Co-driving system permission allocation and planning control method considering driving intention and state

Abstract

The invention belongs to the technical field of vehicle control, and relates to a co-driving system permission allocation and planning control method considering driving intention and state, which comprises the following steps: step 1, constructing a driver intention recognition model, extracting space and time characteristics through multi-modal data acquisition by combining a CNN model and an LSTM model, and recognizing the driving intention in real time after fusion; step 2, trajectory planning decision design: judging lane changing feasibility according to the identified driving intention, generating a safe trajectory cluster, and realizing trajectory planning; 3, the man-machine optimal weight is solved in real time, and smooth cooperation and safe switching between the driver and the co-driving system are achieved; and step 4, solving a motion control quantity, and realizing high-precision tracking control of the vehicle on an expected trajectory by adopting a self-adaptive model prediction control method based on the optimal lane changing trajectory. The method has the advantages that the LSTM model and the CNN model are fused for intention recognition of a driver, and the time features and the space features are fused for ensuring relatively high recognition accuracy.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of vehicle control, and relates to permission allocation, trajectory planning and motion control in a man-machine co-driving vehicle, in particular to a co-driving system permission allocation and planning control method considering driving intention and state. BACKGROUND

[0002] In the man-machine co-driving mode, the driver and the automatic driving system share the driving task to varying degrees and cooperatively complete the planning and control of the vehicle. This mode can not only exert the flexibility of decision-making and intuitive judgment advantages of the driver in complex working conditions, but also utilize the data processing capacity and high-precision control capacity of the automatic driving system, so as to comprehensively improve the safety, comfort and efficiency. However, in the lane changing scene, unreasonable allocation of man-machine weight, unreasonable planning of lane changing trajectory and low trajectory tracking accuracy may all lead to man-machine conflict. Such conflict not only reduces the driving experience, but also may cause instability of vehicle operation, and even cause traffic accidents. Therefore, a man-machine permission allocation and planning control method needs to be designed to ensure the coordination of man-machine intention during lane changing and effective driving assistance according to the real-time state of the driver. SUMMARY

[0003] In view of the above technical problems and defects, the purpose of the present application is to provide a co-driving system permission allocation and planning control method considering driving intention and state. The method is based on the recognition result of the lane changing intention of the driver, and designs lane changing feasibility judgment conditions for trajectory planning and motion control of the co-driving system, and at the same time, allocates the control permission of man and machine based on the real-time state of the driver, effectively realizes the coordination of man-machine driving intention under the premise of vehicle safety, and provides effective support for the development and landing of man-machine co-driving technology.

[0004] To achieve the above purpose, the present application adopts the following technical solutions: A co-driving system permission allocation and planning control method considering driving intention and state, comprising the following steps: Step 1. Driver intention recognition model construction, through multi-modal data acquisition, combining CNN model and LSTM model to extract spatial and temporal features, and fusing to recognize driving intention in real time; Step 2. Trajectory planning decision design, judging the lane changing feasibility according to the recognized driving intention, generating a safe trajectory cluster, and selecting the optimal lane changing trajectory by using multi-objective scoring method to realize trajectory planning; Step 3. Real-time solving of man-machine optimal weight, dynamically adjusting the man-machine weight allocation according to the driver's intention and steering wheel angle fluctuation, realizing smooth cooperation and safe switching between the driver and the co-driving system; Step 4. Motion control amount solving, based on the optimal lane changing trajectory, using an adaptive model predictive control method, combining vehicle speed dynamic adjustment of the prediction and control time domain, to achieve high-precision tracking control of the vehicle to the desired trajectory.

[0005] As a preferred embodiment of the present application, the following steps are further included in step 1: Step 1.1. Real-time multi-modal data acquisition, in vehicle operation, real-time data acquisition using sensors: (1) Time series data: real-time state from the vehicle CAN bus, including position, speed, acceleration, steering wheel angle, and light signals, (2) Spatial feature data: real-time images, lane lines, and relative positions of surrounding vehicles obtained through environmental perception devices, time series data is sliced according to fixed time windows to form sliding time windows; at the same time, spatial feature data is dynamically updated frame by frame to match time series data, wherein the environmental perception devices include cameras, radars or lidars; Step 1.2. Constructing a CNN model to process spatial features; the input image data is preprocessed by size scaling and normalization to adapt to the input format of the CNN model, local spatial features are extracted using convolution layers, including lane line shape, relative position of front or side vehicles, dimension is reduced using pooling layers to retain pre-set key features while reducing computational load to meet real-time requirements, and finally obtain spatial feature vectors; Step 1.3. Constructing an LSTM model to process time series features, according to the sliding time window, input a pre-set period of recent time series data into the LSTM model, use multiple LSTM modules of the LSTM model to process time series sequentially, capture the dynamic trend of driving behavior, output time series feature vectors representing continuous dynamic features of the vehicle; Step 1.4. Real-time feature fusion, at the end of each time window, fuse the features extracted by the LSTM model and the CNN model, combine the time series feature vectors output by the CNN model and the spatial feature vectors output by the CNN model through feature concatenation, input the fused feature vectors into a fully connected layer for further processing and compression to form high-level semantic features; Step 1.5. Classification output of recognition results, based on the fused high-level semantic features, real-time prediction of driving intent, use the fully connected layer to map the high-level semantic features to the classification space, calculate the probability of each driving intent category through the Softmax activation function, the output driving intent categories include: left lane change, right lane change, and straight ahead.

[0006] As a preferred embodiment of the present application, the following steps are further included in step 2: Step 2.1. Intention recognition result input, the driving intention category output by the driving intention recognition model is input to the trajectory planning decision module, wherein the driving intention category includes left lane changing, right lane changing, and keeping straight; Step 2.2. Lane changing feasibility determination, the intention recognition result of keeping straight does not output a lateral control quantity, and the intention recognition results of left lane changing and right lane changing first perform lane changing feasibility determination. In the geodetic coordinate system, the vehicle is assumed to be a rectangle, and there is no intersection between the vehicle and other rectangles during the lane changing process with the maximum lateral acceleration. The lane changing feasibility condition is satisfied. When the lane changing feasibility is satisfied, the shared driving system generates a lane changing trajectory to perform lane changing assistance. If the lane changing feasibility is not satisfied, the following situations occur: Case 1: The driver does not effectively judge the state of the vehicle behind the target lane and takes the lane changing behavior too early, which has a collision risk due to the small longitudinal distance between the vehicle and the vehicle behind the target lane. Case 2: The vehicle in front of the current lane brakes urgently, and the vehicle does not take timely and effective lane changing to avoid obstacles, which has a collision risk due to the small longitudinal distance between the vehicle and the vehicle in front of the original lane or the target lane.

[0007] As a preferred embodiment of the present application, the following steps are further included in step 2: Step 2.3. For case 1 in step 2.2 where the lane changing feasibility is not satisfied, the shared driving system generates a desired trajectory for straight driving and takes over the vehicle control to ensure safety. Step 2.4. For case 2 in step 2.2 where the lane changing feasibility is not satisfied, the shared driving system takes over the vehicle control to ensure safety, controls the vehicle to brake and decelerate to create space for subsequent lane changing actions, so that when the driver's lane changing intention is recognized again, the lane changing feasibility determination can be satisfied, and a safe desired lane changing trajectory can be generated.

[0008] As a preferred embodiment of the present application, the following steps are further included in step 2: Step 2.5. Trajectory cluster construction; when the lane changing feasibility is satisfied, the desired trajectory is designed with longitudinal uniform speed and transverse quintic polynomial. First, the maximum speed of the vehicle in front of the current lane is , and the maximum braking deceleration is Multiple repeated tests are performed in these four lane changing scenarios. The mean characteristics and standard deviation characteristics of the lane changing time are calculated for each scenario. The lower bound is , the upper bound is , and the step size is 0.2s. A series of lane changing times are generated to finally construct a two-dimensional trajectory cluster for each lane changing scenario.

[0009] As a preferred embodiment of the present application, the following steps are further included in step 2: Step 2.6. Optimal lane-changing trajectory solving; set a multi-objective scoring method to select the trajectory with the highest score as the optimal lane-changing trajectory under the current scene, and consider the following four sub-goals in the multi-objective scoring: stability goal, efficiency goal, comfort goal and anthropomorphism goal, and the results are as follows: ; In the formula, is the final score of the trajectory for calculation and comparison,

[0010] respectively, the stability goal, the efficiency goal, the comfort goal and the anthropomorphism goal score results, k 1、 k 2、 k 3、 k 4 is the weight corresponding to each sub-goal score, and the calculation method of each index score is as follows: ; In the formula, is the maximum lateral acceleration under each planned trajectory, is the gravity acceleration, is the lane-changing time required by the evaluated trajectory, that is, the lane-changing time counted for each scene, is the longest lane-changing time under the planned trajectory, that is, , is the maximum lateral jerk under the planned trajectory, is the reference value of the lateral jerk, is the mean characteristic of the lane-changing time counted for each scene, is the standard deviation characteristic of the lane-changing time counted for each scene.

[0011] As preferred in the present application, in step 3, the following steps are further included: Step 3.1. When the driver intention recognition result is to keep straight, the driver weight is set to 1 and the co-driving system weight is set to 0; Step 3.2. When the driver intention recognition result is left lane-changing or right lane-changing, and the lane-changing feasibility condition is not met, the driver weight is set to 0 and the co-driving system weight is set to 1; Step 3.3. When the driver intention recognition result is left lane-changing or right lane-changing, and the lane-changing feasibility condition is met, first, a rolling time window of 0.5s is set to solve the standard deviation of the steering wheel angle of the driver under each window, and based on the statistical result of the logarithmic value of the standard deviation of the steering wheel angle, a lognormal distribution is constructed, and the probability density function is: ; In the formula, is the standard deviation of the steering wheel angle, and is the mean and standard deviation of the logarithm of the steering wheel angle standard deviation, and the expression of the cumulative distribution function thereof is: The driver and the co-driving system under the human-machine co-driving are the driver real-time weight and the co-driving system real-time weight under the current time window and defined as follows: In the formula, is the driver weight under the time window , the weight of the driver at the initial time of the lane change is required to be 1, the weight value of the driver under the adjacent time window changes by no more than 0.2, and in other cases, the weight value of the driver is 1 minus the cumulative distribution function value of the steering wheel angle standard deviation under the corresponding window.

[0012] As a preferred embodiment of the present application, step 4 further comprises the following steps: Step 4.1. Based on the optimal lane changing trajectory obtained by solving, an adaptive model predictive control method is applied to track the expected trajectory, and the model state space is defined , wherein the state quantity is defined as follows: is the lateral velocity and the longitudinal velocity, and is the vehicle yaw angle and the yaw angular velocity, and is the lateral position and the longitudinal position of the vehicle under the geodetic coordinate system, the control quantity input and the output are defined as follows:

[0013] As a preferred embodiment of the present application, step 4 further comprises the following steps: Step 4.2. The co-driving system constructs a quadratic programming optimization problem with the control increment and the output deviation as the target based on the adaptive model predictive control method established in step 4.1. The objective function of the problem comprehensively considers the trajectory tracking accuracy and the control smoothness of the vehicle, and is defined as follows: The weighted sum of the square sum of the deviation between the expected trajectory in the prediction time domain and the actual output and the square sum of the control input increment is taken as the minimum objective function, and the control sequence is solved under the conditions of satisfying the vehicle dynamics constraint, the control quantity constraint and the control increment constraint.​​​​​​​ The quadratic programming problem is expressed as: ; where, is the expected output in the prediction horizon, is the model predicted output, is the control increment of the i-th step, is the control increment of the i-th step, and are the weighting matrices of the output error and the control increment, respectively, is the length of the prediction horizon, is the length of the control horizon; The co-driving system solves the quadratic programming problem in real time using the active set method to obtain a series of control input increments and slack factors in the control horizon. The co-driving system uses the first control quantity in the control sequence as the actual control input at the current time to act on the vehicle, thereby achieving dynamic tracking of the expected lane-changing trajectory. Subsequently, the next time is rolled in optimization until the entire trajectory tracking control process is completed.

[0014] As a preferred embodiment of the present application, step 4 further comprises the following steps: Step 4.3. Adaptability of the adaptive model predictive control method is used to set the upper limit of the prediction horizon and the prediction horizon The relationship between the longitudinal vehicle speed and the change is set as follows: ; where, is a proportional coefficient, , are the longitudinal low-speed threshold and the longitudinal high-speed threshold, respectively, is the upper limit of the prediction horizon.

[0015] Advantages and benefits of the present application: (1) The method provided by the present application fuses LSTM model and CNN model for driver intention recognition, and fuses time features and spatial features to ensure high recognition accuracy.

[0016] (2) The method provided by the present application determines whether the co-driving system will assist and how to assist based on the results of driver lane-changing intention recognition and lane-changing feasibility condition determination. In the lane-changing assistance, a multi-target dynamic weight distribution is set to further reduce human-machine conflict and improve driving experience.

[0017] (3) The method provided by the present application dynamically adjusts the weight between man and machine by constructing a functional relationship between the weight distribution value between man and machine and the standard deviation state of the steering wheel angle of the driver in the window, while ensuring that the weight fluctuation is not too large, ensuring safety while also ensuring driving comfort.

[0018] (4) The method provided by the application sets the prediction time domain and the control time domain in model predictive control as a function of the longitudinal velocity, effectively ensuring the model solving speed and the trajectory tracking accuracy.

[0019] (5) The method provided by the application is mainly divided into four parts: intention recognition, trajectory planning, permission allocation and motion control, each part is relatively independent and organically connected, and personalized lane changing assistance can be provided for different drivers under the premise of ensuring safety, effectively reducing the man-machine conflict. BRIEF DESCRIPTION OF DRAWINGS

[0020] Other objects and results of the application will be more apparent and easy to understand by referring to the following description in conjunction with the accompanying drawings, and with a more comprehensive understanding of the application. In the drawings: Figure 1 The logic block diagram provided for the embodiment; Figure 2 The critical collision time diagram of the left side lane changing and the front vehicle of the original lane prone to scratching collision in the lane changing feasibility determination provided for the embodiment; Figure 3 The critical collision time diagram of the left side lane changing and the front vehicle of the target lane prone to scratching collision in the lane changing feasibility determination provided for the embodiment; Figure 4 The critical collision time diagram of the left side lane changing and the rear vehicle of the target lane prone to scratching collision in the lane changing feasibility determination provided for the embodiment; Figure 5 The flowchart of the trajectory planning decision scheme provided for the embodiment. DETAILED DESCRIPTION

[0021] A co-driving system permission allocation and planning control method considering driving intention and state, comprising the following steps: Step 1. Driver intention recognition model construction, through multi-modal data acquisition (such as vehicle state, environment image), combining CNN model and LSTM model to extract spatial and temporal features, and after fusion, real-time recognition of driving intention (left lane changing, right lane changing or straight driving); Step 1.1. Real-time multi-modal data acquisition, in the process of vehicle running, data is collected in real time by multiple sensors: (1) time series data: real-time state is obtained from the CAN (Controller Area Network) bus of the vehicle, including position, speed, acceleration, steering wheel angle, and light signal, (2) spatial feature data: real-time images, lane lines, and relative positions of surrounding vehicles are obtained by environmental perception devices, time series data is sliced according to a fixed time window (1 second of data, sampled 10 times per second) to form a sliding time window; at the same time, spatial feature data is dynamically updated frame by frame to match the time series data, wherein the environmental perception devices include cameras, radars, or lidars; Step 1.2. Constructing a CNN (Convolutional Neural Network) model to process spatial features; the input image data is preprocessed by size scaling and normalization to adapt to the input format of the CNN model, local spatial features are extracted using convolutional layers, including lane line shape, relative position of front vehicle or side vehicle, dimension is reduced using pooling layers to retain preset key features while reducing computational load to meet real-time requirements, and finally spatial feature vectors are obtained; Step 1.3. Constructing an LSTM (Long Short-Term Memory) model to process time series features, according to the sliding time window, a predetermined period of recent time (1 second in the past) time series data is input into the LSTM model, multiple LSTM modules of the LSTM model are used to process the time series in turn to capture the dynamic trend of driving behavior, and time series feature vectors are output to represent the continuous dynamic features of the vehicle; Step 1.4. Real-time feature fusion, at the end of each time window, the features extracted by the LSTM model and the CNN model are fused, the time series feature vectors output by the CNN model and the spatial feature vectors output by the CNN model are combined through feature concatenation (Concatenate), the fused feature vectors are input into a fully connected layer for further processing and compression to form high-level semantic features; Step 1.5. Classification output of recognition results, based on the fused high-level semantic features, driving intent is predicted in real time, high-level semantic features are mapped to a classification space using a fully connected layer, the probability of each driving intent category is calculated through a Softmax activation function, and the output driving intent categories include: left lane change, right lane change, and straight ahead.

[0022] Step 2. Trajectory planning decision design, according to the recognized driving intent to judge the feasibility of lane changing, a safe trajectory cluster is generated, and the optimal lane changing trajectory is selected by using a multi-objective scoring method to achieve efficient, safe, and anthropomorphic trajectory planning; Step 2.1. Intention recognition result input, the driving intention category output by the driver intention recognition model is input to the trajectory planning decision module, wherein the driving intention category includes left lane changing, right lane changing, and keeping straight; Step 2.2. Lane changing feasibility determination, the shared driving system does not output a lateral control amount under the intention recognition result of keeping straight, under the intention recognition results of left lane changing and right lane changing, first, the vehicle is assumed to be a rectangle in the geodetic coordinate system (XOY), and there is no intersection between the self-vehicle and other rectangles (vehicles) during the lane changing process with the maximum lateral acceleration, thus satisfying the lane changing feasibility condition. When the lane changing feasibility is satisfied, the shared driving system generates a lane changing trajectory to assist lane changing. If the lane changing feasibility is not satisfied, the following situations occur: Case 1: The driver does not effectively judge the state of the vehicle behind the target lane and takes the lane changing behavior too early, which has a collision risk due to the small longitudinal distance between the self-vehicle and the vehicle behind the target lane. Case 2: The vehicle in front of the current lane brakes urgently, and the self-vehicle does not take timely and effective lane changing to avoid obstacles, which has a collision risk due to the small longitudinal distance between the self-vehicle and the vehicle in front of the original lane or the target lane.

[0023] The calculation of the lane changing feasibility condition is as follows, given the coordinates of the center of mass , first, the yaw angle of the center of mass of the vehicle is solved : ; In the formula, is the lateral velocity of the center of mass of the vehicle, is the longitudinal velocity of the center of mass of the vehicle.

[0024] In the process of left lane changing, there are three different dangerous collision situations, and the dangerous scenarios of right lane changing are the same as those of left lane changing, so only the different dangerous scenarios of left lane changing are considered here: The first kind of left lane changing is prone to scraping collision with the vehicle in front of the original lane, and the right upper vertex of the self-vehicle rectangle is most likely to collide with the left rear vertex of the rear of the front vehicle. The calculation formula of the position coordinates of the right front vertex of the self-vehicle during the lane changing process is: ; In the formula, is the distance from the center of mass to the front end of the vehicle, is the distance from the center of mass to the two sides of the vehicle, is the coordinate position of the center of mass in the geodetic coordinate system, and let the coordinate of the left rear vertex of the rectangle of the front vehicle in the original lane be . It can be seen that when the self-vehicle travels to with the maximum lateral acceleration under the current road conditions, the lateral position of the right front vertex of the self-vehicle at this time is ​If it is greater than the lateral position of the left rear corner of the vehicle in front. If the vehicle does not collide with the vehicle in front during the entire lane change process, then the lane change trajectory meets the lane change feasibility conditions. The second type of lane change to the left is prone to collisions with vehicles in front of the target lane. The left front vertex of the vehicle's rectangle is most likely to collide with the right rear vertex of the vehicle in front. The coordinates of the upper left vertex of the vehicle during the lane change process are as follows: X 3, Y 3) The calculation formula is: ; In the formula, let the coordinates of the right rear vertex of the rectangle of the vehicle in front of the target lane be... It can be seen that when the car travels at its maximum lateral acceleration under the current road conditions to... At this time, the lateral position of the left front apex of the vehicle If it is smaller than the right rear corner of the vehicle in front, then the lateral position is... If the vehicle does not collide with the vehicle in front during the entire lane change process, then the lane change trajectory meets the lane change feasibility conditions. The third scenario involves a lane change to the left, where a collision with a vehicle behind in the target lane is likely. The left rear vertices of the vehicle's rectangle are most likely to collide with the right front vertices of the vehicle behind. The coordinates of the left rear vertices of the vehicle during the lane change process are as follows: X 5, Y The calculation formula for 5) is: ; In the formula, Let the distance from the centroid to the rear of the car be denoted as , and let the coordinates of the right front vertex of the rectangle representing the rear of the target lane be . It can be seen that when the car travels at its maximum lateral acceleration under the current road conditions to... At this time, the lateral position of the left rear apex of the vehicle. If it is greater than the lateral position of the right front apex of the rear vehicle If the vehicle does not collide with the vehicle behind it during the entire lane change process, then the lane change trajectory meets the conditions for lane change feasibility.

[0025] Step 2.3. For case 1 where lane changing is not feasible in step 2.2, the co-driving system generates the desired straight-ahead trajectory and takes over vehicle control in a timely manner to ensure safety. Step 2.4. In case 2 where lane change feasibility is not met in step 2.2, the co-driving system promptly takes over vehicle control to ensure safety, controls the vehicle to brake and decelerate to make room for subsequent lane change actions, so that when the driver's intention to change lanes is detected again, the lane change feasibility judgment can be met, and a safe expected lane change trajectory can be generated.

[0026] The selection of maximum lateral acceleration is related to road conditions. The maximum lateral acceleration is set to 0.4g under good road conditions, and the maximum lateral acceleration is set to μ g, μ is the current road adhesion coefficient.

[0027] The early lane-changing trajectory of case 2 is designed as follows. First, the lane-changing time of the set trajectory is calculated The simplest trajectory equation with longitudinal uniform speed and transverse quintic polynomial is: ; In the formula, is the expected longitudinal speed, is the transverse displacement required for lane changing, which is generally set to the standard width of the road, is the longitudinal position of the vehicle, i.e. the longitudinal position in the earth coordinate system, and D is the longitudinal displacement required for lane changing, then Taking the second derivative of time , the equation of lateral acceleration is obtained. Using the chain rule, we get: ; Then let = , the corresponding lane-changing time and the longitudinal displacement required for lane changing are obtained. The obtained is brought into , and and are set, and finally the corresponding critical collision time is obtained. If the driver changes lanes too late, the co-pilot system generates a lane-changing trajectory corresponding to the critical collision time 0.5s in advance, which meets the lane-changing feasibility condition and ensures safety.

[0028] Step 2.5. Trajectory cluster construction; when the lane-changing feasibility is met, the expected trajectory is designed with longitudinal uniform speed and transverse quintic polynomial. First, the maximum speed of the front vehicle in the current lane is , and the maximum braking deceleration is Under these four lane-changing scenarios, multiple repeated tests are performed, and the mean characteristics and standard deviation characteristics of the lane-changing time are calculated. The lower bound is , the upper bound is , and the step size is 0.2s. A series of lane-changing times are generated to finally construct a two-dimensional trajectory cluster under each lane-changing scenario.

[0029] Step 2.6. Optimal lane-changing trajectory solving; set multi-objective scoring method to select the highest scoring trajectory as the optimal lane-changing trajectory under the current scene, consider the following four sub-goals in multi-objective scoring: stability goal, efficiency goal, comfort goal and anthropomorphism goal, and the results are as follows: ; In the formula, is the final score of the trajectory for calculation and comparison,

[0030] respectively, the stability goal, the efficiency goal, the comfort goal and the anthropomorphism goal score results, k 1、 k 2、 k 3、 k 4 is the weight corresponding to each sub-goal score, and the calculation method of each index score is as follows: ; In the formula, is the maximum lateral acceleration under each planned trajectory, is the gravity acceleration, is the required lane-changing time of the evaluated trajectory, i.e. the lane-changing time counted for each scene, is the longest lane-changing time under the planned trajectory, i.e. , is the maximum lateral jerk under the planned trajectory, is the reference value of lateral jerk, is the mean characteristic of lane-changing time counted for each scene, is the standard deviation characteristic of lane-changing time counted for each scene.

[0031] The weight coefficients corresponding to the stability goal, the efficiency goal, the comfort goal and the anthropomorphism goal k 1、 k 2、 k 3、 k 4 are determined by mainly considering the influence of vehicle speed, traffic density and road conditions: In high-speed, low-density scenarios, the efficiency target dominates, ensuring quick lane changes. The stability target is also important, but the comfort and anthropomorphism targets have lower weights, focusing mainly on quick and safe lane changes. The weight distribution is as follows: stability target: 30%, efficiency target: 50%, comfort target: 10%, and anthropomorphism target: 10%. In low-speed, high-density, complex road conditions, the stability and comfort targets are most important, ensuring safe and smooth operation. The efficiency target has a lower weight, and the anthropomorphism target is moderately increased to make the system's operation appear more natural. The weight distribution is as follows: stability target: 35%, efficiency target: 20%, comfort target: 30%, and anthropomorphism target: 15%. In moderate-speed, moderately dense environments, the weights of the four targets are balanced. The system needs to balance efficiency, comfort, and stability while ensuring that the operation meets the driver's expectations and has some anthropomorphic characteristics. The weight distribution is as follows: stability target: 30%, efficiency target: 20%, comfort target: 25%, and anthropomorphism target: 25%.

[0032] This dynamic weight distribution method enables the system to adjust flexibly in different driving scenarios, balancing safety, comfort, and efficiency while simulating the driving behavior of human drivers and providing a more natural driving experience.

[0033] Step 3. Real-time solution of human-machine optimal weights, based on the driver's intention and steering wheel angle fluctuation, dynamically adjusting the human-machine weight distribution to achieve smooth cooperation and safe switching between the driver and the co-driving system; Step 3.1. When the driver's intention recognition result is to maintain straight driving, the driver's weight is set to 1, and the co-driving system's weight is set to 0. Step 3.2. When the driver's intention recognition result is left or right lane change, and the lane change feasibility condition is not met, i.e., in case 1 and case 2 in step 2.2, the driver's weight is set to 0, and the co-driving system's weight is set to 1. Step 3.3. When the driver's intention recognition result is left or right lane change, and the lane change feasibility condition is met, first, a 0.5s rolling time window is used to solve the steering wheel angle standard deviation of the driver in each window, and based on the statistical results of the logarithmic values of the steering wheel angle standard deviation, a lognormal distribution is constructed, with a probability density function : ; where is the steering wheel angle standard deviation, and are the mean and standard deviation of the steering wheel angle standard deviation logarithm, and the cumulative distribution function expression is: ; From the above formula, the larger the steering wheel angle standard deviation in the window, the larger the value of the corresponding cumulative distribution function, and the smaller the weight of the corresponding driver. Therefore, the change in authority between adjacent time windows should not be too large, so that the driver and the co-pilot system under the human-machine co-pilot in the current time window Driver real-time weight and co-pilot system real-time weight are defined as follows: ; In the formula, is the driver weight in the time window , and the driver weight is required to be 1 at the initial time of lane changing. The change in the driver weight value in the adjacent time window is not greater than 0.2, and in other cases, the driver weight value is 1 minus the cumulative distribution function value of the steering wheel angle standard deviation in the corresponding window.

[0034] Step 4. Motion control amount solving, based on the optimal lane changing trajectory, using adaptive model predictive control (MPC) algorithm, combining vehicle speed dynamic adjustment of prediction and control time domain, realizing high-precision tracking control of vehicle to expected trajectory; Step 4.1. Based on the optimal lane changing trajectory obtained by solving, the adaptive model predictive control method is applied to track the control of the expected trajectory, and the model state space is defined as follows: wherein, , is the lateral velocity and longitudinal velocity, and is the vehicle yaw angle and yaw angular velocity, and is the lateral position and longitudinal position of the vehicle in the earth coordinate system (XOY), and the control input and output are defined as follows: , .

[0035] Step 4.2. The co-pilot system is based on the adaptive model predictive control method (adaptive model predictive control framework) established in step 4.1, and constructs a quadratic programming (QP) optimization problem with control increment and output deviation as the target. The objective function of this problem considers the trajectory tracking accuracy and control smoothness of the vehicle, and is defined as follows: A weighted sum of squared deviations between the desired trajectory and the actual output in the prediction horizon and the squared increments of the control input is minimized as an objective function, and the optimal control sequence is solved under the conditions of satisfying the vehicle dynamics constraints, the control constraints and the control increment constraints; Specifically, the quadratic programming problem is expressed as: ; wherein, is the desired output in the prediction horizon, is the model prediction output, is the control increment at the i-th step, is the control increment at the i-th step, and are the weighting matrices of the output error and the control increment respectively, is the length of the prediction horizon, is the length of the control horizon; The co-driving system uses the active set method to solve the quadratic programming problem in real time to obtain a series of optimal control input increments and relaxation factors in the control horizon. The co-driving system takes the first control quantity in the control sequence as the actual control input at the current time to act on the vehicle, thereby realizing dynamic tracking of the desired lane-changing trajectory. Then, the next time is rolled to optimize until the entire trajectory tracking control process is completed.

[0036] Step 4.3. Adaptability of the adaptive model predictive control method is used to set the relationship between the longitudinal vehicle speed and the prediction horizon and the prediction horizon The relationship is set as follows: ; wherein, is a proportional coefficient, , are the longitudinal low-speed threshold and the longitudinal high-speed threshold respectively, is the upper limit of the prediction horizon.

Claims

1. A method for permission allocation and planning control in a co-driving system that considers driving intentions and states, characterized in that, Includes the following steps: Step 1. Driver intent recognition model construction: Through multimodal data collection, spatial and temporal features are extracted by combining CNN and LSTM models, and the driving intent is recognized in real time after fusion. Step 2. Trajectory planning decision design: Based on the identified driving intention, determine the feasibility of lane changing, generate a safe trajectory cluster, and use a multi-objective scoring method to select the optimal lane changing trajectory to achieve trajectory planning; Step 3. Real-time solution of optimal human-machine weights: Based on the driver's intentions and steering wheel angle fluctuations, dynamically adjust the human-machine weight allocation to achieve smooth collaboration and safe switching between the driver and the co-driving system. Step 4. Solve for motion control variables. Based on the optimal lane-changing trajectory, adopt an adaptive model predictive control method, and combine vehicle speed dynamic adjustment prediction and control time domain to achieve high-precision tracking control of the vehicle on the desired trajectory.

2. The method for permission allocation and planning control of a co-driving system considering driving intention and state according to claim 1, characterized in that, Step 1 also includes the following steps: Step 1.

1. Real-time multimodal data acquisition: During vehicle operation, data is collected in real time using sensors: (1) Time series data: real-time status is obtained from the vehicle's CAN bus, including position, speed, acceleration, steering wheel angle, and light signals; (2) Spatial feature data: real-time images, lane lines, and relative positions of surrounding vehicles are obtained through environmental perception devices. The time series data is sliced ​​according to a fixed time window to form a sliding time window; at the same time, the spatial feature data is dynamically updated frame by frame to match the time series data. The environmental perception devices include cameras, radar, or lidar. Step 1.

2. Construct a CNN model to process spatial features; preprocess the input image data by scaling and normalizing to adapt to the input format of the CNN model, use convolutional layers to extract local spatial features, including lane line shape, relative position of vehicles in front or on the side, use pooling layers to reduce dimensionality, retain the preset key features, and reduce the amount of computation to meet real-time requirements, and finally obtain spatial feature vectors. Step 1.

3. Construct an LSTM model to process time series features. According to the sliding time window, input the time series data of the most recent time into the LSTM model. Use multiple LSTM modules of the LSTM model to process the time series in sequence, capture the dynamic change trend of driving behavior, and output the time series feature vector to represent the continuous dynamic features of the vehicle. Step 1.

4. Real-time feature fusion: At the end of each time window, the features extracted by the LSTM model and the CNN model are fused. The time series feature vector output by the CNN model and the spatial feature vector output by the CNN model are combined by feature concatenation. The fused feature vector is then input into a fully connected layer for further processing and compression to form high-level semantic features. Step 1.

5. Classify and output the recognition results. Based on the fused high-level semantic features, predict driving intentions in real time. Use a fully connected layer to map the high-level semantic features to the classification space. Calculate the probability of each driving intention category through the Softmax activation function. The output driving intention categories include: left lane change, right lane change, and keep straight.

3. The method for permission allocation and planning control of a co-driving system considering driving intention and state according to claim 1, characterized in that, Step 2 also includes the following steps: Step 2.

1. Input the intent recognition result: Input the driving intent category output by the driver intent recognition model into the trajectory planning decision module. The driving intent categories include left lane change, right lane change, and keeping straight. Step 2.

2. Lane Change Feasibility Determination: If the intention to maintain straight-ahead is recognized, the co-driving system does not output lateral control. For left and right lane changes, the feasibility of the lane change is first determined. In the geodetic coordinate system, the vehicle is assumed to be a rectangle. If the vehicle's lane-changing process with maximum lateral acceleration has no intersection with other rectangles, then the lane change feasibility condition is met. When lane change feasibility is met, the co-driving system generates a lane-changing trajectory for lane-changing assistance. The following are the cases where lane change feasibility is not met: Scenario 1: The driver fails to effectively assess the situation of vehicles behind in the target lane and changes lanes too early. In this case, the longitudinal distance between the driver and the vehicles behind in the target lane is too small, posing a risk of collision. Scenario 2: The vehicle in front of you in the current lane brakes suddenly, and your vehicle fails to take timely and effective lane-changing avoidance measures, resulting in a risk of collision due to insufficient longitudinal distance between you and the vehicle in front of you in the original lane or the vehicle in front of you in the target lane.

4. The method for permission allocation and planning control of a co-driving system considering driving intention and state according to claim 3, characterized in that, Step 2 also includes the following steps: Step 2.

3. For case 1 where lane changing is not feasible in step 2.2, the co-driving system generates the desired straight-ahead trajectory and takes over vehicle control in a timely manner to ensure safety. Step 2.

4. In case 2 where lane change feasibility is not met in step 2.2, the co-driving system promptly takes over vehicle control to ensure safety, controls the vehicle to brake and decelerate to make room for subsequent lane change actions, so that when the driver's intention to change lanes is detected again, the lane change feasibility judgment can be met, and a safe expected lane change trajectory can be generated.

5. The method for permission allocation and planning control of a co-driving system considering driving intention and state according to claim 4, characterized in that, Step 2 also includes the following steps: Step 2.

5. Trajectory Cluster Construction; When lane-changing feasibility is satisfied, design the desired trajectory with longitudinal constant speed and lateral fifth-order polynomial, first based on the highest speed of the vehicle ahead in the current lane. The maximum braking deceleration is These four lane-changing scenarios were repeatedly tested, and the lane-changing time for each scenario was recorded. Mean characteristics and standard deviation characteristics ,by As the lower boundary, with With 0.2s as the upper bound and 0.2s as the step size, the generated series of lane-changing times ultimately construct a two-dimensional trajectory cluster for each lane-changing scenario.

6. The method for permission allocation and planning control of a co-driving system considering driving intention and state according to claim 5, characterized in that, Step 2 also includes the following steps: Step 2.

6. Solving for the optimal lane-changing trajectory; A multi-objective scoring method is used to select the trajectory with the highest score as the optimal lane-changing trajectory for the current scenario. The multi-objective scoring considers the following four sub-objectives: stability objective, efficiency objective, comfort objective, and anthropomorphism objective. The results are as follows: ； In the formula, To calculate the final score of the compared trajectories, The scores are for stability, efficiency, comfort, and anthropomorphism objectives, respectively. k 1. k 2. k 3. k 4 represents the weights corresponding to the scores of each sub-objective. The calculation methods for the scores of each indicator are as follows: ； In the formula, The maximum lateral acceleration for each planned trajectory. It is the acceleration due to gravity. The lane change time required for the evaluated trajectory, i.e., the lane change time statistically analyzed for each scenario. The longest lane-changing time under the planned trajectory is... , The maximum lateral acceleration under the planned trajectory, This is a reference value for lateral acceleration. Lane change time for each scenario The mean characteristics, Lane change time for each scenario The standard deviation characteristic.

7. The method for permission allocation and planning control of a co-driving system considering driving intention and state according to claim 1, characterized in that, Step 3 also includes the following steps: Step 3.

1. When the driver's intention recognition result is to maintain straight driving, the driver's weight is set to 1, and the weight of the co-driving system is set to 0; Step 3.

2. When the driver's intent recognition result is a left lane change or a right lane change, and the lane change feasibility conditions are not met, the driver's weight is set to 0, and the weight of the co-driving system is set to 1. Step 3.

3. When the driver's intention recognition result is a left lane change or a right lane change, and the lane change feasibility condition is met, firstly, using a rolling time window of 0.5s, calculate the driver's steering wheel angle in each window. The standard deviation of the steering wheel angle is used to construct a log-normal distribution based on the statistical results of the logarithmic values ​​of the standard deviation of the steering wheel angle, and its probability density function is... for: ； In the formula, The standard deviation of the steering wheel angle. and Let the mean and standard deviation of the logarithm of the standard deviation of the steering wheel angle be denoted as , and the expression for its cumulative distribution function be: ； In human-machine co-driving, the driver and the co-driving system are in the current time window. Real-time driver weights Real-time weighting of the shared driving system The definition is as follows: ； In the formula, For time window The driver weights are assigned based on the following conditions: the driver's weight is 1 at the initial moment of lane change, and the change in driver weight within adjacent time windows is no greater than 0.

2. Otherwise, the driver's weight value... It is 1 minus the cumulative distribution function value of the standard deviation of the steering wheel angle under the corresponding window.

8. The method for permission allocation and planning control of a co-driving system considering driving intention and state according to claim 1, characterized in that, Step 4 also includes the following steps: Step 4.

1. Based on the optimal lane-changing trajectory obtained from the solution, apply the adaptive model predictive control method to track the desired trajectory, and define the model state space. Among them, state variables The definition is as follows: ,in, , For lateral velocity and longitudinal velocity, and For the vehicle's yaw angle and yaw rate, and For the lateral and longitudinal positions of the vehicle in the geodetic coordinate system, the control input is... and output The definition is as follows: , .

9. A method for permission allocation and planning control of a co-driving system considering driving intention and state, as described in claim 8, is characterized in that... Step 4 also includes the following steps: Step 4.

2. Based on the adaptive model predictive control method established in Step 4.1, the co-driving system constructs a quadratic programming optimization problem with control increment and output deviation as objectives. The objective function of this problem comprehensively considers the vehicle's trajectory tracking accuracy and control smoothness, and is defined as follows: The control sequence is solved by minimizing the sum of squared deviations between the expected trajectory and the actual output in the predicted time domain and the weighted sum of squared control input increments, while satisfying vehicle dynamics constraints, control quantity constraints and control increment constraints. The quadratic programming problem is expressed as: ； in, To predict the desired output in the time domain, For the model's predicted output, For the first Step control increment, and These are the weighting matrices for output error and control increment, respectively. To predict the length of the time domain, To control the length of the time domain; The co-driving system uses the effective set method to solve the quadratic programming problem in real time, obtaining a series of control input increments and relaxation factors in the control time domain. The co-driving system uses the first control quantity in the control sequence as the actual control input at the current moment to act on the vehicle, realizing dynamic tracking of the desired lane-changing trajectory; then it enters the next moment for rolling optimization until the entire trajectory tracking and control process is completed.

10. A method for permission allocation and planning control of a co-driving system considering driving intention and state, as described in claim 8, is characterized in that... Step 4 also includes the following steps: Step 4.

3. The adaptability of the adaptive model predictive control method is used in the control time domain. and prediction time domain The relationship between longitudinal vehicle speed and speed is set as follows: ； In the formula, This is the proportionality coefficient. , These are the longitudinal low-speed threshold and the longitudinal high-speed threshold, respectively. To predict the upper limit of the time domain.

Images