Dynamic Drift Obstacle Avoidance Control Method for Autonomous Driving Vehicles Based on Driving Condition Recognition

Abstract

The present invention discloses a dynamic drift obstacle avoidance control method for autonomous driving vehicles based on driving condition recognition, belonging to the field of autonomous driving vehicle control. It includes perceiving the multi-dimensional dynamic characteristics of the vehicle and its surrounding environment, predicting the information entropy of historical multi-dimensional dynamic characteristics using a graph neural network, and adaptively adjusting the sensor sampling frequency according to the prediction error; based on the sensor sampling data after adaptive adjustment, using a temporal convolutional neural network enhanced by fuzzy logic to determine whether the vehicle is in the drift critical zone; according to whether the vehicle is in the drift critical zone, adaptively adjusting the state weight matrix and control weight matrix of the model predictive controller, and calculating the control commands for the front wheel steering angle and rear wheel driving force of the vehicle; sending the control commands to the steering actuator and driving actuator to achieve vehicle obstacle avoidance and trajectory tracking. The purpose of the present invention is to improve the handling ability and safety of autonomous driving vehicles under complex driving conditions and enhance the overall dynamic performance of the vehicle.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous vehicle control, and particularly to a dynamic drift obstacle avoidance control method for autonomous vehicles based on driving condition recognition. Background Art

[0002] With the rapid development of autonomous driving technology, the level of vehicle intelligence has been continuously improved, and more and more professional driving behaviors have been successfully developed and applied to autonomous driving control systems. Drift control, as a special type of driving skill, is mostly used in racing cars and stunt performances and is usually regarded as an extreme driving technique. However, with the further development of autonomous driving technology, drift control has gradually been applied to the obstacle avoidance system of autonomous vehicles, providing greater maneuverability and safety guarantee for vehicles in emergency situations.

[0003] Currently, the drift obstacle avoidance technology of autonomous vehicles mainly relies on path planning and obstacle detection under normal driving conditions. However, in complex traffic environments, especially under harsh conditions such as slippery roads, ice, rain or snow, when the vehicle needs to perform emergency obstacle avoidance or trajectory tracking, traditional rule-based drift control strategies are difficult to ensure the stability and safety of vehicle driving, and may even lead to vehicle out of control. This is because traditional drift control methods are usually based on pre-set rules and cannot effectively cope with complex and changeable actual driving environments.

[0004] In addition, traditional drift control methods often ignore the impact of sensor data acquisition on control performance. In extreme driving conditions such as drifting, the vehicle state changes violently, and higher requirements are placed on the real-time and accuracy of sensor data. However, the traditional sensor data acquisition method with a fixed sampling frequency cannot meet the requirements of drift control, which may lead to control delay or instability.

[0005] Therefore, it is necessary to develop a new drift obstacle avoidance control method for autonomous vehicles, which can adaptively adjust the sensor sampling frequency, accurately identify the driving conditions of the vehicle, and adaptively adjust the control strategy according to the driving conditions to improve the maneuverability and safety of the vehicle under extreme driving conditions such as drifting. Summary of the Invention

[0006] The purpose of the present invention is to provide a dynamic drift obstacle avoidance control method for autonomous vehicles based on driving condition recognition, so as to solve the problems of insufficient accuracy and stability of drift obstacle avoidance control in the prior art under complex environments, and improve the maneuverability and safety of the vehicle under different driving conditions.

[0007] To achieve the above object, the present invention provides a dynamic drift obstacle avoidance control method for autonomous vehicles based on driving condition recognition, including the following steps:

[0008] Step S10, perceive the multi-dimensional dynamic features of the vehicle and the surrounding scene domain through in-vehicle sensors;

[0009] Step S20, according to the information entropy of the multi-dimensional dynamic features obtained in S10, predict the current information entropy through a graph neural network, and adaptively adjust the sensor sampling frequency according to the prediction error to implement an adaptive variable-frequency sampling strategy for in-vehicle sensors;

[0010] Step S30, based on the sensor sampling data after the adaptive adjustment in Step S20, use a temporal convolutional neural network enhanced by fuzzy control to perform real-time inference on the current vehicle state to determine whether it is in a drift critical zone or a stable driving condition;

[0011] Step S40, according to the possibility that the vehicle is in the drift critical zone judged in Step S30, adaptively adjust the state weight matrix Q and the control weight matrix R of the model predictive controller to achieve stable control of the vehicle under different working conditions;

[0012] Step S50, calculate the control commands for the front wheel steering angle and the rear wheel driving force of the vehicle according to Step S40, and send these commands to the steering actuator and the driving actuator respectively to achieve safe obstacle avoidance and tracking of the planned trajectory during vehicle driving.

[0013] Preferably, in Step S10, the vehicle multi-dimensional dynamic features include vehicle speed, vehicle longitudinal acceleration, vehicle lateral acceleration, vehicle yaw rate, and front wheel steering angle; the multi-dimensional dynamic features of the surrounding scene domain include the distance to surrounding obstacles, the relative speed of surrounding obstacles, road curvature, lane departure degree, and road adhesion coefficient; the in-vehicle sensors include: speed sensor, acceleration sensor, gyroscope, steering angle sensor, lidar, and camera.

[0014] Preferably, Step S20 includes the following steps:

[0015] Step S201, at each sampling moment t, fuse all in-vehicle sensor data to form a feature vector. Different sensors have different sampling frequencies. Based on the time stamp, perform data interpolation to construct a normalized dynamic feature vector x(t) under a unified time reference t:

[0016]

[0017] where v(t) is the vehicle speed at time t; a x (t) is the vehicle longitudinal acceleration at time t; a y (t) is the vehicle lateral acceleration at time t; is the vehicle yaw rate at time t; δ(t) is the front wheel steering angle at time t; d i (t) is the distance to the i-th surrounding obstacle at time t; vrel,i (t) is the relative velocity with the i-th surrounding obstacle at time t; κ(t) is the road curvature at time t; D lane (t) is the degree of lane line deviation at time t; μ(t) is the road adhesion coefficient at time t;

[0018] Step S202, calculate the information entropy H(t) of the dynamic feature vector x(t) at time t:

[0019]

[0020] In the formula, p(x j (t)) is the probability density function of the j-th component x j (t) of the dynamic feature vector x(t) at time t; n is the dimension of the dynamic feature vector x(t);

[0021] Step S203, construct a graph neural network GNN model, whose input is the historical information entropy sequence H history :

[0022] H history = {H(t - m), H(t - m + 1), …, H(t - 1)};

[0023] In the formula, m is the number of samples of the historical information entropy sequence;

[0024] Step S204, input the historical information entropy sequence H history into the GNN model to predict the information entropy at time t

[0025]

[0026] Step S205, calculate the absolute error e(t) between the actual perceived information entropy H(t) and the predicted information entropy :

[0027]

[0028] Step S206, based on the prediction error e(t), adaptively adjust the sampling frequency f i (t) of the i-th vehicle-mounted sensor:

[0029]

[0030] In the formula, f i,min is the minimum sampling frequency of the i-th sensor; f i,max is the maximum sampling frequency of the i-th sensor; e i,max is the preset maximum error threshold of the i-th sensor.

[0031] Preferably, step S30 includes the following steps:

[0032] Step S301, perform normalization processing on the time series data obtained after adjusting the adaptive sensor frequency conversion sampling strategy in step S20. The formula is as follows:

[0033]

[0034] In the formula, x i (t) is the i-th original feature data at time t, including vehicle speed and vehicle longitudinal acceleration; is the normalized value of the i-th feature at time t; min(x i ) and max(x i ) respectively represent the minimum and maximum values of the i-th feature in the entire time series;

[0035] Step S302, combine the normalized feature data at time t into a feature vector:

[0036]

[0037] In the formula, n is the number of features; construct a feature vector sequence in chronological order where N is the time window length;

[0038] Step S303, perform fuzzification processing on the normalized feature vector sequence X:

[0039] ① Convert each component of each feature vector in the feature vector sequence X into a fuzzy membership degree, and use a triangular membership function for fuzzification:

[0040]

[0041] In the formula, is the membership degree of the variable belonging to the fuzzy set A; a, b, and c are the parameters defining the triangular membership function respectively;

[0042] ② Perform fuzzy inference using Mamdani fuzzy rules:

[0043]

[0044] In the formula, A i is the fuzzy set corresponding to the i-th feature;

[0045] ③ Perform defuzzification using the centroid method to obtain the fuzzy inference result:

[0046]

[0047] In the formula, p fuzzy is the output value after defuzzification, representing the possibility that the vehicle is in the drift critical zone at time t; c j is the output center value of the j-th fuzzy rule;

[0048] ④ Add the fuzzy inference result p fuzzy (t) to the feature vector to obtain the feature enhanced vector and construct an enhanced feature vector sequence:

[0049]

[0050] Step S304: Input the feature enhanced vector sequence X′ into the Temporal Convolutional Network (TCN) model to predict the possibility that the vehicle is in the drift critical zone:

[0051] p tcn (t) = TCN(X′);

[0052] In the formula, p tcn (t) is the output result of the TCN model, representing the possibility that the vehicle is in the drift critical zone at time t;

[0053] Step S305: Use the weighted average method to comprehensively obtain the possibility that the vehicle is in the drift critical zone based on the prediction result of the TCN model and the fuzzy inference result:

[0054] p final (t) = (w tcn ·p tcn (t)) + (w fuzzy ·p fuzzy (t));

[0055] In the formula, p final (t) is the finally obtained possibility that the vehicle is in the drift critical zone; w tcn and w fuzzy are the weights of the TCN model and the fuzzy rule respectively;

[0056] Step S306: Determine whether the vehicle is in the drift critical zone according to the fused possibility p final (t):

[0057]

[0058] In the formula, θ is the critical zone judgment threshold.

[0059] Preferably, step S40 includes the following steps:

[0060] Step S401: Based on the three-degree-of-freedom model of the vehicle, construct the lateral error e of the vehicle centroidr A dynamic model with the lateral error $e$ and the heading angle error $\xi$ as state variables:

[0061]

[0062] In the formula, is the derivative of the lateral error; is the derivative of the heading angle error; is the derivative of the longitudinal velocity; is the derivative of the sideslip angle at the center of mass; is the derivative of the yaw rate; $V$ is the longitudinal velocity; $\beta$ is the sideslip angle at the center of mass; $\xi$ is the heading angle error; $r$ is the yaw rate; $k$ is the road curvature; $e$ r is the lateral error; $m$ is the vehicle mass; $F$ yf , $F$ yr are the lateral forces on the front and rear wheels respectively; $F$ xr is the driving force on the rear wheel; $\delta$ is the front wheel steering angle; $I$ z is the moment of inertia of the vehicle body about the $z$-axis; $l$ r , $l$ f are the distances from the center of mass to the front and rear axles of the vehicle respectively;

[0063] Step S402, linearize and discretize the nonlinear dynamic model : Linearization is performed using the Taylor expansion, and discretization is performed using the forward Euler method:

[0064]

[0065] In the formula, is the discretized state matrix; is the discretized control matrix; The state vector is $x(k)=[e$ r (k),\xi(k),V(k),\beta(k),r(k)] T , and $u(k)=[\delta(k),F$ xr (k)] T is the control vector at time $k$;

[0066] Step S403, construct the optimization problem of the model predictive controller. The optimization goal is to find the optimal control input within the prediction horizon to minimize the following objective function:

[0067]

[0068] In the formula, $NP$ is the prediction horizon length of the model predictive controller; $x$ sg is the desired state vector; $x(k)$ is the state vector at time $k$; $\Delta u(k)$ is the change in the control increment; $Q$ is the state weight matrix; $R$ is the control weight matrix;

[0069] During the optimization process, the constraint conditions for the model state variables and control variables are as follows:

[0070]

[0071] In the formula, Δx(k) is the change in the state variable; Δx min , Δx max are the minimum and maximum values of the change in the state variable respectively; Δu min , Δu max are the minimum and maximum values of the change in the control variable respectively; u min , u max are the minimum and maximum values of the control variable respectively.

[0072] Step S404, under the condition of satisfying the constraint conditions, solve the optimization problem to obtain the optimal control sequence of the vehicle within the prediction horizon:

[0073] U * =[u(k), u(k + 1),..., u(k + NP - 1)] T .

[0074] Preferably, adaptively adjust the state weight matrix Q and the control weight matrix R:

[0075] ① Determine the basic state weight matrix Q base and the control weight matrix R base as:

[0076]

[0077] In the formula, is the weight coefficient of the lateral error; q ξ is the weight coefficient of the heading angle error; q V is the weight coefficient of the vehicle longitudinal speed; q β is the weight coefficient of the vehicle sideslip angle at the center of mass; q r is the weight coefficient of the vehicle yaw rate; r δ is the weight coefficient of the front wheel steering angle; is the weight coefficient of the rear wheel driving force;

[0078] ② According to the final possibility p final of the vehicle being in the drift critical area obtained in step S30, adaptively adjust the state weight matrix Q and the control weight matrix R:

[0079]

[0080] In the formula, K Q is the proportional coefficient matrix for adjusting the weight of the Q matrix; K R is the proportional coefficient matrix for adjusting the weight of the R matrix.

[0081] A system for a dynamic drift obstacle avoidance control method of an autonomous driving vehicle based on driving condition recognition according to the present invention includes

[0082] Feature perception and adaptive sensor variable-frequency sampling module: Perceive the multi-dimensional dynamic features of the vehicle and the surrounding scene domain through in-vehicle sensors, predict the information entropy based on the graph neural network, and adaptively adjust the sensor sampling frequency;

[0083] Driving condition recognition module: Use a temporal convolutional neural network enhanced by fuzzy logic to judge the driving conditions of the vehicle, such as drifting or stable driving;

[0084] Driving condition adaptive model predictive control module: According to the driving conditions of the vehicle, adaptively adjust the parameters of the model predictive controller and calculate the control quantity, and output the control instruction to the steering and drive actuators.

[0085] Therefore, the above-mentioned dynamic drift obstacle avoidance control method of an autonomous driving vehicle based on driving condition recognition according to the present invention has the following beneficial effects:

[0086] (1) The present invention proposes an adaptive sensor variable-frequency sampling strategy based on information entropy prediction, uses the graph neural network to learn the relationship between historical multi-dimensional dynamic features and information entropy, and predicts the change trend of future information entropy. According to the prediction error, the sensor sampling frequency is adaptively adjusted to achieve increasing the sampling frequency to obtain more key information features when the multi-dimensional dynamic features of the vehicle and the surrounding scene domain change violently, and reducing the sampling frequency to save computing resources when the vehicle state is stable, ultimately improving the control accuracy.

[0087] (2) The present invention adopts a driving condition recognition method based on a temporal convolutional neural network enhanced by fuzzy logic. By combining fuzzy control and temporal convolutional neural network, it is judged whether the vehicle is in the drift critical area. The introduction of fuzzy logic enhances the model's ability to process uncertainty and fuzzy information, and improves the robustness of driving condition recognition.

[0088] (3) According to the driving conditions (drifting or stable driving) of the vehicle, the present invention adaptively adjusts the state weight matrix Q and the control weight matrix R of the model predictive controller. This strategy enables the controller to dynamically adjust according to the real-time state of the vehicle, so as to achieve stable control of the vehicle under different driving conditions.

[0089] (4) The present invention proposes a new integration method of a drift obstacle avoidance control system, including a feature perception and adaptive sensor variable-frequency sampling module, a driving condition recognition module, and a driving condition adaptive model predictive control module, realizing a closed-loop control from perception to control, and improving the overall performance of the system.

[0090] The technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Description of the Drawings

[0091] Figure 1 It is a flowchart of the control method according to an embodiment of the present invention;

[0092] Figure 2 It is a flowchart of the adaptive sensor sampling based on information entropy prediction according to an embodiment of the present invention;

[0093] Figure 3 It is a flowchart of the working condition recognition method of the temporal convolutional neural network enhanced by fuzzy logic according to an embodiment of the present invention;

[0094] Figure 4 It is a flowchart of the model predictive controller method based on adaptation according to an embodiment of the present invention. Detailed Embodiments

[0095] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention, and are not used to limit the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in this application without creative efforts belong to the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0096] It should be noted that the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or server that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0097] Similar reference numerals and letters denote similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0098] In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "upper", "lower", "inner", "outer", etc. is based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the product of the present invention is usually placed during use. It is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention.

[0099] In the description of the present invention, it should also be noted that unless otherwise clearly specified and defined, the terms "set", "installed", and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two components. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0100] Embodiment

[0101] As Figure 1 shown, a dynamic drift obstacle avoidance control method for an autonomous vehicle based on working condition recognition according to the present invention includes the following steps:

[0102] Step S10, perceiving the multi-dimensional dynamic characteristics of the vehicle and the surrounding scene domain through in-vehicle sensors.

[0103] In step S10, the multi-dimensional dynamic characteristics of the vehicle include vehicle speed, vehicle longitudinal acceleration, vehicle lateral acceleration, vehicle yaw rate, front wheel angle, etc.; the multi-dimensional dynamic characteristics of the surrounding scene domain include the distance to surrounding obstacles, the relative speed of surrounding obstacles, road curvature, lane departure degree, and road adhesion coefficient, etc.; the in-vehicle sensors include: speed sensor, acceleration sensor, gyroscope, steering angle sensor, lidar, camera, etc.

[0104] As Figure 2 shown, step S20, according to the information entropy of the multi-dimensional dynamic characteristics obtained in S10, predicting the current information entropy through a graph neural network, and adaptively adjusting the sensor sampling frequency according to the prediction error to implement an adaptive variable-frequency sampling strategy for in-vehicle sensors.

[0105] Step S20 includes the following steps:

[0106] Step S201, at each sampling moment t, fusing all in-vehicle sensor data to form a feature vector. Considering that different sensors may have different sampling frequencies, data interpolation needs to be performed based on timestamps, and then a normalized dynamic feature vector x(t) is constructed under a unified time base t:

[0107]

[0108] where v(t) is the vehicle speed at moment t; a x (t) is the vehicle longitudinal acceleration at moment t; a y (t) is the vehicle lateral acceleration at moment t; is the vehicle yaw rate at moment t; δ(t) is the front wheel angle at moment t; d iThe distance of the i-th surrounding obstacle at time t is (t); v rel,i The relative velocity with the i-th surrounding obstacle at time t is v(t); the road curvature at time t is κ(t); D lane The lane line deviation degree at time t is D(t); the road adhesion coefficient at time t is μ(t);

[0109] Step S202: Calculate the information entropy H(t) of the dynamic feature vector x(t) at time t:

[0110]

[0111] In the formula, p(x j (t)) is the probability density function of the j-th component x j (t) of the dynamic feature vector x(t) at time t; n is the dimension of the dynamic feature vector x(t);

[0112] Step S203: Construct a graph neural network (GNN) model, and its input is the historical information entropy sequence H history :

[0113] H history ={H(t - m), H(t - m + 1),..., H(t - 1)};

[0114] In the formula, m is the number of samples of the historical information entropy sequence;

[0115] Step S204: Input the historical information entropy sequence H history into the GNN model to predict the information entropy at time t

[0116]

[0117] Step S205: Calculate the absolute error e(t) between the actual perceived information entropy H(t) and the predicted information entropy at time t:

[0118]

[0119] Step S206: Based on the prediction error e(t), adaptively adjust the sampling frequency f i (t) of the i-th vehicle-mounted sensor:

[0120]

[0121] In the formula, f i,min is the minimum sampling frequency of the i-th sensor; f i,max is the maximum sampling frequency of the i-th sensor; e i,max is the preset maximum error threshold of the i-th sensor.

[0122] As Figure 3 shown, in step S30, based on the sensor sampling data after adaptive adjustment in step S20, a temporal convolutional neural network enhanced by fuzzy control is used to perform real-time inference on the current vehicle state to determine whether it is in the drift critical zone or the stable driving condition.

[0123] Step S30 includes the following steps:

[0124] In step S301, the temporal data obtained after adjusting the adaptive sensor variable-frequency sampling strategy in step S20 is normalized, and the formula is as follows:

[0125]

[0126] where x i (t) is the i-th original feature data at time t, including vehicle speed, vehicle longitudinal acceleration, etc.; is the normalized value of the i-th feature at time t; min(x i ) and max(x i ) respectively represent the minimum and maximum values of the i-th feature in the entire time series;

[0127] In step S302, the normalized feature data at time t are combined into a feature vector:

[0128]

[0129] where n is the number of features. Then, a feature vector sequence is constructed in chronological order where N is the time window length;

[0130] In step S303, the normalized feature vector sequence X is fuzzified:

[0131] ① Each component of each feature vector in the feature vector sequence X is converted into a fuzzy membership degree, and triangular membership degree function is used for fuzzification:

[0132]

[0133] where is the membership degree of the variable belonging to the fuzzy set A; a, b, and c are the parameters defining the triangular membership degree function respectively;

[0134] ② Mamdani fuzzy rules are used for fuzzy inference:

[0135]

[0136] where Ai is the fuzzy set corresponding to the i-th feature;

[0137] ③ Defuzzification is performed using the centroid method to obtain the fuzzy inference result:

[0138]

[0139] In the formula, p fuzzy is the output value after defuzzification, representing the possibility that the vehicle is in the drift critical zone at time t; c j is the output center value of the j-th fuzzy rule;

[0140] ④ Add the fuzzy inference result p fuzzy (t) to the feature vector to obtain the feature enhanced vector and construct the enhanced feature vector sequence:

[0141]

[0142] Step S304, input the feature enhanced vector sequence X′ into the Temporal Convolutional Network (TCN) model to predict the possibility that the vehicle is in the drift critical zone:

[0143] p tcn (t) = TCN(X′);

[0144] In the formula, p tcn (t) is the output result of the TCN model, representing the possibility that the vehicle is in the drift critical zone at time t;

[0145] Step S305, use the weighted average method to comprehensively obtain the possibility that the vehicle is in the drift critical zone according to the prediction result of the TCN model and the fuzzy inference result:

[0146] p final (t) = (w tcn · p tcn (t)) + (w fuzzy · p fuzzy (t));

[0147] In the formula, p final (t) is the finally obtained possibility that the vehicle is in the drift critical zone; w tcn and w fuzzy are the weights of the TCN model and the fuzzy rule respectively;

[0148] Step S306, judge whether the vehicle is in the drift critical zone according to the fused possibility p final (t):

[0149]

[0150] In the formula, θ is the critical area judgment threshold.

[0151] As Figure 4 shown, in step S40, according to the possibility that the vehicle is in the drift critical area judged in step S30, adaptively adjust the state weight matrix Q and the control weight matrix R of the model predictive controller to achieve stable control of the vehicle under different working conditions (drifting or stable driving).

[0152] Step S40 includes the following steps:

[0153] Step S401, based on the three-degree-of-freedom model of the vehicle, construct a dynamic model with the lateral error e r of the vehicle's center of mass and the heading angle error ξ as state variables:

[0154]

[0155] In the formula, is the derivative of the lateral error; is the derivative of the heading angle error; is the derivative of the longitudinal speed; is the derivative of the sideslip angle of the center of mass; is the derivative of the yaw rate; V is the longitudinal speed; β is the sideslip angle of the center of mass; ξ is the heading angle error; r is the yaw rate; k is the road curvature; e r is the lateral error; m is the vehicle mass; F yf , F yr are the lateral forces of the front and rear wheels respectively; F xr is the driving force of the rear wheel; δ is the front wheel steering angle; I z is the moment of inertia of the vehicle body along the z-axis; l r , l f are the distances from the center of mass to the front and rear axles of the vehicle respectively;

[0156] Step S402, linearize and discretize the nonlinear dynamic model ; use the Taylor expansion for linearization and the forward Euler method for discretization:

[0157]

[0158] In the formula, is the discretized state matrix; is the discretized control matrix; the state vector is x(k) = [e r (k), ξ(k), V(k), β(k), r(k)] T , u(k) = [δ(k), F xr (k)] T is the control vector at time k;

[0159] Step S403: Construct the optimization problem of the model predictive controller. The optimization objective is to find the optimal control input within the prediction horizon to minimize the following objective function:

[0160]

[0161] where \(N_P\) is the prediction horizon length of the model predictive controller; \(x_d\) sg is the desired state vector; \(x(k)\) is the state vector at time \(k\); \(\Delta u(k)\) is the change in the control increment; \(Q\) is the state weight matrix; \(R\) is the control weight matrix;

[0162] During the optimization process, consider the following constraint conditions for the model state variables and control variables:

[0163]

[0164] where \(\Delta x(k)\) is the change in the state variable; \(\Delta x_{min}\) min , \(\Delta x_{max}\) max are the minimum and maximum values of the change in the state variable respectively; \(\Delta u_{min}\) min , \(\Delta u_{max}\) max are the minimum and maximum values of the change in the control variable respectively; \(u_{min}\) min , \(u_{max}\) max are the minimum and maximum values of the control variable respectively.

[0165] Step S404: Solve the optimization problem under the condition of satisfying the constraint conditions to obtain the optimal control sequence of the vehicle within the prediction horizon:

[0166] \(U\) * = [u(k), u(k + 1),..., u(k + N_P - 1)] T .

[0167] Adaptively adjust the state weight matrix \(Q\) and the control weight matrix \(R\):

[0168] ① Determine the basic state weight matrix \(Q\) base and the control weight matrix \(R\) base as:

[0169]

[0170] where is the weight coefficient of the lateral error; \(q_{\psi}\) ξ is the weight coefficient of the heading angle error; \(q_v\) V is the weight coefficient of the vehicle longitudinal speed; \(q_{\delta}\) β is the weight coefficient of the vehicle sideslip angle at the center of mass; \(q_r\) r is the weight coefficient of the vehicle yaw rate; \(r_{\delta}\) δ is the weight coefficient of the front wheel steering angle; is the weight coefficient of the rear-wheel driving force;

[0171] ② According to the final possibility p that the vehicle is in the drift critical zone obtained in step S30 final , adaptively adjust the state weight matrix Q and the control weight matrix R:

[0172]

[0173] In the formula, K Q is the proportional coefficient matrix for adjusting the weight of the Q matrix; K R is the proportional coefficient matrix for adjusting the weight of the R matrix.

[0174] Step S50: Calculate the control instructions for the front-wheel steering angle and the rear-wheel driving force of the vehicle according to step S40, and send these instructions to the steering actuator and the driving actuator respectively to achieve safe obstacle avoidance during vehicle driving and accurately track the planned trajectory.

[0175] A system for a dynamic drift obstacle avoidance control method of an autonomous driving vehicle based on working condition recognition according to the present invention includes

[0176] Feature perception and adaptive sensor frequency conversion sampling module: Sense the multi-dimensional dynamic features of the vehicle and the surrounding scene domain through on-vehicle sensors, predict the information entropy based on the graph neural network, and adaptively adjust the sensor sampling frequency;

[0177] Working condition recognition module: Use a temporal convolutional neural network enhanced by fuzzy logic to judge the working conditions of vehicle drifting or stable driving;

[0178] Working condition adaptive model predictive control module: According to the vehicle working conditions, adaptively adjust the parameters of the model predictive controller and calculate the control quantity, and output the control instructions to the steering and driving actuators.

[0179] Therefore, the present invention adopts the above-mentioned dynamic drift obstacle avoidance control method for an autonomous driving vehicle based on working condition recognition to solve the problems of insufficient accuracy and stability of drift obstacle avoidance control in the prior art in complex environments, and improve the handling performance and safety of the vehicle under different working conditions. Through multi-sensor fusion, the multi-dimensional dynamic features of the vehicle and its surrounding environment are sensed in real time, and the change of feature information entropy is predicted by using the graph neural network, and then the sensor sampling frequency is adaptively adjusted. Then, the present invention uses a temporal convolutional neural network enhanced by fuzzy logic to identify the vehicle driving working conditions in real time. According to the recognition results, the parameters of the model predictive controller are adaptively adjusted, the vehicle control instructions are calculated and executed, and stable control of the vehicle under different working conditions is achieved, including steady-state drift and accurate trajectory tracking.

[0180] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions of the present invention or make equivalent replacements, and these modifications or equivalent replacements do not enable the modified technical solutions to depart from the spirit and scope of the technical solutions of the present invention.

Claims

1. A dynamic drift obstacle avoidance control method for autonomous driving vehicles based on driving condition recognition, characterized in that: It includes the following steps: Step S10, perceiving the multi-dimensional dynamic features of the vehicle and the surrounding scene domain through in-vehicle sensors; Step S20, according to the information entropy of the multi-dimensional dynamic features obtained in S10, predicting the current information entropy through a graph neural network, and adaptively adjusting the sensor sampling frequency according to the prediction error to implement an adaptive variable-frequency sampling strategy for in-vehicle sensors; Step S30, based on the sensor sampling data after the adaptive adjustment in Step S20, using a temporal convolutional neural network enhanced by fuzzy control to perform real-time inference on the current vehicle state to determine whether it is in a drift critical zone or a stable driving condition; Step S40: adaptively adjust the state weight matrix of the model predictive controller according to the possibility that the vehicle is in the drift critical area determined in step S30 Q and the control weight matrix R to achieve stable control of the vehicle under different working conditions; Step S50, according to the vehicle front-wheel steering angle and rear-wheel driving force control commands calculated in Step S40, and sending these commands to the steering actuator and the driving actuator respectively to achieve safe obstacle avoidance during vehicle driving and track the planned trajectory; Step S30 includes the following steps: Step S301, performing normalization processing on the temporal data obtained after the adaptive sensor variable-frequency sampling strategy adjustment in Step S20, and the formula is as follows: ； In the formula, is the th original feature data at time t, including vehicle speed and vehicle longitudinal acceleration; is the normalized value of the th feature at time t; and respectively represent the minimum and maximum values of the th feature in the entire time series; Step S302, combining the normalized feature data at moments into a feature vector: ； In the formula, is the number of features; a feature vector sequence is constructed in chronological order , where N is the time window length; Step S303, perform fuzzification processing on the normalized eigenvector sequence X : ①Convert each feature vector in the feature vector sequence X into a fuzzy membership degree, and use a triangular membership function for fuzzification: for each component of each feature vector ​ ； wherein, is a variable subordinate to the fuzzy set A membership degree; , , are the parameters for defining the triangular membership function, respectively; ② Performing fuzzy inference using Mamdani fuzzy rules: ； wherein, is the fuzzy set corresponding to the th feature; ③ Performing defuzzification using the centroid method to obtain the fuzzy inference result: ； In the formula, is the output value after defuzzification, indicating the possibility that the vehicle is in the drift critical area at the th output center value of the fuzzy rule; ④ Incorporate the fuzzy inference result into the feature vector to obtain the feature enhanced vector , and construct an enhanced feature vector sequence: ； Step S304, input the feature enhancement vector sequence into the temporal convolutional network (TCN) model to predict the possibility that the vehicle is in the drift critical zone: ； In the formula, is the output result of the TCN model, indicating the possibility that the vehicle is in the drift critical zone at moment. Step S305, using a weighted average method to comprehensively obtain the possibility that the vehicle is in the drift critical zone according to the TCN model prediction result and the fuzzy inference result; ； In the formula, is the possibility that the finally obtained vehicle is in the drift critical zone; and are the weights of the TCN model and the fuzzy rules respectively; Step S306, based on the fused possibility , determine whether the vehicle is in the drift critical zone: ； In the formula, is the critical region judgment threshold value; Step S40 includes the following steps: Step S401, based on the three-degree-of-freedom model of the vehicle, construct a dynamic model with the lateral error of the vehicle's center of mass and the heading angle error as state variables: ； In the formula, is the derivative of the lateral error; is the derivative of the heading angle error; is the derivative of the longitudinal speed; is the derivative of the sideslip angle at the center of mass; is the derivative of the yaw rate; V is the longitudinal speed; is the sideslip angle at the center of mass; is the heading angle error; is the yaw rate; is the road curvature; is the lateral error; is the vehicle mass; , are the lateral forces of the front and rear wheels respectively; is the driving force of the rear wheel; is the steering angle of the front wheel; is the moment of inertia of the vehicle body about the z-axis; are the distances from the center of mass to the front and rear axles of the vehicle respectively; Step S402, linearize and discretize the non-linear dynamics model ; for linearization, use the Taylor expansion formula, and for discretization, use the forward Euler method: ； In the formula, is the discretized state matrix; is the discretized control matrix; the state vector is , is the control vector at time Step S403, constructing an optimization problem of the model predictive controller, and the optimization goal is to find the optimal control input within the prediction horizon to minimize the following objective function: ； wherein, is the prediction time domain length of the model predictive controller; is the desired state vector; is the state vector at time is the change in the control increment; is the state weight matrix; R is the control weight matrix; During the optimization process, the constraint conditions of the model state quantity and the control quantity are as follows: ； In the formula, is the change amount of the state variable; , are the minimum value and the maximum value of the change amount of the state variable respectively; , are the minimum value and the maximum value of the change amount of the control variable respectively; , are the minimum value and the maximum value of the control amount respectively; Step S404, solving the optimization problem under the condition of satisfying the constraint conditions to obtain the optimal control sequence of the vehicle within the prediction horizon; 。 2. The dynamic drift obstacle avoidance control method for an autonomous driving vehicle based on driving condition recognition according to claim 1, wherein: In Step S10, the vehicle multi-dimensional dynamic features include vehicle speed, vehicle longitudinal acceleration, vehicle lateral acceleration, vehicle yaw rate, and front-wheel steering angle; the multi-dimensional dynamic features of the surrounding scene domain include the distance to surrounding obstacles, the relative speed of surrounding obstacles, road curvature, lane departure degree, and road adhesion coefficient; The in-vehicle sensors include: speed sensor, acceleration sensor, gyroscope, steering angle sensor, lidar, camera.

3. The dynamic drift obstacle avoidance control method for an autonomous driving vehicle based on driving condition recognition according to claim 2, wherein: Step S20 includes the following steps: Step S201, at each sampling moment , fuse all vehicle-mounted sensor data to form a feature vector. Different sensors have different sampling frequencies. Based on timestamps, data interpolation is performed to construct a normalized dynamic feature vector under a unified time reference : ； Wherein, is the vehicle speed at time is the longitudinal acceleration of the vehicle at time is the lateral acceleration of the vehicle at time is the yaw rate of the vehicle at time is the front wheel angle at time is the distance to the th peripheral obstacle at time is the relative speed to the th peripheral obstacle at time is the road curvature at time is the degree of lane departure at time is the road adhesion coefficient at time Step S202, calculate the information entropy of the dynamic feature vector at the moment : ； In the formula, is the probability density function of the th dimensional component of the dynamic feature vector at time ; is the dimension of the dynamic feature vector Step S203, construct a graph neural network GNN model, whose input is the historical information entropy sequence :[[]]END]] ； In the formula, is the number of samples of the historical information entropy sequence; Step S204, input the historical information entropy sequence into the GNN model for prediction of the information entropy at a moment : ； Step S205, calculate the actual perceived information entropy at the moment and the predicted information entropy the absolute error between them : ； Step S206, based on the prediction error , adaptively adjust the sampling frequency of the th vehicle-mounted sensor : ； In the formula, is the minimum sampling frequency of the th sensor; is the maximum sampling frequency of the th sensor; is the maximum error threshold preset for the th sensor.

4. The method for dynamically drifting obstacle avoidance control of an autonomous vehicle based on working condition recognition according to claim 3, wherein: Adaptive adjustment of the state weight matrix and the control weight matrix R： ① Determine the state weight matrix of the basis and the control weight matrix as follows: ； wherein, is the weight coefficient of the lateral error; is the weight coefficient of the heading angle error; is the weight coefficient of the longitudinal speed of the vehicle; is the weight coefficient of the sideslip angle of the vehicle's center of mass; is the weight coefficient of the yaw rate of the vehicle; is the weight coefficient of the front wheel steering angle; is the weight coefficient of the rear wheel driving force; ②According to the final possibility that the vehicle is in the drift critical area obtained in step S30 , adaptively adjust the state weight matrix and the control weight matrix R： ； In the formula, is the proportional coefficient matrix for matrix weight adjustment; is R the proportional coefficient matrix for matrix weight adjustment.

Images