stability of the learning process, and improve the efficiency of data reuse.

[0106] Experience replay can improve the reuse rate of samples and help the system avoid overfitting problems caused by sample correlation through sampling. When the environment changes during the training process, the system can still learn decision-making strategies with strong generalization from past experience, thereby improving the sample efficiency and training stability of the learning process.

[0107] The path planning method is combined with the state transition probability function in the Markov decision process to dynamically adjust the path planning strategy. The transition probability function is:

[0108]

[0109] Among them, s is the current state, s′ is the next state, and o env (t) is the position of the obstacle, α is the adjustment coefficient, x(t) is the state vector of the vehicle at time t,

[0110] P(s′|s, a) represents the probability of transitioning to the next state s′ given the current state s and action a.

[0111] The vehicle path planning optimization is a multi-stage decision-making process. The path strategy is updated in each stage according to the real-time status of the vehicle, and the optimal action in the current stage is calculated by the following formula:

[0112]

[0113] where a * (t) is the optimal control action determined at time t,

[0114] Q(s(t), a) is the Q value of executing action a in state s(t),

[0115] Action a that maximizes Q(s(t), a) is the optimal control strategy under the current state.

[0116] By combining the path planning method with the state transition probability function of the Markov decision process, the system can dynamically adjust the path planning strategy according to the vehicle's current state, action, and the location of obstacles in the environment. The state transition probability function enables path planning to quantify the impact of different states and actions on the next state, providing a scientific basis for real-time decision-making in dynamic environments.

[0117] The transition probability function dynamically adjusts the probability of state transition by calculating the distance between the vehicle's current state and the obstacle's position. If an obstacle approaches, the system can identify potential risks and reduce the transition probability associated with high-risk states, significantly improving the safety of path planning in dynamic environments, effectively avoiding collision risks, and providing robustness for the parking process.

[0118] The multi-stage decision-making process is used to decompose the path planning into multiple continuous stages, and each stage dynamically updates the path strategy based on the current state. By calculating the optimal control action in real time, the system can quickly determine the optimal control instruction at the current stage. Staged path optimization reduces the complexity of calculation, can improve the real-time response speed of the system in a dynamic environment, and make path planning efficient.

[0119] The multi-stage decision-making process achieves local optimization of the current state by maximizing the Q value at each stage, while considering the benefits of the future state to ensure the global optimality of path planning. By decomposing the global goal into multi-stage goals, path planning can accurately respond to complex parking environments and improve the quality of the overall parking strategy.

[0120] The system's control command generation is combined with a multi-task learning algorithm to share learned features between multiple groups of tasks, allowing the system to quickly adapt and optimize paths in complex environments. Specifically, multiple groups of tasks are jointly trained in the following ways:

[0121]

[0122] Among them, L i (x, u) represents the loss function of the i-th task, N is the total number of tasks, L(x, u) is the total loss function, x is the state of the vehicle, and u is the control input of the vehicle.

[0123] By combining multi-task learning algorithms and jointly training the loss functions of multiple groups of tasks, the system can achieve multi-objective optimization and coordination in complex environments. By sharing learned features, the system can efficiently complete path planning, speed control, and obstacle avoidance, improving the adaptability, efficiency, and accuracy of the system. Multi-task learning accelerates the training process, reduces the demand for computing resources, enhances the system's generalization ability in dynamic environments, and ensures the optimality, safety, and stability of the path during automatic parking.

[0124] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A parking control method based on artificial intelligence, characterized in that: include: Step 1: Modeling the motion of a vehicle using a vehicle dynamics model, wherein the vehicle motion model describes the relationship between the position, velocity, direction angle and acceleration of the vehicle, and the vehicle dynamics model is a nonlinear differential equation; Step 2: After determining the dynamic model of the vehicle, construct an objective function for path planning, which includes the distance between the current position of the vehicle and the target parking space and the smoothness of the control input. The objective function aims to minimize the path length from the starting position to the target parking space and the cost of the control input to achieve the optimal motion trajectory. Step 3: Combined with the Lagrange multiplier method, the optimal path solution problem is established by means of constraint optimization, wherein the constraints include dynamic constraints, obstacle constraints, boundary constraints, starting point and end point constraints, control input constraints, smoothness constraints and dynamic environment constraints; Step 4: Based on the objective function and constraints obtained in the previous steps, the deep Q learning algorithm in reinforcement learning is used to achieve real-time optimization of the path. When the vehicle plans the path in a dynamic environment, reinforcement learning updates the strategy by interacting with the environment. In each interaction, the reward is calculated according to the current state and action, and the strategy is adjusted; Step 5: Generate real-time control instructions based on the path planning results obtained by the deep Q learning algorithm. The control instructions include the acceleration and steering angle of the vehicle. The control instructions will be implemented through the vehicle control system to adjust the vehicle's motion state in real time.

2. The parking control method based on artificial intelligence according to claim 1, characterized in that: The control input weighted term β in the path planning objective function is adjusted to a dynamic adjustment term to perform adaptive adjustment according to different environmental conditions, obstacle density and vehicle speed. The path planning objective function is adjusted to: Where x(t) is the position vector of the vehicle at time t, x goal is the position vector of the target parking space, α is the weight coefficient of the position error in the control path planning, u(t) is the control input vector of the vehicle at time t, β(t) is the dynamic weighted term of the vehicle control input, t0 and t1 are the start time and end time of the path planning, where β(t) is calculated as: Among them, β(t) is the weighting coefficient of the control input, γ env is the environmental response coefficient, o env (t) represents the influence of environmental obstacles.

3. The parking control method based on artificial intelligence according to claim 2, characterized in that: The dynamic adjustment term β(t) is affected by the vehicle speed. If the vehicle speed is higher than a certain threshold, β(t) increases, which reduces the smoothness of the path and the cost of control input. The calculation formula of β(t) is: Among them, β(t) is the weighting coefficient of the control input, γ env is the environmental response coefficient, o env (t) represents the influence of environmental obstacles, γ v is the adjustment coefficient related to the vehicle speed, and v(t) is the vehicle speed.

4. The parking control method based on artificial intelligence according to claim 1, characterized in that: The control input constraint takes into account the dynamic response time limit of the vehicle and the acceleration of the control input. and steering angle The rate of change of is subject to the following constraints: Among them, κ a is the maximum rate of change of acceleration, κ φ is the maximum rate of change of the steering angle; The change rate is adaptively adjusted according to the change of the vehicle speed v(t) to optimize the dynamic response time of the vehicle. The adaptive adjustment rule is: Among them, γ v is the adjustment coefficient related to vehicle speed, is the acceleration, is the steering angle, κ a is the maximum rate of change of acceleration, κ φ is the maximum rate of change of the steering angle.

5. The parking control method based on artificial intelligence according to claim 4, characterized in that: The constraints include a time window limit based on a prediction model of the vehicle, the prediction model is a real-time prediction of the vehicle path, and the future state prediction of the vehicle is represented by the following equation: Among them, f(x(t), u(t)) is the dynamic equation of the vehicle, Δt is the prediction time window, x(t) is the state vector of the vehicle at time t, and u(t) is the control input of the vehicle at time t. x(t+Δt) is the state vector of the vehicle at time t+Δt.

6. The parking control method based on artificial intelligence according to claim 1, characterized in that: The reinforcement learning algorithm adopts a double Q learning structure to reduce the bias of overestimating the Q value and updates the Q value in the following form: Where Q1(s, a) and Q2(s, a) are the values ​​of the Q network, η is the learning rate, γ is the discount factor, s is the current state, a is the current action, s′ is the next state, a′ is the next action, R is the immediate reward, and represents the Q-value of the optimal action in the next state s′.

7. The parking control method based on artificial intelligence according to claim 6, characterized in that: In the Q-value updating process, an experience replay mechanism is added, which stores the four-tuple (s, a, R, s′) of historical state, action, reward, and next state, and samples small batches of data for training.

8. The parking control method based on artificial intelligence according to claim 1, characterized in that: The path planning method is combined with the state transition probability function in the Markov decision process to dynamically adjust the path planning strategy. The transition probability function is: Among them, s is the current state, s′ is the next state, and o env (t) is the position of the obstacle, α is the adjustment coefficient, x(t) is the state vector of the vehicle at time t, P(s′|s, a) represents the probability of transitioning to the next state s′ given the current state s and action a.

9. The parking control method based on artificial intelligence according to claim 8, characterized in that: The vehicle path planning optimization is a multi-stage decision-making process. The path strategy is updated in each stage according to the real-time status of the vehicle, and the optimal action of the current stage is calculated by the following formula: where a * (t) is the optimal control action determined at time t, Q(s(t), a) is the Q value of executing action a in state s(t), Action a that maximizes Q(s(t), a) is the optimal control strategy under the current state.

10. The parking control method based on artificial intelligence according to claim 1, characterized in that: The control command generation of the system is combined with a multi-task learning algorithm to share learned features among multiple groups of tasks, so that the system can quickly adapt and optimize the path in complex environments. Specifically, multiple groups of tasks are jointly trained in the following forms: Among them, L i (x, u) represents the loss function of the i-th task, N is the total number of tasks, L(x, u) is the total loss function, x is the state of the vehicle, and u is the control input of the vehicle.

Images