Risk avoidance control method and system and vehicle

Abstract

The invention provides a risk avoidance control method and system and a vehicle, and relates to the technical field of auxiliary driving, and the method comprises the steps: responding to a driving and parking mode switching request of the vehicle, and judging whether the vehicle has a collision risk through a risk avoidance model; when the collision risk exists, controlling the vehicle to perform risk avoidance through the risk avoidance model; after avoiding is finished, whether the vehicle has the collision risk or not is judged again through the risk avoiding model; and when the collision risk does not exist, controlling the vehicle to carry out mode switching so as to enhance the control continuity when the auxiliary driving mode switching is carried out.

Description

Technical Field

[0001] This application relates to the field of driver assistance technology, and more specifically, to a risk avoidance and control method, system, and vehicle. Background Technology

[0002] Currently, driver assistance systems are developing rapidly and are gradually being applied to mass-produced vehicles. After the driver selects their destination and activates the function, driving assistance is provided throughout the journey. However, switching assistance modes is often necessary when traveling from public roads to the destination. There may be a certain risk of collision during the mode switching process.

[0003] In related technologies, when a vehicle detects a collision risk, it immediately applies emergency braking and disengages the driver assistance function, resulting in a poor driving experience and directly affecting the system's control effectiveness, safety, and usability. Summary of the Invention

[0004] The problem addressed in this application is how to enhance control continuity when switching between driver assistance modes.

[0005] To address the aforementioned issues, this application provides a risk avoidance and control method, system, and vehicle.

[0006] This application provides a risk avoidance and control method, including: In response to a vehicle's parking mode switching request, a risk avoidance model is used to determine whether the vehicle poses a collision risk. When the collision risk exists, the vehicle is controlled to avoid the risk through the risk avoidance model. After the avoidance is completed, the risk avoidance model is used again to determine whether the vehicle still has the risk of collision. When there is no collision risk, control the vehicle to switch modes.

[0007] Optionally, controlling the vehicle to avoid risks through the risk avoidance model includes: The current state vector of the vehicle is obtained, and the current state vector is input into the risk avoidance model to obtain the risk avoidance action of the vehicle. The vehicle is then controlled to execute the risk avoidance action in order to avoid risks.

[0008] Optionally, the risk avoidance and control method further includes: A reinforcement learning model is constructed, wherein a deep Q-network is used as the agent of the reinforcement learning model, the state space of the reinforcement learning model is constructed based on the state vector, the state vector is composed of at least one of driving data, environmental state and displacement data, the action space of the reinforcement learning model is constructed based on the action vector, the action vector is composed of the change of the vehicle's system acceleration over time, and a multi-objective reward function is used as the reward function of the reinforcement learning model, the multi-objective reward function is obtained by weighted fusion of collision avoidance reward function, comfort reward function, behavior habit reward function and delay penalty function; The current state vector at the current training moment is input into the deep Q-network, which outputs the target action. Control the vehicle to perform the target action, obtain the state vector at the next training moment, and determine the reward value based on the multi-objective reward function; The Q-function in the deep Q-network is optimized based on the reward value. The state vector of the next training time is used as the state vector of the current training time. The step of inputting the state vector of the current training time into the deep Q-network is returned. The training is iterated until the deep Q-network meets the preset convergence condition, and the risk avoidance model is obtained.

[0009] Optionally, the risk aversion control method further includes a self-learning model, and the state vector further includes a state trigger threshold; the risk aversion control method further includes: The self-learning model determines the state trigger threshold based on the vehicle's driving data and the environmental state, and the state trigger threshold is used to participate in the optimization of the multi-objective reward function.

[0010] Optionally, the construction of the collision avoidance reward function includes: Based on the kinematic properties of the vehicle and the obstacle, the predicted relative distance between the vehicle and the obstacle after a preset future time is determined by a kinematic model, wherein the kinematic properties include at least one of position, velocity and acceleration; When the predicted relative distance is greater than or equal to the collision threshold, a reward value positively correlated with the predicted relative distance is generated through an exponential decay term, wherein the state trigger threshold includes the collision threshold; When the predicted relative distance is less than the collision threshold, a negative reward is applied to the predicted relative distance through a collision indicator.

[0011] Optionally, the construction of the comfort reward function includes: The change in acceleration is determined based on the vehicle's acceleration at the current moment and the previous moment; When the absolute value of the acceleration change increases, a negative reward that is positively correlated with the absolute value of the acceleration change is applied through a smoothness penalty term; When the magnitude of the acceleration exceeds the preset acceleration range, a negative reward positively correlated with the degree of exceeding the limit is applied through a reasonableness penalty term; When the amplitude of the acceleration does not exceed the preset acceleration range, a positive reward negatively correlated with the amplitude is applied through the rationality penalty item, wherein the state trigger threshold includes the preset acceleration range.

[0012] Optionally, the construction of the behavioral habit reward function includes: Based on the scene features of the current driving scenario and the scene records in the historical driving database, the target historical scene with the highest similarity to the current driving scenario is determined. A reward value negatively correlated with the acceleration difference value is applied through a behavioral consistency term, wherein the acceleration difference value is determined by the absolute value of the difference between the system acceleration currently selected by the reinforcement learning model and the historical acceleration recorded in the target historical scene.

[0013] Optionally, the construction of the delayed punishment function includes: Based on the risk assessment value of the current driving state and the acceleration of the vehicle, the state of the vehicle is determined by a preset risk threshold; When the vehicle is in a high-risk state and the magnitude of the acceleration is less than the braking response threshold, a negative reward positively correlated with the risk assessment value is applied through the insufficient response term; When the vehicle is in a low-risk state and the magnitude of the acceleration is greater than the aggressive operation threshold, a negative reward negatively correlated with the risk assessment value is applied through a misoperation item, wherein the state trigger threshold includes the braking response threshold and the aggressive operation threshold, and the braking response threshold is less than the aggressive operation threshold.

[0014] Secondly, this application also provides a risk avoidance control system, including: The instruction activation module is used to respond to the vehicle's parking mode switching request and determine whether the vehicle has a collision risk through a risk avoidance model; The risk assessment module is used to control the vehicle to avoid the risk through the risk avoidance model when the collision risk exists. The risk avoidance module is used to determine whether the vehicle still has the collision risk after the avoidance is completed, by using the risk avoidance model again. The mode switching module is used to control the vehicle to switch modes when there is no collision risk.

[0015] Thirdly, this application also provides a vehicle, including a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the risk avoidance and control method described above when executing the computer program.

[0016] The beneficial effects of the risk avoidance and control method in this application are: Before switching modes, a collision risk assessment is performed to avoid blindly executing state transitions in scenarios with potential collision risks, and to prevent function interruption or emergency exit due to unsafe environmental conditions. When a collision risk exists, a risk avoidance model is used to generate matching dynamic avoidance actions, replacing the traditional hard strategy of detection and exit, allowing the vehicle to maintain the continuity of the assisted driving function without interrupting its operation. After the avoidance is completed, a second assessment is performed to determine whether the risk has truly been eliminated, preventing misjudgments due to insufficient judgment in a single instance or dynamic changes in the environment; if the risk still exists, a new round of avoidance can be triggered to avoid safety hazards caused by hasty switching, or to interrupt the assisted driving function to ensure safety; the parking mode switch is only executed when the collision risk is no longer present, ensuring that the switching action is based on safety, avoiding control conflicts or functional failures caused by forced switching, and enhancing the control continuity when switching assisted driving modes. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the risk avoidance and control method according to an embodiment of this application; Figure 2 This is a flowchart illustrating the risk avoidance and control method of this application after responding to a berthing mode switching request; Figure 3 This is a model architecture diagram of the risk avoidance and control method according to an embodiment of this application; Figure 4 This is an example diagram of a vehicle according to an embodiment of this application. Detailed Implementation

[0018] To make the above-mentioned objects, features, and advantages of this application more apparent and understandable, specific embodiments of this application are described in detail below with reference to the accompanying drawings. Although some embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this application. It should be understood that the accompanying drawings and embodiments of this application are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0019] It should be understood that the steps described in the method embodiments of this application may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of this application is not limited in this respect.

[0020] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to"; the term "based on" means "at least partially based on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; and the term "optionally" means "optional embodiments". Definitions of other terms will be given in the description below. It should be noted that the concepts of "first," "second," etc., mentioned in this application are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0021] It should be noted that the terms "one" and "more" used in this application are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0022] The names of messages or information exchanged between multiple devices in the embodiments of this application are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0023] like Figure 1 As shown in the embodiment of this application, a risk avoidance and control method includes: Step S100: In response to the vehicle's parking mode switching request, determine whether the vehicle has a collision risk through a risk avoidance model.

[0024] In the embodiments of this application, the driving mode switching request refers to the need to switch from driving state to parking state, or vice versa, triggered by the driver or automatically identified by the driver assistance system.

[0025] The risk avoidance model is used to assess the risks around the vehicle when a switching request is made. In subsequent steps, avoidance strategies can be determined based on the risk level. By using the risk avoidance model to assess risks, a response mechanism can be determined based on the overall risk during the switching process, rather than stopping or directly exiting the assisted driving mode, thus ensuring the continuity of the driving mode switching.

[0026] Collision risk refers to the potential for a vehicle to collide with surrounding road users or static obstacles during a transition from one driving mode to another, due to changes in path, speed, or control logic switching. It is based on a comprehensive assessment of the current environmental state, obstacle dynamics, and predicted trajectory using a risk avoidance model.

[0027] Step S200: When the collision risk exists, the vehicle is controlled to avoid the risk through the risk avoidance model.

[0028] In embodiments of the present invention, the risk avoidance model is trained using reinforcement learning. This risk avoidance method represents a learning framework based on a trial-and-error mechanism. By continuously interacting with the driving environment, it selects control actions such as acceleration under different states, thereby continuously optimizing long-term decision-making strategies. By introducing a risk avoidance model trained using reinforcement learning, it can autonomously adjust its response intensity according to the current situation, effectively handling complex scenarios with ambiguous boundaries or multiple interferences.

[0029] Step S300: After the avoidance is completed, the risk avoidance model is used again to determine whether the vehicle still has the collision risk.

[0030] The end of avoidance means that the vehicle has completed a risk avoidance control action generated by the risk avoidance model, such as deceleration or collision avoidance, and the vehicle's motion state tends to be stable.

[0031] After the avoidance action is completed, the risk avoidance model is called again for judgment to achieve closed-loop monitoring of the driving environment.

[0032] Initial evasive maneuvers may only mitigate part of the risk. The behavior of surrounding road users and the traffic environment in which the vehicle is located are uncertain, necessitating a new round of assessment to confirm the safety of the current traffic environment. If a collision risk still exists, a new evasive strategy can be triggered to prevent subsequent actions from becoming dangerous due to insufficient judgment in a single instance. Through multiple risk reassessments, the user experience is prevented from being interrupted by a one-time disengagement of the assisted driving function, ensuring a smooth parking process.

[0033] Step S400: When there is no collision risk, control the vehicle to switch modes.

[0034] Once a collision risk is determined to be nonexistent, the vehicle switches modes automatically, ensuring seamless functionality while maintaining safety. Driver intervention is unnecessary; the vehicle autonomously switches modes based on real-time environmental conditions, enhancing overall usability.

[0035] In one embodiment, if a collision risk still exists when re-switching modes, the vehicle is brought to a stop, an HMI prompt is generated, mode switching is stopped, and the end-to-end mode is exited, returning vehicle control to the driver. This is used to proactively terminate automated tasks in persistent high-risk scenarios, preventing the assisted driving function from operating under uncontrollable conditions and ensuring basic driving safety.

[0036] In one embodiment, in response to an end-to-end activation command triggered by the driver, an initial driving mode is determined based on the starting location type and vehicle control commands. The starting location type includes parking spaces or road areas, and the driving mode correspondingly includes parking mode or driving mode. By identifying the attributes of the task's starting point, the system automatically selects the matching control logic, avoiding the need for the driver to manually switch functions in the vehicle's infotainment system, thus reducing operational burden and the risk of accidental touches.

[0037] like Figure 2 and Figure 3 As shown, the driving mode is switched when the vehicle meets the mode switching conditions. Mode switching conditions include the vehicle entering a drivable road after completing a parking maneuver, or the vehicle approaching the target parking space and meeting the parking preparation state. The construction of mode switching conditions is used to achieve seamless connection between driving and parking scenarios, enabling end-to-end function usage without manual intervention in mode selection, improving usability. For example, after the vehicle is powered on, the driver activates the end-to-end function and sets the driving destination. If both the starting point and destination are parking spaces, the system enters parking mode and controls the vehicle to park out of the parking space. The mode switching condition at this time is the completion of parking. When the mode switching condition is met, the driver assistance system needs to switch to driving mode and automatically plan the optimal navigation route. At this time, the risk avoidance model is used to avoid collision risks. If the starting point is not a parking space and the destination is a parking space, the system enters driving mode, automatically plans the optimal navigation route, and then controls the vehicle to drive to the destination parking space. The mode switching condition at this time is approaching the target parking space. When the mode switching condition is met, the risk avoidance model is used to avoid collision risks, ensuring that there is no collision risk when the system switches modes.

[0038] When the vehicle meets the mode switching conditions, the risk avoidance model determines the target action the vehicle should take. At this time, the risk avoidance model combines the current environmental state, obstacle dynamics, and the risk characteristics unique to mode switching to output an acceleration command that balances safety and smoothness.

[0039] Optionally, the mode switching conditions include the first mode switching condition and the second mode switching condition; When the starting location type is non-road type and the vehicle control command is parking (including parking in and parking out), the vehicle is controlled to enter parking mode until the vehicle meets the first mode switching condition, and then the vehicle is controlled to enter driving mode. When the starting location type is road type and the vehicle control command is driving, the vehicle is controlled to enter the driving mode until the vehicle meets the second mode switching conditions, and then the vehicle is controlled to enter the parking mode.

[0040] In one embodiment, such as Figure 2 As shown, when the starting location type is non-road type, such as within a parking space, and the vehicle control command is to park the vehicle, the vehicle is controlled to enter parking mode and execute the parking action. After the vehicle meets the first mode switching condition, the vehicle is controlled to enter driving mode and begin driving along the navigation path. This allows the vehicle to autonomously recognize the starting point semantics and seamlessly transition to the driving task after completing the initial parking, avoiding the need for the driver to manually activate the driving function after parking. The first mode switching condition includes the vehicle completing the parking maneuver in parking mode.

[0041] When the starting point type is road and the vehicle control command is driving, the vehicle enters driving mode to perform path tracking and traffic interaction. After the vehicle meets the second mode switching conditions, it enters parking mode and initiates the parking process. This automatically prepares for parking before arriving at the destination, eliminating the need for the driver to perform additional function switching when approaching the parking space. The second mode switching conditions include completing road driving actions in driving mode.

[0042] The logic connecting two typical scenarios—driving and parking—is integrated into the mode switching conditions. By matching the starting location type with the corresponding switching conditions, the entire end-to-end task process is automated. This reduces manual intervention required for function switching and lowers the probability of misoperation. Because mode switching is triggered by preset conditions and the risk avoidance model guides the mode switching process, the entire process has good safety and behavioral consistency, taking into account the effectiveness of assisted driving control, safety, and usability.

[0043] Optionally, controlling the vehicle to avoid risks through the risk avoidance model includes: The current state vector of the vehicle is obtained, and the current state vector is input into the risk avoidance model to obtain the risk avoidance action of the vehicle. The vehicle is then controlled to execute the risk avoidance action in order to avoid risks.

[0044] In one embodiment, the state vector is composed of at least one of driving data, environmental state, state trigger threshold, and displacement data. Driving data includes the vehicle's position, speed, acceleration, etc., relative to the target; environmental state includes road conditions, weather, etc.; the state trigger threshold is derived from a self-learning model and reflects personalized driving preferences; displacement data represents the relative relationship between the vehicle and the target position or obstacle. The state vector provides comprehensive and dynamic data to the decision-making model, enabling it to select target actions by integrating the vehicle's own state, the external environment, and user habits.

[0045] The motion vector is the change in the vehicle's system acceleration over time. The configuration consists of a command combination that includes lateral and longitudinal acceleration, where 'a' represents acceleration, 'k' represents the vehicle's acceleration, and 't' represents time. , where x represents the vertical axis and y represents the horizontal axis.

[0046] After identifying a collision risk, a state vector is formed by acquiring multi-source information reflecting the current driving situation. This vector is used as input to the mode switching model, which infers and generates the optimal avoidance action. The vehicle is then driven to execute this action, thus completing a closed-loop risk response.

[0047] Alternatively, risk avoidance and control methods may also include: The reinforcement learning model is constructed in which a deep Q-network is used as the agent of the reinforcement learning model, the state space of the reinforcement learning model is constructed based on the state vector, the state vector is composed of at least one of driving data, environmental state and displacement data, the action space of the reinforcement learning model is constructed based on the action vector, the action vector is composed of the change of the vehicle's system acceleration over time, and a multi-objective reward function is used as the reward function of the reinforcement learning model. The current state vector at the current training moment is input into the deep Q-network, which outputs the target action. Control the vehicle to perform the target action, obtain the state vector at the next training moment, and determine the reward value based on the multi-objective reward function; The Q-function in the deep Q-network is optimized based on the reward value. The state vector of the next training time is used as the state vector of the current training time. The step of inputting the state vector of the current training time into the deep Q-network is returned. The training is iterated until the deep Q-network meets the preset convergence condition, and the trained reinforcement learning model, i.e., the risk avoidance model, is obtained.

[0048] Deep Q-networks take state vectors as input and action vectors as outputs, and iteratively optimize using reward signals provided by a multi-objective reward function. This multi-objective reward function integrates dimensions such as collision avoidance, comfort, behavioral habits, and delayed penalties to learn strategies that balance safety and driving quality.

[0049] During the training phase, after obtaining the state vector at the current training moment, it is input into the deep Q-network, which outputs an action vector that maximizes the expected cumulative reward in the future as the target action. The vehicle is controlled to execute this target action, and the state vector for the next training moment is collected. The reward value brought by this action is calculated according to the multi-objective reward function and used to update the Q-function parameters in the deep Q-network. The above steps are repeated, that is, the newly obtained state vector is continuously used as the state vector at the current training moment, until the deep Q-network meets the preset convergence condition.

[0050] During the runtime phase, the deep Q-network selects the future duration based on the current state vector. The action vector that maximizes the expected cumulative reward is output as the target action. This is used to transform complex driving tasks into Markov decision processes, automatically extracting the optimal control strategy from a large amount of interaction experience using deep reinforcement learning. A state trigger threshold is introduced into the state vector to enable the general strategy to be personalized and adaptable. This ensures that the target action meets both basic safety requirements and user operating habits, thus balancing the effectiveness of driver assistance control, safety, and usability.

[0051] Optionally, such as Figure 3 As shown, the multi-objective reward function is obtained by weighted fusion of the collision avoidance reward function, comfort reward function, behavioral habit reward function, and delayed punishment function.

[0052] The multi-objective reward function is expressed as: , in, These are weighting coefficients used to balance rewards from various aspects. This represents the real-time dynamic acceleration of the vehicle, where t represents time. Indicate scene characteristics, This represents the collision avoidance reward function. This represents the comfort reward function. This represents a reward function for behavioral habits. This represents a delayed penalty function.

[0053] Optionally, such as Figure 3 As shown, the risk aversion control method further includes a self-learning model, and the state vector further includes a state trigger threshold; the risk aversion control method also includes: The self-learning model determines the state trigger threshold based on the vehicle's driving data and the environmental state, and the state trigger threshold is used to participate in the optimization of the multi-objective reward function.

[0054] The state trigger threshold refers to a dynamic parameter used to define the conditions for initiating a risk response or to serve as the reward switching boundary for the reward function. Its value is not fixed but is adjusted according to the driver's habits. As a component of the state vector, the state trigger threshold participates in the state representation of the risk avoidance model, enabling the decision-making process to perceive individual driving style and environmental characteristics.

[0055] The self-learning model represents a system that uses historical vehicle driving data, such as pedal operation, steering wheel torque, and speed changes, as well as real-time environmental conditions, such as road type, weather, and traffic density, for online adjustments. It extracts the driver's behavioral tendencies in different scenarios and generates state trigger thresholds adapted to the current situation. For example, in congested city traffic, drivers tend to decelerate earlier, so the corresponding braking trigger threshold is increased; while in highway scenarios, the threshold is relaxed accordingly.

[0056] State trigger thresholds are used to optimize the multi-objective reward function, specifically affecting the calculation boundaries of sub-reward items such as collision avoidance, comfort, or behavioral habits. Examples include the collision threshold in collision avoidance rewards, the preset acceleration range in comfort rewards, and the braking response threshold and aggressive operation threshold in delayed penalty rewards. Therefore, the multi-objective reward function does not rely on general settings but possesses personalized and contextual adjustment capabilities, achieving a leap from general design to human-vehicle adaptation. While ensuring a safety baseline, it significantly improves the smoothness of human-machine co-driving and user acceptance.

[0057] Optionally, the construction of the collision avoidance reward function includes: Based on the kinematic properties of the vehicle and the obstacle, the predicted relative distance between the vehicle and the obstacle after a preset future time is determined by a kinematic model, wherein the kinematic properties include at least one of position, velocity and acceleration; When the predicted relative distance is greater than or equal to the collision threshold, a reward value positively correlated with the predicted relative distance is generated through an exponential decay term, wherein the state trigger threshold includes the collision threshold; When the predicted relative distance is less than the collision threshold, a negative reward is applied to the predicted relative distance through a collision indicator.

[0058] In one embodiment, the collision avoidance reward function is constructed by determining the predicted relative distance between the vehicle and the obstacle after a preset future time period based on the position, velocity, and acceleration of the vehicle and the obstacle using a kinematic model.

[0059] The predicted relative distance characterizes the spatial proximity of the vehicle and the obstacle at future moments under the current motion trend. This distance serves as a risk feedforward indicator, enabling the reinforcement learning model to predict collision avoidance capabilities, rather than reacting based on the current instantaneous state.

[0060] When the predicted relative distance is greater than or equal to the collision threshold, a reward value positively correlated with the predicted relative distance is generated through an exponential decay term. The collision threshold, a type of state trigger threshold, is used to define the boundary between safe and risky areas. This encourages the reinforcement learning model to actively maintain a large safety margin, and the reward growth slows down as distance increases, consistent with the generally low sensitivity of drivers to long-distance risks.

[0061] When the predicted relative distance is less than or equal to the collision threshold, a negative reward is applied to the predicted relative distance via a collision indicator. This is used to penalize the behavior of entering a high-risk area, prompting the reinforcement learning model to prioritize deceleration, lane changing, or other avoidance actions to avoid potential conflicts.

[0062] The collision avoidance reward function combines dynamic prediction with risk boundaries, making risk avoidance behavior predictable and continuous. At the same time, the collision threshold, as part of the state trigger threshold, can be dynamically adjusted according to driver habits or scenario characteristics, so that the safety strategy has personalized adaptability while maintaining the bottom line constraints, thereby taking into account the effect of assisted driving control, safety and usage consistency.

[0063] In one embodiment, the predicted relative distance can be expressed as: , in, This represents the position of the i-th obstacle at time t. Indicates the duration in the future, typically ranging from 0 to 3 seconds. This represents the velocity vector of the i-th obstacle at time t. Let represent the acceleration vector of the i-th obstacle at time t.

[0064] For vehicle k, its predicted future position is expressed as: , in, This indicates the position of vehicle k at time t. This represents the velocity vector of vehicle k. This represents the velocity vector of vehicle k at the initial moment. Let represent the acceleration vector of vehicle k at time t.

[0065] The collision reward function can then be constructed as follows: , in, This represents the collision reward function value, where n represents the number of obstacles. This represents the reward factor related to safe distance. This represents the distance decay factor, used to adjust the impact of distance on reward. Indicates the collision penalty factor. Indicates the future duration of vehicle k. The relative distance between the i-th obstacle and the i-th obstacle.

[0066] in, , , d represents the predicted relative distance. Represents the exponentially decaying term. Indicates the safe distance threshold. Indicates collision indication items, This represents the collision threshold. Therefore, when the predicted relative distance is less than the safe distance threshold but greater than the collision threshold, the reward applied is determined by the predicted relative distance d itself; when the predicted relative distance d is greater than the safe distance threshold... At that time, the applied reward coefficient is capped at the safe distance threshold. When the predicted relative distance is less than the collision threshold, the penalty coefficient is 1, which is the maximum penalty; when the predicted relative distance is greater than or equal to the collision threshold but less than the safe distance threshold, the penalty is applied. When the penalty coefficient is 0, it means that there is neither punishment nor encouragement.

[0067] In this embodiment, , and All are set to positive values.

[0068] Optionally, the construction of the comfort reward function includes: The change in acceleration is determined based on the vehicle's acceleration at the current moment and the previous moment; When the absolute value of the acceleration change increases, a negative reward that is positively correlated with the absolute value of the acceleration change is applied through a smoothness penalty term; When the magnitude of the acceleration exceeds the preset acceleration range, a negative reward positively correlated with the degree of exceeding the limit is applied through a reasonableness penalty term; When the amplitude of the acceleration does not exceed the preset acceleration range, a positive reward negatively correlated with the amplitude is applied through the rationality penalty item, wherein the state trigger threshold includes the preset acceleration range.

[0069] In one embodiment, the acceleration change is used to reflect the degree of abrupt change in the vehicle's longitudinal or lateral motion. It serves as a core indicator for comfort evaluation, measuring the smoothness of control commands. When the absolute value of the acceleration change increases, a negative reward positively correlated with the absolute value of the acceleration change is applied through a smoothness penalty term. This aims to suppress drastically changing acceleration commands from the control strategy, prompting a more gradual transition in action selection and improving ride comfort.

[0070] When the acceleration amplitude exceeds the preset acceleration range, a negative reward positively correlated with the degree of exceedance is applied through a reasonableness penalty term. The preset acceleration range is a type of state trigger threshold used to define a reasonable acceleration interval that conforms to normal driving habits. It is used to prevent the reinforcement learning model from selecting excessive acceleration or braking behavior, ensuring that the control output matches human expectations of normal driving dynamics.

[0071] By constructing a comfort reward function, the action space is guided from the two dimensions of smoothness and rationality, effectively reducing driving discomfort caused by abrupt control. The preset acceleration range is used as one of the state trigger thresholds, which can be dynamically adjusted according to the driver's operating characteristics. While ensuring basic comfort, it adapts to individual preferences, thereby taking into account the effect of assisted driving control, safety and usage consistency.

[0072] In one embodiment, the comfort reward function is expressed as: , in, This represents the value of the comfort reward function. Indicates the smoothness reward weight, This represents the smoothness penalty term. Indicates the weight of reasonable reward. Indicates a reasonable penalty.

[0073] The smoothness penalty term is expressed as: , in, The weighting coefficients for the smoothness penalty term are used to adjust the degree of emphasis placed on smoothness. This represents the change in longitudinal acceleration of vehicle k at time t. This represents the change in lateral acceleration of vehicle k at time t.

[0074] Taking x-axis acceleration as an example, the rationality penalty term is expressed as follows: , in, A coefficient representing the reasonableness of the acceleration reward, used to control the intensity of the reward. This represents the maximum and minimum values ​​of longitudinal acceleration. Typically, based on the vehicle's performance design, both positive and negative values ​​have upper limits.

[0075] The penalty for y-axis acceleration rationality is the same as that for x-axis acceleration rationality.

[0076] Optionally, the construction of the behavioral habit reward function includes: Based on the scene features of the current driving scenario and the scene records in the historical driving database, the target historical scene with the highest similarity to the current driving scenario is determined. A reward value negatively correlated with the acceleration difference value is applied through a behavioral consistency term, wherein the acceleration difference value is determined by the absolute value of the difference between the system acceleration currently selected by the reinforcement learning model and the historical acceleration recorded in the target historical scene.

[0077] In one embodiment, scene features are used to characterize the overall situation of the current traffic environment and vehicle status, including relative vehicle distance, vehicle speed, road type, and distribution of surrounding obstacles; the historical driving database stores past driver operation records in similar scenarios. By measuring the geometric proximity between scenes using Euclidean distance, the degree of matching between the current situation and historical experience can be quantified.

[0078] The acceleration difference value is composed of the absolute value of the difference between the system acceleration currently selected by the reinforcement learning model and the historical acceleration recorded in the most similar historical scenario. It is used to reflect the degree of deviation between the current action and the typical behavior of the driver in similar situations.

[0079] By applying a reward value negatively correlated with the acceleration difference value to the behavioral consistency term, the reinforcement learning model is encouraged to output control commands that closely resemble driver habits. This makes automated decision-making more predictable and reduces driver distrust or frequent takeovers caused by overly mechanical or conservative strategies.

[0080] In one embodiment, it is assumed that the scene feature is a vector. If the similarity to historical scene i is measured by calculating the Euclidean distance, then the similarity quantification process can be expressed as: , , in, This represents the Euclidean distance between the current scene s(t) and the i-th historical scene. This indicates the current scene characteristics of the vehicle. Represents the scene features of the i-th record in the historical driving database, ∥ ∥ represents the Euclidean norm. argmin represents the most similar historical scene to the current scene s(t), i.e., the target historical scene. The input variable i when the minimum value is obtained.

[0081] The habit reward function is expressed as: , , , , in, This represents the habitual reward function value. The weighting factor represents the similarity between the current scene and historical scenes, and exp represents the exponential function operation. This represents a coefficient indicating the reward for behavioral habits, controlling for the impact of acceleration differences on the reward. This represents the difference between the current longitudinal acceleration and the longitudinal acceleration in the i-th historical scene. This represents the difference between the current lateral acceleration and the lateral acceleration in the i-th historical scene. These represent the current longitudinal acceleration and lateral acceleration, respectively. In historical context The longitudinal acceleration selected by the driver In historical context The lateral acceleration selected by the driver.

[0082] Optionally, the construction of the delayed punishment function includes: Based on the risk assessment value of the current driving state and the acceleration of the vehicle, the state of the vehicle is determined by a preset risk threshold; When the vehicle is in a high-risk state and the magnitude of the acceleration is less than the braking response threshold, a negative reward positively correlated with the risk assessment value is applied through the insufficient response term; When the vehicle is in a low-risk state and the magnitude of the acceleration is greater than the aggressive operation threshold, a negative reward negatively correlated with the risk assessment value is applied through a misoperation item, wherein the state trigger threshold includes the braking response threshold and the aggressive operation threshold, and the braking response threshold is less than the aggressive operation threshold.

[0083] In one embodiment, the risk assessment value reflects the potential danger level under the current traffic environment and vehicle dynamics, and is generated by a reinforcement learning model based on factors such as obstacle distance, relative speed, and road structure. A preset risk threshold is used to delineate the boundary between high-risk and low-risk states.

[0084] When a vehicle is in a high-risk state and the magnitude of its acceleration is less than the braking response threshold, a negative reward positively correlated with the risk assessment value is applied through an insufficient response term. The braking response threshold, a type of state trigger threshold, represents the minimum braking intensity that should be achieved in an emergency situation. It is used to penalize slow reactions or insufficient braking force in dangerous scenarios, prompting the reinforcement learning model to take timely and effective avoidance actions.

[0085] When the vehicle is in a low-risk state and the magnitude of acceleration exceeds the aggressive operation threshold, a negative reward negatively correlated with the risk assessment value is applied through a misoperation item. The aggressive operation threshold, a type of state trigger threshold, is used to define a reasonable upper limit for acceleration during normal driving. This is used to suppress unnecessary aggressive operations in safe scenarios, preventing excessive intervention from affecting passenger comfort or causing misjudgments by other road users.

[0086] The delayed penalty function couples the evaluation of risk level with action range, enabling the reinforcement learning model to avoid negative responses under high risk and prevent overreactions under low risk. The braking response threshold and aggressive operation threshold are dynamically adjusted according to driving habits or environmental characteristics, improving behavioral adaptability while maintaining decision rationality, thereby taking into account the assisted driving control effect, safety and usage consistency.

[0087] The delayed penalty function is expressed as: , in, This represents the value of the delayed penalty function. This represents the vehicle's risk assessment value at the current moment; a higher value indicates a higher risk. high This represents the aggressive operation threshold, used to define the upper limit of acceleration that should not be exceeded in safe scenarios. low The braking response threshold is used to define the minimum braking intensity that should be achieved in high-risk scenarios. low Risk represents the upper bound threshold of a low-risk state. high λ1 and λ2 represent the lower bound threshold for high-risk states, λ3 and λ4 represent the penalty intensity for excessive acceleration in low-risk scenarios, and λ4 represent the penalty intensity for insufficient acceleration in high-risk scenarios.

[0088] Optionally, controlling the vehicle to avoid the risk through the risk avoidance model when the collision risk exists includes: When a high collision risk is determined, the risk avoidance model outputs an emergency braking action, and the vehicle is controlled to perform the emergency braking action. When a low collision risk is determined, the risk avoidance model outputs a comfort braking action, and the vehicle is controlled to perform the comfort braking action.

[0089] High collision risk refers to a situation where the relative distance and relative speed between the vehicle and surrounding obstacles have entered a critical range, and contact may occur within a short time if immediate intervention is not taken. The emergency braking action output by the risk avoidance model is the maximum feasible deceleration command, used to reduce the vehicle speed within the shortest distance to avoid a collision.

[0090] Low collision risk means there is a potential conflict, but there is sufficient time and space for a smooth intervention. The comfort braking action output by the risk avoidance model is a moderate deceleration command with a small rate of acceleration change to avoid abruptness and maintain driving comfort.

[0091] When it is determined that there is no risk of collision, only a warning is given to remind the driver to be more vigilant.

[0092] Optionally, the risk avoidance model includes a collision time calculation model; The method of determining whether the vehicle poses a collision risk using a risk aversion model includes: The predicted collision time between the vehicle and the obstacle is determined using the collision time calculation model. When the predicted collision time is less than the first preset collision time, it is determined that there is a high risk of collision. When the predicted collision time is greater than or equal to the first preset collision time and less than or equal to the second preset collision time, it is determined that there is a low collision risk. If the predicted collision time is greater than the second preset collision time, it is determined that there is no collision risk.

[0093] In one embodiment, a time-to-collision (TTC) calculation model is used to determine the state trigger threshold and measure the collision risk.

[0094] In one embodiment, an index characterizing future collision risk is used through a TTC calculation model, which is expressed as: , , in, Indicates the vehicle's speed. Indicates the speed of the target object.

[0095] when When the collision risk reaches a threshold (corresponding to a high risk threshold), the system immediately applies maximum braking; when When necessary, the system applies moderate braking to avoid sudden braking while maintaining driving comfort; when When the collision risk is low (the threshold corresponding to low risk), only a warning is issued to remind the driver to be more vigilant.

[0096] Optionally, , and The value of is determined by the self-learning model.

[0097] Optionally, the deep Q-network is updated iteratively using the Bellman equation.

[0098] In Deep Q-Networks (DQNs), a deep neural network approximates the Q-function, predicting future rewards based on states and actions, and using an approximately optimal decision-making strategy. The long-term reward (Q-function) is the cumulative reward the system can expect to obtain after taking a certain action in a given state. According to the Bellman equation, the Q-function is defined as: , in, The Q-value represents the current state and action. This represents the overall state reward function. This indicates the system acceleration selected by the current system. This represents the discount factor, used to measure the importance of future rewards. This represents the Q value corresponding to the optimal action among all possible actions at the next time step (t+1).

[0099] By decomposing long-term rewards into a weighted sum of current rewards and potential future gains, decision-making considers not only immediate outcomes but also subsequent impacts. By training a deep neural network to approximate the Q-function, a mapping from states to optimal actions can be achieved. In the application phase, the deep Q-network can continuously update its parameters based on historical interaction data. The output Q-value is used to compare the merits of different actions, ultimately selecting the action that maximizes the Q-value as the target action.

[0100] A risk avoidance control system according to an embodiment of this application includes: The instruction activation module is used to respond to the vehicle's parking mode switching request and determine whether the vehicle has a collision risk through a risk avoidance model; The risk assessment module is used to control the vehicle to avoid the risk through the risk avoidance model when the collision risk exists. The risk avoidance module is used to determine whether the vehicle still has the collision risk after the avoidance is completed, by using the risk avoidance model again. The mode switching module is used to control the vehicle to switch modes when there is no collision risk.

[0101] like Figure 4As shown in the embodiment of this application, a vehicle 400 includes a memory 410 and a processor 420; the memory 410 is used to store a computer program; the processor 420 is used to implement the risk avoidance and control method described above when the computer program is executed.

[0102] Alternatively, a vehicle 400 includes a memory 410 and a processor 420 coupled to the memory 410; the memory 410 is configured to store a computer program; the processor 420 is configured to perform the following operations when the computer program is executed: In response to a vehicle's parking mode switching request, a risk avoidance model is used to determine whether the vehicle poses a collision risk. When the collision risk exists, the vehicle is controlled to avoid the risk through the risk avoidance model. After the avoidance is completed, the risk avoidance model is used again to determine whether the vehicle still has the risk of collision. When there is no collision risk, control the vehicle to switch modes.

[0103] This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the risk avoidance and control method described above.

[0104] Alternatively, a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the following operations: In response to a vehicle's parking mode switching request, a risk avoidance model is used to determine whether the vehicle poses a collision risk. When the collision risk exists, the vehicle is controlled to avoid the risk through the risk avoidance model. After the avoidance is completed, the risk avoidance model is used again to determine whether the vehicle still has the risk of collision. When there is no collision risk, control the vehicle to switch modes.

[0105] The vehicle 400 in this embodiment is an example of a hardware device that can be applied to various aspects of this application. Vehicle 400 includes various forms of digital electronic computer equipment, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Vehicle 400 may also include various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the application described and / or claimed herein.

[0106] Vehicle 400 includes a computing unit that can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) or a computer program loaded from a storage unit into random access memory (RAM). The RAM can also store various programs and data required for device operation. The computing unit, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0107] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc. In this application, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this application according to actual needs. Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units can be implemented in hardware or as software functional units.

[0108] Although the above disclosure is provided, the scope of protection of this application is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of this application, and all such changes and modifications will fall within the scope of protection of this application.

Claims

1. A risk aversion control method, characterized in that, include: In response to a vehicle's parking mode switching request, a risk avoidance model is used to determine whether the vehicle poses a collision risk. When the collision risk exists, the vehicle is controlled to avoid the risk through the risk avoidance model. After the avoidance is completed, the risk avoidance model is used again to determine whether the vehicle still has the risk of collision. When there is no collision risk, control the vehicle to switch modes.

2. The risk avoidance and control method according to claim 1, characterized in that, The step of controlling the vehicle to avoid risks through the risk avoidance model includes: The current state vector of the vehicle is obtained, and the current state vector is input into the risk avoidance model to obtain the risk avoidance action of the vehicle. The vehicle is then controlled to execute the risk avoidance action in order to avoid risks.

3. The risk avoidance and control method according to claim 2, characterized in that, Also includes: A reinforcement learning model is constructed, wherein a deep Q-network is used as the agent of the reinforcement learning model, the state space of the reinforcement learning model is constructed based on the state vector, the state vector is composed of at least one of driving data, environmental state and displacement data, the action space of the reinforcement learning model is constructed based on the action vector, the action vector is composed of the change of the vehicle's system acceleration over time, and a multi-objective reward function is used as the reward function of the reinforcement learning model, the multi-objective reward function is obtained by weighted fusion of collision avoidance reward function, comfort reward function, behavior habit reward function and delay penalty function; The current state vector at the current training moment is input into the deep Q-network, which outputs the target action. Control the vehicle to perform the target action, obtain the state vector at the next training moment, and determine the reward value based on the multi-objective reward function; The Q-function in the deep Q-network is optimized based on the reward value. The state vector of the next training time is used as the state vector of the current training time. The step of inputting the state vector of the current training time into the deep Q-network is returned. The training is iterated until the deep Q-network meets the preset convergence condition, and the risk avoidance model is obtained.

4. The risk avoidance and control method according to claim 3, characterized in that, It also includes a self-learning model, and the state vector further includes a state trigger threshold; the risk avoidance control method further includes: The self-learning model determines the state trigger threshold based on the vehicle's driving data and the environmental state, and the state trigger threshold is used to participate in the optimization of the multi-objective reward function.

5. The risk avoidance and control method according to claim 3, characterized in that, The construction of the collision avoidance reward function includes: Based on the kinematic properties of the vehicle and the obstacle, the predicted relative distance between the vehicle and the obstacle after a preset future time is determined by a kinematic model, wherein the kinematic properties include at least one of position, velocity and acceleration; When the predicted relative distance is greater than or equal to the collision threshold, a reward value positively correlated with the predicted relative distance is generated through an exponential decay term, wherein the state trigger threshold includes the collision threshold; When the predicted relative distance is less than the collision threshold, a negative reward is applied to the predicted relative distance through a collision indicator.

6. The risk avoidance and control method according to claim 3, characterized in that, The construction of the comfort reward function includes: The change in acceleration is determined based on the vehicle's acceleration at the current moment and the previous moment; When the absolute value of the acceleration change increases, a negative reward that is positively correlated with the absolute value of the acceleration change is applied through a smoothness penalty term; When the magnitude of the acceleration exceeds the preset acceleration range, a negative reward positively correlated with the degree of exceeding the limit is applied through a reasonableness penalty term; When the amplitude of the acceleration does not exceed the preset acceleration range, a positive reward negatively correlated with the amplitude is applied through the rationality penalty item, wherein the state trigger threshold includes the preset acceleration range.

7. The risk avoidance and control method according to claim 3, characterized in that, The construction of the behavioral habit reward function includes: Based on the scene features of the current driving scenario and the scene records in the historical driving database, the target historical scene with the highest similarity to the current driving scenario is determined. A reward value negatively correlated with the acceleration difference value is applied through a behavioral consistency term, wherein the acceleration difference value is determined by the absolute value of the difference between the system acceleration currently selected by the reinforcement learning model and the historical acceleration recorded in the target historical scene.

8. The risk avoidance and control method according to claim 3, characterized in that, The construction of the delayed penalty function includes: Based on the risk assessment value of the current driving state and the acceleration of the vehicle, the state of the vehicle is determined by a preset risk threshold; When the vehicle is in a high-risk state and the magnitude of the acceleration is less than the braking response threshold, a negative reward positively correlated with the risk assessment value is applied through the insufficient response term; When the vehicle is in a low-risk state and the magnitude of the acceleration is greater than the aggressive operation threshold, a negative reward negatively correlated with the risk assessment value is applied through a misoperation item, wherein the state trigger threshold includes the braking response threshold and the aggressive operation threshold, and the braking response threshold is less than the aggressive operation threshold.

9. A risk aversion control system, characterized in that, include: The instruction activation module is used to respond to the vehicle's parking mode switching request and determine whether the vehicle has a collision risk through a risk avoidance model; The risk assessment module is used to control the vehicle to avoid the risk through the risk avoidance model when the collision risk exists. The risk avoidance module is used to determine whether the vehicle still has the collision risk after the avoidance is completed, by using the risk avoidance model again. The mode switching module is used to control the vehicle to switch modes when there is no collision risk.

10. A vehicle, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to implement the risk avoidance control method as described in any one of claims 1-8 when executing the computer program.

Images