A vehicle lane-changing control method integrating multimodal perception and DRL-MPC

Abstract

A vehicle lane-changing control method integrating multimodal perception and DRL-MPC is used to ensure the safety of vehicle lane-changing in complex traffic environments. The present invention relates to the field of intelligent driving. The present invention comprises two layers, the upper layer comprising an environment perception module, a driving intention recognition module, and a multimodal trajectory prediction module, and the lower layer comprising a DRL decision module and an MPC module. The upper layer utilizes the environment perception module to extract traffic environment information from the driving environment to generate BEV feature information, the driving intention recognition module generates driving intention probability based on the BEV feature information, and the multimodal trajectory prediction module generates multimodal trajectory information based on the BEV feature information and the driving intention probability; the lower layer DRL decision module generates the expected vehicle speed, expected front wheel turning angle, target weight, and control quantity weight based on the environment state information and the multimodal trajectory information, and finally the MPC module solves and determines the optimal control action to achieve lane-changing control of the autonomous driving vehicle.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent driving, and specifically provides a vehicle lane-changing control method integrating multimodal perception with DRL-MPC. Background Art

[0002] With the rapid development of intelligent and connected vehicles, intelligent driving technology has become a key approach to improving vehicle safety and energy efficiency. Within this context, the stability of vehicle lane-changing control, a core challenge in intelligent driving, directly impacts driving safety. Model predictive control (MPC) is widely used in autonomous lane-changing control by predicting the vehicle's future state and optimizing the control sequence. However, MPC relies on physical models, and its computational complexity increases with the prediction step size and constraint complexity, making it difficult to adapt to complex road environments in real time. Deep reinforcement learning (DRL) can generate control policies through interactive learning with the environment, but its lack of generalization makes it susceptible to loss of control in novel scenarios. When controlling the motion of autonomous vehicles, MPC can leverage its advantages in considering safety constraints and actuator physical constraints, acting as a safety shield to ensure that the control actions output by DRL guarantee vehicle driving safety. Furthermore, DRL's ability to continuously interact with the environment for global optimization can be leveraged, providing a priori control guidance for MPC and significantly reducing the computational burden of online MPC solutions. Therefore, the DRL-MPC fusion vehicle lane-changing control method can effectively cope with complex dynamic traffic scenarios and achieve stable and efficient lane-changing control.

[0003] Currently, in the field of vehicle lane change control using DRL-MPC fusion, patent CN117922567A studies longitudinal trajectory errors based on MPC. The DRL module collects the spacing, velocity, and acceleration errors of each vehicle in the formation in real time, learns and adjusts the control weights in MPC online, and achieves multi-objective optimization of the desired acceleration of the intelligent connected vehicles in the formation, which is converted from the lower layer into control instructions to achieve safe and stable coordinated control of the queue. However, this method only focuses on longitudinal dynamics and lacks in-depth modeling of lateral behavior and road semantics, making it difficult to ensure robustness in complex traffic scenarios. Patent CN118605160A, based on the MPC framework, designs a quadratic performance indicator that includes trajectory error and stability error, and uses the Deep Deterministic Policy Gradient (DDPG) algorithm to update the weight matrix in real time to coordinate lateral tracking accuracy and lateral stability. However, this method fails to consider the multi-source environmental information from the BEV's perspective and the driver's driving intentions, making it difficult to adapt to complex traffic scenarios. Patent CN117360544A uses MPC to solve the initial optimal turning angle sequence and uses DRL combined with tracking error to adjust and optimize the MPC prediction results in real time, significantly improving lateral control accuracy and anti-interference capabilities. However, this method uses a simplified two-degree-of-freedom bicycle model and fails to fully utilize the complex interactions between vehicles, which can easily lead to safety hazards in complex traffic environments with multiple vehicles interacting. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a vehicle lane change control method that integrates multimodal perception and DRL-MPC. This method uses the hierarchical Transformer in the environmental perception module to extract traffic environment information from the driving environment to generate BEV feature information. The driving intention recognition module outputs the driving intention probability based on the BEV feature information. The multimodal trajectory prediction module generates multimodal trajectory information based on the driving intention probability and the BEV feature information. The DRL decision module generates the expected vehicle speed, expected front wheel angle, target weight, and control quantity weight end-to-end. Finally, the MPC online optimization solver determines the optimal control action to achieve lane change control of the autonomous driving vehicle. This method effectively improves the foresight, risk assessment ability, and generalization ability of the autonomous driving vehicle in selecting lane change plans while ensuring safety.

[0005] The technical solutions adopted by the present invention to solve the technical problems are as follows:

[0006] The present invention provides a vehicle lane-changing control method integrating multimodal perception and DRL-MPC. The method comprises two layers, the upper layer comprising a multimodal environment perception module, a driving intention recognition module, and a multimodal trajectory prediction module, and the lower layer comprising a DRL decision module and an MPC module. The upper layer utilizes the environment perception module to extract traffic environment information from the driving environment to generate BEV feature information, the driving intention recognition module generates a driving intention probability based on the BEV feature information, the multimodal trajectory prediction module generates multimodal trajectory information based on the BEV feature information and the driving intention probability and transmits it to the lower layer, the lower layer DRL decision module generates an expected vehicle speed, an expected front wheel turning angle, a target weight, and a control quantity weight based on the environment state information and the multimodal trajectory information, and finally the MPC module solves and determines the optimal control action to realize lane-changing control of the autonomous driving vehicle.

[0007] The method comprises the following steps:

[0008] Step 1: Traffic environment information perception:

[0009] The environment perception module extracts BEV feature information from the driving environment through a hierarchical Transformer network and adopts a dynamic sparse attention mechanism to focus on key areas.

[0010] Step 2: Driving intention recognition:

[0011] The driving intention recognition module generates the driving intention probability of the surrounding vehicle based on the BEV feature information. Specifically, the driving intention recognition module receives the BEV feature information generated by the environment perception module and inputs it into the graph neural network in the driving intention recognition module, constructs the interactive relationship between the intelligent agent and road information, uses the Transformer to output the intention time series features and scores the intention time series features, and introduces the mask suppression term at the same time. Finally, the normalization is performed to obtain the driving intention probability P. The scoring definition of the driving intention recognition module is as shown in Equations (1) to (3).

[0012] Z i =W (i) h t +b (i) (1),

[0013]

[0014]

[0015] Among them, Z i is the intention score, h t is the input feature, W and b are learnable parameters, β is the mask penalty strength parameter, ρ is the nonlinear suppression coefficient, is the intention score after the mask suppression term is introduced, m iis the mask identifier, and ε is the probability smoothing factor to prevent division by zero.

[0016] Step 3: Multimodal trajectory prediction:

[0017] The multimodal trajectory prediction module generates multimodal trajectory information based on the driving intention probability and BEV feature information. Specifically, the Transformer in the multimodal trajectory prediction module processes the BEV feature information, generates a sequence of future trajectory points through the LSTM decoder, calculates the semantic score and geometric score of each trajectory, and finally obtains the trajectory confidence Conf K , the multimodal trajectory confidence score is defined as Equation (4) to Equation (7),

[0018]

[0019]

[0020]

[0021]

[0022] in, is the semantic consistency score, w i is a learnable parameter, h k is the state feature of the k-th trajectory, is the geometric matching score, is the feature of the corresponding area in the BEV space, IOU is the intersection-over-union ratio of the trajectory and the lane line, λ is the road deviation penalty factor, H is the indicator function, τ is the threshold, when IOU ≤ τ, H is 1, otherwise H is 0.

[0023] Step 4: Design the DRL decision module:

[0024] Step 4.1, state space design:

[0025] The state space of the lower-level DRL decision module is defined as s t =[X ego ,X sur ,r road ,r topo ], where X ego is the vehicle status information, X sur is the weekly vehicle status information, r road is the road curvature, r topo is the road topology; the action space of the DRL decision module is defined as v f is the expected vehicle speed, δ f is the desired front wheel turning angle, is the target weight, and ψ is the control weight.

[0026] Step 4.2, reward function design:

[0027] The present invention designs a reward function based on safety and comfort, such as equations (8) to (10),

[0028] r t =r safe +r comfort (8),

[0029]

[0030]

[0031] Where r safe is the security reward function, r comfort is the comfort reward function, p1, p2, p3 are constants, ΔT d is the driving torque variation, F b For braking force.

[0032] Step 5. Design the MPC module:

[0033] Step 5.1: Build a prediction model:

[0034] The prediction model is established using the vehicle kinematics model, vehicle dynamics model and nonlinear tire model, as shown in formula (11):

[0035]

[0036] Where x, y are the horizontal and vertical position coordinates of the vehicle, v x and v y are the longitudinal and lateral speeds of the vehicle, θ is the heading angle of the vehicle, and F y,f is the front wheel lateral force, F y,r is the rear wheel lateral force, m is the vehicle mass, γ is the vehicle yaw rate, a, b is the distance from the vehicle center of mass to the front and rear axles, I z The vehicle's yaw moment of inertia.

[0037] Step 5.2: Model discretization:

[0038] The above model is discretized using the fourth-order Runge-Kutta method, with a sampling time of T s , the input-output relationship of the incremental discrete system model can be expressed as formula (12),

[0039]

[0040] Where, u(s)=[δ f ],

[0041] Among them, ξ(s) is the state variable of the vehicle at the current time s, u(s) and Δu(s) are the control input and control input increment of the system at the current time s, respectively, and y c (s) is the predicted output of the system at the current time s, and C is the coefficient matrix used to determine the number of predicted outputs of the system.

[0042] Step 5.3, optimization problem construction:

[0043] The optimization problem is constructed by taking into account the driving comfort and the stability of the control action, imposing hard constraints on the control action, and defining the optimization objectives of the MPC module, such as Equations (13) to (15).

[0044]

[0045]

[0046] L(x(k),u(k))=ω r (v(k)-v f ) 2 +ω r (δ(k)-δ f ) 2 +ω c P(k)(15),

[0047] Where J is the objective function representing the weighted combination of speed and front wheel angle tracking deviation, x = [v h ,δ h ] is the system state variable, v h (N+1) is the vehicle speed at the N+1th moment, δ h (N+1) is the front wheel turning angle at the N+1th moment, ω r is the power performance parameter, ω c is the comfort weight parameter, and P(k) in the objective function represents the comfort performance function, as shown in formula (16),

[0048] P(k)=(T f (k)I g (k)-T f (k-1)I g (k-1)) 2 +F b 2 (k) (16),

[0049] Among them, T f (k) is the driving motor torque at the kth moment, I g (k) is the transmission gear ratio at the kth moment, F b (k) is the braking force at the kth moment.

[0050] Step 5.4, dynamic optimization solution:

[0051] Dynamic optimization is to solve the optimization problem and obtain the optimal control action u0 at the current moment, including the front wheel angle δ o , longitudinal speed v0; finally, the optimal control action u o Input to the autonomous vehicle to implement lane change maneuvers.

[0052] The beneficial effects of the present invention are as follows: the present invention relates to a vehicle lane-changing control method integrating multimodal perception and DRL-MPC, the method comprising two layers, the upper layer constructing the traffic environment information in the driving environment as BEV feature information from a bird's-eye view, and combining the driving intention recognition module and the multimodal trajectory prediction module to generate multimodal trajectory information. On this basis, the DRL decision module interacts with the environment to output the desired vehicle speed, desired front wheel turning angle, target weight, and control quantity weight, and finally the MPC solver determines the optimal control action to realize lane-changing control of the autonomous driving vehicle. This method can improve the generalization and robustness of the vehicle lane-changing control strategy while ensuring safety, and has good application prospects in the field of intelligent vehicle lane-changing control. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a schematic diagram of the hierarchical control of DRL-MPC based on multimodal perception of the present invention. DETAILED DESCRIPTION

[0054] The present invention will be described in detail below with reference to the accompanying drawings.

[0055] The present invention proposes a vehicle lane change control method that integrates multimodal perception and DRL-MPC. Figure 1 The schematic diagram specifically includes the following steps:

[0056] Step 1: Environmental Perception Module

[0057] The environmental perception module is responsible for extracting traffic environment information from the driving environment and generating BEV feature information. BEV feature information includes obstacle type, vehicle position, speed, acceleration, and road environment characteristics. The environmental perception module is composed of a layered Transformer structure. The lower layer extracts geometric features, the middle layer extracts semantic features, and the upper layer uses a dynamic sparse attention mechanism to focus on key areas, ultimately forming an efficient and concise feature representation system. By collecting traffic environment information in real time, the system can provide accurate data support for the control algorithm.

[0058] Step 2: Driving intention recognition module:

[0059] The driving intention recognition module is used to obtain the driving intention probability of the surrounding vehicle based on the BEV feature information. Specifically, the driving intention recognition module receives the BEV feature information generated by the environment perception module and inputs it into the graph neural network in the driving intention recognition module, constructs the interactive relationship between the intelligent agent and road information, uses the Transformer to output the intention time series features and score the intention time series features, and introduces the mask suppression term at the same time. Finally, the normalization is performed to obtain the driving intention probability P. The scoring definition of the driving intention recognition module is as shown in Equations (17) to (19).

[0060] Z i =W (i) h t +b (i) (17),

[0061]

[0062]

[0063] Among them, Z i is the intention score, h t is the intention temporal feature, W and b are learnable parameters, β is the mask penalty strength parameter, ρ is the nonlinear suppression coefficient, is the intention score after the mask suppression term is introduced, m i is the mask identifier, and ε is the probability smoothing factor to prevent division by zero.

[0064] Step 3: Multimodal trajectory prediction module:

[0065] The multimodal trajectory prediction module generates multimodal trajectory information based on the driving intention probability and BEV feature information. Specifically, the Transformer in the multimodal trajectory prediction module processes the BEV feature information to form spatiotemporal dependencies, and generates a sequence of future trajectory points through the LSTM decoder. Next, the semantic score and geometric score of each trajectory are calculated, and finally the trajectory confidence Conf is obtained. K , the multimodal trajectory confidence score is defined as Equation (20) to Equation (23),

[0066]

[0067]

[0068]

[0069]

[0070] in, is the semantic consistency score, w i is a learnable parameter, h kis the state feature of the k-th trajectory, is the geometric matching score, is the feature of the corresponding area in the BEV space, IOU is the intersection-over-union ratio of the trajectory and the lane line, λ is the road deviation penalty factor, H is the indicator function, τ is the threshold, when IOU ≤ τ, H is 1, otherwise H is 0.

[0071] Step 4: Design the DRL decision module:

[0072] Step 4.1, state space design:

[0073] The state space of the DRL decision module is defined as s t =[X ego ,X sur ,r road ,r topo ], where X ego is the vehicle status information, X sur is the weekly vehicle status information, r road is the road curvature, r topo is the road topology; the action space of the DRL decision module is defined as v f is the expected vehicle speed, δ f is the desired front wheel angle, is the target weight, and ψ is the control weight.

[0074] Step 4.2, action space design:

[0075] The action space of the DRL decision module is defined as This action outputs the desired front wheel steering angle, desired speed, target weight, and control weight of the vehicle through a deep reinforcement learning algorithm to ensure that the vehicle can achieve good lane changing results in different driving scenarios. The output of the front wheel steering angle and speed is based on the vehicle state information X ego , weekly vehicle status information X sur , road curvature r road and the road topology r topo Dynamic adjustment, where the vehicle status information X ego Including vehicle speed, vehicle position coordinates and front wheel angle, surrounding vehicle status information X sur Including vehicle speed, vehicle position coordinates and driving intention probability. Target weight The control weight ψ is calculated based on the speed error, front wheel angle error, and road curvature r road And the weekly vehicle status information X sur Dynamic adjustment.

[0076] Step 4.3, reward function design:

[0077] The present invention designs a reward function based on safety and comfort, such as equations (24) to (26).

[0078] r t =r safe +r comfort (twenty four),

[0079]

[0080]

[0081] Where r safe is the security reward function, r comfort is the comfort reward function, p1 is 100, p2 is 200, p3 is 100, ΔT d is the driving torque variation, F b For braking force.

[0082] Step 4.4, experience playback:

[0083] In deep reinforcement learning methods, experience replay is used to improve sample utilization and accelerate policy convergence. This mechanism breaks down temporal correlations between data by storing and recollecting historical interaction data, thereby improving training stability and generalization. This solution defines an experience pool, into which training data at each moment is stored. When the number of experiences in the experience pool reaches a preset value, a certain number of experience samples are randomly drawn from the pool to form a batch for training. Specifically, at the kth moment in the tth round of training, N experiences are randomly selected from the experience pool to form a batch of experiences for training, where N is set to 64.

[0084] Step 4.5. Select a deep reinforcement learning algorithm:

[0085] Different reinforcement learning methods have their own advantages and disadvantages. The present invention selects the SAC algorithm as the deep reinforcement learning algorithm in step 4. The SAC algorithm can select the optimal action in the continuous action space and, by introducing the maximum entropy mechanism, makes the strategy learned by the agent more random, thereby helping to improve the generalization ability of the control strategy.

[0086] Step 4.6, DRL decision module training:

[0087] Set the regularization parameter α = 0.01, the soft update parameter τ = 0.003, the discount factor γ = 0.95, and initialize the neural network parameters in the SAC algorithm, including the Actor network parameter θ actor 、Critic1 network parameters θ critic1 、Critic2 network parameters θ critic2 、Target Critic1 network parameters θtargetcritic1 and target Critic2 network parameters θ targetcritic2 In the tth training round, the environmental state information is obtained from the driving environment, namely the state of the vehicle, the state of the surrounding vehicles, the road topology, and the road curvature, and the current state s in step 4.1 is formed. t This state is input to the Actor network, and after global optimization, the action a in step 4.2 is output. t , the action is within the predetermined action space. Execute action a t After that, the information collection module returns the reward value r according to the current state t and the new state s t+1 Then, the state s t 、Action a t , reward value r t and the new state s t+1 Store it in the experience pool in step 4.4. When the number of experiences in the experience pool reaches 512, randomly extract 128 experiences from the pool. Input these experiences into the Critic1 network, Critic2 network, target Critic1 network and target Critic2 network, and calculate Q respectively. critic1 , Q critic2 , target value Q targetcritic1 and Q targetcritic2 , and use formula (27) to calculate the timing difference (TD) target value.

[0088] η t =r t +γmin{Q targetcritic1 ,Q targetcritic2}-αlogπ(a t+1 |s t+1 ) (27),

[0089]

[0090]

[0091] By minimizing the loss functions in Equations (28) and (29), the Critic1 network and the Critic2 network are updated.

[0092] Sampling actions using reparameterization

[0093]

[0094] The Actor network is updated using the loss function in Equation (30).

[0095] The target Critic1 network and the target Critic2 network are updated using equations (31) and (32).

[0096] τQ critic1 +(1-τ)Q targetcritic1 →Q targetcritic1 (31),

[0097] τQ critic2 +(1-τ)Q targetcritic1 →Q targetcritic1 (32),

[0098] The above training process will continue for multiple iterations until the algorithm converges. After the algorithm converges, the optimal network parameters are selected and loaded into the Actor network, thus completing the training process of the reinforcement learning module in the present invention.

[0099] Step 5. Design the MPC module:

[0100] Step 5.1: Build a prediction model:

[0101] The prediction model is established using the vehicle kinematics model, vehicle dynamics model and nonlinear tire model, as shown in Equation (33):

[0102]

[0103] Where, F y,f is the front wheel lateral force, F y,r is the rear wheel lateral force, m is the vehicle mass, γ is the vehicle yaw rate, a, b is the distance from the vehicle center of mass to the front and rear axles, I z The vehicle's yaw moment of inertia.

[0104] Tire lateral force calculation, as shown in formula (34),

[0105] F y =μD y sin(C y arctan(B y α-E y (B y α-arctan(B y α))))+S vy (34),

[0106] Where, F y is the tire lateral force, μ is the road adhesion coefficient, α is the tire slip angle, B y is the tire characteristic curve stiffness factor, C y is the tire characteristic curve shape factor, D y is the peak factor of the tire characteristic curve, E y is the tire characteristic curve curvature factor, S vy is the vertical offset of the tire characteristic curve, except C yIn addition, other tire characteristic curve parameters are related to the vertical load of the tire.

[0107] The equivalent slip angles of the front and rear axle tires of the vehicle are as shown in equations (35) and (36):

[0108]

[0109]

[0110] Among them, α f , α r Represent the equivalent sideslip angles of the front and rear axle tires respectively.

[0111] The vertical loads on the front and rear axle tires of the vehicle are as shown in Equations (37) and (38).

[0112]

[0113]

[0114] The above model is discretized using the fourth-order Runge-Kutta method, with a sampling time of T s , the input-output relationship of the incremental discrete system model can be expressed as Equation (39),

[0115]

[0116] Where, u(s)=[δ f ],

[0117] Among them, ξ(s) is the state variable of the vehicle at the current time s, u(s) and Δu(s) are the control input and control input increment of the system at the current time s, respectively, and y c (s) is the predicted output of the system at the current time s, and C is the coefficient matrix used to determine the number of predicted outputs of the system.

[0118] The optimization problem is constructed by taking into account the driving comfort and the stability of the control action, imposing hard constraints on the control action, and defining the optimization objectives of the MPC module, such as Equations (40) to (42).

[0119]

[0120]

[0121] L(x(k),u(k))=ω r (v(k)-v f ) 2 +ω r (δ(k)-δ f )2 +ω c P(k)(42),

[0122] Where J is the objective function representing the weighted combination of speed and front wheel angle tracking deviation, x = [v h ,δ h ] is the system state variable, v h (N+1) is the vehicle speed at the N+1th moment, δ h (N+1) is the front wheel turning angle at the N+1th moment, ω r is the power performance parameter, ω c is the comfort weight parameter, and P(k) in the objective function represents the comfort performance function, as shown in formula (43),

[0123] P(k)=(T f (k)I g (k)-T f (k-1)I g (k-1)) 2 +F b 2 (k) (43),

[0124] Among them, T f (k) is the driving motor torque at the kth moment, I g (k) is the transmission gear ratio at the kth moment, F b (k) is the braking force at the kth moment.

[0125] Dynamic optimization is to solve the optimization problem and obtain the optimal control action u0 at the current moment, including the front wheel angle δ o , longitudinal speed v0; finally, the optimal control action u o Input to the autonomous vehicle to implement lane changing maneuvers.

[0126] In summary, the present invention relates to a vehicle lane-changing control method that integrates multimodal perception and DRL-MPC. This method constructs traffic environment information in the driving environment as BEV feature information from a bird's-eye view, and combines a driving intention recognition module and a multimodal trajectory prediction module to generate multimodal trajectory information. On this basis, the DRL decision module interacts with the environment to output the desired vehicle speed, desired front wheel steering angle, target weight, and control quantity weight. At the same time, the optimal control action is determined based on the MPC solution to achieve lane-changing control of the autonomous vehicle. This method can improve the generalization and robustness of the vehicle lane-changing control strategy while ensuring safety, and has good application prospects in the field of intelligent vehicle lane-changing control.

Claims

1. A vehicle lane-changing control method integrating multimodal perception with DRL-MPC, characterized by: The method is implemented through two layers of modules. The upper layer includes an environment perception module, a driving intention recognition module, and a multimodal trajectory prediction module, and the lower layer includes a DRL decision module and an MPC module. The upper layer environment perception module extracts traffic environment information from the driving environment by integrating a hierarchical Transformer network with dynamic sparse attention to generate BEV feature information. The driving intention recognition module processes BEV feature information through a graph neural network and Transformer and introduces a mask mechanism to generate driving intention probability. The multimodal trajectory prediction module processes BEV feature information and driving intention probability based on Transformer and LSTM and introduces a scoring mechanism to generate multimodal trajectory information. The lower-level DRL decision module generates the desired vehicle speed, desired front wheel steering angle, target weight, and control weight based on environmental state information and multimodal trajectory information. Finally, the MPC module solves and determines the optimal control action to achieve lane change control for the autonomous vehicle. The lower-level DRL decision module generates the expected vehicle speed, expected front wheel angle, target weight and control amount weight according to the environmental state information and multimodal trajectory information; the state space of the DRL decision module is defined as s t =[X ego ,X sur ,r road ,r topo ], where X ego is the vehicle status information, X sur is the weekly vehicle status information, r road is the road curvature, r topo is the road topology; the action space of the DRL decision module is defined as v f is the expected vehicle speed, δ f is the desired front wheel angle, is the target weight, ψ is the control weight; target weight The control weight ψ is calculated based on the speed error, front wheel angle error, and road curvature r road And the weekly vehicle status information X sur Dynamic adjustment; design reward functions based on safety and comfort, such as equations (1) to (3), r t ＝r safe +r comfort (1)， Where r safe is the security reward function, r comfort is the comfort reward function, p1, p2, p3 are constants, ΔT d is the driving torque variation, F b For braking force; The lower-layer MPC module uses the desired vehicle speed, desired front wheel steering angle, target weight, and control quantity weight output by the DRL decision module to output the optimal control action; the MPC module includes three parts: prediction model, optimization problem construction, and dynamic optimization. The prediction model is established using the vehicle kinematic model, vehicle dynamics model, and nonlinear tire model, as shown in formula (4): Where x, y are the horizontal and vertical position coordinates of the vehicle, v x and v y are the longitudinal and lateral speeds of the vehicle, θ is the heading angle of the vehicle, and F y,f is the front wheel lateral force, F y,r is the rear wheel lateral force, m is the vehicle mass, γ is the vehicle yaw rate, a, b are the distances from the vehicle center of mass to the front and rear axles, I z Vehicle yaw moment of inertia; The above model is discretized using the fourth-order Runge-Kutta method, with a sampling time of T s , the input-output relationship of the incremental discrete system model is expressed as formula (5), Where, u(s)=[δ f ], Among them, ξ(s) is the state variable of the vehicle at the current moment s, f N is the state update equation, x0, y0 are the horizontal and vertical position coordinates of the vehicle at the current time s, θ0 is the heading angle of the vehicle at the current time s, u(s) and Δu(s) are the control input and control input increment of the system at the current time s, y c (s) is the predicted output of the system at the current time s, and C is the coefficient matrix used to determine the number of predicted outputs of the system; The optimization problem is constructed by taking into account the driving comfort and the stability of the control action, imposing hard constraints on the control action, and defining the optimization objectives of the MPC module, such as Equations (6) to (8). L(x(k),u(k))=ω r (v(k)-v f ) 2 +oh r (δ(k)-δ f ) 2 +oh c P(k)(8), Where J is the objective function representing the weighted combination of speed and front wheel angle tracking deviation, N p For the prediction time domain, x=[v h ,δ h ] is the system state variable, v h (N+1) is the vehicle speed at the N+1th moment, δ h (N+1) is the front wheel turning angle at the N+1th moment, ω r is the power performance parameter, ω c is the comfort weight parameter, and P(k) in the objective function represents the comfort performance function, as shown in formula (9), P(k)=(T f (k)I g (k)-T f (k-1)I g (k-1)) 2 +F b 2 (k)(9), Among them, T f (k) is the driving motor torque at the kth moment, I g (k) is the transmission gear ratio at the kth moment, F b (k) is the braking force at the kth moment; Dynamic optimization is to solve the optimization problem and obtain the optimal control action u0 at the current moment, including the front wheel angle δ o , longitudinal speed v0; finally, the optimal control action u o Input to the autonomous vehicle to implement lane change maneuvers.

2. The vehicle lane change control method integrating multimodal perception and DRL-MPC according to claim 1 is characterized by: The upper-layer environment perception module extracts traffic environment information from the driving environment through a hierarchical Transformer network to generate BEV feature information, and uses a dynamic sparse attention mechanism to focus on key areas; The upper-layer driving intention recognition module receives the BEV feature information generated by the environmental perception module and inputs it into the graph neural network in the driving intention recognition module, constructs the interactive relationship between the intelligent agent and road information, uses Transformer to output the intention time series features and scores the intention time series features, and introduces the mask suppression term. Finally, the normalization is performed to obtain the driving intention probability P. The scoring definition of the driving intention recognition module is as shown in Equations (10) to (12). Z i =W (i) h t +b (i) (10), Among them, Z i is the intention score, h t is the input feature, W and b are learnable parameters, β is the mask penalty strength parameter, ρ is the nonlinear suppression coefficient, is the intention score after the mask suppression term is introduced, m i is the mask identifier, ε is the probability smoothing factor to prevent division by zero; The upper multimodal trajectory prediction module processes BEV feature information and driving intention probability according to Transformer, generates future trajectory point sequence using LSTM decoder, calculates semantic score and geometric score of each trajectory, and finally obtains trajectory confidence Conf K , the multimodal trajectory confidence score is defined as Equation (13) to Equation (16), in, is the semantic consistency score, P i is the probability of the i-th driving intention, w i is a learnable parameter, h k is the state feature of the k-th trajectory, is the geometric matching score, is the feature of the corresponding area in the BEV space, IOU is the intersection-over-union ratio of the trajectory and the lane line, λ is the road deviation penalty factor, H is the indicator function, τ is the threshold, when IOU ≤ τ, H is 1, otherwise H is 0.

Images