Emergency obstacle avoidance method, system and device and storage medium

Abstract

The invention relates to an emergency obstacle avoidance method, system and device and a storage medium, and the method comprises the steps: inputting a longitudinal control instruction into an acceleration and deceleration optimization strategy in an execution stage, and obtaining an optimized longitudinal control instruction, the acceleration and deceleration optimization strategy being obtained by learning an optimization target including tracking precision and energy consumption cost; the transverse control instruction is input into a steering optimization strategy, an optimized transverse control instruction is obtained, and the steering optimization strategy is obtained by learning an optimization target including steering precision and stability cost; and controlling the vehicle to perform emergency obstacle avoidance according to the optimized transverse control instruction and longitudinal control instruction. By implementing the emergency obstacle avoidance method provided by the invention, the problem of poor emergency obstacle avoidance effect in the existing obstacle avoidance technology can be improved.

Description

Technical Field

[0001] The present application relates to the field of intelligent driving technology, and in particular to an emergency obstacle avoidance method, system, device and storage medium. Background Art

[0002] As one of the core functions of the vehicle's active safety system, emergency obstacle avoidance is widely used in autonomous driving and advanced driver assistance systems, and is of great significance to improving driving safety. The emergency obstacle avoidance system includes various stages such as perception, planning and execution. Through a series of complex and sophisticated operations, the emergency obstacle avoidance system can quickly identify and avoid obstacles in an emergency, thereby reducing the occurrence of traffic accidents.

[0003] However, although this technology has made significant progress, in practical applications, emergency obstacle avoidance still has the problem of poor effect when facing complex and changing environments. Summary of the invention

[0004] Based on this, the present application provides an emergency obstacle avoidance method, system, device and storage medium, which can improve the problem of poor emergency obstacle avoidance effect existing in existing obstacle avoidance technology.

[0005] In the first aspect, the present application provides an emergency obstacle avoidance method, which includes in the execution stage: inputting the longitudinal control instruction into the acceleration and deceleration optimization strategy to obtain the optimized longitudinal control instruction, wherein the acceleration and deceleration optimization strategy is obtained by learning the optimization objectives including tracking accuracy and energy consumption cost; inputting the lateral control instruction into the steering optimization strategy to obtain the optimized lateral control instruction, wherein the steering optimization strategy is obtained by learning the optimization objectives including steering accuracy and stability cost; controlling the vehicle for emergency obstacle avoidance according to the optimized lateral control instruction and longitudinal control instruction.

[0006] In combination with the first aspect, in a first possible implementation manner of the first aspect, the step of inputting the longitudinal control instruction into the acceleration and deceleration optimization strategy to obtain the optimized longitudinal control instruction includes: inputting the longitudinal control instruction into the acceleration and deceleration optimization strategy obtained by inverse reinforcement learning to obtain the optimized longitudinal control instruction, wherein the reward function of the inverse reinforcement learning includes tracking accuracy and energy consumption cost; the step of inputting the lateral control instruction into the steering optimization strategy to obtain the optimized lateral control instruction includes: inputting the lateral control instruction into the steering optimization strategy obtained by deep reinforcement learning to obtain the optimized lateral control instruction, wherein the reward function of the deep reinforcement learning includes steering accuracy and stability cost.

[0007] In combination with the first aspect, in a second feasible implementation of the first aspect, before the execution stage, it also includes a planning stage, and the emergency obstacle avoidance method also includes in the planning stage: based on the environmental model, using a deep deterministic policy gradient algorithm to solve the first optimization objective function to obtain a path adjustment strategy, and converting the path adjustment strategy into a continuous obstacle avoidance path, wherein the first optimization objective function includes obstacle avoidance cost, obstacle avoidance efficiency and energy efficiency; obtaining longitudinal control instructions and lateral control instructions according to the obstacle avoidance path, and sending the longitudinal control instructions and lateral control instructions to the execution layer to achieve control instruction optimization in the execution stage.

[0008] In combination with the second feasible implementation method of the first aspect, in the third feasible implementation method of the first aspect, the aforementioned step of converting the path adjustment strategy into a continuous obstacle avoidance path includes: inputting the path adjustment strategy into the neural differential equation framework to obtain a continuous initial path; adjusting the initial path using dynamic constraints so that the adjusted obstacle avoidance path satisfies the continuity constraints while satisfying the vehicle dynamics constraints.

[0009] In combination with the second possible implementation method of the first aspect, in a fourth possible implementation method of the first aspect, the aforementioned step of obtaining longitudinal control instructions and lateral control instructions based on the obstacle avoidance path includes: using a distributed model predictive control framework to perform parallel optimization on the obstacle avoidance path based on multiple sub-optimization objective functions, wherein each sub-optimization objective function is obtained by splitting the second optimization objective function, and the second optimization objective function includes path tracking error, vehicle stability, and energy consumption cost; obtaining corresponding longitudinal control instructions and lateral control instructions according to the optimized obstacle avoidance path.

[0010] In combination with the first aspect, in a fifth possible implementation manner of the first aspect, before the execution stage, it also includes a perception stage and a planning stage, and the emergency obstacle avoidance method also includes in the perception stage: obtaining a task weight matrix based on vehicle speed, obstacle dynamic characteristics and road characteristics, wherein the task weight matrix is ​​used to indicate the importance of each perception task for the emergency obstacle avoidance decision; processing the environment feature matrix and the vehicle feature matrix respectively through the two heads of the multi-head attention mechanism using the task weight matrix, and fusing the processing results to obtain an environment model; measuring the uncertainty of the environment model to obtain a confidence vector, and using the confidence vector to optimize the environment model to obtain an optimized environment model, wherein the optimized environment model is used to plan the obstacle avoidance path in the planning stage.

[0011] On the second aspect, the present application provides an emergency obstacle avoidance system, including an execution layer, which is used to: input the longitudinal control instruction into the acceleration and deceleration optimization strategy to obtain the optimized longitudinal control instruction, wherein the acceleration and deceleration optimization strategy is obtained by learning the optimization objectives including tracking accuracy and energy consumption cost; input the lateral control instruction into the steering optimization strategy to obtain the optimized lateral control instruction, wherein the steering optimization strategy is obtained by learning the optimization objectives including steering accuracy and stability cost; control the vehicle for emergency obstacle avoidance according to the optimized lateral control instruction and longitudinal control instruction.

[0012] In combination with the second aspect, in a first possible implementation mode of the second aspect, the emergency obstacle avoidance system also includes at least one of a perception layer, a planning layer, and a recovery layer, wherein: the perception layer is used to construct an environmental model based on the task weight matrix of emergency obstacle avoidance, and use the confidence vector of the environmental model to optimize the environmental model to obtain an optimized environmental model; the planning layer is used to obtain a path adjustment strategy based on the environmental model sent by the perception layer, using a deep deterministic policy gradient algorithm, and convert the path adjustment strategy into a continuous obstacle avoidance path, and then obtain the vehicle's longitudinal control instructions and lateral control instructions according to the planned obstacle avoidance path; the recovery layer is used to use an automatic encoder to analyze the state deviation of the vehicle during obstacle avoidance after the obstacle avoidance is completed to obtain a smooth recovery strategy, and gradually restore the vehicle to a normal driving state.

[0013] In a third aspect, the present application also provides an emergency obstacle avoidance device, which includes a processor, a transceiver and a memory, wherein the processor, the transceiver and the memory are connected via a bus; the processor is used to execute multiple instructions; the transceiver is used to exchange data with other devices; the memory is used to store multiple instructions, and the instructions are suitable for being loaded by the processor and executed as the emergency obstacle avoidance method of the first aspect or any one of the embodiments of the first aspect.

[0014] In a fourth aspect, the present application also provides a computer-readable storage medium, in which a plurality of instructions are stored, and the instructions are suitable for being loaded by a processor and executed by the emergency obstacle avoidance method such as the first aspect or any one embodiment of the first aspect.

[0015] In summary, the present application provides an emergency obstacle avoidance method, system, device, and storage medium, which can improve the problem of poor emergency obstacle avoidance effect in the prior art. The present application first improves the execution stage of emergency obstacle avoidance:

[0016] On the one hand, considering that the real-time changes in the environment may make the control instructions in the planning stage inapplicable, the present application optimizes the control instructions in the execution stage of emergency obstacle avoidance to obtain control instructions that are more suitable for the current real-time environment, thereby improving the emergency obstacle avoidance effect;

[0017] On the other hand, considering the real-time requirements of the execution phase and the needs of subsequent iterative updates, the present application divides the control instructions into lateral control instructions and longitudinal control instructions, and independently optimizes the lateral control instructions and longitudinal control instructions, so that the optimization of the lateral control instructions and the optimization of the longitudinal control instructions can be performed in parallel during the execution phase, thereby improving the efficiency of instruction optimization and meeting the real-time requirements of the execution phase. In addition, any optimization model of the lateral control instructions or the longitudinal control instructions can be designed and optimized separately in the future, simplifying the complexity of the emergency obstacle avoidance system and facilitating debugging and upgrading.

[0018] On the other hand, considering the different lateral and longitudinal motion characteristics and control requirements of the vehicle, the present application optimizes the longitudinal control and lateral control instructions according to different optimization objectives, focusing on balancing tracking accuracy and energy consumption costs in lateral control, and balancing steering accuracy and stability costs in longitudinal control, so that lateral control and longitudinal control focus on their respective optimization objectives, so as to further improve the learning efficiency and obstacle avoidance effect of the model, thereby achieving timely response, stable and low-energy emergency obstacle avoidance;

[0019] In general, this application optimizes the lateral control instructions and longitudinal control instructions in a targeted manner during the execution phase, so that the vehicle can cope with various real-time conditions and make correct responses in the face of complex and changing environments, thereby achieving real-time dynamic obstacle avoidance and improving the problem of poor emergency obstacle avoidance effect in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A flowchart of an emergency obstacle avoidance method in an embodiment of the present application;

[0021] Figure 2 An improved Transformer perception network in an embodiment of the present application;

[0022] Figure 3 Schematic diagram of the flow of an emergency obstacle avoidance method in another embodiment of the present application. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0024] In practical applications, existing emergency obstacle avoidance methods still have the problem of poor obstacle avoidance effect when facing complex and changing environments. For example, real-time changes in the environment may cause the control instructions generated by the planning layer to be inapplicable. In this regard, the present application optimizes the lateral control instructions and longitudinal control instructions respectively during the execution phase, so that the vehicle can cope with various real-time conditions and make correct responses when facing complex and changing environments, thereby realizing real-time dynamic obstacle avoidance and improving the obstacle avoidance effect.

[0025] like Figure 1 As shown, the present application provides an embodiment of an emergency obstacle avoidance method. Next, the present application takes the emergency obstacle avoidance system as the execution subject to describe the emergency obstacle avoidance method in detail:

[0026] 110: inputting the longitudinal control command into the acceleration / deceleration optimization strategy to obtain an optimized longitudinal control command, wherein the acceleration / deceleration optimization strategy is obtained by learning optimization objectives including tracking accuracy and energy consumption cost;

[0027] 120: inputting the lateral control command into the steering optimization strategy to obtain an optimized lateral control command, wherein the steering optimization strategy is obtained by learning an optimization target including steering accuracy and stability cost;

[0028] 130: Control the vehicle to perform emergency obstacle avoidance according to the optimized lateral control instructions and longitudinal control instructions.

[0029] For steps 110 and 120, the execution layer of the emergency obstacle avoidance system optimizes the lateral control instructions by taking tracking accuracy and energy consumption cost as optimization targets, and optimizes the longitudinal control instructions by taking steering accuracy and stability cost as optimization targets, thereby achieving targeted optimization. During the optimization, the same type of optimization algorithm can be used to optimize the two control instructions, or different types of optimization algorithms can be preferably used for optimization. Next, this application will illustrate these two methods with examples:

[0030] In the first possible implementation, the execution layer of the emergency obstacle avoidance system uses a type of algorithm such as Deep Reinforcement Learning (DRL) to optimize the longitudinal control instructions and the lateral control instructions. Compared with other algorithms, the reinforcement learning algorithm can autonomously learn the optimal strategy through interaction with the environment in the emergency obstacle avoidance task, and has unique advantages such as adapting to complex and dynamic environments. Therefore, the control instructions optimized by the reinforcement algorithm can improve the emergency obstacle avoidance effect;

[0031] In the second possible implementation, the emergency obstacle avoidance system adopts an optimization algorithm that is more suitable for different control instructions, that is, the inverse reinforcement learning (IRL) algorithm is used to optimize the longitudinal control instructions, and the deep reinforcement learning (DRL) algorithm is used to optimize the lateral control instructions. Specifically:

[0032] The aforementioned step 110 includes: inputting the longitudinal control instruction into the acceleration and deceleration optimization strategy obtained by inverse reinforcement learning to obtain an optimized longitudinal control instruction, wherein the reward function of the inverse reinforcement learning includes tracking accuracy and energy consumption cost;

[0033] The aforementioned step 120 includes: inputting the lateral control instruction into the steering optimization strategy obtained by deep reinforcement learning to obtain an optimized lateral control instruction, wherein the reward function of deep reinforcement learning includes steering accuracy and stability cost.

[0034] Among them, the acceleration and deceleration optimization strategy and the steering optimization strategy are inferred by the inverse reinforcement learning and deep reinforcement learning models respectively based on the reward function and with the goal of maximizing the reward. The acceleration and deceleration optimization strategy and the steering optimization strategy can output the optimal longitudinal control instructions and lateral control instructions in the current state according to the input state data such as the longitudinal control instructions and lateral control instructions.

[0035] This embodiment uses two models to solve different problems of lateral control and longitudinal control. Longitudinal control needs to be directly related to the speed regulation of the vehicle, which is crucial to safety. Therefore, this application uses IRL to learn safe acceleration and deceleration strategies from expert demonstrations to avoid overly aggressive or conservative speed regulation. Lateral control needs to cope with complex dynamic environments (such as the position, shape, and motion state of obstacles), and has high requirements for flexibility and real-time performance. Therefore, this application uses DRL to interact with the environment to autonomously learn steering strategies that adapt to complex environments, thereby improving steering flexibility and stability.

[0036] In short, in emergency obstacle avoidance scenarios, longitudinal control using inverse reinforcement learning can imitate expert behavior and improve safety and explainability; lateral control using deep reinforcement learning can adapt to complex environments and achieve real-time decision-making and flexibility. Therefore, by adopting two types of reinforcement learning, combining expert knowledge with autonomous learning, and giving full play to the advantages of both, the performance and reliability of the emergency obstacle avoidance system can be effectively improved.

[0037] In addition, it can be seen that the reward function plays a vital role in reinforcement learning, which is directly related to the setting and realization of the learning objectives of reinforcement learning. In order to achieve different optimization objectives of lateral control and longitudinal control, this application provides the determination process of the reward function of inverse reinforcement learning and the reward function of deep reinforcement learning, as well as the inference process of acceleration and deceleration optimization strategy and steering optimization strategy:

[0038] S11. For longitudinal control, first define the reward function R x (s, a), as shown in Formula 1, the reward function reflects the desired behavior characteristics, and the inverse reinforcement learning model is used to infer the implicit features of the reward function from the observed expert behavior or data, that is, those features that are not explicitly defined but affect the behavior selection;

[0039] Formula 1:

[0040] Among them, a x (t) is the longitudinal control command generated by the planning layer, i.e., the actual acceleration and deceleration. To optimize acceleration and deceleration, Indicates the deviation between actual acceleration and deceleration and optimized acceleration and deceleration, E x (t) is the energy consumption cost during acceleration and deceleration, λ1 and λ2 are weight coefficients used to balance tracking accuracy and energy consumption cost;

[0041] S12. Construct a cumulative reward function R based on the vertical reward function x , such as formula 2, which takes into account the discount effect of future rewards, and then maximizes the cumulative reward function R through reinforcement learning algorithms (such as Q learning, policy gradient method, etc.) x , thus deriving the acceleration and deceleration optimization strategy

[0042] Formula 2:

[0043] Among them, γ1 is the discount factor, which controls the decay of future rewards, R x (s t ,a t ) represents the reward value at time t, s t is the state of the vehicle at time t, including the speed v x (t), position P actual (t) and actual acceleration / deceleration a x (t), a t The acceleration and deceleration action of the vehicle;

[0044] S21. For lateral control, first define the reward function R y (s,a), for example, formula 3:

[0045] Formula 3: R y (s,a)=-[β1·‖δ(t)-δ * (t)|| 2 +β2·S y (t)];

[0046] Among them, δ(t) is the lateral control command generated in the planning stage, that is, the actual steering angle, δ * (t) is the optimized steering angle, ||δ(t)-δ * (t)|| is the deviation between the actual steering angle and the optimized steering angle, S y (t) is the lateral stability cost, β1 and β2 are weight coefficients used to balance steering accuracy and stability;

[0047] S22. Constructing state-action value function Q based on horizontal reward function y (s, a), as shown in Formula 4, and then the neural network of the deep reinforcement learning model is used to approximate the state-action value function, and the deep reinforcement learning model is trained so that the deep reinforcement learning model learns how to adjust the strategy to maximize the cumulative reward, thereby deriving the steering optimization strategy for lateral control

[0048] Formula 4: Q y (s,a)=R y (s,a)+γ·E s ·[Q y (s',a')];

[0049] Among them, s' is the state at the next moment after executing action a, a' is the action taken at the next moment, and E s Q y (s′, a′)] is the expected value based on the current strategy, Q y The optimization of (s,a) can be achieved through a deep Q network;

[0050] S31, the execution layer of the emergency obstacle avoidance system combined with the acceleration and deceleration optimization strategy and steering optimization strategies The final execution instruction u(t) including the longitudinal and lateral control of the vehicle is obtained. u(t) is also the comprehensive control instruction at time t, such as Formula 5. The execution layer controls the car to perform emergency obstacle avoidance by executing the comprehensive control instruction. After the obstacle avoidance is completed, the automatic encoder can also be used to analyze the state deviation of the vehicle during the obstacle avoidance period, obtain a smooth recovery strategy, and gradually restore the vehicle to a normal driving state:

[0051] Formula 5:

[0052] Among them, u(t) contains the acceleration and deceleration instructions a of the acceleration and deceleration optimization strategy at time t. x The execution layer of the emergency obstacle avoidance system executes u(t) based on the angular velocity command δ of the steering optimization strategy at time t to achieve precise obstacle avoidance, stable control and energy consumption optimization of the vehicle on low adhesion coefficient roads and complex working conditions.

[0053] In summary, the acceleration and deceleration optimization strategy inferred by the inverse reinforcement learning method in this application realizes the rapid response and smooth control of the vehicle in the emergency obstacle avoidance process. At the same time, the steering optimization strategy inferred by the deep reinforcement learning method optimizes the vehicle's handling performance on low adhesion coefficient roads and complex working conditions. Compared with the traditional control method based on fixed rules, the emergency obstacle avoidance method of this application can adjust the control instructions according to the real-time working conditions, which significantly improves the stability and safety of the vehicle under extreme conditions. Experimental data show that the control error of this method on low adhesion coefficient roads is reduced by 30%, and the obstacle avoidance success rate in complex dynamic environments is increased to more than 96%, enhancing the adaptability and safety of the vehicle.

[0054] In addition, the planning stage of the existing obstacle avoidance method is still poor in optimizing target balance, computational efficiency or trajectory smoothness, and it is difficult to adapt to dynamic changes in complex environments in real time. In this regard, the present application also improves the planning stage. Next, the planning layer of the emergency obstacle avoidance system is used as the execution body for explanation:

[0055] 210: The planning layer of the emergency obstacle avoidance system is based on the environment model and uses the Deep Deterministic Policy Gradient (DDPG) algorithm to solve the first optimization objective function to obtain the path adjustment strategy π θ (a|s), as shown in Formula 6, wherein the first optimization objective function includes obstacle avoidance cost, obstacle avoidance efficiency and energy efficiency, and the first optimization objective function is as shown in Formula 7:

[0056] Formula 6: π θ (a|s)=argmax a Q Φ (s,a);

[0057] Among them, π θ (a|s) is the path adjustment strategy output by the DDPG policy network, S = {M * , C is the state of the current environment, Q Φ (s,a) is the value network, which is used to estimate the value of performing action a in state s. θ and Φ are the trainable parameters of the policy network and value network of DDPG.

[0058] Formula 7:

[0059] Among them, P path (t) is the path position of the vehicle at time t, P o (t) is the position of the obstacle at time t, |P path (t)-P o (t)‖ 2 represents the obstacle avoidance cost; C dyn (t) is the dynamic constraint cost of the path, which represents the obstacle avoidance efficiency of the vehicle in a dynamic environment, such as C dyn (t) = α1·(a lat (t)) 2 +α2·(r yaw (t)) 2 , a lat (t) is the instantaneous lateral acceleration, r yaw (t) is the yaw angular acceleration, α1 and α2 represent the priority weights of the instantaneous lateral acceleration and yaw angular acceleration; E eff (t) is the path energy cost, which represents the energy efficiency in path planning, for example, E eff (t) = ∫P(t)dt, P(t) is the power consumption of the motor; λ1, λ2, λ3 are the weight coefficients of obstacle avoidance cost, obstacle avoidance efficiency and energy efficiency, which are used to dynamically balance the priorities of different goals in path planning;

[0060] 220: Convert the path adjustment strategy into a continuous obstacle avoidance path;

[0061] For step 220, since the DDPG output result is a continuous action (acceleration, deceleration and steering angle) rather than a continuous path, the planning layer can use at least two methods to obtain a continuous obstacle avoidance path:

[0062] One way is to transform the path adjustment strategy into a continuous obstacle avoidance path through a dynamic model;

[0063] Another way is to constrain the continuity of the path through the neural differential equation framework to make the obstacle avoidance path smoother to improve the smoothing effect. The specific implementation process of this method may include:

[0064] 221: Input the path adjustment strategy into the neural differential equation framework to obtain a continuous initial path to obtain an initial path, such as formula 8. The initial path can be directly used as an obstacle avoidance path, or step 222 can be executed to further constrain the logistics feasibility of the path;

[0065] Formula 8:

[0066] Among them, P path (t) is the initial path, f ode(·) is the dynamic function of the neural differential equation framework, θ ode are neural network parameters used to ensure that the path remains smooth and physically feasible during the acquisition process;

[0067] 222: Using the dynamic constraint condition (e.g., Formula 9) to adjust the initial path, so that the adjusted obstacle avoidance path satisfies the continuity constraint and the vehicle dynamics constraint;

[0068] Formula 9:

[0069] Where v(t) is the vehicle’s position on the obstacle avoidance path P. path (t) is the velocity on the surface, a(t) is the acceleration, F long (t) and F lat (t) are the longitudinal force and lateral force of the vehicle, v max 、a max 、F long,max 、F lat,max are the upper limits of vehicle dynamics constraints, namely the maximum speed, acceleration, longitudinal force and lateral force respectively.

[0070] 230: Obtain longitudinal control instructions and lateral control instructions according to the obstacle avoidance path, and send the longitudinal control instructions and lateral control instructions to the execution layer to optimize the control instructions in the execution stage;

[0071] With respect to step 230, the present application may directly obtain the longitudinal control command and the lateral control command of the vehicle according to the planned obstacle avoidance path, or may further optimize the obstacle avoidance path and then obtain the corresponding longitudinal control command and the lateral control command according to the optimized obstacle avoidance path. Further optimization is because the obstacle avoidance path may deviate from the expectation after the continuity constraint and the dynamic constraint. Therefore, the path may be further optimized through multi-objective optimization. The latter method specifically includes:

[0072] 231: The obstacle avoidance path is optimized in parallel based on multiple sub-optimization objective functions using a distributed model predictive control framework (DMPC), wherein each sub-optimization objective function is obtained by splitting the second optimization objective function, and the second optimization objective function includes path tracking error, vehicle stability and energy consumption cost. The second optimization objective function is shown in Formula 10, and each sub-optimization objective function is shown in Formula 11:

[0073] Formula 10:

[0074] Formula 11:

[0075] Among them, J is the second optimization objective function, J1, J2, and J3 are sub-optimization objective functions, and P actual (t) is the actual position of the vehicle, P path (t) is the position of the vehicle on the obstacle avoidance path, S(t) is the vehicle stability cost function, which can be combined with the yaw rate Ψ(t) and the lateral acceleration a y (t) correlation, s(t) = k1|Ψ(t)|+k2|a y (t)|, k1 and k2 are the optimizable weight coefficients of yaw rate and lateral acceleration, E(t) is the vehicle energy cost function, which can be combined with the longitudinal driving force F x (t) and braking force B x (t) related, E(t) = η1|F x (t)|+η1|B x (t)|, k1 and k2 are the optimizable weight coefficients of the longitudinal driving force and braking force; α1, α2, α3 are the weight coefficients of the objective function, which can be optimized through self-supervised learning. The optimization process is shown in Formula 12:

[0076] Formula 12:

[0077] Among them, α i (t) is the α at the current time t i The weight value, α i (t+1) is α i In the updated weight value, γ is the learning rate, which is used to control the step size of weight adjustment. Represents the function of weight α i The partial derivative of is used to reflect the contribution of the current weight to the optimization effect;

[0078] 232: Obtain corresponding longitudinal control instructions and lateral control instructions according to the optimized obstacle avoidance path. The longitudinal control instruction a x (t) is as shown in Formula 13, and the lateral control command δ(t) is as shown in Formula 14;

[0079] Formula 13:

[0080] Formula 14: δ(t) = K p ·Δy(t)+K d Ψ(t),

[0081] Among them, F x (t) is the longitudinal driving force, m is the vehicle mass, g is the acceleration of gravity, θ is the road slope angle, a x (t) represents acceleration and deceleration; K p and K dare proportional gain and differential gain, respectively, used to adjust the lateral control accuracy, Δy(t) is the target path deviation, Δy(t)=y path (t)-y actual (t), is the lateral deviation between the obstacle avoidance path and the actual path, Ψ(t) is the yaw rate, and δ(t) represents the steering angle.

[0082] In summary, in the planning stage, this application obtains a real-time, dynamic and continuous obstacle avoidance path through a deep deterministic policy gradient algorithm and a neural differential equation framework, and introduces dynamic constraints in the path generation process to ensure the physical feasibility of the trajectory. It also uses distributed model predictive control and self-supervised learning strategies to perform multi-objective optimization of path tracking error, vehicle stability and energy consumption to dynamically adjust control parameters, thereby achieving efficient acquisition and execution of control instructions, effectively solving the problem of disconnection between traditional path planning and execution, and significantly improving the real-time and smoothness of the path. The vehicle's yaw angular velocity deviation in complex road conditions is reduced by about 20%, the trajectory tracking error is reduced to below 0.2 meters, and the overall energy consumption is reduced by about 12%.

[0083] In addition, the perception stage of existing obstacle avoidance technology has limited ability to process uncertainty in sensor data and cannot provide high-confidence perception results in a rapidly changing environment. When obstacles suddenly change, the existing algorithm may cause a decrease in perception accuracy due to noise interference or data delay, thereby affecting the real-time and reliability of path planning. In this regard, the present application also provides an implementable method for the perception stage:

[0084] This application provides an improved Transformer perception network, such as Figure 2 As shown in FIG. 1 , the network includes an embedding layer, a priority perception module, an encoder (including a multi-head self-attention module, a cross-attention module, and a fusion feature matrix module), a decoder, a confidence module, and a feature guidance module, wherein:

[0085] 310: The embedding layer receives the environment data set and the vehicle status data set acquired in real time by the multimodal sensor, and transforms the environment data set E = {e1, e2, …e n} and vehicle state dataset V = {v1,v2,…v m} Perform high-dimensional feature embedding processing to obtain the embedded feature matrix F E and F V , F E and F V The mathematical expression of is as shown in Formula 15, wherein the environmental data set includes the relative position, dynamic characteristics and road characteristics of the obstacles, and the vehicle state data set includes the real-time torque output, braking force distribution, tire grip and vehicle speed of the vehicle;

[0086] Formula 15: F E =ΦE (E) and F V =Φ V (V),

[0087] Among them, Φ E (·) and Φ V (·) are the embedding functions of the environment data and vehicle status data, respectively, which are used to map the data into a d-dimensional feature space and capture the nonlinear relationship between the environment features and the vehicle status;

[0088] 320: The task priority dynamic allocation network in the priority perception module obtains the task weight matrix W according to the vehicle speed, the dynamic characteristics of the obstacle and the road characteristics, as shown in Formula 16, wherein the task weight matrix is ​​used to represent the importance of each perception task for the emergency obstacle avoidance decision, and the task priority dynamic allocation network can obtain the task priority weight matrix according to the real-time requirements of the emergency obstacle avoidance scenario;

[0089] Formula 16: W = softmax(U·Ψ(V c ,V o ,D)+b);

[0090] Among them, softmax(·) is the activation function, and is the trainable weight matrix and bias vector, Ψ(V c ,V o ,D) is the feature extraction function based on the vehicle speed V c , obstacle dynamic characteristics V o The h-dimensional task feature vector obtained by the road feature D is used to calculate the priority of each task in the current scene and highlight the perception tasks that are key to obstacle avoidance decision-making;

[0091] 330: The environment feature matrix and the vehicle feature matrix are processed by the two heads of the multi-head attention mechanism respectively using the task weight matrix, and the processing results are integrated to obtain the environment model;

[0092] This step is completed by the encoder and decoder. The encoder is used to embed the feature matrix F using the weighted multi-head self-attention mechanism. E and F V Perform interactive calculations to obtain the fusion feature matrix F fusion , F fusion =concat(Attention(F E ,W),Attention(F V ,W)), the decoder fused feature matrix F fusion Decoding is performed to obtain the environment model M containing obstacle position, velocity, acceleration and road curvature by mapping the fused feature matrix to the feature space of the environment model, where M = Φdec (F fusion ), where Φ dec (·) is the decoding function, where concat(·) represents the matrix concatenation operation, which weights the embedded features through the priority weight matrix W, emphasizes the key task features, and enhances the attention to the emergency obstacle avoidance related information;

[0093] 340: The confidence module measures the uncertainty of the environment model to obtain a confidence vector, and uses the confidence vector to optimize the environment model to obtain an optimized environment model, wherein the optimized environment model is used to plan an obstacle avoidance path in the planning stage.

[0094] The confidence module may include a dynamic Bayesian network, which quantifies the uncertainty of the initial model M and calculates the confidence vector C of each model feature;

[0095] 350: The feature guidance module performs confidence weighting processing on the environment model M to obtain an optimized environment model in, The feature f i The historical average or default safety value, c i The feature f i confidence level.

[0096] In summary, this implementation example uses an improved Transformer perception network to fuse the environmental data set and the vehicle status data set to obtain an environmental model, and quantifies the uncertainty of the environmental data based on a dynamic Bayesian network to provide confidence information for the environmental model. Therefore, compared with the limitations of traditional perception methods that rely on a single data feature, this implementation greatly improves the confidence of the perception results, and can quantify and optimize the confidence of the environmental model in real time, so that the system can accurately capture the dynamic characteristics of obstacles and environmental changes in a dynamic and complex environment, significantly improving the accuracy and reliability of perception, effectively solving the problem of fluctuations in perception results caused by sensor noise and dynamic environmental changes, and providing high-confidence data support for path planning. Experimental results show that the accuracy of dynamic obstacle detection has increased by about 15%, providing more reliable data support for subsequent real-time path planning.

[0097] Based on the above several possible implementation methods, it can be seen that the present application provides improvements in all stages of the emergency obstacle avoidance process, including the perception stage, planning stage, execution stage, and recovery stage. Figure 3 , an overall introduction to the entire emergency obstacle avoidance process of this application is given.

[0098] 410: Emergency obstacle avoidance data acquisition: The perception layer of the emergency obstacle avoidance system acquires emergency obstacle avoidance data through multimodal sensors. The emergency obstacle avoidance data includes an environmental data set and a vehicle status data set.

[0099] 420: Improved Transformer Perception Network Fusion: The perception layer of the emergency obstacle avoidance system inputs the emergency obstacle avoidance data into the improved Transformer perception network to fuse and obtain a reliable environment model;

[0100] 430: Path planning based on deep policy gradient: The planning stage of the emergency obstacle avoidance system is based on the environment model, and the path is planned according to the DDPG algorithm to obtain the obstacle avoidance path;

[0101] 440: Distributed Optimization Control Instruction Generation: In the planning stage of the emergency obstacle avoidance system, the distributed model predictive control framework is used to perform multi-objective optimization on the obstacle avoidance path to obtain the optimized obstacle avoidance path, and the corresponding longitudinal control instructions and lateral control instructions are obtained according to the optimized obstacle avoidance path;

[0102] 450: Dynamic acceleration and deceleration and corner control: In the execution phase of the emergency obstacle avoidance system, the lateral control instructions are optimized with tracking accuracy and energy consumption cost as the optimization targets, and the longitudinal control instructions are optimized with steering accuracy and stability cost as the optimization targets, and then the two optimized control instructions are executed;

[0103] 460: State recovery after obstacle avoidance: After obstacle avoidance is completed, the recovery phase of the emergency obstacle avoidance system uses an autoencoder to analyze the state deviation of the vehicle during obstacle avoidance, obtain a smooth recovery strategy, and gradually restore the vehicle to a normal driving state.

[0104] The present application also provides an emergency obstacle avoidance system, which includes an execution layer, which is used to: input a longitudinal control instruction into an acceleration / deceleration optimization strategy to obtain an optimized longitudinal control instruction, wherein the acceleration / deceleration optimization strategy is obtained by learning an optimization target including tracking accuracy and energy consumption cost; input a lateral control instruction into a steering optimization strategy to obtain an optimized lateral control instruction, wherein the steering optimization strategy is obtained by learning an optimization target including steering accuracy and stability cost; and control the vehicle for emergency obstacle avoidance according to the optimized lateral control instruction and longitudinal control instruction.

[0105] In one practicable manner, the emergency obstacle avoidance system also includes at least one of a perception layer, a planning layer, and a recovery layer, wherein: the perception layer is used to construct an environmental model based on the task weight matrix of emergency obstacle avoidance, and optimize the environmental model using the confidence vector of the environmental model to obtain an optimized environmental model; the planning layer is used to obtain a path adjustment strategy based on the environmental model sent by the perception layer using a deep deterministic policy gradient algorithm, and convert the path adjustment strategy into a continuous obstacle avoidance path, and then obtain the longitudinal control instructions and lateral control instructions of the vehicle according to the planned obstacle avoidance path; the recovery layer is used to analyze the state deviation of the vehicle during obstacle avoidance using an automatic encoder after the obstacle avoidance is completed to obtain a smooth recovery strategy, and gradually restore the vehicle to a normal driving state.

[0106] In one implementable manner, the execution layer is specifically used to: input the longitudinal control instruction into the acceleration and deceleration optimization strategy obtained by inverse reinforcement learning to obtain the optimized longitudinal control instruction, wherein the reward function of inverse reinforcement learning includes tracking accuracy and energy consumption cost; input the lateral control instruction into the steering optimization strategy obtained by deep reinforcement learning to obtain the optimized lateral control instruction, wherein the reward function of deep reinforcement learning includes steering accuracy and stability cost.

[0107] In one implementable manner, the planning layer is specifically used for: based on the environmental model, using a deep deterministic policy gradient algorithm to solve the first optimization objective function to obtain a path adjustment strategy, and converting the path adjustment strategy into a continuous obstacle avoidance path, wherein the first optimization objective function includes obstacle avoidance cost, obstacle avoidance efficiency, and energy efficiency; obtaining longitudinal control instructions and lateral control instructions according to the obstacle avoidance path, and sending the longitudinal control instructions and lateral control instructions to the execution layer to achieve control instruction optimization in the execution stage.

[0108] In one implementable manner, the planning layer is specifically used to: input the path adjustment strategy into the neural differential equation framework to obtain a continuous initial path; and adjust the initial path using dynamic constraints so that the adjusted obstacle avoidance path satisfies the continuity constraints while satisfying the vehicle dynamics constraints.

[0109] In one implementable manner, the planning layer is specifically used to: utilize a distributed model predictive control framework to perform parallel optimization of obstacle avoidance paths based on multiple sub-optimization objective functions, wherein each sub-optimization objective function is obtained by splitting a second optimization objective function, and the second optimization objective function includes path tracking error, vehicle stability, and energy consumption cost; obtain corresponding longitudinal control instructions and lateral control instructions according to the optimized obstacle avoidance path.

[0110] In one implementable manner, the perception layer is specifically used to: obtain a task weight matrix based on vehicle speed, obstacle dynamic characteristics and road characteristics, wherein the task weight matrix is ​​used to indicate the importance of each perception task for emergency obstacle avoidance decision-making; process the environment feature matrix and the vehicle feature matrix respectively through the two heads of the multi-head attention mechanism using the task weight matrix, and fuse the processing results to obtain an environment model; measure the uncertainty of the environment model to obtain a confidence vector, and use the confidence vector to optimize the environment model to obtain an optimized environment model, wherein the optimized environment model is used to plan obstacle avoidance paths in the planning stage.

[0111] The present application also provides an emergency obstacle avoidance device, which may include: a processor, a transceiver and a memory. The processor and the memory are connected via a bus. The processor is used to execute multiple instructions; the transceiver is used to exchange data with other devices; the memory is used to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor for executing the aforementioned emergency obstacle avoidance method, specifically for executing any of the aforementioned emergency obstacle avoidance methods that can be implemented, or for realizing the functions of any of the perception layer, planning layer, execution layer and recovery layer in the emergency obstacle avoidance system. Among them, the processor can be an electronic control unit (Electronic Control Unit, ECU), a central processing unit (central processing unit, CPU), a general-purpose processor, a coprocessor, a digital signal processor (digital signal processor, DSP), an application-specific integrated circuit (application-specific integrated circuit, ASIC), a field programmable gate array (field programmable gate array, FPGA) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. The processor can also be a combination that implements a computing function, such as a combination of one or more microprocessors, a combination of 5SP and a microprocessor, and the like. In this embodiment, the processor can adopt a single-chip microcomputer, and various control functions can be realized by programming the single-chip microcomputer. The processor has the advantages of powerful computing power and fast processing. Specifically: the processor is used to realize the functions of the execution layer, and the processor is used to: input the longitudinal control instruction into the acceleration and deceleration optimization strategy to obtain the optimized longitudinal control instruction, wherein the acceleration and deceleration optimization strategy is obtained by learning the optimization objectives including tracking accuracy and energy consumption cost; input the lateral control instruction into the steering optimization strategy to obtain the optimized lateral control instruction, wherein the steering optimization strategy is obtained by learning the optimization objectives including steering accuracy and stability cost; control the vehicle for emergency obstacle avoidance according to the optimized lateral control instruction and longitudinal control instruction.

[0112] In one implementable manner, the processor can also be used to implement the functions of at least one of the perception layer, planning layer and recovery layer, and the processor is used to: construct an environmental model based on the task weight matrix of emergency obstacle avoidance, and use the confidence vector of the environmental model to optimize the environmental model to obtain an optimized environmental model; based on the environmental model sent by the perception layer, use a deep deterministic policy gradient algorithm to obtain a path adjustment strategy, and convert the path adjustment strategy into a continuous obstacle avoidance path, and then obtain the longitudinal control instructions and lateral control instructions of the vehicle according to the planned obstacle avoidance path; after the obstacle avoidance is completed, use an automatic encoder to analyze the state deviation of the vehicle during the obstacle avoidance to obtain a smooth recovery strategy, and gradually restore the vehicle to a normal driving state.

[0113] In one implementable manner, the processor is specifically used to: input the longitudinal control instruction into the acceleration and deceleration optimization strategy obtained by inverse reinforcement learning to obtain the optimized longitudinal control instruction, wherein the reward function of the inverse reinforcement learning includes tracking accuracy and energy consumption cost; input the lateral control instruction into the steering optimization strategy obtained by deep reinforcement learning to obtain the optimized lateral control instruction, wherein the reward function of the deep reinforcement learning includes steering accuracy and stability cost.

[0114] In one implementable manner, the processor is also used to implement the functions of the planning layer, and the processor is used to: based on the environmental model, use the deep deterministic policy gradient algorithm to solve the first optimization objective function to obtain a path adjustment strategy, and convert the path adjustment strategy into a continuous obstacle avoidance path, wherein the first optimization objective function includes obstacle avoidance cost, obstacle avoidance efficiency and energy efficiency; obtain longitudinal control instructions and lateral control instructions according to the obstacle avoidance path, and send the longitudinal control instructions and lateral control instructions to the execution layer to realize the optimization of control instructions in the execution stage.

[0115] In one implementable manner, the processor is specifically used to: input the path adjustment strategy into the neural differential equation framework to obtain a continuous initial path; and adjust the initial path using dynamic constraints so that the adjusted obstacle avoidance path satisfies the continuity constraints while satisfying the vehicle dynamics constraints.

[0116] In one implementable manner, the processor is specifically used to: utilize a distributed model predictive control framework to perform parallel optimization on the obstacle avoidance path based on multiple sub-optimization objective functions, wherein each sub-optimization objective function is obtained by splitting a second optimization objective function, and the second optimization objective function includes path tracking error, vehicle stability, and energy consumption cost; obtain corresponding longitudinal control instructions and lateral control instructions according to the optimized obstacle avoidance path.

[0117] In one implementable manner, the processor is also used to implement the functions of the perception layer, and the processor is used to: obtain a task weight matrix based on vehicle speed, obstacle dynamic characteristics and road characteristics, wherein the task weight matrix is ​​used to indicate the importance of each perception task for emergency obstacle avoidance decision-making; process the environment feature matrix and the vehicle feature matrix respectively through the two heads of the multi-head attention mechanism using the task weight matrix, and fuse the processing results to obtain an environment model; measure the uncertainty of the environment model to obtain a confidence vector, and use the confidence vector to optimize the environment model to obtain an optimized environment model, wherein the optimized environment model is used to plan obstacle avoidance paths in the planning stage.

[0118] In one embodiment, the present application further provides a computer-readable storage medium storing a plurality of instructions, which are suitable for being loaded by a processor and executing the method in any of the foregoing embodiments. The processor is used to execute the plurality of instructions; the memory is used to store the plurality of instructions, which are loaded by the processor and executed as the emergency obstacle avoidance method in the foregoing embodiments.

[0119] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0120] The above embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. An emergency obstacle avoidance method, characterized in that: The emergency obstacle avoidance method comprises the following steps during the execution phase: Inputting the longitudinal control command into the acceleration / deceleration optimization strategy to obtain an optimized longitudinal control command, wherein the acceleration / deceleration optimization strategy is obtained by learning optimization objectives including tracking accuracy and energy consumption cost; Inputting the lateral control command into the steering optimization strategy to obtain an optimized lateral control command, wherein the steering optimization strategy is obtained by learning an optimization target including steering accuracy and stability cost; The vehicle is controlled to perform emergency obstacle avoidance according to the optimized lateral and longitudinal control instructions.

2. The method according to claim 1, characterized in that: The step of inputting the longitudinal control instruction into the acceleration / deceleration optimization strategy to obtain the optimized longitudinal control instruction comprises: inputting the longitudinal control instruction into the acceleration / deceleration optimization strategy obtained by inverse reinforcement learning to obtain the optimized longitudinal control instruction, wherein the reward function of the inverse reinforcement learning comprises tracking accuracy and energy consumption cost; The step of inputting the lateral control instruction into the steering optimization strategy to obtain the optimized lateral control instruction includes: inputting the lateral control instruction into the steering optimization strategy obtained by deep reinforcement learning to obtain the optimized lateral control instruction, wherein the reward function of the deep reinforcement learning includes steering accuracy and stability cost.

3. The method according to claim 1, characterized in that Before the execution phase, a planning phase is also included. The emergency obstacle avoidance method further includes: Based on the environment model, a first optimization objective function is solved by using a deep deterministic policy gradient algorithm to obtain a path adjustment strategy, and the path adjustment strategy is converted into a continuous obstacle avoidance path, wherein the first optimization objective function includes obstacle avoidance cost, obstacle avoidance efficiency and energy efficiency; A longitudinal control instruction and a lateral control instruction are obtained according to the obstacle avoidance path, and the longitudinal control instruction and the lateral control instruction are sent to the execution layer to achieve control instruction optimization in the execution stage.

4. The method according to claim 3, characterized in that: The step of converting the path adjustment strategy into a continuous obstacle avoidance path comprises: Inputting the path adjustment strategy into the neural differential equation framework to obtain a continuous initial path; The initial path is adjusted using the dynamic constraint condition so that the adjusted obstacle avoidance path satisfies the continuity constraint and the vehicle dynamics restriction at the same time.

5. The method according to claim 3, characterized in that: The step of obtaining a longitudinal control instruction and a lateral control instruction according to the obstacle avoidance path comprises: The obstacle avoidance path is optimized in parallel based on a plurality of sub-optimization objective functions using a distributed model predictive control framework, wherein each sub-optimization objective function is obtained by splitting a second optimization objective function, and the second optimization objective function includes a path tracking error, a vehicle stability, and an energy consumption cost; The corresponding longitudinal control instructions and lateral control instructions are obtained according to the optimized obstacle avoidance path.

6. The method according to claim 1, characterized in that The execution phase also includes a perception phase and a planning phase. The emergency obstacle avoidance method further includes the following steps in the perception phase: Obtaining a task weight matrix according to vehicle speed, obstacle dynamic characteristics and road characteristics, wherein the task weight matrix is ​​used to represent the importance of each perception task for emergency obstacle avoidance decision-making; The environment feature matrix and the vehicle feature matrix are processed respectively by two heads of the multi-head attention mechanism using the task weight matrix, and the processing results are integrated to obtain the environment model; The uncertainty of the environmental model is measured to obtain a confidence vector, and the environmental model is optimized using the confidence vector to obtain an optimized environmental model, wherein the optimized environmental model is used to plan an obstacle avoidance path in the planning stage.

7. An emergency obstacle avoidance system, characterized in that: The invention comprises an execution layer, wherein the execution layer is used to: Inputting the longitudinal control command into the acceleration / deceleration optimization strategy to obtain an optimized longitudinal control command, wherein the acceleration / deceleration optimization strategy is obtained by learning optimization objectives including tracking accuracy and energy consumption cost; Inputting the lateral control command into the steering optimization strategy to obtain an optimized lateral control command, wherein the steering optimization strategy is obtained by learning an optimization target including steering accuracy and stability cost; The vehicle is controlled to perform emergency obstacle avoidance according to the optimized lateral and longitudinal control instructions.

8. The emergency obstacle avoidance system according to claim 7, characterized in that: The emergency obstacle avoidance system further includes at least one of a perception layer, a planning layer, and a recovery layer, wherein: The perception layer is used to construct an environment model based on the task weight matrix of emergency obstacle avoidance, and optimize the environment model using the confidence vector of the environment model to obtain an optimized environment model; The planning layer is used to obtain a path adjustment strategy based on the environment model sent by the perception layer using a deep deterministic policy gradient algorithm, and convert the path adjustment strategy into a continuous obstacle avoidance path, and then obtain the longitudinal control command and lateral control command of the vehicle according to the obstacle avoidance path obtained by the planning; The recovery layer is used to analyze the state deviation of the vehicle during obstacle avoidance using an automatic encoder to obtain a smooth recovery strategy after obstacle avoidance is completed, and gradually restore the vehicle to a normal driving state.

9. An emergency obstacle avoidance device, characterized in that: The emergency obstacle avoidance device includes a processor, a transceiver and a memory, and the processor, transceiver and memory are connected via a bus; the processor is used to execute multiple instructions; the transceiver is used to exchange data with other devices; the memory is used to store the multiple instructions, and the instructions are suitable for being loaded by the processor and executing the emergency obstacle avoidance method described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor and executing the emergency obstacle avoidance method according to any one of claims 1 to 6.

Images