Human-machine shared collision avoidance control system and method for intelligent vehicle in critical scenario
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2025-03-26
- Publication Date
- 2026-08-13
Smart Images
Figure CN2025085173_13082026_PF_FP_ABST
Abstract
Description
A human-machine shared collision avoidance control system and method for intelligent vehicles under critical conditions
[0001] Cross-references to related applications
[0002] This application is based on and claims priority to Chinese Patent Application No. 202510146478.3, filed on February 10, 2025, the entire contents of which are incorporated herein by reference. Technical Field
[0003] This application relates to the field of active safety in automobiles, and in particular to an intelligent vehicle human-machine shared collision avoidance control system and method under critical conditions. Background Technology
[0004] According to a report by the World Health Organization, road traffic injuries are the leading cause of death among young people aged 5 to 29, causing approximately 1.3 million deaths annually. Statistics show that in high-risk driving scenarios, complex operational requirements, tight reaction times, and decreased decision-making ability due to cognitive overload make timely collision avoidance a significant challenge. In this context, intelligent vehicles need to assist drivers in performing rapid and precise avoidance maneuvers to address potential safety risks.
[0005] Traditional human-machine collision avoidance technologies typically generate collision avoidance trajectories through path replanning and trajectory tracking, and have proven highly effective in ensuring vehicle safety. Their basic principle is to achieve higher precision obstacle avoidance control through collaborative operation between the intelligent system and the driver. However, these methods often overlook the driver's real-time intentions in emergency scenarios. When the vehicle needs to perform rapid or extreme maneuvers, forcing it to follow the system-generated trajectory may trigger driver resistance or misoperation, thereby increasing the risk of new collisions.
[0006] To reduce human-machine conflict, driver-centric shared control methods have gradually emerged. These methods, by introducing the concept of dynamic safety range and model predictive control techniques, provide necessary interventions while respecting driver intent. However, treating environmental constraints as soft constraints may reduce overall safety in certain scenarios. Furthermore, reinforcement learning-based collision avoidance control methods have gained attention in recent years. These methods, by directly generating collision avoidance actions, avoid the limitations of traditional path planning and achieve robust collision avoidance control in complex dynamic environments. These methods still primarily rely on distance-based risk assessment, intervening near manually set thresholds. Within this framework, a theoretical basis for effective system intervention has not yet been established; that is, it is impossible to determine when machine intervention is truly necessary.
[0007] Currently, while significant progress has been made in human-machine shared collision avoidance control, including trajectory tracking-based frameworks, human-centered flexible constraint methods, and recent reinforcement learning-based non-explicit collision avoidance path strategies, these technologies still face several challenges. Traditional trajectory tracking strategies typically require drivers to follow pre-defined paths, which may lead to human-machine conflict in dynamic or emergency scenarios, reducing driver acceptance. Attempts to improve cooperation by softening obstacle constraints may result in decreased safety, and distance- or artificial potential field (APF)-based risk assessment methods lack rigorous theoretical safeguards regarding the necessity of intervention. Furthermore, current research lacks consideration of the driver's own collision avoidance intentions.
[0008] To address the above issues, there is an urgent need to establish a human-machine co-driving control method and system for collision avoidance scenarios that has theoretical machine intervention triggering standards and minimizes human-machine conflict. Summary of the Invention
[0009] This application aims to at least partially address one of the technical problems in the related art.
[0010] Therefore, the first objective of this application is to propose a human-machine shared collision avoidance control method for intelligent vehicles under critical conditions, so as to theoretically assess the collision risk of the vehicle state and the collision avoidance intention of the driver's actions, and minimize human-machine conflict to the greatest extent possible while ensuring that no collision occurs.
[0011] The second objective of this application is to propose an intelligent vehicle human-machine shared collision avoidance control system for critical operating conditions.
[0012] The third objective of this application is to propose an electronic device.
[0013] The fourth objective of this application is to provide a computer-readable storage medium.
[0014] The fifth objective of this application is to provide a computer program product.
[0015] To achieve the above objectives, the first aspect of this application proposes a human-machine shared collision avoidance control method for intelligent vehicles under critical conditions, comprising:
[0016] The vehicle collects and transmits information about its current status and surrounding environment through a sensing system, which includes a camera, millimeter-wave radar, a combined inertial navigation IMU, and related network facilities.
[0017] The set of unreachable states for collision avoidance is approximated using offline learning and large-scale vehicle data. The Hamilton-Jacobi reachability value function and action-value function are updated offline using a large amount of pre-collected real vehicle collision avoidance data to obtain the reachability value function and action-value function.
[0018] Based on the reachability value function and the action-value function, the driver's collision avoidance ability and collision avoidance intention are quantified, and the human-machine control weights are dynamically allocated according to the quantified collision avoidance ability and collision avoidance intention.
[0019] Machine operations are generated based on an accessibility-inspired reinforcement learning algorithm, which combines the driver's collision avoidance intention, collision avoidance capability, and accessibility information into the state space.
[0020] By combining the human-machine control weights, driver operations, and machine operations, a final operation to be executed is generated. This final operation is then sent to each actuator in the vehicle in real time via a communication system to perform the collision avoidance operation.
[0021] In some embodiments, offline learning and large-scale vehicle data are used to approximate the set of inaccessible states for collision avoidance, and the Hamilton-Jacobi reachability value function and action-value function are iteratively updated offline using a large amount of pre-collected real-vehicle collision avoidance data to obtain the reachability value function and action-value function, including:
[0022] The obstacle is represented as an elliptical envelope or a T-shaped collision surface, where the parameters of the target obstacle are denoted as [X0,Y0,a,b], which represent the center coordinates and elliptical shape parameters of the obstacle, respectively.
[0023] The target obstacle is represented as a set of safety constraint states T. Vehicle states within this set are considered dangerous states where a collision has already occurred, and are defined as follows:
[0024] Where x=[X,Y,φ,v] x ,v y [r] represents the vehicle's state and kinematic information, X and Y are the vehicle's global position coordinates, φ is the yaw angle, and v x ,v y These are the longitudinal velocity and the lateral velocity, respectively, and r represents the yaw rate;
[0025] The vehicle state and driver operation from a large amount of pre-collected real vehicle collision avoidance data are substituted into the reachability value function network and action value function network corresponding to the safety constraint state set T to calculate the sample reachability value function value and sample action value function value.
[0026] The sample reachability value function and sample action value are updated offline iteratively to further optimize the accuracy of collision avoidance decisions. After the iteration is completed, the updated reachability value function V is obtained. h Network and Action-Value Function Q h The network will be deployed in intelligent vehicles and advanced driver assistance systems.
[0027] In some embodiments, based on the reachability value function and the action-value function, the driver's collision avoidance capability and collision avoidance intention are quantified, and human-machine control weights are dynamically allocated according to the quantified collision avoidance capability and collision avoidance intention, including:
[0028] The current vehicle state, obstacle envelope, and driver actions collected by the sensor system are substituted into the updated reachability value function V corresponding to the constraint state set T. h Network and Action-Value Function Q h The network yields the reachability value V. h (x) and action-value Q h (x,u d );
[0029] Through reachability value V h (x) represents the optimal collision avoidance distance in the current state. The driver's collision avoidance capability (CAA) is calculated using the following formula:
[0030] Where, α CAA C is the sensitivity adjustment parameter. CAA It is the offset constant;
[0031] Based on the driver's current operation d Corresponding action-value Q h (x,u d The reachability value V corresponding to the optimal action h (x), calculate the driver's collision avoidance intention (CAI), which is obtained by the following formula:
[0032] Based on the collision avoidance capability (CAA) and collision avoidance intention (CAI), the human-machine control weight γ is dynamically allocated, and the control weight is calculated using the following formula: γ = max(γ min ,(1-s CAI (1-s) CAA ))
[0033] Among them, s CAI With s CAA The normalization functions are defined by the following formulas:
[0034] Where, k CAI1 and k CAA1 Control s respectively CAI and s CAA Sensitivity to changes in CAI and CAA, k CAI2 and k CAA2 Define s CAI and s CAAReaching a threshold of 0.5 represents the balance point between collision avoidance intent and capability.
[0035] In some embodiments, machine operations are generated based on an reachability-inspired reinforcement learning algorithm, which incorporates the driver's collision avoidance intention, collision avoidance capability, and reachability information into a state space, including:
[0036] The driver's collision avoidance intention, collision avoidance capability, accessibility information, and other vehicle dynamic states are explicitly integrated into the state space of the reinforcement learning algorithm. This state space includes: the vehicle's dynamic state x, and the accessibility value function V. h (x) Driver's collision avoidance capability (CAA), driver's collision avoidance intent (CAI), human-machine control weights (γ), and driver and machine control operations (u) d and u m , where the action space u m Includes only the front wheel steering angle δ mf ;
[0037] In the state space, by combining the vehicle's dynamics and kinematics, a driver action generation model is constructed to simulate the driver's collision avoidance operation during training. The driver action generation model simulates the driver's collision avoidance decision by dynamically adjusting the relative relationship between the driver's input and the obstacle.
[0038] A reinforcement learning reward function is constructed based on the driver's collision avoidance intention. The reward function is designed based on the goal of reducing human-machine conflict and maintaining the original task performance. The reward function combines the obstacle envelope as a constraint condition to guide the generation of machine control actions during the reinforcement learning process.
[0039] During reinforcement learning training, by maximizing the above reward function, the machine can prioritize safety while reducing human-machine conflict, and generate the final control action.
[0040] In some embodiments, in the state space, by combining the vehicle's dynamics and kinematics, a driver action generation model is constructed to simulate the driver's collision avoidance maneuvers during training. The driver action generation model simulates the driver's collision avoidance decisions by dynamically adjusting the relative relationship between the driver's input and the obstacle, including:
[0041] The current vehicle state, obstacle envelope, and driver control inputs are fed into the driver action generation model. By calculating the relative position and velocity between the vehicle and the obstacle, the driver's steering input is calculated and adjusted to ensure that the driver's preview direction is aligned with the tangent of the obstacle boundary, thereby ensuring that the vehicle can avoid the obstacle. The preview angle θ... c Calculate using the following formula:
[0042] Among them, (X) trg ,Y trg () represents the point of tangency on the ellipse, located on the boundary of the ellipse as defined below:
[0043] Where λ is a scaling factor, which is based on the reachability value V. h (x) and the collision avoidance intention CAI adjust the ellipse size, defined as:
[0044] Where, k λ This is a hyperparameter.
[0045] In some embodiments, the construction of a reinforcement learning reward function based on the driver's collision avoidance intention, wherein the reward function is designed based on the goal of reducing human-machine conflict and maintaining the original task performance, and wherein the reward function incorporates obstacle envelopes as constraints to guide the generation of machine control actions during reinforcement learning, includes:
[0046] The cooperation between the driver and the machine is optimized by using a reinforcement learning reward function to reduce human-machine conflict. This reward function includes a safety reward R. sf and collaboration reward R co ;
[0047] The security reward R sf The distance between the vehicle state and the set of unreachable states for collision avoidance is calculated and defined as:
[0048] Where d0 is the scaling parameter, k sf1 To emphasize the importance of the boundary of advanced driver assistance systems (ADAS) for intelligent vehicles, V h (x)>k sf2 A state that results in severe penalties is designed to ensure that agents learn to avoid unsafe states.
[0049] The collaboration reward R co Human-machine collaboration is optimized by quantifying the deviation between machine and driver behavior, defined as: R co =-k co ·γ·(u M -u D ) 2
[0050] Among them, the penalty item (u M -u D ) 2 k represents the deviation between machine and driver behavior. co It is a scaling factor, and the weighting factor γ reflects the necessity of machine intervention based on the driver's collision avoidance ability and intention.
[0051] In some embodiments, the step of combining the human-machine control weights, driver operations, and machine operations to generate a final operation to be executed, and sending the final operation to be executed to each actuator of the vehicle in real time via a communication system to execute a collision avoidance operation, includes:
[0052] Based on the human-machine control weight γ and machine operation u m and driver operation u d Generate the final operation to be executed u f The formula is: u f =γ·u m +(1-γ)·u d
[0053] The final operation to be executed is transmitted via the communication system. f It sends real-time signals to all actuators in the vehicle, including the active differential steering system, drive system, and braking system, to perform the corresponding collision avoidance maneuvers.
[0054] During execution, the vehicle's status is monitored in real time through continuous sensor feedback, and the final operation is dynamically adjusted according to the relative position of the vehicle and obstacles to ensure that the vehicle successfully avoids obstacles and returns to the normal driving trajectory.
[0055] To achieve the above objectives, a second aspect of this application provides an intelligent vehicle human-machine shared collision avoidance control system for critical situations, comprising:
[0056] The data acquisition module is used to collect and transmit information about the vehicle's current status and surrounding environment through a sensing system, which includes a camera, millimeter-wave radar, a combined inertial navigation IMU, and related networking facilities.
[0057] The reachability analysis module is used to approximate the set of unreachable states for collision avoidance using offline learning and large-scale vehicle data. It also performs offline iterative updates on the Hamilton-Jacobi reachability value function and action-value function using a large amount of pre-collected real vehicle collision avoidance data to obtain the reachability value function and action-value function.
[0058] The driver intention and ability assessment module is used to quantify the driver's collision avoidance ability and collision avoidance intention based on the reachability value function and the action-value function, and dynamically allocate human-machine control weights according to the quantified collision avoidance ability and collision avoidance intention.
[0059] The reinforcement learning control module is used to generate machine operations based on an reachability-inspired reinforcement learning algorithm, which combines the driver's collision avoidance intention, collision avoidance capability, and reachability information into the state space.
[0060] The control decision module is used to combine the human-machine control weights, driver operations, and machine operations to generate the final operation to be executed, and send the final operation to be executed to each actuator of the vehicle in real time through the communication system to execute the collision avoidance operation.
[0061] To achieve the above objectives, a third aspect of this application provides an electronic device, including: a processor, and a memory communicatively connected to the processor;
[0062] The memory stores computer-executed instructions;
[0063] The processor executes computer execution instructions stored in the memory to implement the method as described in any one of the first aspects above.
[0064] To achieve the above objectives, a fourth aspect of this application provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, are used to implement the method as described in any one of the first aspects above.
[0065] To achieve the above objectives, a fifth aspect of this application provides a computer program product including a computer program that, when executed by a processor, implements the method as described in any one of the first aspects above.
[0066] The technical solutions provided by the embodiments of this application bring at least the following beneficial effects:
[0067] (1) The Hamilton-Jacobi accessibility method is used to theoretically assess collision risk. By calculating the accessibility value, the system can make timely intervention decisions when danger is imminent. This method enables the machine to determine whether intervention is necessary based on real-time collision risk, effectively avoiding collisions and reducing human-machine conflict, thus ensuring vehicle driving safety in complex environments.
[0068] (2) A human-machine shared control system is adopted, which can accurately control the vehicle trajectory and achieve smoother yaw angle adjustment. During obstacle avoidance, the system's real-time adjustment capability can effectively cope with sudden obstacles or complex working conditions, thereby improving the vehicle's stability and safety, especially its responsiveness in emergency situations. Whether in emergency situations or during daily driving, the vehicle can maintain a higher level of safety.
[0069] (3) By dynamically quantifying the driver's collision avoidance ability and intentions, and combining the driver's real-time operations with the system's calculations, the system intelligently allocates human-machine control weights. At critical moments, the system can effectively take over vehicle control, while respecting the driver's operations at other times to maintain the driver's sense of control. This intelligent mechanism not only optimizes the driving experience but also balances human-machine collaboration in complex driving environments, improving driving safety and comfort.
[0070] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description
[0071] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:
[0072] Figure 1 is a flowchart illustrating a human-machine shared collision avoidance control method for intelligent vehicles under critical conditions provided in an embodiment of this application.
[0073] Figure 2 is a schematic diagram of the architecture of a human-machine shared collision avoidance control method for intelligent vehicles under critical conditions provided in an embodiment of this application;
[0074] Figure 3 is a schematic diagram illustrating the influence of the driver's collision avoidance ability and collision avoidance intention on the machine control weights provided in the embodiments of this application;
[0075] Figure 4 is a comparison diagram of the deployment effects of this application and other prior art solutions provided in the embodiments of this application;
[0076] Figure 5 is a schematic diagram of the structure of an intelligent vehicle human-machine shared collision avoidance control system under critical conditions provided in an embodiment of this application. Detailed Implementation
[0077] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.
[0078] To address the shortcomings of existing technologies, this application provides a human-machine shared collision avoidance control method for intelligent vehicles under critical conditions. It utilizes large-scale data to approximate a set of unreachable collision avoidance states, and uses the degree of similarity between the vehicle's state and this set as a theoretical basis for assessing collision risk, thereby determining the necessity of machine intervention. This method can theoretically assess the collision risk of the vehicle's state and the driver's collision avoidance intentions, minimizing human-machine conflict while ensuring that a collision does not occur.
[0079] Referring to Figures 1 and 2, the method includes the following steps:
[0080] Step S1: Collect and transmit information about the vehicle's current status and surrounding environment through a sensing system, which includes a camera, millimeter-wave radar, a combined inertial navigation IMU, and related networking facilities.
[0081] In this embodiment, this step is performed by module 10 of FIG2. Referring to FIG2, the sensing system includes a camera, millimeter-wave radar, a combined inertial navigation IMU, and associated networking facilities. These sensors can acquire and transmit in real time dynamic state data such as vehicle position, speed, and acceleration, as well as information such as the relative position and size of obstacles in the surrounding environment. This data is crucial for subsequent collision avoidance decisions.
[0082] Considering that a large amount of historical collision avoidance data is needed when solving for the collision avoidance unreachable state set, the sensing system should have strong data integration, collection, and storage capabilities to ensure that it can store and quickly retrieve the necessary collision avoidance data, thereby supporting subsequent analysis and computation tasks.
[0083] Furthermore, in this embodiment, the collected data is transmitted to the autonomous driving system and the advanced driver assistance system (ADAS) of the intelligent vehicle via a communication network. Specifically, Ethernet and CAN FD (Controller Area Network Flexible Data Transmission) protocols are used for data fusion transmission to ensure timely data delivery in a high-speed, reliable environment. This efficient data transmission method ensures real-time information sharing, thereby providing the necessary data support for subsequent collision avoidance operations.
[0084] Through this data acquisition and transmission method, the embodiments of this application realize real-time and accurate vehicle status monitoring and surrounding environment perception, providing strong support for intelligent vehicles to make collision avoidance decisions in complex traffic situations.
[0085] Step S2: Use offline learning and large-scale vehicle data to approximate the set of unreachable states for collision avoidance, and use a large amount of pre-collected real vehicle collision avoidance data to perform offline iterative updates on the Hamilton-Jacobi reachability value function and action-value function to obtain the reachability value function and action-value function.
[0086] In this embodiment, module 20 obtains obstacle information, vehicle kinematics, and dynamics information from module 10. After substituting the relevant vehicle and environmental information into the pre-trained neural network, it can obtain the value function V of the current state. h (x), combined with the driver's current operation, yields Q. h (x,u d ).
[0087] Specifically, in this embodiment, the obstacle is represented as an elliptical envelope or a T-shaped collision surface. The parameters of the target obstacle are denoted as [X0, Y0, a, b], which represent the center coordinates and elliptical shape parameters of the obstacle, respectively. These parameters are used to describe the geometric features of the obstacle, ensuring that the obstacle is represented more accurately in the system.
[0088] Furthermore, in this embodiment, the target obstacle is represented as a set of safety constraint states T, and the vehicle state within this set is considered a dangerous state where a collision has already occurred, defined as follows:
[0089] in, This represents the vehicle's state and kinematic information, where X and Y are the vehicle's global position coordinates. It's the yaw angle, v x ,v y These are the longitudinal speed and the lateral speed, respectively, and r represents the yaw rate.
[0090] Furthermore, in the solution process, the embodiments of this application use approximation techniques to approximate the Hamiltonian-Jacobi reachability value function V. h and action-value function Q h The solution process is simplified, and it is updated offline using a large amount of pre-collected real vehicle collision avoidance data.
[0091] The specific training process is as follows: First, the vehicle state and driver operation from a large amount of pre-collected real vehicle collision avoidance data are substituted into the reachability value function network and action value function network corresponding to the safety constraint state set T, and the sample reachability value function value and sample action value function value are calculated. These sample data provide a basis for subsequent optimization.
[0092] Next, the reachability value function and action-value function of these samples are updated offline iteratively to optimize the accuracy of collision avoidance decisions. Through continuous updates, this application can achieve more accurate collision avoidance decisions in various driving scenarios.
[0093] After the iteration is complete, the updated reachability value function V is obtained. h Network and Action-Value Function Q h The network will be deployed in intelligent vehicles and in advanced driver assistance systems (ADAS) for subsequent applications.
[0094] Through this process, the embodiments of this application effectively improve the human-machine shared control system's ability to assess collision avoidance risks and the accuracy of collision avoidance decisions, providing strong support for the human-machine shared control system, ensuring timely safety decisions in complex environments, avoiding collisions and improving vehicle driving safety.
[0095] In one possible embodiment, module 10 detects a sudden accident ahead when the vehicle speed is 50 km / h and transmits the relevant information to module 20 for driving state identification during the collision avoidance process. Simultaneously, the information is sent to module 32 for temporary storage. The target obstacle in this embodiment can be described as an elliptical envelope with a long axis of 8m and a short axis of 6m. To ensure a human-centric strategy, the embodiments of this application implement shared control for the steering system, while the drive and braking systems are under full driver control. For further specific scenarios, the embodiments of this application assume that the driver wishes to avoid the collision at the original speed.
[0096] Step S3: Based on the reachability value function and action-value function, quantify the driver's collision avoidance ability and collision avoidance intention, and dynamically allocate human-machine control weights according to the quantified collision avoidance ability and collision avoidance intention.
[0097] Step S3 is a key step in the embodiments of this application. It aims to optimize the cooperation between the driver and the machine by quantifying the driver's collision avoidance ability and collision avoidance intention, and dynamically allocating human-machine control weights according to the quantification results, thereby ensuring that the best collision avoidance decision is made in complex environments.
[0098] Specifically, the following steps are included:
[0099] First, the current vehicle state, obstacle envelope, and driver operation data collected by the sensing system are substituted into the updated reachability value function V corresponding to the constraint state set T. h Network and Action-Value Function Q h The network is then used to calculate the reachability value V in the current state. h (x) and action-value Q h (x,u d These values provide the necessary foundation for subsequent calculations of collision avoidance capabilities and collision avoidance intentions.
[0100] Next, through the reachability value V h (x) represents the optimal collision avoidance distance in the current state, and further calculates the driver's collision avoidance capability (CAA). The collision avoidance capability is calculated using the following formula:
[0101] Where, α CAA C is a sensitivity adjustment parameter used to control the sensitivity of collision avoidance capability to changes in reachability values; CAA This is an offset constant, used as a benchmark value to adjust collision avoidance capability. This formula effectively quantifies the driver's collision avoidance capability under current conditions, thus providing a basis for subsequent control weight allocation.
[0102] Based on the calculation of collision avoidance capability, the embodiments of this application further consider the driver's current control operation. d Corresponding action-value Q h(x,u d The reachability value V corresponding to the optimal action h (x), further calculate the driver's collision avoidance intention (CAI). The collision avoidance intention is calculated using the following formula:
[0103] The Collision Avoidance Intent (CAI) reflects the driver's degree of intention to avoid a collision, with a value ranging from 0 to 1. The closer the value is to 1, the stronger the driver's intention to avoid a collision. This indicator can quantify the driver's collision avoidance motivation in the current driving state.
[0104] Furthermore, this application also proposes a method for allocating weights γ in a human-machine controller that considers the driver's collision avoidance capability (CAA) and collision avoidance intention (CAI). This method should allow the weights to fluctuate between 0 and 1 and be inversely proportional to the collision avoidance capability and collision avoidance intention. This method should be decoupled from machine motion generation to ensure the interpretability and transparency of the final collision avoidance operation.
[0105] Specifically, module 20 can update the driver collision avoidance capability (CAA) and driver collision avoidance intention (CAI) in real time and transmit them to module 30. The human-machine weight allocation mechanism in module 30 calculates the control weight γ of the machine control according to the following method: γ = max(γ min ,(1-s CAI (1-s) CAA ))
[0106] Where, γ min It is the minimum control weight, ensuring that machine intervention has minimal effect under any circumstances. Through this formula, the control roles of the driver and the machine can be intelligently balanced, ensuring that the machine can effectively take over control at critical moments and reduce the risk of collision.
[0107] Furthermore, to ensure more accurate calculation of the control weights, the embodiments of this application also use a normalization function s. CAI With s CAA This adjusts the sensitivity to collision avoidance intentions and capabilities. The specific formula is as follows:
[0108] Where, k CAI1 and k CAA1 Control s respectively CAI and s CAA Sensitivity to changes in CAI and CAA, k CAI2 and k CAA2 Define s CAI and s CAA Reaching a threshold of 0.5 represents the balance point between collision avoidance intent and capability.
[0109] Figure 3 is a schematic diagram illustrating the influence of the driver's collision avoidance ability and collision avoidance intention on the machine control weights provided in the embodiments of this application. Through these calculations, the embodiments of this application can dynamically adjust the balance of human-machine control in real time, enabling effective coordination between the driver and the machine in dangerous situations, ensuring that the vehicle can smoothly avoid obstacles and reduce the risk of collision.
[0110] Step S4: Generate machine operations based on reachability-inspired reinforcement learning algorithm. The reinforcement learning algorithm combines the driver's collision avoidance intention, collision avoidance ability and reachability information into the state space.
[0111] This application also proposes a method for generating shared driving machine actions using reachability-inspired reinforcement learning. First, the reachability value function, driving ability, and driving intention are explicitly integrated into the state space of the reinforcement learning (which should also include vehicle dynamics and motion habit states). Second, a driver action generation model is constructed by combining collision avoidance intention (CAI) to simulate and generate driver collision avoidance operations during training. Finally, a reward function is designed based on reducing human-machine conflict and maintaining original task performance, with obstacle envelopes as constraints. An online reinforcement learning method is then used to solve for the machine action u. m .
[0112] Specifically, step S4 also includes:
[0113] Step S41: Explicitly integrate the driver's collision avoidance intention, collision avoidance capability, accessibility information, and other vehicle dynamic states into the reinforcement learning state space. The state space includes: the vehicle's dynamic state x, and the accessibility value function V. h (x) Driver's collision avoidance capability (CAA), driver's collision avoidance intent (CAI), human-machine control weights (γ), and driver and machine control operations (u) d and u m , where the action space u m Includes only the front wheel steering angle δ mf .
[0114] In this embodiment of the application, the state space can be represented as: s=[x,V h (x),CAA,CAI,γ,u d ,u m The state space comprises several key components.
[0115] Specifically, firstly, the vehicle's dynamic state x can fully describe the vehicle's kinematic state at the current moment, providing a basis for subsequent collision avoidance decisions.
[0116] Secondly, the reachability value function V h(x) can measure the likelihood of avoiding a collision in the current vehicle state, reflecting whether the vehicle is close to the critical state of a collision in the current state, and predict its long-term safety in that state.
[0117] The driver's collision avoidance ability (CAA) and collision avoidance intention (CAI) are also incorporated into the state space, quantifying the driver's ability and intention to avoid collisions in the current situation, respectively. The driver's collision avoidance ability is calculated using an reachability value function, reflecting whether the driver can take timely collision avoidance actions. The collision avoidance intention is quantified using the driver's current operation and state action values, reflecting whether the driver is willing to take measures to avoid a collision in the current situation, enabling the reinforcement learning agent to evaluate and supplement driver behavior.
[0118] The human-machine control weight γ determines the balance between machine and driver control. Through this weight, this application can dynamically adjust the degree of machine intervention based on the driver's collision avoidance capabilities and intentions, ensuring that the machine can take over control in a timely manner when necessary to avoid a collision.
[0119] It is important to emphasize that the machine's control operation δ mf Only front wheel steering angle δ mf Because in this application, the machine mainly avoids obstacles by controlling the steering, without involving other control operations (such as braking or acceleration).
[0120] By integrating this information into the state space of reinforcement learning, this application can comprehensively assess the current driving situation and make appropriate responses. Furthermore, based on these inputs, embodiments of this application can more accurately predict the driver's intentions and determine whether machine intervention is necessary, thereby achieving more intelligent collision avoidance control. Overall, this method improves the collaboration between the driver and the machine, ensuring optimal collision avoidance decisions are made in various complex environments.
[0121] Step S42: In the state space, by combining the vehicle's dynamics and kinematics, a driver action generation model is constructed to simulate the driver's collision avoidance operation during training. The driver action generation model simulates the driver's collision avoidance decision by dynamically adjusting the relative relationship between the driver's input and the obstacle.
[0122] In this embodiment of the application, module 30 uses a reinforcement learning algorithm to generate machine operation u m .
[0123] It should be noted that reinforcement learning algorithms require an accurate environmental interaction to train the policy network. To address this process, this application provides a driver action generation model, as shown in Figure 4. This model mainly consists of two stages: (1) obstacle avoidance stage, where the driver avoids obstacles; and (2) recovery stage, where the driver readjusts to the original trajectory after avoiding the obstacles. In the first stage, the driver's steering behavior can be represented as maneuvering around an elliptical envelope slightly larger than the obstacle, as shown in the upper part of Figure 4.
[0124] During calculation, this embodiment incorporates the current vehicle state, obstacle envelope, and driver control input into the driver action generation model. By calculating the relative position and speed between the vehicle and the obstacle, the driver's steering operation is calculated and adjusted. The goal of the adjustment is to align the driver's preview direction with the tangent of the obstacle boundary, thereby ensuring that the vehicle can smoothly avoid the obstacle and prevent a collision.
[0125] To achieve this goal, the embodiments of this application will use the driver's preview angle θ c Defined as the angular difference between the vehicle's current position and the obstacle, the calculation formula is as follows:
[0126] Among them, (X) trg ,Y trg (x, y) represents the point of tangency on the ellipse, and the current position of the vehicle is (x, y). Using this formula, the embodiments of this application can calculate in real time the steering angle that the driver should take to ensure that the vehicle's driving direction is aligned with the boundary tangent of the obstacle, thereby avoiding a collision.
[0127] To accurately describe the obstacle, this application represents the target obstacle as an ellipse, the boundary of which is described by the following equation:
[0128] Where λ is a scaling factor, which is based on the reachability value V. h (x) and the collision avoidance intention CAI adjust the ellipse size, defined as:
[0129] Where, k λ This is a hyperparameter used to control the sensitivity of the scaling factor. Through this scaling factor, the size of the obstacle will be dynamically adjusted based on the current collision avoidance intention and the vehicle's accessibility information, ensuring more accurate collision avoidance decisions.
[0130] The recovery process can be viewed as tracking the original trajectory, and there are well-documented studies on this topic. Therefore, the embodiments described in this application will not be described in detail here.
[0131] Through the above process, it is ensured that the advanced driver assistance system (ADAS) of intelligent vehicles can accurately simulate collision avoidance behavior when the driver encounters obstacles. By simulating the driver's collision avoidance decisions, the ADAS not only reflects how the driver controls the vehicle to avoid obstacles, but also continuously optimizes the model during training, thus providing reliable support for collision avoidance decisions in actual driving.
[0132] Step S43: Construct a reinforcement learning reward function based on the driver's collision avoidance intention. The reward function is designed based on the goal of reducing human-machine conflict and maintaining the original task performance. The reward function combines the obstacle envelope as a constraint condition to guide the generation of machine control actions during the reinforcement learning process.
[0133] In this embodiment, the reward design of the reinforcement learning algorithm used to generate machine operations comprehensively considers collision avoidance safety and driver cooperation. The reward function is designed based on two main objectives: reducing human-machine conflict and maintaining the performance of the original task. The reward function plays a crucial role in the reinforcement learning process, guiding the generation of machine control actions and ensuring that the driver and machine can work together to avoid collisions.
[0134] First, the reward function considers two main aspects: security rewards and collaboration rewards.
[0135] Safety Reward R sf The aim is to ensure that vehicles remain within a safe range and avoid entering an inaccessible state set, thus effectively preventing collisions. The safety bonus is calculated using the following formula:
[0136] Among them, V h (x) represents the reachability value of the vehicle's current state, reflecting whether the vehicle is in a safe state; d0 in the formula is a scaling parameter that adjusts the sensitivity of the reward function; k sf1 A negative weighted factor, used to emphasize the importance of moving away from the boundaries of intelligent vehicles and advanced driver assistance systems (ADAS), is employed to highlight the importance of moving away from unsafe states. If V h (x)>k sf2 If a vehicle is in a state that is unreachable due to a collision, a significant penalty will be imposed to ensure that the reinforcement learning model can guide the vehicle to avoid entering unsafe areas.
[0137] Secondly, the collaboration reward R co The aim is to optimize collaboration between machines and drivers, reducing conflicts caused by excessive or insufficient intervention. Collaboration rewards are calculated by quantifying the deviation between machine and driver behavior, using the following formula: R co =-k co ·γ·(u M -u D) 2
[0138] Among them, the penalty item (u M -u D ) 2 The penalty is used to guide machine control to align with driver behavior as closely as possible to avoid excessive intervention, as it represents the deviation between machine and driver actions; k co This is a scaling factor used to adjust the weight of the cooperative reward, while the weighting factor γ reflects the necessity of machine intervention based on the driver's collision avoidance ability and intention. When the driver's collision avoidance ability and intention are weak, the value of γ is larger, indicating a stronger necessity for machine intervention, thus prompting the machine to intervene more; conversely, when the driver's ability is strong, the machine will intervene relatively less.
[0139] Step S44: During the reinforcement learning training process, by maximizing the above reward function, the machine can prioritize safety while reducing human-machine conflict, and generate the final control action.
[0140] In the reinforcement learning process of this application embodiment, the training objective is to enable the machine to make optimal control decisions in a real-world environment. To this end, the machine learns based on previously defined safety and cooperation rewards, optimizing the interaction between the driver and the machine by maximizing the total reward. Specifically, the machine continuously adjusts its control strategy to reduce deviations from the driver's actions, ensuring effective collision avoidance during human-machine collaboration while maintaining driver comfort and avoiding excessive intervention.
[0141] In this way, the control weights between the machine and the driver can be dynamically adjusted during reinforcement learning. This allows the machine to quickly take over control when needed, ensuring the vehicle can smoothly avoid obstacles in potentially dangerous situations. Conversely, when the driver's collision avoidance ability is strong, the machine reduces intervention, respecting the driver's intentions to the greatest extent possible. This dynamic adjustment process reduces human-machine conflict while ensuring that safety is the highest priority in every decision.
[0142] Ultimately, through reinforcement learning training, the machine can generate final control actions that fit the current environment and situation, thereby achieving more intelligent, precise, and safe collision avoidance. This process greatly improves the system's performance in complex driving situations, enabling intelligent vehicles to ensure driving safety in various traffic conditions and optimize the driving experience.
[0143] In practical applications, a policy network that can be deployed in an ADAS system can be trained based on the principles and demonstrations provided in the above embodiments. This network can be used to generate machine steering actions during human-machine co-driving in emergency scenarios. m
[0144] Step S5: Combining human-machine control weights, driver operations, and machine operations, the final operation to be executed is generated and sent to each actuator in the vehicle in real time through the communication system to perform the collision avoidance operation.
[0145] In this embodiment of the application, based on the aforementioned calculation results, combined with human-machine control weights, driver operations, and machine operations, the final operation to be executed is generated, and the operation is transmitted in real time to each actuator of the vehicle through the communication system to ensure that the corresponding collision avoidance operation is executed.
[0146] Specifically, firstly, based on the human-machine control weight γ and the machine operation u... m and driver operation u d Generate the final operation to be executed u f The formula is: u f =γ·u m +(1-γ)·u d
[0147] This formula allows for the dynamic adjustment of the degree of machine intervention based on the weight ratios in the current context, ensuring a balance in collaboration between the machine and the driver.
[0148] Once the final operation to be executed is u f Once calculated, the advanced driver assistance system (ADAS) for intelligent vehicles will send the operation in real time to the various actuators in the vehicle via the communication system. These actuators include the active differential steering system, drive system, and braking system, which can perform corresponding collision avoidance actions based on the received operation signals. For example, when facing an obstacle, the ADAS may instruct the steering system to make necessary adjustments or instruct the braking system to decelerate to avoid the obstacle.
[0149] Furthermore, during execution, this embodiment continuously monitors sensor feedback, acquires vehicle status information in real time, and continuously assesses the relative position of the vehicle and obstacles based on this real-time feedback. It then dynamically adjusts the final operation according to real-time changes to ensure the vehicle successfully avoids obstacles and returns to its normal driving trajectory. This adjustment ensures the vehicle can successfully avoid obstacles and return to its normal driving trajectory during collision avoidance, preventing secondary risks caused by excessive intervention or errors.
[0150] In the tests conducted according to the embodiments of this application, the results shown in Figure 4 reflect the outstanding performance of the embodiments of this application in the human-machine sharing control system. The first sub-figure of Figure 4 shows that during obstacle avoidance, the human-machine sharing control system of this application is less affected by obstacle interference than the baseline driving mode, and the vehicle travels a distance of up to 4.5 meters within the same 6 seconds. This indicates that this application can effectively improve the driving efficiency of the vehicle and maintain good original driving task performance without sacrificing driving safety.
[0151] The second sub-figure in Figure 4 illustrates the effect of yaw angle adjustment. The yaw angle adjustment using the method described in this application is smoother, avoiding the overshoot phenomenon that may occur in traditional human-controlled modes. This means that when encountering obstacles, the vehicle can adjust its trajectory more stably, thereby reducing the driver's workload and improving driving comfort and safety.
[0152] The third subplot in Figure 4 illustrates the change in reachability values during obstacle avoidance. Using the method described in this application, the reachability value steadily increases and stabilizes at a low peak as the vehicle approaches the obstacle (approximately at t = 2 seconds), indicating that the human-machine shared control system can respond to obstacles in advance and ensure safety. Compared to the baseline method, this application significantly reduces response latency and avoids higher peak values, demonstrating a more accurate and timely collision avoidance response.
[0153] In summary, the test results demonstrate that this application, in the application of a human-machine shared control system, can improve obstacle avoidance performance while ensuring vehicle safety superior to traditional driving modes.
[0154] To achieve the above embodiments, this application also proposes an intelligent vehicle human-machine shared collision avoidance control system for critical situations. Figure 5 is a schematic diagram of the structure of an intelligent vehicle human-machine shared collision avoidance control system for critical situations provided in an embodiment of this application. As shown in Figure 5, the system includes:
[0155] The data acquisition module 100 collects and transmits information about the vehicle's current status and surrounding environment through a sensing system, which includes a camera, millimeter-wave radar, a combined inertial navigation IMU, and related networking facilities.
[0156] The reachability analysis module 200 uses offline learning and large-scale vehicle data to approximate the set of unreachable states for collision avoidance, and performs offline iterative updates on the Hamilton-Jacobi reachability value function and action-value function through a large amount of pre-collected real vehicle collision avoidance data to obtain the reachability value function and action-value function.
[0157] The driver intention and ability assessment module 300 quantifies the driver's collision avoidance ability and collision avoidance intention based on the reachability value function and action-value function, and dynamically allocates human-machine control weights according to the quantified collision avoidance ability and collision avoidance intention.
[0158] The reinforcement learning control module 400 generates machine operations based on an reachability-inspired reinforcement learning algorithm. The reinforcement learning algorithm combines the driver's collision avoidance intention, collision avoidance ability, and reachability information into the state space.
[0159] The control decision module 500 combines human-machine control weights, driver operations, and machine operations to generate the final operation to be executed. The final operation to be executed is then sent to each actuator in the vehicle in real time through the communication system to perform the collision avoidance operation.
[0160] Regarding the system in the above embodiments, the specific ways in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated here.
[0161] To implement the above embodiments, this application also proposes an electronic device, including: a processor and a memory communicatively connected to the processor; the memory stores computer execution instructions; the processor executes the computer execution instructions stored in the memory to implement the method provided in the foregoing embodiments.
[0162] To implement the above embodiments, this application also proposes a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods provided in the foregoing embodiments.
[0163] To implement the above embodiments, this application also proposes a computer program product, including a computer program that, when executed by a processor, implements the methods provided in the foregoing embodiments.
[0164] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.
[0165] It should be noted that personal information collected from users should be used for legitimate and reasonable purposes and should not be shared or sold outside of these legitimate uses. Furthermore, such collection / sharing should only be conducted after receiving the user's informed consent, including but not limited to notifying the user to read the user agreement / user notice and sign an agreement / authorization that includes authorization of relevant user information before the user uses the function. In addition, any necessary steps must be taken to protect and safeguard access to such personal information data and ensure that others with access to personal information data comply with their privacy policies and procedures.
[0166] This application is intended to provide an implementation scheme for users to selectively prevent the use or access to their personal information data. Specifically, this application is intended to provide hardware and / or software to prevent or block access to such personal information data. Once personal information data is no longer needed, risks can be minimized by restricting data collection and deleting data. Furthermore, where applicable, such personal information is de-identified to protect user privacy.
[0167] In the foregoing descriptions of the embodiments, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. 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. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0168] 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 application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.
[0169] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.
[0170] 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 embodied 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.
[0171] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using 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.
[0172] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.
[0173] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.
[0174] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.
[0175] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this application can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this application can be achieved, and this is not limited herein.
[0176] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.
Claims
1. A method for intelligent vehicle human-machine shared collision avoidance control under critical conditions, comprising: The vehicle collects and transmits information about its current status and surrounding environment through a sensing system, which includes a camera, millimeter-wave radar, a combined inertial navigation IMU, and related network facilities. The set of unreachable states for collision avoidance is approximated using offline learning and large-scale vehicle data. The Hamilton-Jacobi reachability value function and action-value function are updated offline using a large amount of pre-collected real vehicle collision avoidance data to obtain the reachability value function and action-value function. Based on the reachability value function and the action-value function, the driver's collision avoidance ability and collision avoidance intention are quantified, and the human-machine control weights are dynamically allocated according to the quantified collision avoidance ability and collision avoidance intention. Machine operations are generated based on an accessibility-inspired reinforcement learning algorithm, which combines the driver's collision avoidance intention, collision avoidance capability, and accessibility information into the state space. By combining the human-machine control weights, driver operations, and machine operations, a final operation to be executed is generated. This final operation is then sent to each actuator in the vehicle in real time via a communication system to perform the collision avoidance operation.
2. The method according to claim 1, wherein, The set of inaccessible states for collision avoidance is approximated using offline learning and large-scale automotive data. The Hamilton-Jacobi reachability value function and action-value function are iteratively updated offline using a large amount of pre-collected real-vehicle collision avoidance data to obtain the reachability value function and action-value function, including: The obstacle is represented as an elliptical envelope or a T-shaped collision surface, where the parameters of the target obstacle are denoted as [X0,Y0,a,b], which represent the center coordinates and elliptical shape parameters of the obstacle, respectively. The target obstacle is represented as a set of safety constraint states T. Vehicle states within this set are considered dangerous states where a collision has already occurred, and are defined as follows: in, This represents the vehicle's state and kinematic information, where X and Y are the vehicle's global position coordinates. It's the yaw angle, v x ,v y These are the longitudinal velocity and the lateral velocity, respectively, and r represents the yaw rate; The vehicle state and driver operation from a large amount of pre-collected real vehicle collision avoidance data are substituted into the reachability value function network and action value function network corresponding to the safety constraint state set T to calculate the sample reachability value function value and sample action value function value. The sample reachability value function and sample action value are updated offline iteratively to further optimize the accuracy of collision avoidance decisions. After the iteration is completed, the updated reachability value function V is obtained. h Network and Action-Value Function Q h The network will be deployed in intelligent vehicles and advanced driver assistance systems.
3. The method according to claim 2, wherein, Based on the reachability value function and the action-value function, the driver's collision avoidance ability and collision avoidance intention are quantified, and human-machine control weights are dynamically allocated according to the quantified collision avoidance ability and collision avoidance intention, including: The current vehicle state, obstacle envelope, and driver actions collected by the sensor system are substituted into the updated reachability value function V corresponding to the constraint state set T. h Network and Action-Value Function Q h The network yields the reachability value V. h (x) and action-value Q h (x,u d ); Through reachability value V h (x) represents the optimal collision avoidance distance in the current state. The driver's collision avoidance capability (CAA) is calculated using the following formula: Where, α CAA C is the sensitivity adjustment parameter. CAA It is the offset constant; Based on the driver's current operation d Corresponding action-value Q h (x,u d The reachability value V corresponding to the optimal action h (x), calculate the driver's collision avoidance intention (CAI), which is obtained by the following formula: Based on the collision avoidance capability (CAA) and collision avoidance intention (CAI), the human-machine control weight γ is dynamically allocated, and the control weight is calculated using the following formula: γ=max(γ min ,(1-s CAI )(1-s CAA )) Among them, s CAI With s CCA The normalization functions are defined by the following formulas: Where, k CAI1 and k CA11 Control s respectively CAI and s CAA Sensitivity to changes in CAI and CAA, k CAI2 and k CAA2 Define s CAI and s CAA Reaching a threshold of 0.5 represents the balance point between collision avoidance intent and capability.
4. The method according to claim 3, wherein, Machine operations are generated based on an reachability-inspired reinforcement learning algorithm. This algorithm integrates the driver's collision avoidance intention, collision avoidance capability, and reachability information into a state space, including: The driver's collision avoidance intention, collision avoidance capability, accessibility information, and other vehicle dynamic states are explicitly integrated into the state space of the reinforcement learning algorithm. This state space includes: the vehicle's dynamic state x, and the accessibility value function V. h (x) Driver's collision avoidance capability (CAA), driver's collision avoidance intent (CAI), human-machine control weights (γ), and driver and machine control operations (u) d and u m , where the action space u m Includes only the front wheel steering angle δ mf ; In the state space, by combining the vehicle's dynamics and kinematics, a driver action generation model is constructed to simulate the driver's collision avoidance operation during training. The driver action generation model simulates the driver's collision avoidance decision by dynamically adjusting the relative relationship between the driver's input and the obstacle. A reinforcement learning reward function is constructed based on the driver's collision avoidance intention. The reward function is designed based on the goal of reducing human-machine conflict and maintaining the original task performance. The reward function combines the obstacle envelope as a constraint condition to guide the generation of machine control actions during the reinforcement learning process. During reinforcement learning training, by maximizing the above reward function, the machine can prioritize safety while reducing human-machine conflict, and generate the final control action.
5. The method according to claim 4, wherein, In the state space, combining the vehicle's dynamics and kinematics, a driver action generation model is constructed to simulate the driver's collision avoidance maneuvers during training. This model dynamically adjusts the relative relationship between the driver's input and the obstacle to simulate the driver's collision avoidance decisions, including: The current vehicle state, obstacle envelope, and driver control inputs are fed into the driver action generation model. By calculating the relative position and velocity between the vehicle and the obstacle, the driver's steering input is calculated and adjusted to ensure that the driver's preview direction is aligned with the tangent of the obstacle boundary, thereby ensuring that the vehicle can avoid the obstacle. The preview angle θ... c Calculate using the following formula: Among them, (X) trg ,Y trg () represents the point of tangency on the ellipse, located on the boundary of the ellipse as defined below: Where λ is a scaling factor, which is based on the reachability value V. h (x) and the collision avoidance intention CAI adjust the ellipse size, defined as: Where, k λ This is a hyperparameter.
6. The method according to claim 5, wherein, The step involves constructing a reinforcement learning reward function based on the driver's collision avoidance intention. This reward function is designed with the goal of reducing human-machine conflict and maintaining the original task performance. The reward function incorporates obstacle envelopes as constraints to guide the generation of machine control actions during the reinforcement learning process, including: The cooperation between the driver and the machine is optimized by using a reinforcement learning reward function to reduce human-machine conflict. This reward function includes a safety reward R. sf and collaboration reward R co ; The security reward R sf The distance between the vehicle state and the set of unreachable states for collision avoidance is calculated and defined as: Where d0 is the scaling parameter, k sf1 To emphasize the importance of the boundary of advanced driver assistance systems (ADAS) for intelligent vehicles, V h (x)>k sf2 A state that results in severe penalties is designed to ensure that agents learn to avoid unsafe states. The collaboration reward R co Optimizing human-machine collaboration by quantifying the deviation between machine and driver behavior is defined as: R co =-k co ·γ·(in M -in D ) 2 Among them, the penalty item (u M -u D ) 2 k represents the deviation between machine and driver behavior. co It is a scaling factor, and the weighting factor γ reflects the necessity of machine intervention based on the driver's collision avoidance ability and intention.
7. The method according to claim 6, wherein, The process involves combining the human-machine control weights, driver operations, and machine operations to generate a final operation to be executed. This final operation is then sent in real-time to each actuator in the vehicle via a communication system to perform collision avoidance maneuvers, including: Based on the human-machine control weight γ and machine operation u m and driver operation u d Generate the final operation to be executed u f The formula is: in f =γ·u m +(1-γ)·u d The final operation to be executed is transmitted via the communication system. f It sends real-time signals to all actuators in the vehicle, including the active differential steering system, drive system, and braking system, to perform the corresponding collision avoidance maneuvers. During execution, the vehicle's status is monitored in real time through continuous sensor feedback, and the final operation is dynamically adjusted according to the relative position of the vehicle and obstacles to ensure that the vehicle successfully avoids obstacles and returns to the normal driving trajectory.
8. A human-machine shared collision avoidance control system for intelligent vehicles under critical conditions, comprising: The data acquisition module is used to collect and transmit information about the vehicle's current status and surrounding environment through a sensing system, which includes a camera, millimeter-wave radar, a combined inertial navigation IMU, and related networking facilities. The reachability analysis module is used to approximate the set of unreachable states for collision avoidance using offline learning and large-scale vehicle data. It also performs offline iterative updates on the Hamilton-Jacobi reachability value function and action-value function using a large amount of pre-collected real vehicle collision avoidance data to obtain the reachability value function and action-value function. The driver intention and ability assessment module is used to quantify the driver's collision avoidance ability and collision avoidance intention based on the reachability value function and the action-value function, and dynamically allocate human-machine control weights according to the quantified collision avoidance ability and collision avoidance intention. The reinforcement learning control module is used to generate machine operations based on an reachability-inspired reinforcement learning algorithm, which combines the driver's collision avoidance intention, collision avoidance capability, and reachability information into the state space. The control decision module is used to combine the human-machine control weights, driver operations, and machine operations to generate the final operation to be executed, and send the final operation to be executed to each actuator of the vehicle in real time through the communication system to execute the collision avoidance operation.
9. An electronic device, comprising: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1-7.
10. A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-7.