A steering decoupling vehicle intelligent driving data set construction method based on an expert algorithm

Abstract

The application is a kind of steering decoupling vehicle intelligent driving data set construction method based on expert algorithm. It includes the following steps: constructing a simulation scene library; constructing a static road potential field and a dynamic obstacle Gaussian risk field to generate a comprehensive risk cost map; calculating the yaw stability margin, tire grip force margin and normalized speed permission factor to calculate the dynamic reconfigurable index; defining an extended state space, respectively constructing the kinematics differential equations of Ackerman, crab, in-place steering and transition mode, constructing the comprehensive edge cost function of the integrated basic trajectory, environmental risk and dynamic reconfiguration cost to generate discrete motion primitives; designing a mode-aware adaptive heuristic function, selecting heuristic items, and backtracking to generate a rough reference trajectory containing motion mode labels; solving the optimal control sequence; and obtaining the steering decoupling vehicle intelligent driving data set. The application fills the gap in the field of steering decoupling vehicle intelligent driving data set.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving simulation and dataset construction, specifically involving a method for constructing a steering decoupled vehicle intelligent driving dataset based on expert algorithms. Background Technology

[0002] With the deep integration of autonomous driving and drive-by-wire chassis technology, intelligent vehicle platforms with multi-axle independent steering capabilities are attracting increasing attention. These platforms offer multiple driving modes, including Ackerman steering, diagonal steering, and stationary steering, significantly improving maneuverability and passability in complex road conditions. Meanwhile, data-driven end-to-end control methods are demonstrating promising potential in complex scenarios, although their performance is highly dependent on the quality and coverage of the training data.

[0003] However, existing mainstream open-source datasets for end-to-end autonomous driving are mainly designed for traditional Ackerman steering vehicles. The trajectories and control commands contained therein are strictly limited to a single kinematic model, lacking multimodal maneuvering behavior data such as diagonal translation, stationary rotation, and mode switching during travel, which restricts the full potential of advanced chassis hardware.

[0004] Furthermore, current data acquisition schemes mostly focus on pose and environmental perception information, lacking simultaneous recording of the deep physical mechanisms of the chassis. For vehicles with decoupled steering, mode switching decisions depend on dynamic boundary conditions such as yaw stability margin, tire grip margin, and dynamic reconfigurability index. Existing datasets lack explicit steering mode labels and do not align expert strategies with physical feasibility indicators, making it difficult for algorithms to learn safe and stable mode switching logic.

[0005] Therefore, there is an urgent need for a method to construct autonomous driving datasets that can systematically generate multimodal motion behaviors, synchronously record physical mechanism characteristics and expert decision-making truth values, and ensure spatiotemporal alignment of multimodal data, so as to support the research and development and verification of steering-decoupled vehicle intelligent driving systems. Summary of the Invention

[0006] The purpose of this invention is to provide a method for constructing a steering decoupled vehicle intelligent driving dataset based on expert algorithms. Its technical objective is to solve the problems of existing autonomous driving datasets having limited coverage of multi-degree-of-freedom vehicle steering modes and lacking physical mechanism characteristics and expert decision-making logic.

[0007] The technical solution of this invention to solve the above problems is as follows: A method for constructing a steering decoupled vehicle intelligent driving dataset based on expert algorithms, comprising the following steps:

[0008] S1: Build a simulation scene library, import a three-axis vehicle model with independent steering capability and bind a multimodal sensor kit, initialize the vehicle pose and start the simulation to obtain a snapshot of the simulation world;

[0009] S2: Obtain the vehicle state and privileged map grid initialization based on the simulation world snapshot of the current frame, construct the static road potential field and the dynamic obstacle Gaussian risk field respectively, and generate a comprehensive risk cost map through linear weighted superposition and numerical saturation processing;

[0010] S3: Obtain the vehicle's physical state and road adhesion coefficient, calculate the yaw stability margin, tire grip margin and normalized speed allowance factor, and use the multiplication gating mechanism to calculate the dynamic reconfigurability index.

[0011] S4: Define an extended state space covering Ackerman, slant, in-situ turning and transition modes, construct the kinematic differential equations for Ackerman, slant, in-situ turning and transition modes respectively, construct a comprehensive edge cost function that integrates the basic trajectory, environmental risk and dynamic reconstruction cost, and generate discrete motion primitives using the four sets of kinematic differential equations;

[0012] S5: A hybrid A* algorithm is used to perform a global search in a rasterized privileged map. A modality-aware adaptive heuristic function is designed. The minimum value among the Reeds-Shepp distance in the Ackerman mode, the Euclidean distance in the slant mode, and the heading time in the stationary turn mode is selected as the heuristic term. A rough reference trajectory containing motion mode labels is generated by backtracking.

[0013] S6: Perform spline interpolation on the coarse reference trajectory to construct a nonlinear objective functional containing tracking error, control quantity minimization, and comfort terms, and use sequential quadratic programming to solve for the optimal control sequence;

[0014] S7: Collect sensor data, and perform post-processing on the sensor data, the coarse reference trajectory obtained in S5, the optimal control sequence obtained in S6, the yaw stability margin, tire grip margin and normalized speed allowance factor in S3, and the dynamic reconfigurability index to obtain the steering decoupled vehicle intelligent driving dataset.

[0015] Furthermore, the multimodal sensor suite in the S1 includes six surround-view cameras, a top-mounted lidar and front and rear millimeter-wave radars, and a GPS navigation system and inertial navigation system at the center of the vehicle.

[0016] Furthermore, the static road potential field constructed by S2 is represented as follows:

[0017] ,

[0018] in, For nodes The static risk value, These represent the horizontal and vertical positions in the global coordinate system, respectively. The potential energy gain coefficient at the road boundary. This represents the lateral Euclidean distance of the node from the current lane centerline. This refers to the lane width of this road section. The order of the potential field;

[0019] The constructed dynamic obstacle risk field is represented as follows:

[0020] ,

[0021] in, For all dynamic obstacles at grid points The total dynamic risk value superimposed at the point, The peak value constant for obstacle risk. , The first The standard deviation of a dynamic obstacle along its own longitudinal and lateral directions , These are the grid points relative to the first one. The longitudinal and lateral distances between the centers of the obstacles This represents the total number of dynamic obstacles within the current sensing range. For indexing dynamic obstacles;

[0022] The generated comprehensive risk cost map is shown below:

[0023] ,

[0024] in, Grid points The overall risk value at the location, This is the risk saturation threshold. These are the weighting coefficients for the static road potential field. These are the weighting coefficients for the dynamic obstacle risk field.

[0025] Furthermore, S3 specifically refers to:

[0026] S31: Read the vehicle's physical state information and road surface adhesion coefficient;

[0027] S32: Calculate the yaw stability margin:

[0028] ,

[0029] in, For yaw stability margin, The yaw rate is angular velocity. The critical yaw rate, The road surface adhesion coefficient, It is the acceleration due to gravity. This refers to the vehicle's longitudinal speed. It is a smoothing factor;

[0030] S33: Calculate tire grip margin:

[0031] ,

[0032] in, This refers to the tire's grip margin. , These are the vehicle's longitudinal acceleration and lateral acceleration, respectively.

[0033] S34: Calculate the maximum no-slip steering angle, then compare the maximum no-slip steering angle with the mechanical limit steering angle of the corner module vehicle to obtain the normalized speed allowance factor, expressed as follows:

[0034] ,

[0035] in, The maximum no-slip steering angle, Wheelbase As the speed permitting factor, For mechanical limit angle, Sensitive factors;

[0036] S35: The dynamic reconfigurability index is calculated by using a multiplicative gating mechanism to fuse the yaw stability margin of S32, the tire grip margin of S33, and the speed allowance factor of S34, as follows:

[0037] ,

[0038] in, It is a dynamically reconfigurable index. These are the weighting coefficients.

[0039] Furthermore, the physical state information of the vehicle in step S31 includes the velocity vector, yaw rate, longitudinal acceleration, and lateral acceleration.

[0040] Furthermore, S4 specifically refers to:

[0041] S41: An extended state space encompassing Ackermann, slant, in-place turning, and transition modes, represented as follows:

[0042] ,

[0043] in, Indicates the current state. For heading angle, To unify the kinematic steering angle, For sports mode tags, These represent Ackermann steering, swerving, stationary steering, and transition modes, respectively.

[0044] S42: Construct the kinematic differential equations for the Ackermann steering, as follows:

[0045] ,

[0046] in, The state derivative function for Ackermann turning. , These represent the longitudinal and lateral velocity components of the vehicle's geometric center in the global coordinate system. The yaw rate is angular velocity. The longitudinal linear velocity is the rate of change. For acceleration control commands, To unify the rate of change of kinematic steering angle, This refers to the steering angular velocity;

[0047] The kinematic differential equations for the oblique motion are constructed as follows:

[0048] ,

[0049] in, The state derivative function for the diagonal;

[0050] The kinematic differential equations for turning in place are constructed as follows:

[0051] ,

[0052] in, For the in-place turning mode, This is the angular velocity of turning in place.

[0053] The kinematic differential equations for the transition modes are constructed as follows:

[0054] ,

[0055] in, Let be the state derivative function of the transition mode. For braking command, This is the steering drag coefficient. For symbolic functions, To unify the kinematic steering angle for the target, This refers to the rated angular velocity of the steering motor.

[0056] S43: For connecting parent nodes and child nodes any directed edge The ternary comprehensive cost function is calculated as follows:

[0057] ,

[0058] in, For the ternary comprehensive cost function, The cost of the basic trajectory, As a cost to environmental risks, This is the cost of dynamic reconstruction;

[0059] S44: The same-modal extended primitive is generated using the fourth-order Runge-Kutta numerical integration method, as shown below:

[0060] ,

[0061] in, This is the state vector of the child node. The parent node's state vector. , , , All are slope terms. Let be the state derivative function under the current motion mode. The integral time step for state expansion. For control input vectors;

[0062] Generate the belt speed switching primitive, represented as follows:

[0063] ,

[0064] in, The time cost required for the reconstruction process.

[0065] Generate the parking switching primitive, represented as follows:

[0066] ,

[0067] in, for Motion mode tags at any time The target motion pattern.

[0068] Furthermore, S5 specifically refers to:

[0069] S51: A hybrid A* approach is used for global path search. The comprehensive evaluation function for the search nodes is defined as follows:

[0070] ,

[0071] in, For nodes The comprehensive evaluation function value, To represent the distance from the starting node to the current node The actual cumulative cost, These are heuristic weighting coefficients. To start from the current node Heuristic cost estimation to the target node;

[0072] The cumulative cost is the sum of the cumulative cost of the parent node and the edge cost calculated by S43, recursively expressed as:

[0073] ,

[0074] in, child node The actual cumulative cost, parent node The actual cumulative cost;

[0075] S52: Design a modality-aware adaptive heuristic function, expressed as follows:

[0076] ,

[0077] in, Estimate the cost of Ackermann steering mode. Estimate the cost for the diagonal pattern. Estimate the cost of the stationary turning mode. To estimate the reconstruction cost constant, This is a function for calculating the length of Reeds-Shepp curves. The target node state, For the minimum turning radius, ( , ) represents the coordinates of the current node's position, ( , () represents the coordinates of the target location. The maximum permissible linear speed of the vehicle. The heading angle at the target position. The heading angle of the current node. This represents the maximum angular velocity for turning in place.

[0078] S53: Execute node expansion, backtrack to generate a rough reference trajectory, as shown below:

[0079] ,

[0080] in, This is the final set of upper-level coarse reference trajectories. This represents the total number of discrete path points contained in the trajectory. For the first in the trajectory The vehicle extended state vector corresponding to each discrete point. For the first in the trajectory Motion pattern labels corresponding to discrete points.

[0081] Furthermore, S6 specifically refers to:

[0082] S61: Perform cubic spline interpolation on the coarse reference trajectory;

[0083] S62: Construct a nonlinear objective functional containing tracking error, control quantity minimization, and comfort terms, expressed as follows:

[0084] ,

[0085] in, The total algebraic value of the optimization objective function for model predictive control. The prediction time step for model predictive control. To predict the discrete time step index in the time domain, To predict the first in the time domain The vehicle's predicted state vector at each step. For the first The reference trajectory state vector corresponding to the step, To predict the first in the time domain The control input vector of the step, , , These are the weight matrices for the state tracking error, the control input, the control input rate of change, and the terminal state error, respectively. To predict the vehicle's predicted state vector at the end of the time domain, The reference trajectory state vector for predicting the time-domain endpoint;

[0086] S63: The kinematic characteristics, physical limits, dynamic stability, and environmental safety requirements of the vehicle are constructed into a simultaneous set of constraint equations, which serve as the boundary of the feasible region of the optimization problem, as follows:

[0087] ,

[0088] in, Let be the state derivative function under the current motion mode. , These are the lower and upper threshold vectors controlling the input vector, respectively. In the state The dynamic reconfigurability index under the following conditions The preset dynamic stability safety threshold constant. For time-varying risk fields generated based on obstacle prediction trajectories, This is a preset environmental collision risk threshold constant;

[0089] S64: The above nonlinear programming problem is solved using the sequential quadratic programming method to obtain the optimal control input sequence, and the first element in the optimal control sequence is selected. Vehicle control is performed based on the actual control commands given at the current moment.

[0090] Compared with the prior art, the significant advantages of this invention are:

[0091] This application addresses the issues of missing multimodal steering modes and misaligned physical mechanism features in existing datasets by constructing a simulation data generation architecture based on an independent steering chassis. In the decision-making and planning phase, a comprehensive cost map integrating the static road potential field and the dynamic obstacle risk field is built. A dynamic reconfigurable index based on multiplicative gating is designed to accurately define the physical switching boundaries of Ackerman, swerve, and stationary steering. For multimodal motion control, an extended state space is defined, and a hybrid A* algorithm is used to generate discrete motion primitives covering same-modal extension, speed switching, and stop switching. The optimal trajectory is solved by nonlinear model predictive control combining environmental risk and dynamic constraints. Finally, raw sensor data and expert decision ground truth containing physical mechanism features are synchronously integrated to construct a fully stack-aligned multimodal dataset for perception, decision-making, and control. This fills the gap in intelligent driving datasets for steering-decoupled vehicles and provides high-quality data support for iterative verification of learning models for new vehicle configurations. Attached Figure Description

[0092] Figure 1 This is a flowchart of a method for constructing a steering decoupling vehicle intelligent driving dataset based on expert algorithms.

[0093] Figure 2 This is a flowchart of an expert algorithm.

[0094] Figure 3 A schematic diagram of a three-axle vehicle model with independent steering capability and its multimodal sensor suite. Detailed Implementation

[0095] The present invention will now be described in further detail with reference to the accompanying drawings.

[0096] Figure 1 This is a method for constructing a steering decoupling vehicle intelligent driving dataset based on expert algorithms, such as... Figure 1 As shown, the method for constructing a steering decoupling vehicle intelligent driving dataset based on expert algorithms includes:

[0097] S1: Build a simulation scenario library, import a three-axis vehicle model with independent steering capability and bind a multimodal sensor kit, initialize the vehicle pose and start the simulation.

[0098] S2: Initialize the grid space based on the simulation snapshot, construct the static road potential field and the dynamic obstacle Gaussian risk field respectively, and generate a comprehensive risk cost map through linear weighted superposition and numerical saturation processing.

[0099] S3: Obtain the vehicle's physical state and road adhesion coefficient, calculate the yaw stability margin, tire grip margin and normalized speed allowance factor, and calculate the dynamic reconfigurability index using a multiplication gating mechanism.

[0100] S4: Defines an extended state space and kinematic differential equations covering Ackerman, slant, in-situ turning and transition modes, constructs a comprehensive edge cost function that integrates the basic trajectory, environmental risk and dynamic reconstruction cost, and generates discrete motion primitives using four sets of kinematic differential equations.

[0101] S5: A hybrid A* algorithm is used for global search. An adaptive heuristic function for modality awareness is designed. The minimum value among the Reeds-Shepp distance in the Ackerman mode, the Euclidean distance in the slant mode, and the heading time in the stationary turn mode is selected as the heuristic term. A rough reference trajectory containing motion mode labels is generated by backtracking.

[0102] S6: Perform cubic spline interpolation on the coarse reference trajectory to construct a nonlinear objective functional containing tracking error, control quantity minimization, and comfort terms. Under kinematic, dynamic reconfigurability exponent, and environmental safety constraints, use sequential quadratic programming to solve for the optimal control sequence.

[0103] S7: Enable synchronous data acquisition, record raw sensor data, expert decision truth values, and physical mechanism characteristic data including dynamic reconfigurability index, and post-process the raw data to obtain the steering decoupled vehicle intelligent driving dataset.

[0104] In step S1, a simulation scene library is built, a three-axis vehicle model with independent steering capability is imported and a multimodal sensor suite is bound, the vehicle pose is initialized, and the simulation is started.

[0105] The three-axle vehicle, equipped with independent steering capability, is symmetrical front and rear with decoupled steering for each wheel. The multimodal sensor suite includes six surround-view cameras, a lidar on top and millimeter-wave radars at the front and rear, and a global positioning navigation system and inertial navigation system at the center of the vehicle.

[0106] In step S2, the grid space is initialized based on the simulation snapshot, and the static road potential field and the dynamic obstacle Gaussian risk field are constructed respectively. Through linear weighted superposition and numerical saturation processing, a comprehensive risk cost map is generated.

[0107] S21: Communication is established with the simulation server via the simulation platform's Python API interface to synchronously acquire a snapshot of the simulation world for vehicle state acquisition, privileged map parsing, and grid initialization. The vehicle state includes its pose and velocity vectors in the simulation world coordinate system. The privileged map includes lane centerlines, lane widths, and road types. Grid initialization refers to creating a two-dimensional grid map centered on the vehicle, with a vertical dimension of 80 meters and a horizontal dimension of 40 meters, setting the grid resolution to 0.1 meters. Each grid corresponds to a specific simulation world coordinate system. .

[0108] S22: Utilize the waypoint lookup function provided by the Python API of the simulation platform to calculate each grid point. The static risk value is used to construct a static road potential field, which is specifically represented as follows:

[0109] ,

[0110] in, For nodes The static risk value, These represent the horizontal and vertical positions in the global coordinate system, respectively. The potential energy gain coefficient at the road boundary. This represents the lateral Euclidean distance of the node from the current lane centerline. This refers to the lane width of this road section. The order of the potential field.

[0111] S23: Obtain the truth value list of all dynamic obstacles within the local perception range described in S21, and construct a dynamic obstacle risk field.

[0112] S231: Iterate through the list of Actors and filter for dynamic obstacles within the 2D grid map. For the... For a dynamic obstacle, read its transformation matrix to obtain its position and heading, and read its physical properties to obtain its velocity vector and collider size.

[0113] S232: First, calculate the standard deviation of the dynamic obstacle distribution along the vehicle's front and sides to establish the range of the risk field that dynamically changes with speed. Then, transform the grid points from the world coordinate system to the local coordinate system of the obstacle. Finally, use a Gaussian function to calculate the independent risk value of each dynamic obstacle at the grid point, and superimpose the risk values ​​of all dynamic obstacles to generate the dynamic obstacle risk field, as shown below:

[0114] The formula for calculating the standard deviation of the distribution is as follows:

[0115] ,

[0116] in, , The first The standard deviation of a dynamic obstacle along its own longitudinal and lateral directions , The first The length and width of the collision body dimensions of a dynamic obstacle. As a time predictor, For the first The velocity vector of a dynamic obstacle. This represents a lateral safety margin.

[0117] The transformation from the world coordinate system to the local coordinate system of the obstacle is represented as follows:

[0118] ,

[0119] in, , These are the grid points relative to the first one. The longitudinal and lateral distances between the centers of the obstacles This refers to the position of the dynamic obstacle in the world coordinate system. The heading angle of the dynamic obstacle.

[0120] The risk values ​​of all dynamic obstacles are superimposed to generate a dynamic obstacle risk field, as shown below:

[0121] ,

[0122] in, For all dynamic obstacles at grid points The total dynamic risk value superimposed at the point, Let N be the peak obstacle risk constant, and N be the total number of dynamic obstacles within the current perception range. For indexing dynamic obstacles.

[0123] S24: The static road potential field from S22 and the dynamic obstacle risk field from S23 are linearly weighted and superimposed, and then numerical saturation processing is performed to generate a comprehensive risk cost map, as shown below:

[0124] ,

[0125] in, For grid points The overall risk value at the location, This is the risk saturation threshold. These are the weighting coefficients for the static road potential field. These are the weighting coefficients for the dynamic obstacle risk field.

[0126] In step S3, the vehicle's physical state and road surface adhesion coefficient are obtained, the yaw stability margin, tire grip margin and normalized speed allowance factor are calculated, and the dynamic reconfigurability index is calculated using a multiplication gating mechanism.

[0127] S31: Use the Python API of the simulation platform to directly read the vehicle's physical state information and road adhesion coefficient. The vehicle's physical state information includes velocity vector, yaw rate, longitudinal acceleration and lateral acceleration.

[0128] S32: Calculate the yaw stability margin to assess whether the vehicle is currently at risk of fishtailing or sharp turns, as shown below:

[0129] ,

[0130] in, For yaw stability margin, The yaw rate is angular velocity. The critical yaw rate, The road surface adhesion coefficient, It is the acceleration due to gravity. This refers to the vehicle's longitudinal speed. This is a smoothing factor.

[0131] S33: Calculate the tire grip margin, expressed as follows:

[0132] ,

[0133] in, This refers to the tire's grip margin. , These are the vehicle's longitudinal acceleration and lateral acceleration, respectively.

[0134] S34: Calculate the maximum no-slip steering angle, then compare the maximum no-slip steering angle with the mechanical limit steering angle of the corner module vehicle to obtain the normalized speed permission factor. By introducing the speed permission factor, the mode switching cost of the vehicle under extreme conditions such as high speed or large steering increases dramatically, thereby suppressing dangerous dynamic mode switching, as shown below:

[0135] ,

[0136] in, The maximum no-slip steering angle, Wheelbase As the speed permitting factor, For mechanical limit angle, It is a sensitive factor.

[0137] S35: A multiplicative gating mechanism is used to fuse the yaw stability margin of S32, the tire grip margin of S33, and the speed permission factor of S34 to calculate the dynamic reconfigurability index. This index is used to ensure that the ability to switch modes is limited when any physical boundary is triggered, as shown below:

[0138] ,

[0139] in, It is a dynamically reconfigurable index. These are the weighting coefficients.

[0140] In step S4, an extended state space and kinematic differential equations covering Ackerman, slant, in-situ turning and transition modes are defined. A comprehensive edge cost function that integrates the basic trajectory, environmental risk and dynamic reconstruction cost is constructed. Discrete motion primitives are generated using four sets of kinematic differential equations.

[0141] S41: Extend the traditional vehicle pose state into a high-dimensional state space that includes steering pattern labels, as shown below:

[0142] ,

[0143] in, Indicates the current state. For heading angle, To unify the kinematic steering angle, For sports mode tags, These represent Ackermann steering, diagonal steering, stationary steering, and transition modes, respectively.

[0144] in, Defined as the uniform kinematic steering angle of the vehicle chassis steering system, specifically:

[0145] When in Ackermann steering mode This refers to the deflection angle of the virtual front wheel on the first axle; when driving at an angle, Refers to the uniform deflection angle of all wheels; when in a transitional state, It characterizes the instantaneous angular position of the steering actuator during mode switching. The unified kinematic steering angle is used to calculate the time cost of mode switching and the instantaneous direction of motion during the transition.

[0146] S42: Based on the different starting node motion modes in S41, construct kinematic differential equations for different motion modes.

[0147] S421: Construct the kinematic differential equations for Ackermann steering. In this mode, the vehicle performs normal cruise or turning. The Ackermann steering is defined as follows: the first and third axles have opposite rotation angles, the second axle is locked and its rotation angle is zero, and the instantaneous center of rotation is located on the extension line of the second axle, as shown below:

[0148] ,

[0149] in, The state derivative function for Ackermann turning. , These represent the longitudinal and lateral velocity components of the vehicle's geometric center in the global coordinate system. The yaw rate is angular velocity. The rate of change of longitudinal travel speed. For acceleration control commands, To unify the rate of change of kinematic steering angle, This refers to the steering angular velocity.

[0150] S422: Establish the kinematic differential equations for the oblique movement. In this mode, all six wheels maintain the same deflection angle, and the vehicle translates along the wheel direction while maintaining a constant vehicle attitude, as shown below:

[0151] ,

[0152] in, It is the state derivative function of the diagonal line.

[0153] S423: Establish the kinematic differential equations for stationary steering. In this mode, the wheel steering is in a geometrically locked state, the steering angle of each wheel is locked at a specific tangential angle, the steering mechanism does not move, and the rotational power comes from the differential drive of the left and right wheels. The instantaneous center of rotation is located at the geometric center of the vehicle body, as shown below:

[0154] ,

[0155] in, For the in-place turning mode, This is the angular velocity of turning in place.

[0156] S424: Establish the kinematic differential equations for the transition mode. In this mode, the vehicle switches from the initial mode angle to the target mode angle. The steering action consumes longitudinal kinetic energy and is accompanied by a slight lateral drift. Unlike switching after braking, this equation reflects the physical characteristics of speed-based switching, as shown below:

[0157] ,

[0158] in, Let be the state derivative function of the transition mode. For braking command, This is the steering drag coefficient. For symbolic functions, To unify the kinematic steering angle for the target, This is the rated angular velocity of the steering motor.

[0159] S43: For connecting parent nodes and child nodes any directed edge The ternary comprehensive cost function is calculated as follows:

[0160] ,

[0161] in, For the ternary comprehensive cost function, The cost of the basic trajectory, As a cost to environmental risks, This is the cost of dynamic reconstruction.

[0162] S431: Calculate the basic trajectory cost, which guides the vehicle to the target in the shortest time, along the shortest path, and with the fewest unnecessary operations. Its discretized form is as follows:

[0163] ,

[0164] in, , , , These represent the weighting coefficients for the time, distance, reversing, and turning items, respectively. The integral time step for state expansion. This represents the travel arc length corresponding to that edge. This is the reversing indicator function. It takes the value 1 when the longitudinal speed is less than zero, and 0 otherwise.

[0165] S432: Calculate the environmental risk cost using the comprehensive risk cost map in S34, as shown below:

[0166] ,

[0167] in, As a cost to environmental risks, For environmental risk weighting coefficients, This represents the total number of discrete sampling points on the currently calculated path segment. For the first The coordinates of each grid point This represents the tiny distance interval between sampling points.

[0168] S433: The dynamic reconfiguration cost is calculated using the dynamic reconfigurability exponent of S35, as shown below:

[0169] ,

[0170] in, The cost of dynamic reconstruction, Based on the switching cost (constant). The kinetic energy penalty coefficient, , They are respectively and Motion patterns at all times This is a mechanical execution complexity factor used to differentiate the physical cost of switching between different modes.

[0171] S44: Based on S42, four sets of kinematic differential equations are established to generate discrete motion primitives, which are used for the parent node. To child nodes State expansion.

[0172] S441: Generates a modal extension primitive to describe the vehicle's normal driving trajectory generated by different combinations of control inputs while keeping the current steering mode unchanged.

[0173] First, based on the current steering mode, a control input vector is selected from the discrete control space.

[0174] If the current steering mode is Ackermann or diagonal, the control input vector is taken as the Cartesian product of the set of longitudinal linear velocities and the set of steering angles, expressed as:

[0175] ,

[0176] in, For the control input vector, This is the longitudinal linear velocity control variable. For steering angle control, For the set of longitudinal linear velocities, This is the set of steering angles.

[0177] If the current steering mode is stationary steering, the control input vector is taken from the set of stationary steering angular velocities, and the longitudinal linear velocity is forced to be zero, as shown below:

[0178] ,

[0179] in, This is the in-situ steering angle control value. It is the set of angular velocities for turning in place.

[0180] Furthermore, based on the selected control input vector Based on the current parent node state, the acceleration control command and steering angular velocity required to calculate the kinematic differential equation described in step S42 are as follows:

[0181] When in Ackermann turn or diagonal driving, proportional control logic is used to convert the target speed and target turning angle into derivative control quantities, as shown below:

[0182] ,

[0183] in, This is a truncation function used to ensure that control commands do not exceed the vehicle's physical limits. The current longitudinal linear velocity of the parent node. These are the upper and lower limits of vehicle acceleration, respectively. These are the upper and lower limits of the vehicle's steering angular velocity, respectively. The uniform kinematic steering angle of the parent node.

[0184] When in stationary turning mode, directly set the stationary turning angular velocity in the kinematic differential equation to be equal to the stationary turning angle control quantity.

[0185] Secondly, based on the corresponding kinematic differential equations in S42, the fourth-order Runge-Kutta numerical integration method is used for state update to obtain the child node states, as shown below:

[0186] ,

[0187] in, This is the state vector of the child node. The parent node's state vector. , , , All are slope terms. This is the state derivative function under the current motion mode.

[0188] S442: Dynamic transition primitives are generated using the transition mode differential equations in S424 to connect nodes in two different modes. The switching execution strategy is uniquely determined by the dynamic reconfigurability exponent.

[0189] First, the target unified kinematic steering angle is determined according to the target motion mode. If the target is in a stationary turning mode, the target unified kinematic steering angle is the geometric locking tangential angle calculated based on the vehicle's geometric dimensions. If the target is Ackerman turning or diagonal driving, the target unified kinematic steering angle is the steering angle control value selected in S441 for generating the sub-node.

[0190] Secondly, the time cost required for the reconstruction process is calculated, as follows:

[0191] ,

[0192] in, The time cost required for the reconstruction process.

[0193] Finally, based on the comparison between the dynamic reconfigurability index and the preset threshold, either the belt speed switching state generation or the stop switching state generation method is selected to generate child nodes. Specifically, when the dynamic reconfigurability index is greater than or equal to the preset threshold, a belt speed switching primitive is generated, and the vehicle maintains longitudinal movement while adjusting the steering angle, as shown below:

[0194] ,

[0195] When the dynamic reconfigurability index is less than a preset threshold, a parking switching primitive is generated, and the vehicle brakes to a stop to complete the switching, as shown below:

[0196] ,

[0197] in, The target motion pattern.

[0198] In step S5, a hybrid A* algorithm is used for global search. An adaptive heuristic function for modal awareness is designed. The minimum value among the Reeds-Shepp distance in the Ackerman mode, the Euclidean distance in the slant mode, and the heading time in the stationary turn mode is selected as the heuristic term. A rough reference trajectory containing motion mode labels is generated by backtracking.

[0199] S51: Multi-mode kinematic graph and cost function constructed based on S4, using hybrid A* for global path search, and defining a comprehensive evaluation function for search nodes, as follows:

[0200] ,

[0201] in, For nodes The comprehensive evaluation function value, To represent the distance from the starting node to the current node The actual cumulative cost, These are heuristic weighting coefficients. To start from the current node Heuristic cost estimation to the target node.

[0202] The cumulative cost is the sum of the cumulative cost of the parent node and the edge cost calculated by S33, recursively expressed as:

[0203] ,

[0204] in, child node The actual cumulative cost, parent node The actual cumulative cost.

[0205] S52: Design an adaptive heuristic function for modal perception. To accommodate the vehicle's motion characteristics and omnidirectional movement capabilities under nonholonomic constraints, the estimated cost is calculated and the minimum value is taken for different potential motion modes of the current node, as shown below:

[0206] ,

[0207] in, Estimate the cost of Ackermann steering mode. Estimate the cost for the diagonal pattern. Estimate the cost of the stationary turning mode. To estimate the reconstruction cost constant, This is a function for calculating the length of Reeds-Shepp curves. The target node state, For the minimum turning radius, ( , ) represents the coordinates of the current node's position, ( , () represents the coordinates of the target location. The maximum permissible linear speed of the vehicle. The heading angle at the target position. The heading angle of the current node. This represents the maximum angular velocity for turning in place.

[0208] S53: Execute node expansion, select the node with the smallest comprehensive evaluation function value from the priority queue as the current parent node, and generate a set of child nodes using discrete motion primitives.

[0209] S54: Backtracking to generate a coarse reference trajectory. The search stops when a search node enters the allowable error neighborhood of the target point. The search proceeds backward from the target node, tracing the parent node pointer back to the starting point. The extracted state sequence and corresponding control mode sequence are the upper-level coarse planning results, as shown below:

[0210] ,

[0211] in, This is the final set of upper-level coarse reference trajectories. This represents the total number of discrete path points contained in the trajectory. For the first in the trajectory The vehicle state corresponding to each discrete point.

[0212] In step S6, cubic spline interpolation is performed on the coarse reference trajectory to construct a nonlinear objective functional containing tracking error, control quantity minimization, and comfort terms. Under kinematic, dynamic reconfigurability exponent, and environmental safety constraints, sequential quadratic programming is used to solve for the optimal control sequence.

[0213] S61: Perform cubic spline interpolation on the coarse reference trajectory output from S54 to generate a prediction time-domain length of... The local reference trajectory is obtained, where spline interpolation is performed on continuous state variables, and zero-order preservation is used to augment the discrete mode labels. The generated mode sequence will serve as a fixed index for selecting the dynamic equation at each time step in the prediction time domain, and will not participate in gradient optimization. The system state vector is defined. , control input vector .

[0214] S62: Construct a nonlinear optimization objective functional in the finite time domain. This function consists of three parts: a trajectory tracking error term, a control quantity minimization term, and a comfort smoothing term, as shown below:

[0215] ,

[0216] in, The total algebraic value of the optimization objective function for model predictive control. The prediction time step for model predictive control. To predict the discrete time step index in the time domain, To predict the first in the time domain The vehicle's predicted state vector at each step. For the first The reference trajectory state vector corresponding to the step, To predict the first in the time domain The control input vector of the step, , , These are the weight matrices for the state tracking error, the control input, the control input rate of change, and the terminal state error, respectively. To predict the vehicle's predicted state vector at the end of the time domain, This is the reference trajectory state vector for predicting the endpoint in the time domain.

[0217] S63: In the optimization process, the vehicle's kinematic characteristics, physical limits, dynamic stability, and environmental safety requirements are constructed into a simultaneous set of constraint equations, which serve as the boundary of the feasible region of the optimization problem, as follows:

[0218] ,

[0219] in, Let be the state derivative function under the current motion mode. , These are the lower and upper threshold vectors controlling the input vector, respectively. In the state The dynamic reconfigurability index under the following conditions The preset dynamic stability safety threshold constant. For time-varying risk fields generated based on obstacle prediction trajectories, This is a preset environmental collision risk threshold constant.

[0220] S64: The above nonlinear programming problem is solved using the sequential quadratic programming method to obtain the optimal control input sequence, as shown below:

[0221] ,

[0222] in, To optimize the calculated optimal control sequence over the entire prediction time domain, To represent the first in the sequence The optimal control input vector for the step. This represents the total number of output control sequences.

[0223] S65: Select the first element in the optimal control sequence. Vehicle control is performed based on the actual control commands given at the current moment.

[0224] In step S7, synchronous data acquisition is started, and the original sensor data, expert decision truth values, and physical mechanism feature data containing dynamic reconfigurability index are recorded. The original data is post-processed to obtain the steering decoupled vehicle intelligent driving dataset.

Claims

1. A method for constructing a steering decoupling vehicle intelligent driving dataset based on an expert algorithm, characterized in that, The method comprises the following steps: S1: Constructing a simulation scene library, importing a three-axle vehicle model with independent steering capability and binding a multi-modal sensor suite, initializing the vehicle pose, starting the simulation, and obtaining a simulation world snapshot; S2: According to the simulation world snapshot of the current frame, the self-vehicle state and the privileged map grid are initialized, a static road potential field and a dynamic obstacle Gaussian risk field are constructed respectively, a comprehensive risk cost map is generated through linear weighting superposition and numerical saturation processing; S3: Obtain the physical state of the ego vehicle and the road adhesion coefficient, calculate the yaw stability margin, tire grip margin and normalized speed permission factor, and calculate the dynamic reconfigurable index using the multiplication gating mechanism; S4: Define an extended state space covering Ackermann, crab, in-place steering and transition modes, construct the kinematic differential equations of Ackermann, crab, in-place steering and transition modes respectively, construct a comprehensive edge cost function that integrates basic trajectories, environmental risks and dynamic reconstruction costs, and generate discrete motion primitives using four sets of kinematic differential equations; S5: A hybrid A* algorithm is used to perform global search in the grid-based privileged map, an adaptive heuristic function is designed, the minimum value of the Ackermann mode Reeds-Shepp distance, the crab mode Euclidean distance and the in-place steering mode heading time is selected as the heuristic item, and a rough reference trajectory containing a motion mode label is generated by backtracking; S6: Perform spline interpolation on the rough reference trajectory, construct a nonlinear objective functional containing tracking error, control amount minimization and comfort items, and solve the optimal control sequence using sequential quadratic programming; S7: Collect sensor data, process the sensor data, the rough reference trajectory obtained in S5, the optimal control sequence obtained in S6, the yaw stability margin, tire grip margin and normalized speed permission factor, and the dynamic reconfigurable index in S3, to obtain the steering decoupling vehicle intelligent driving dataset.

2. The method of claim 1, wherein, The multi-modal sensor suite in S1 includes six surround-view cameras, a laser radar on the top and millimeter wave radars in front and back, as well as a global positioning navigation system and an inertial navigation system in the center of the vehicle body.

3. The method of claim 2, wherein, The static road potential field constructed in S2 is represented as follows: ， in, For nodes The static risk value, These represent the horizontal and vertical positions in the global coordinate system, respectively. The potential energy gain coefficient at the road boundary. This represents the lateral Euclidean distance of the node from the current lane centerline. Lane width, The order of the potential field; The dynamic obstacle risk field constructed is represented as follows: ， wherein, is the total dynamic risk value superimposed at the grid point by all dynamic obstacles, is the obstacle risk peak constant, , are the standard deviations of the first dynamic obstacle along its longitudinal and lateral directions, respectively, , are the longitudinal and lateral distances of the grid point relative to the center of the first obstacle, respectively, is the total number of dynamic obstacles within the current perception range, is the index of the dynamic obstacle. The comprehensive risk cost map generated is represented as follows: ， wherein, grid points a combined risk value at the grid point, is a risk saturation threshold, is a weight coefficient of the static road potential field, is a weight coefficient of the dynamic obstacle risk field.

4. The method of claim 3, wherein, S3 is specifically: S31: Read the physical state information of the ego vehicle and the road adhesion coefficient; S32: Calculate the yaw stability margin: ， wherein is a yaw stability margin, is a yaw angle velocity, is a critical yaw angle velocity, is a road surface adhesion coefficient, is a gravitational acceleration, is a longitudinal driving speed of the vehicle, is a smoothing factor; S33: Calculate the tire grip margin: ， wherein, is the tire grip margin, , are the longitudinal and lateral acceleration of the vehicle, respectively; S34: Calculate the maximum no-sideslip steering angle, then compare it with the mechanical limit angle of the angular module vehicle to obtain the normalized speed permission factor, which is represented as follows: ， wherein, is the maximum cornering angle without side slip, is the wheel base, is the speed permission factor, is the mechanical limit angle, is the sensitivity factor; S35: Use the multiplication gating mechanism to fuse the yaw stability margin of S32, the tire grip margin of S33 and the speed permission factor of S34 to calculate the dynamic reconfigurable index, which is represented as follows: ， wherein, is a dynamic reconfigurable index, is a weight coefficient.

5. The method of claim 4, wherein, The physical state information of the ego vehicle in step S31 includes the velocity vector, yaw angular velocity, longitudinal acceleration and lateral acceleration of the vehicle.

6. The method of claim 4, wherein, S4 is specifically: S41: The extended state space covering Ackermann, crab, in-place steering and transition modes is represented as follows: ， wherein, denotes the current state, is the heading angle, is the unified kinematic steering angle, is the motion mode label, representing Ackermann steering, crabbing, spin steering and transition mode, respectively; S42: Construct the kinematic differential equations of Ackerman steering, which are shown as follows: ， wherein, is a state derivative function for Ackerman steering, , are longitudinal and lateral velocity components of the vehicle geometric center in the global coordinate system, respectively, is the yaw rate, is the longitudinal linear velocity change rate, is the acceleration control command, is the uniform kinematic steering angle change rate, is the steering angle velocity; Construct the kinematic differential equations of crabbing, which are shown as follows: ， wherein is the derivative of the state function with respect to the slant. Construct the kinematic differential equations of spin turning, which are shown as follows: ， wherein, is a state derivative function for the in-place turning mode, is an in-place turning angular velocity; Construct the kinematic differential equations of transition mode, which are shown as follows: ， wherein is a state derivative function for the transition mode, is a braking command, is a steering resistance coefficient, is a sign function, is a target unified kinematic steering angle, is a nominal execution angular velocity of the steering motor; S43: For connecting parent nodes and child nodes any directed edge The ternary comprehensive cost function is calculated as follows: ， wherein, is a ternary comprehensive cost function, is a base trajectory cost, is an environmental risk cost, is a dynamic reconstruction cost; S44: Generate the same-mode extension primitives by using the fourth-order Runge-Kutta numerical integration method, which are shown as follows: ， wherein is the child state vector, is the parent state vector, , , , are all slope terms, is the state derivative function in the current motion mode, is the integration time step of the state expansion, is the control input vector; Generate the speed switching primitives, which are shown as follows: ， wherein, time cost required for the reconstruction process; Generate the parking switching primitives, which are shown as follows: ， wherein, is a motion pattern label at the moment, is the target motion pattern.

7. The method of claim 6, wherein, S5 is specifically: S51: Perform global path search by using hybrid A*, define the comprehensive evaluation function of search nodes, which are shown as follows: ， wherein, is a combined evaluation function value for the node , is an actual accumulated cost representing the path from the start node to the current node , is a heuristic weight coefficient, is a heuristic estimated cost from the current node to the goal node. The cumulative cost is the sum of the cumulative cost of the parent node and the edge cost calculated by S43, which is recursively shown as follows: ， in, child node The actual cumulative cost, parent node The actual cumulative cost; S52: Design the adaptive heuristic function of mode perception, which is shown as follows: ， in, Estimate the cost of Ackermann steering mode. Estimate the cost for the diagonal pattern. Estimate the cost of the stationary turning mode. To estimate the reconstruction cost constant, This is a function for calculating the length of Reeds-Shepp curves. The target node state, For the minimum turning radius, ( , ) represents the coordinates of the current node's position, ( , () represents the coordinates of the target location. The maximum permissible linear speed of the vehicle. The heading angle at the target position. The heading angle of the current node. This represents the maximum angular velocity for turning in place. S53: Perform node expansion, backtrack to generate a rough reference trajectory, which is shown as follows: ， wherein, is the final generated set of upper-level coarse reference trajectories, is the total number of discrete path points contained in the trajectory, is the vehicle extended state vector corresponding to the discrete point in the trajectory, is the motion pattern label corresponding to the discrete point in the trajectory.

8. The method of claim 7, wherein, S6 is specifically: S61: Perform cubic spline interpolation on the rough reference trajectory; S62: Construct a nonlinear objective functional containing tracking error, control quantity minimization and comfort term, which is shown as follows: ， wherein, is the total value of the optimization objective function of the model predictive control, is the prediction horizon step size of the model predictive control, is the discrete time step index within the prediction horizon, is the vehicle predicted state vector at the step within the prediction horizon, is the reference trajectory state vector corresponding to the step, is the control input vector at the step within the prediction horizon, , , are the weight matrix of the state tracking error, the weight matrix of the control input, the weight matrix of the control input rate of change, the weight matrix of the terminal state error, respectively, is the vehicle predicted state vector at the end of the prediction horizon, is the reference trajectory state vector at the end of the prediction horizon; S63: Construct the kinematic characteristics, physical limits, dynamic stability and environmental safety requirements of the vehicle as a set of simultaneous constraint equations, which are shown as follows: ， wherein, is a state derivative function under the current motion mode, , are lower and upper threshold vectors of the control input vector, respectively, is a dynamically reconfigurable index under the state , is a preset dynamic stability safety threshold constant, is a time-varying risk field generated based on the predicted trajectory of the obstacle, is a preset environmental collision risk threshold constant; S64: Adopting the sequence quadratic programming method, solving the nonlinear objective functional of S62 under the constraint equation group of S63, obtaining the optimal control input sequence, and selecting the first element in the optimal control sequence Vehicle control is performed as the actual control command at the current time.

Images