Method and device for personalized adaptation of automatic parking, vehicle and medium

Abstract

This invention discloses a personalized adaptation method, device, vehicle, and medium for automatic parking in the field of intelligent driving technology. The method includes: acquiring multi-source data containing user historical parking operation data and vehicle status information; constructing a fusion feature based on the multi-source data to characterize user parking preferences and the vehicle's current parking capability; inputting the fusion feature and environmental perception information as input to a reinforcement learning decision model; outputting a parking decision matching the user's parking preferences based on the current parking environment; generating a parking trajectory matching the parking decision using a path planning algorithm based on the parking decision and combined with vehicle kinematic constraints; and using a predictive control algorithm to track the parking trajectory in real time and generate vehicle control operations to achieve automatic parking. This method enables the automatic parking strategy to adaptively adjust to user operating habits and vehicle status, improving parking success rate and parking experience in complex parking scenarios.

Description

Technical Field

[0001] This invention relates to the field of intelligent driving technology, and in particular to an automatic parking personalized adaptation method, device, vehicle, and medium. Background Technology

[0002] With the development of intelligent driving technology, automatic parking systems have been gradually applied in passenger vehicles, reducing the driver's workload to some extent. However, in practical applications, these systems still suffer from inconsistent user experience and insufficient adaptability to different scenarios. On one hand, most automatic parking systems use preset parameters or fixed control strategies, ignoring the differences in operating habits among different users during parking. For example, users may have different preferences for parking speed, steering rhythm, and parking accuracy, which can easily lead to some users feeling uncomfortable or even distrustful of the system's behavior. On the other hand, these systems have limited ability to comprehensively utilize the vehicle's real-time operating status and user operating characteristics during parking, making it difficult to adjust parking strategies in a timely manner when vehicle status changes or environmental complexity increases, thus affecting parking success rate and safety.

[0003] At the path planning level, related technologies for automated parking mainly include rule-based path planning methods, complex curve model-based path planning methods, and optimal control-based problem-solving model-based path planning methods. Rule-based path planning typically simplifies the parking trajectory to a combination of straight lines and arcs. While computationally efficient, it is highly dependent on the size and shape of the parking space and prone to curvature discontinuities at path transition points, affecting parking smoothness. Path planning using polynomials or spline curves improves trajectory smoothness to some extent, but its computational complexity is high, requiring significant sensor accuracy and computational resources, and it has limitations in real-time performance. Optimal control-based path planning methods can comprehensively consider vehicle dynamics constraints and various environmental constraints, but their solution process is usually complex and difficult to achieve rapid response in dynamic environments or when vehicle states are uncertain.

[0004] Furthermore, in complex parking scenarios such as narrow parking spaces, areas with dense obstacles, or poor lighting conditions, the automatic parking systems in related technologies are insufficiently adaptable to environmental perception errors, vehicle state fluctuations, and the generalization ability of the algorithms, easily leading to parking failures or inconsistent actions. In particular, parking strategies trained in simulation environments often perform unstably in real-world scenarios, and the gap between simulation and reality further limits the reliability and consistency of automatic parking systems in complex scenarios. Summary of the Invention

[0005] This invention aims to at least partially address one of the technical problems in related technologies. Therefore, the purpose of this invention is to propose an automatic parking personalized adaptation method, device, vehicle, and medium to improve the adaptability of the automatic parking system to differentiated user needs and changes in vehicle operating status, thereby improving parking performance and user experience in complex parking scenarios.

[0006] To achieve the above objectives, a first aspect of the present invention provides an automatic parking personalization adaptation method, comprising: Acquire multi-source data containing user historical parking operation data and vehicle status information, and construct a fusion feature based on the multi-source data to characterize user parking preferences and vehicle current parking capabilities; The fused features and environmental perception information are used together as input to a preset reinforcement learning decision model, which outputs a parking decision that matches the user's parking preferences based on the current parking environment. Based on the parking decision and combined with vehicle kinematic constraints, a parking trajectory matching the parking decision is generated through a path planning algorithm. The parking trajectory is tracked in real time using a predictive control algorithm to generate vehicle control operations, thereby enabling automatic parking of the vehicle.

[0007] In addition, the method of the above embodiments of the present invention may also have the following additional technical features: According to one embodiment of the present invention, constructing a fusion feature representing user parking preferences and the vehicle's current parking capability based on the multi-source data includes: Feature extraction is performed on the user's historical parking operation data to form user preference features that characterize the user's parking style and operating habits; The vehicle state information is used to extract features to form vehicle capability features that characterize the vehicle's current motion capabilities and response characteristics; The user preference features and the vehicle capability features are weighted and fused to form a fused feature for reinforcement learning decision-making.

[0008] According to one embodiment of the present invention, the step of inputting the fused features and environmental perception information together as model input states into a preset reinforcement learning decision model includes: Visual spatial features are extracted from environmental perception data, and the fused features are processed using a multilayer perceptron. The visual spatial features are concatenated with the processed fusion features and input into the reinforcement learning decision model for action value function evaluation to output the parking decision.

[0009] According to one embodiment of the present invention, generating a parking trajectory matching the parking decision using a path planning algorithm based on the parking decision and in conjunction with vehicle kinematic constraints includes: An initial parking trajectory is generated by performing an initial path search in the current parking environment using a path search algorithm. Using the parking decision as a reference constraint, the initial parking trajectory is optimized and smoothed based on a quadratic programming algorithm to generate a parking trajectory that satisfies the constraints of curvature continuity and executability.

[0010] According to one embodiment of the present invention, the real-time tracking of the parking trajectory using a predictive control algorithm includes: Based on the current vehicle status information and the vehicle's dynamic characteristic parameters, a vehicle dynamics prediction model is constructed. Based on the vehicle dynamics prediction model, the vehicle motion state in the prediction time domain is predicted, and the parking trajectory is used as a reference trajectory. Under the conditions of satisfying vehicle dynamics constraints and actuator constraints, the corresponding target control quantity is solved by the model predictive control algorithm in a rolling manner. Based on the target control quantity, the vehicle control operation is dynamically updated to achieve accurate tracking of the parking trajectory.

[0011] According to one embodiment of the present invention, the method further includes: During the vehicle parking process, the parking decision is made according to the current parking environment, and the corresponding new input state and reward signal are obtained based on the execution result of the parking decision to form an interaction sample; The interaction samples are stored in the experience replay pool, and samples are randomly selected from the experience replay pool to optimize and train the reinforcement learning decision model, so as to reduce the impact of sample correlation on model convergence.

[0012] According to one embodiment of the present invention, the method further includes: Based on the Q-value corresponding to the parking decision output by the reinforcement learning decision model, and the target Q-value output by the target network under the same input state, a loss function is constructed and the model parameters of the reinforcement learning decision model are updated through backpropagation; the target network has the same structure as the reinforcement learning decision model and synchronizes the model parameters of the reinforcement learning decision model through periodic updates.

[0013] To achieve the above objectives, a second aspect of the present invention provides an automatic parking personalization adaptation device, comprising: The feature fusion construction module is used to acquire multi-source data containing user historical parking operation data and vehicle status information, and to construct fusion features based on the multi-source data to characterize user parking preferences and vehicle current parking capabilities. The parking decision output module is used to input the fused features and environmental perception information as input states into a preset reinforcement learning decision model, and output a parking decision that matches the user's parking preferences based on the current parking environment. The parking trajectory planning module is used to generate a parking trajectory that matches the parking decision based on the parking decision and in combination with vehicle kinematic constraints, through a path planning algorithm. The parking execution module is used to track the parking trajectory in real time using predictive control algorithms and generate vehicle control operations to achieve automatic parking of the vehicle.

[0014] To achieve the above objectives, a third aspect of the present invention provides a vehicle including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described automatic parking personalized adaptation method.

[0015] To achieve the above objectives, a fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the steps of the above-described automatic parking personalized adaptation method.

[0016] The automatic parking personalized adaptation method, device, vehicle, and medium of this invention comprehensively incorporate user's historical parking operation data, vehicle status information, and environmental perception information during the automatic parking process. This allows for a unified representation of user parking preferences and the vehicle's current parking capability. Based on a reinforcement learning decision model, it outputs parking decisions that match user habits, enabling dynamic adjustments to the parking strategy according to different users and vehicle states. Simultaneously, it generates a parking trajectory consistent with the decision result by incorporating vehicle kinematic constraints and achieves real-time tracking of the parking trajectory through predictive control, thus improving the smoothness and stability of the parking process. This method effectively enhances the adaptability and success rate of the automatic parking system in complex parking environments, narrowing the performance gap between simulation training and practical applications, thereby improving the overall parking experience. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an automatic parking personalization adaptation method in one embodiment; Figure 2 This is a schematic diagram illustrating the specific process of enhanced decision generation in one embodiment; Figure 3 This is a schematic diagram of the process for determining the parking trajectory in one embodiment; Figure 4 This is a schematic diagram of the real-time tracking process for parking trajectories in one embodiment; Figure 5 This is a schematic diagram of the optimization process of a reinforcement learning decision model in one embodiment; Figure 6 This is a schematic diagram of the execution flow for personalized adaptation of automatic parking in one embodiment; Figure 7 This is a structural block diagram of an automatic parking personalization adaptation device in one embodiment. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] The implementation details of the technical solutions of the embodiments of the present invention are described in detail below.

[0020] In one embodiment, such as Figure 1 The diagram illustrates a flowchart of a personalized adaptation method for automatic parking, which may include the following steps: Step S101: Obtain multi-source data containing user's historical parking operation data and vehicle status information, and construct a fusion feature based on the multi-source data to characterize the user's parking preferences and the vehicle's current parking capabilities.

[0021] The driver assistance system records user actions during parking, including vehicle speed changes, steering timing, and parking position deviation. Simultaneously, it collects user preference information, such as desired parking speed and preferred parking space type, via in-vehicle display or mobile device. At the same time, it collects current vehicle status data, including vehicle position, attitude, steering angle, gear position, and braking status, recording them chronologically to ensure each data point includes a timestamp and vehicle identification information for subsequent feature analysis and historical tracking.

[0022] To ensure data reliability and availability, the collected multi-source data is stored and managed. Specifically, user historical operation data is categorized according to user identifiers, with each user's data stored in an independent data record table for efficient querying and management. Vehicle status data is recorded in time series, with each record including a timestamp, vehicle ID, and various status parameters. Data backup and recovery mechanisms are also configured to ensure data security and traceability.

[0023] In practical applications, the acquired multi-source data is preprocessed, including data cleaning to remove duplicates, correct errors and fill missing values, normalization to unify various data scales, and feature engineering to extract valuable features, such as calculating vehicle speed and steering angle change rate to form acceleration / deceleration features and steering flexibility features, which are used to characterize user operation preferences and vehicle response capabilities.

[0024] In the feature fusion stage, multi-source data is comprehensively analyzed and processed to generate fused features that represent user parking preferences and the vehicle's current mobility, providing support for the input of subsequent reinforcement learning models.

[0025] In one embodiment, to achieve a comprehensive analysis of a user's personalized parking habits and the vehicle's current state, a feature representation for reinforcement learning decision-making is constructed, which can be achieved in the following way: Feature extraction is performed on users' historical parking operation data, including calculating vehicle speed change characteristics, steering operation timing characteristics, and parking completion position deviation characteristics, to characterize the user's operating style and behavioral habits during the parking process. In this process, subjective preference information provided by the user through in-vehicle displays or mobile terminals, such as desired parking speed or preferred parking space type, can also be incorporated to form a more comprehensive set of user preference characteristics.

[0026] Feature extraction is performed on vehicle state information, including calculating indicators such as vehicle attitude, steering angle change, braking status, and vehicle position change, to characterize the vehicle's current motion capability and response characteristics.

[0027] After feature extraction, user preference features and vehicle capability features are weighted and fused according to preset weights to generate the fused feature representation required for reinforcement learning decision-making, so as to provide effective input data support in subsequent parking strategy optimization and path planning processes.

[0028] Throughout the processing, the generation of fused features follows the principles of making full use of multi-source data, unifying scales, and eliminating anomalies, ensuring that the final features can effectively characterize user personalized preferences and vehicle dynamic performance, providing a reliable basis for personalized adaptation of automatic parking.

[0029] Step S102: The fused features and environmental perception information are used as input states to a preset reinforcement learning decision model, and a parking decision matching the user's parking preferences is output based on the current parking environment.

[0030] Environmental perception information includes images acquired by cameras and information about the vehicle's surroundings provided by lidar or ultrasonic sensors, which is used to characterize the location of parking spaces, the distribution of obstacles, and spatial constraints.

[0031] Subsequently, the obtained fused features and environmental perception information are synchronized over time to form a complete input state, ensuring that the input state simultaneously reflects user operation preferences, vehicle mobility, and current environmental conditions. In practical applications, the input state can be represented as a multi-dimensional feature vector or matrix, where the fused features provide information on user operation preferences and vehicle mobility, and the environmental perception information provides information on obstacle locations, parking space boundaries, and spatial constraints. This input state, after unified encoding processing, serves as the state input for the reinforcement learning decision model.

[0032] After the input state is fed into the reinforcement learning decision model, the model selects possible parking actions based on the current state. The output parking decision matches the user's historical parking preferences while considering the limitations and constraints of the current parking environment. The parking decision includes forward, reverse, and steering operations to guide the vehicle in completing personalized parking actions in the current parking environment, achieving dynamic adaptation to different parking space shapes and obstacle distributions, while taking into account user preferences and safety constraints.

[0033] In one embodiment, Figure 2 The diagram illustrates the specific process of enhancing decision generation, which may include the following steps: Step S201: Extract visual spatial features from environmental perception data and process and fuse the features using a multilayer perceptron.

[0034] The environmental perception data is processed. Images acquired by cameras are used to extract spatial features through convolution operations to characterize obstacle positions, parking space boundaries, and depth information; distance information and obstacle distance information provided by other sensors (such as lidar or ultrasonic sensors) can also be combined to form an environmental feature matrix.

[0035] Simultaneously, the obtained fused features are input into a multilayer perceptron for nonlinear mapping, resulting in a high-dimensional representation suitable for reinforcement learning models. This operation transforms the fused features into a representation suitable for combining with visual space features, providing a unified feature space for subsequent decision inputs.

[0036] Step S202: The visual spatial features are spliced ​​with the processed fusion features and input into the reinforcement learning decision model for action value function evaluation to output parking decision.

[0037] The processed fusion features are combined with visual spatial features to form a complete decision input state. This input state simultaneously includes user operation preferences, vehicle dynamic capabilities, and environmental information, comprehensively representing the current parking situation.

[0038] In reinforcement learning decision-making models, each candidate parking action (e.g., moving forward, backward, or turning) is evaluated using an action value function (such as the Q-function). The action value function is used to calculate the expected effect of each candidate parking action under the current input state, thereby measuring the merits of the action in the current environment and user preferences.

[0039] By evaluating all candidate actions, we can compare their performance in terms of safety, efficiency, and alignment with user preferences. The evaluation results can guide the model to select the optimal action (such as choosing the candidate action with the highest score), generate actual parking decisions, and achieve dynamic adaptation to complex parking space and obstacle environments.

[0040] Ultimately, the parking decisions output by the reinforcement learning decision model include forward, reverse, and steering maneuvers. These decisions not only match the user's historical operational preferences but also comprehensively consider the limitations and constraints of the current parking environment, ensuring that the vehicle can complete parking safely, efficiently, and in a personalized manner.

[0041] Step S103: Based on the parking decision and combined with vehicle kinematic constraints, a parking trajectory matching the parking decision is generated through a path planning algorithm.

[0042] The parking decision determines the sequence of actions the vehicle needs to perform, including forward, reverse, and steering operations. It also takes into account the vehicle's current motion state, position, and available space to plan the vehicle's driving mode as a whole, in order to generate a parking trajectory that can guide the vehicle to smoothly drive from its current position to the target parking space.

[0043] During the parking trajectory generation process, the vehicle's minimum turning radius, wheelbase parameters, and forward / backward direction switching capability are incorporated as kinematic constraints into the path planning process. This ensures that the generated parking trajectory conforms to the vehicle's basic steering and driving characteristics, enabling the vehicle to park successfully under various parking space shapes and obstacle layouts. Simultaneously, the trajectory can be appropriately adjusted based on the user's personalized parking preferences, such as desired parking speed and parking position accuracy. This ensures that the final generated trajectory, while guaranteeing safety and feasibility, also reflects the user's operating habits and preferences.

[0044] The generated parking trajectory can be used as input for vehicle control, guiding the vehicle to perform forward, reverse, and steering operations during actual parking, achieving safe, efficient, and personalized parking results.

[0045] In one embodiment, Figure 3 The diagram illustrates the process of determining the parking trajectory, which may include the following steps: Step S301: Perform an initial path search in the current parking environment using a path search algorithm to generate an initial parking trajectory.

[0046] An initial parking trajectory is generated using a path search algorithm (such as a fast exploratory random tree algorithm) within the current parking environment. This path search, based on the vehicle's current position, attitude, speed, and environmental constraints (such as obstacle positions and parking space boundaries), quickly explores the vehicle's drivable space and generates an initial trajectory for the vehicle to move. The generated initial trajectory ensures that the vehicle will not collide with surrounding obstacles when performing parking maneuvers, while also satisfying the vehicle's basic movement capabilities. The current parking environment includes the parking space range where the vehicle is located, the distribution of surrounding obstacles, and spatial constraint information, used to limit the vehicle's drivable space.

[0047] Step S302: Using parking decision as a reference constraint, the initial parking trajectory is optimized and smoothed based on a quadratic programming algorithm to generate a parking trajectory that satisfies the constraints of curvature continuity and executability.

[0048] Here, by adjusting the key points of the initial parking trajectory, the trajectory curvature changes smoothly and the speed changes continuously, thereby ensuring the operability of the vehicle on the entire trajectory.

[0049] During the optimization process, it is necessary to consider the vehicle's dynamic constraints, including the maximum steering angle, minimum turning radius, and vehicle acceleration and deceleration capabilities, to ensure that the trajectory is feasible within the vehicle's maneuverability range and that there are no sharp turns or excessive speeds that would render the trajectory unexecutable.

[0050] The quadratic programming algorithm is used to optimize the position and direction of key points on the trajectory, making the trajectory smooth and continuous while maintaining safety, and balancing operational feasibility and efficiency. By using the vehicle action sequence corresponding to the parking decision as a reference constraint for trajectory optimization, the optimized parking trajectory is kept consistent with the parking decision in terms of geometry and movement sequence, thereby ensuring that forward, backward, and turning actions can be implemented continuously and executablely.

[0051] In the process of optimizing the initial parking trajectory based on the quadratic programming algorithm, the user's personalized parking preferences can be incorporated into the optimization constraints or optimization objectives. For example, the desired parking speed can be mapped to the trajectory speed constraint range, and the parking position accuracy can be mapped to the termination pose error constraint. This allows for targeted adjustment of the trajectory shape and termination state while optimizing trajectory smoothness and curvature continuity.

[0052] In this way, the optimized parking trajectory not only meets the requirements of vehicle dynamics and executability, but also better suits the user's personalized parking habits, and serves as a reference trajectory for the path tracking stage in subsequent control execution.

[0053] Step S104: The parking trajectory is tracked in real time using a predictive control algorithm to generate vehicle control operations, thereby realizing automatic parking of the vehicle.

[0054] After obtaining the parking trajectory that matches the parking decision, a predictive control algorithm is used to track the parking trajectory in real time to generate vehicle control operations. Specifically, during the parking process, based on the vehicle's current position, attitude, speed, and steering state, the vehicle's motion state in the prediction time domain is predicted, and combined with the parking trajectory as a reference, the vehicle's control operations are dynamically calculated.

[0055] Among them, the predictive control algorithm is used to adjust the vehicle's control operation in real time according to the deviation between the vehicle's current state and the parking trajectory, so that the vehicle's operating state continuously converges to the parking trajectory, thereby achieving stable and smooth trajectory tracking control.

[0056] In practical applications, predictive control takes into account both the vehicle's dynamic characteristics and the actuator's physical constraints, ensuring that the generated control operations vary within the executable range, thereby guaranteeing that the vehicle runs smoothly, safely, and reliably while tracking the parking trajectory.

[0057] For example, when the distance between the vehicle and an obstacle is detected to be less than a preset safety threshold, the predictive control algorithm can automatically reduce the vehicle's speed and adjust the steering angle accordingly to avoid a collision. When sudden situations such as road bumps or abnormal changes in vehicle posture are detected, the throttle and braking force can be dynamically adjusted to make the vehicle run more smoothly.

[0058] In addition, predictive control algorithms can work in conjunction with the vehicle's Electronic Stability Program (ESP) and Automatic Emergency Braking (AEB) system to trigger stability control or emergency braking operations when necessary, thereby further improving the safety and reliability of the parking process.

[0059] Through the above methods, the vehicle can continuously correct its motion state in complex parking environments, accurately track the parking trajectory, and finally complete the automatic parking operation, thereby providing users with a safe, convenient and comfortable automatic parking experience.

[0060] In one embodiment, Figure 4 A schematic diagram of the processing flow for real-time tracking of parking trajectories may include the following steps: Step S401: Based on the current vehicle state information and the vehicle's dynamic characteristic parameters, construct a vehicle dynamics prediction model.

[0061] Before performing real-time tracking of parking trajectories, it is necessary to model the vehicle's motion behavior under control inputs to provide a foundation for subsequent prediction and optimization of control quantities. Therefore, a vehicle dynamics prediction model is first constructed based on the vehicle's real-time operating state and its own dynamic characteristic parameters.

[0062] Vehicle status information includes the vehicle's current position, attitude, speed, steering angle, and braking status. Dynamic characteristic parameters are used to describe the motion constraints of the vehicle during steering, acceleration, and deceleration, so that the vehicle dynamics prediction model can reflect the vehicle's motion response characteristics under different control inputs.

[0063] The above modeling method enables the constructed vehicle dynamics prediction model to depict the intrinsic mapping relationship between vehicle control input and changes in vehicle motion state, and to predict the motion state of the vehicle at multiple subsequent moments given a control input sequence, thereby providing a predictive basis for subsequent trajectory tracking control.

[0064] Step S402: Based on the vehicle dynamics prediction model, predict the vehicle motion state in the prediction time domain, and take the parking trajectory as the reference trajectory. Under the conditions of satisfying vehicle dynamics constraints and actuator constraints, solve the corresponding target control quantity by rolling through the model predictive control algorithm.

[0065] After constructing the vehicle dynamics prediction model, the trajectory tracking and control phase begins. Within each control cycle, based on the current vehicle state information, the vehicle dynamics prediction model is used to predict the future motion state of the vehicle within the prediction time domain. The prediction time domain is a preset time interval extending into the future from the current control moment, used to describe the predicted motion process of the vehicle within this time interval.

[0066] Using the parking trajectory as a reference trajectory, and based on the vehicle dynamics prediction model, this method comprehensively considers the vehicle's dynamic constraints as well as actuator constraints such as steering and braking. A model predictive control (MPC) algorithm is used to perform rolling optimization to solve for the target control quantity. Dynamic constraints include the vehicle's maximum steering angle, minimum turning radius, and acceleration / deceleration capabilities, while actuator constraints include limits on the range and rate of change of control inputs such as steering, driving, and braking.

[0067] Among them, the model predictive control algorithm constructs an optimization problem with trajectory tracking error and control cost as objectives within a finite prediction time domain, and executes only the optimal control quantity corresponding to the current moment in each control cycle, and then re-predicts and optimizes based on the latest vehicle state, thereby achieving rolling solution.

[0068] During the rolling solution process, the predicted vehicle motion state in the time domain is used to evaluate the impact of different candidate control input sequences on the vehicle motion trajectory in the future period of time, and serves as the basis for optimizing the objective function and constraint conditions. By updating the current state of the vehicle and re-predicting the future motion state in each control cycle, the corresponding target control quantity is continuously calculated, so that the vehicle gradually reduces the deviation from the reference parking trajectory in the prediction time domain, thereby achieving accurate tracking of the parking trajectory.

[0069] The target control quantity obtained is the set of vehicle control commands solved in the current control cycle, which is used to characterize the vehicle's control requirements in steering, driving and braking, etc., and may include steering angle commands, acceleration and deceleration commands or equivalent control quantities.

[0070] Step S403: Dynamically update vehicle control operations based on the target control quantity to achieve accurate tracking of the parking trajectory.

[0071] After obtaining the target control quantity, the target control quantity is converted into a low-level control command that the vehicle can execute, and sent to the vehicle's steering actuator, drive actuator and braking actuator to drive the vehicle to move in accordance with the target control quantity.

[0072] Specifically, by issuing steering angle commands, drive torque commands, or braking force commands, the vehicle executes the optimal control strategy calculated by the model predictive control algorithm within the current control cycle, thereby gradually bringing the vehicle's actual motion state closer to the parking trajectory.

[0073] In the next control cycle, the real-time status information of the vehicle is acquired again, and the prediction and rolling optimization solution process is executed again based on the updated vehicle status. By continuously looping the above prediction, optimization and execution closed-loop control process, the vehicle can stably and continuously track the parking trajectory throughout the parking process.

[0074] Through the above methods, the vehicle can effectively suppress the accumulation of deviations and control oscillations during the trajectory tracking process, while meeting the dynamic constraints and actuator constraints, ensuring that the vehicle travels smoothly along the parking trajectory, and ultimately achieving safe and efficient automatic parking operation.

[0075] It should be noted that the parking trajectory generation stage primarily relies on vehicle kinematic constraints to plan a reference trajectory, which is used to determine the vehicle's geometric driving path. The precise execution of the parking trajectory, however, is accomplished by a subsequent model predictive control algorithm based on a vehicle dynamics prediction model, ensuring the vehicle's stability and executability during actual driving.

[0076] In one embodiment, Figure 5 To illustrate the optimization process of a reinforcement learning decision model, the following steps may be included: Step S501: During the vehicle parking process, according to the parking decision in the current parking environment, and based on the parking decision execution result, obtain the corresponding new input state and reward signal to form an interaction sample.

[0077] During the parking process, the parking operation is executed according to the parking decision output by the reinforcement learning decision model in the current parking environment. The parking decision may include operation commands such as moving forward, reversing, and turning.

[0078] After making a parking decision, new vehicle state information is acquired, including the vehicle's current position, attitude, speed, and steering status. This new input state reflects the vehicle's real-time operating state after the parking decision and its relationship with the parking objective, serving as a crucial basis for the state transitions in the subsequent reinforcement learning decision model.

[0079] Based on this, the effectiveness of the current parking decision is comprehensively evaluated by combining factors such as safety, trajectory deviation, parking efficiency, and the degree of matching with the user's personalized preferences during the parking process, generating a corresponding reward signal. The magnitude of the reward signal is used to characterize the comprehensive contribution of the parking decision to safety, smoothness, efficiency, and the degree of personalized matching in the current parking environment.

[0080] By integrating the input state, the parking decision made, the generated reward signal, and the new input state, a complete interaction sample can be formed. This interaction sample is used to characterize the entire process of a vehicle transitioning from its original state to a new state under a single parking decision, providing fundamental data for policy optimization and parameter updates in reinforcement learning decision-making models.

[0081] Step S502: Store the interaction samples in the experience replay pool, and randomly select samples from the experience replay pool to optimize the training of the reinforcement learning decision model, so as to reduce the impact of sample correlation on model convergence.

[0082] Interaction samples are stored in an experience replay pool, which is used to cache historical interaction samples continuously generated by the vehicle during parking, so as to achieve unified management and reuse of different parking scenarios and different decision results.

[0083] Interaction samples were randomly selected from the experience replay pool to optimize the training of the reinforcement learning decision model. By randomly selecting interaction samples corresponding to different times and parking stages for training, the model avoids relying solely on local samples during continuous parking, thus effectively reducing the impact of sample temporal correlation on the model training process.

[0084] In this way, the reinforcement learning decision-making model can continuously use historical interaction samples to iterate and update during actual parking, gradually correcting the action selection tendency of the parking decision strategy under different states, and improving the model's stability, generalization ability and personalized adaptation ability in complex parking environments.

[0085] In one embodiment, during model training, a loss function is constructed to measure the difference between the Q-values ​​of the reinforcement learning decision model outputting each candidate parking decision under the current input state and the target Q-values ​​output by the target network under the same input state.

[0086] The Q-value is calculated by the reinforcement learning decision model under a given input state. It represents the expected cumulative reward of each candidate parking decision in the current state and is used to select the action with the largest Q-value as the current parking decision.

[0087] In practical applications, the mean squared error loss function can be used to measure the difference between the Q-value and the target Q-value to construct the loss function. By performing backpropagation on the loss function, the model parameters of the reinforcement learning decision model can be updated, allowing the model to gradually approximate the target Q-value function, thereby optimizing the parking decision strategy.

[0088] The target network and the reinforcement learning decision model use the same network structure. The model parameters are obtained synchronously by the reinforcement learning decision model through periodic replication or updating, so as to provide a relatively smooth and stable reference for Q-value updates during training, reduce training oscillations caused by target value fluctuations, and improve model convergence speed and training stability.

[0089] During the optimization of model parameters, an adaptive learning rate strategy can be used to dynamically adjust the step size of model parameter updates: a larger learning rate is used in the early stage of training to speed up the convergence speed, and a smaller learning rate is used in the later stage of training to improve the convergence accuracy of parameters; at the same time, a gradient clipping strategy can be introduced to limit the gradient magnitude in backpropagation to prevent gradient explosion or gradient vanishing problems.

[0090] Through the above optimizations, the reinforcement learning decision model can stably learn the mapping relationship between the input state and the parking decision within a limited training time, forming a robust parking strategy applicable to various parking scenarios. This improves the reliability of decision-making, the consistency of execution, and the ability to adapt to users' personalized preferences during the automatic parking process.

[0091] In one application embodiment, a schematic diagram of the execution flow for personalized adaptation of automatic parking is provided. For example... Figure 6 As shown, the entire automatic parking process begins with the perception and input phase of multi-source information. Sensors such as cameras, LiDAR, and ultrasonic radar on the vehicle acquire real-time information about the surrounding environment, including parking space boundaries, obstacle positions, and dynamic environmental elements. Simultaneously, the CAN bus collects real-time data on the vehicle's current position, attitude, speed, and steering angle.

[0092] In addition, user-personalized preference data has been introduced. This data includes driving habits (such as acceleration and deceleration preferences and steering timing) recorded and analyzed over a long period of time through the driver assistance system, as well as subjective needs set by users through the in-vehicle terminal (such as desired parking speed), providing data support and optimization basis for subsequent personalized parking decisions.

[0093] After acquiring multi-source input information, the core stage of path planning begins. Based on the parking decision output by the reinforcement learning decision model, and combined with vehicle dynamics constraints and user preferences, a sampling search and quadratic programming optimization strategy is employed to generate a personalized parking trajectory. First, a fast exploration random tree algorithm is used to quickly sample and expand the current parking environment, rapidly searching for a feasible initial path that avoids all known obstacles in complex parking conditions. Subsequently, to eliminate potential curvature discontinuities or unstable driving issues in the initial path, a quadratic programming algorithm is invoked to refine and optimize the path. During the optimization process, dynamic constraints such as the vehicle's maximum steering angle, minimum turning radius, and acceleration / deceleration capabilities, as well as the user's personalized preferences for parking speed and parking accuracy, are considered to smoothly adjust key path points, thereby generating a personalized parking trajectory that conforms to both vehicle physical constraints and the user's individual style.

[0094] After trajectory generation, the execution control phase begins. The generated parking trajectory is tracked in real time using the MPC algorithm. Based on the vehicle dynamics prediction model, the MPC controller predicts the future vehicle state in real time and, combined with the current pose feedback, calculates the optimal control increments such as steering wheel angle, throttle opening, and braking force through a rolling optimization method.

[0095] In actual control, it has dynamic adjustment capabilities, enabling it to correct strategies in real time according to environmental changes. For example, when sensors detect sudden obstacles or severe road bumps, the vehicle will automatically reduce its speed or make minor steering adjustments, and work with underlying safety modules such as ESP and AEB to ensure the robustness and safety of the parking process.

[0096] Ultimately, guided by the calculated optimal control values, the vehicle's actuators complete a series of continuous and precise actions, enabling the vehicle to smoothly and steadily park in the target parking space, achieving safe, efficient, and user-customized automatic parking operations.

[0097] In the above embodiments, user historical parking operation data and vehicle status information can be fused from multiple sources to generate comprehensive features reflecting user parking preferences and the vehicle's current parking capabilities. These features are then combined with environmental perception information input into a reinforcement learning decision model, enabling the vehicle to output parking decisions highly aligned with user preferences in different parking environments. Based on the parking decisions, a parking trajectory matching the decisions is generated through a path planning algorithm, and real-time tracking is performed using a predictive control algorithm. This allows for precise control of the vehicle's motion state, ensuring the safety, smoothness, and efficiency of the parking process. Using this method, the vehicle, when performing automatic parking operations, not only adheres to dynamic constraints and actuator constraints but also dynamically adapts to different parking environments and user preferences, achieving personalized, stable, and efficient automatic parking operations, thereby significantly improving user experience and parking safety.

[0098] In one embodiment, an automatic parking personalization adaptation device is provided, with reference to Figure 7 As shown, the automatic parking personalization adaptation device 700 may include: a feature fusion construction module 701, a parking decision output module 702, a parking trajectory planning module 703, and a parking execution module 704. Among them, The feature fusion construction module 701 is used to acquire multi-source data containing user historical parking operation data and vehicle status information, and to construct fusion features based on the multi-source data to characterize user parking preferences and vehicle current parking capabilities. The parking decision output module 702 is used to input the fused features and environmental perception information as input states into a preset reinforcement learning decision model, and output a parking decision that matches the user's parking preferences based on the current parking environment. The parking trajectory planning module 703 is used to generate a parking trajectory that matches the parking decision based on the parking decision and combined with vehicle kinematic constraints through a path planning algorithm. The parking execution module 704 is used to track the parking trajectory in real time using a predictive control algorithm and generate vehicle control operations to achieve automatic parking of the vehicle.

[0099] In one embodiment, the feature fusion construction module 701 is specifically used for: Feature extraction is performed on users' historical parking operation data to form user preference features that characterize users' parking styles and operating habits; Features are extracted from vehicle status information to form vehicle capability features that characterize the vehicle's current motion capabilities and response characteristics; User preference features and vehicle capability features are weighted and fused to form fused features for reinforcement learning decision-making.

[0100] In one embodiment, the parking trajectory planning module 703 is specifically used for: Extract visual spatial features from environmental perception data and use a multilayer perceptron to process and fuse these features; The visual spatial features are concatenated with the processed fused features and then input into a reinforcement learning decision model for action value function evaluation to output a parking decision.

[0101] In one embodiment, the parking decision output module 702 is specifically used for: An initial parking trajectory is generated by performing an initial path search in the current parking environment using a path search algorithm. Using parking decisions as a reference constraint, the initial parking trajectory is optimized and smoothed based on a quadratic programming algorithm to generate a parking trajectory that satisfies the constraints of curvature continuity and executability.

[0102] In one embodiment, the parking execution module 704 is specifically used for: Based on the current vehicle status information and the vehicle's dynamic characteristic parameters, a vehicle dynamics prediction model is constructed. Based on the vehicle dynamics prediction model, the vehicle motion state in the prediction time domain is predicted, and the parking trajectory is used as the reference trajectory. Under the conditions of satisfying vehicle dynamics constraints and actuator constraints, the corresponding target control quantity is solved by the model predictive control algorithm in a rolling manner. Based on the target control quantity, the vehicle control operation is dynamically updated to achieve accurate tracking of the parking trajectory.

[0103] In one embodiment, the parking decision output module 702 is further configured to: During the vehicle parking process, the parking decision is made according to the current parking environment, and the corresponding new input state and reward signal are obtained based on the parking decision execution result to form an interaction sample; Interaction samples are stored in the experience replay pool, and samples are randomly drawn from the experience replay pool to optimize the training of the reinforcement learning decision model, so as to reduce the impact of sample correlation on model convergence.

[0104] In one embodiment, the parking decision output module 702 is further configured to: Based on the Q-value corresponding to the parking decision output by the reinforcement learning decision model, and the target Q-value output by the target network under the same input state, a loss function is constructed and the model parameters of the reinforcement learning decision model are updated through backpropagation; the target network has the same structure as the reinforcement learning decision model, and the model parameters of the reinforcement learning decision model are synchronized through periodic updates.

[0105] Specific limitations regarding the automatic parking personalization adaptation device 700 can be found in the limitations of the automatic parking personalization adaptation method described above, and will not be repeated here. Each module in the aforementioned automatic parking personalization adaptation device 700 can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0106] In one embodiment, a vehicle is provided, including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement an automatic parking personalization adaptation method.

[0107] In one embodiment, a computer storage medium is provided on which a computer program is stored, which, when executed by a processor, implements an automatic parking personalization adaptation method.

[0108] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0109] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0110] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0111] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0112] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. An automatic parking individualization adaptation method, characterized in that, include: Acquire multi-source data containing user historical parking operation data and vehicle status information, and construct a fusion feature based on the multi-source data to characterize user parking preferences and vehicle current parking capabilities; The fused features and environmental perception information are used together as input to a preset reinforcement learning decision model, which outputs a parking decision that matches the user's parking preferences based on the current parking environment. Based on the parking decision and combined with vehicle kinematic constraints, a parking trajectory matching the parking decision is generated through a path planning algorithm. The parking trajectory is tracked in real time using a predictive control algorithm to generate vehicle control operations, thereby enabling automatic parking of the vehicle.

2. The automatic parking personalized adaptation method according to claim 1, wherein constructing a fusion feature representing the user's parking preferences and the vehicle's current parking capability based on the multi-source data includes: Feature extraction is performed on the user's historical parking operation data to form user preference features that characterize the user's parking style and operating habits; The vehicle state information is used to extract features to form vehicle capability features that characterize the vehicle's current motion capabilities and response characteristics; The user preference features and the vehicle capability features are weighted and fused to form a fused feature for reinforcement learning decision-making.

3. The automatic parking personalized adaptation method according to claim 1, wherein the step of inputting the fused features and environmental perception information together as model input states into a preset reinforcement learning decision model includes: Visual spatial features are extracted from environmental perception data, and the fused features are processed using a multilayer perceptron. The visual spatial features are concatenated with the processed fusion features and input into the reinforcement learning decision model for action value function evaluation to output the parking decision.

4. The automatic parking personalization adaptation method according to claim 1, characterized in that, The step of generating a parking trajectory that matches the parking decision based on the parking decision and in conjunction with vehicle kinematic constraints, using a path planning algorithm, includes: An initial parking trajectory is generated by performing an initial path search in the current parking environment using a path search algorithm. Using the parking decision as a reference constraint, the initial parking trajectory is optimized and smoothed based on a quadratic programming algorithm to generate a parking trajectory that satisfies the constraints of curvature continuity and executability.

5. The automatic parking personalization adaptation method according to claim 1, characterized in that, The real-time tracking of the parking trajectory using a predictive control algorithm includes: Based on the current vehicle status information and the vehicle's dynamic characteristic parameters, a vehicle dynamics prediction model is constructed. Based on the vehicle dynamics prediction model, the vehicle motion state in the prediction time domain is predicted, and the parking trajectory is used as a reference trajectory. Under the conditions of satisfying vehicle dynamics constraints and actuator constraints, the corresponding target control quantity is solved by the model predictive control algorithm in a rolling manner. Based on the target control quantity, the vehicle control operation is dynamically updated to achieve accurate tracking of the parking trajectory.

6. The automatic parking personalization adaptation method according to claim 1, characterized in that, The method further includes: During the vehicle parking process, the parking decision is made according to the current parking environment, and the corresponding new input state and reward signal are obtained based on the execution result of the parking decision to form an interaction sample; The interaction samples are stored in the experience replay pool, and samples are randomly selected from the experience replay pool to optimize and train the reinforcement learning decision model, so as to reduce the impact of sample correlation on model convergence.

7. The automatic parking personalization adaptation method according to claim 1 or 6, characterized in that, The method further includes: Based on the Q-value corresponding to the parking decision output by the reinforcement learning decision model, and the target Q-value output by the target network under the same input state, a loss function is constructed and the model parameters of the reinforcement learning decision model are updated through backpropagation; the target network has the same structure as the reinforcement learning decision model and synchronizes the model parameters of the reinforcement learning decision model through periodic updates.

8. An automatic parking personalization adaptation device, characterized in that, include: The feature fusion construction module is used to acquire multi-source data containing user historical parking operation data and vehicle status information, and to construct fusion features based on the multi-source data to characterize user parking preferences and vehicle current parking capabilities. The parking decision output module is used to input the fused features and environmental perception information as input states into a preset reinforcement learning decision model, and output a parking decision that matches the user's parking preferences based on the current parking environment. The parking trajectory planning module is used to generate a parking trajectory that matches the parking decision based on the parking decision and in combination with vehicle kinematic constraints, through a path planning algorithm. The parking execution module is used to track the parking trajectory in real time using predictive control algorithms and generate vehicle control operations to achieve automatic parking of the vehicle.

9. A vehicle comprising a memory and a processor, said memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the automatic parking personalization adaptation method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the automatic parking personalization adaptation method according to any one of claims 1 to 7.

Images