Automatic driving lane changing decision and control method based on risk trust perception

Abstract

The invention relates to the technical field of automatic driving, and discloses an automatic driving lane changing decision and control method based on risk trust awareness, comprising the following steps: S1, an upper-layer lane changing command: receiving an upper-layer lane changing instruction from a vehicle path planning module, and triggering a lane changing opportunity judgment process of a decision-making layer; step S2, a decision-making layer judges the lane changing opportunity based on a dominant function: the decision-making layer determines when to change a lane, introduces a conditional value-at-risk to measure a tail collision risk, constructs the dominant function in combination with an adjacent vehicle intention trust degree, and realizes a risk aversive decision for determining when to execute a lane changing operation; and S3, the adjustment layer and the execution layer adaptively control the risk and constraint security. According to the method, the CVaR is introduced to measure the tail collision risk, and an extreme risk scene is accurately captured; cBF-QP safety hard constraint is adopted, it is ensured that the dynamic safety boundary is not broken through in the lane changing process, and collision is avoided; and an adjacent vehicle intention trust punishment mechanism actively avoids non-cooperative vehicles and reduces the interaction risk.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic driving, in particular to an automatic driving lane changing decision and control method based on risk trust perception. BACKGROUND

[0002] With the deployment of automatic driving systems approaching, the safety and robustness of these systems running in the real world are attracting more and more attention. To achieve SAE level 4 or 5 automatic driving, the vehicle needs to learn when and how to make the right decision, and how to safely perform the operation, especially when the vehicle is in an interactive environment, such as a lane changing scene, the actions of surrounding vehicles can be very difficult to predict. A study shows that nearly 10% of car accidents on the highway are caused by lane changing operations, therefore, safe, smooth and efficient lane changing operation is a basic task for autonomous vehicles.

[0003] In the prior art, the schemes for lane changing decision and control have the following shortcomings: (1) Traditional pre-defined rule model is difficult to deal with non-rational and unpredictable behaviors in traffic environment, such as sudden acceleration and deceleration of neighboring vehicles, non-cooperative lane changing, resulting in low decision efficiency or high safety risk; (2) Existing deep reinforcement learning driven decision scheme has the problems of low data utilization rate and poor training stability, and does not fully integrate risk perception and dynamic environment adaptation ability, making it difficult to balance safety and traffic efficiency; (3) Traditional pure pursuit control only focuses on geometric trajectory tracking, lacks safety hard constraints, and cannot dynamically adjust safety boundaries to cope with the uncertainty of neighboring vehicle speed and distance, which is easy to break through the safety threshold.

[0004] In view of the above problems, we provide an automatic driving lane changing decision and control method based on risk trust perception. SUMMARY

[0005] Technical problems solved In view of the shortcomings of the prior art, the present application provides an automatic driving lane changing decision and control method based on risk trust perception, mainly to solve the problem of complex decision-making task faced by autonomous vehicles in complex traffic environment.

[0006] (II) Technical scheme To achieve the above purpose, the present application provides the following technical scheme: The automatic driving lane changing decision and control method based on risk trust perception comprises the following steps: Step S1, receiving the upper lane changing instruction from the vehicle path planning module, triggering the lane changing time judgment process of the decision layer; Step S2, decision layer judges the lane changing opportunity based on the advantage function: the decision layer decides when to change lanes, and introduces the conditional risk value Measure the risk of rear-end collision, combined with the trust degree of adjacent vehicle intention Construct the advantage function to realize risk-averse decision-making for determining when to perform lane changing operation; Step S3, the adjustment layer and the execution layer adaptively control the risk and constrain safety: the adjustment layer is responsible for how to change lanes, which is divided into longitudinal continuous control and reward function optimization; the execution layer introduces the control barrier function CBF- quadratic programming QP safety constraint, replacing the single geometric tracking of traditional pure pursuit control.

[0007] Preferably, the advantage function formula is: Wherein: is the normalized score of the front / rear safety gap of the target lane; is the lane changing waiting cost; is the risk measure of rear-end collision; Wherein, Wherein, Take 0.95-0.99, is the collision cost Quantile function, Given by the adjacent vehicle speed and distance uncertainty model; is the trust penalty of adjacent vehicle intention; Wherein, Wherein, is the set of adjacent vehicles, is the confidence of cooperative intention to the th adjacent vehicle; is the intention classification; is the weight of each index; Action selection logic: If and meet the safety barrier constraint of the execution layer, trigger lane changing; otherwise, keep the current lane and enter the following mode.

[0008] Preferably, the adjustment layer in step S3 adopts a risk-weighted quadratic Q approximator to realize risk-aware dynamic weight structure and adaptive control, wherein the risk-weighted quadratic Q approximator is: Wherein: is the state value; is the value function of the state is the average of the optimal action in the state is a diagonal matrix of longitudinal / lateral control gains; wherein, corresponds to longitudinal acceleration, corresponds to lateral yaw rate; is a risk weighting matrix, is mapped by is mapped by The state vector is extended, and the risk indicator is concatenated to the original state to realize joint learning of risk and action.

[0009] Preferably, the adjustment layer in step S3 adopts a multi-objective reward function, including a following reward function and a lane-changing reward function; The following reward function is used for the multi-objective reward mechanism of the automatic driving adjustment layer, and quantifies the comprehensive performance of the vehicle in the following scenario, and the formula is: wherein: is the reward value of the following scenario; is a driving efficiency term, representing the ratio of the vehicle speed to the speed limit, reflecting the traffic efficiency; is the road speed limit; is a longitudinal comfort term, which punishes the absolute value of the longitudinal jerk to reduce the feeling of jolt; is the longitudinal ; is an environmental emission term, which is an emission proxy function based on the vehicle speed and acceleration, to reduce fuel / energy waste; is an emission proxy term; is a neighbor vehicle intention trust term, which punishes the trust degree of the "non-cooperative" intention of the neighbor vehicle, to avoid risky vehicles; ​​​​​​For the tail collision risk term, the conditional value at risk CVaR is used Quantify extreme collision risk, improve safety Trust and risk weight in following scenario The lane change reward function explicitly includes tail risk, trust penalty and lateral smoothness, including lane change basic reward formula and lane change total reward formula Lane change basic reward formula, used to quantify lane change process, including Lane change total reward formula, on the basis of reward, supplement longitudinal efficiency gain, that is, the speed matching degree of target lane traffic after lane change, including Wherein: The weight of target lane gap score The normalized score of target lane gap score The yaw angle The lateral acceleration Weight The lateral The time of merging into target lane The weight of merging into target lane time The tail collision risk weight in lane change scenario The tail collision risk quantified by conditional value at risk CVaR The neighbor car intention trust weight in lane change scenario The trust penalty of neighbor car intention The longitudinal efficiency gain weight The longitudinal speed difference of target lane traffic after lane change , The risk and trust weight in lane change scenario

[0010] Preferably, the execution layer in step S3 adopts safety filtering combined with control barrier function CBF and quadratic programming QP, and the pursuit control is used for reference trajectory generation, the CBF safety hard constraint is embedded into the control solution, and it is ensured that the safety boundary is not broken ​The control barrier CBF safety function formula is used to constrain the safe distance between the vehicle and adjacent vehicles, ensuring that the safety boundary is not exceeded. This includes: in: For the first The safety barrier function value corresponding to each adjacent vehicle; For self-vehicle and the first The actual distance between two adjacent vehicles; To and The safe distance threshold between adjacent vehicles; Based on dynamic adjustment of relative velocity , For minimum safe distance, For reaction time, The longitudinal / lateral relative velocity.

[0011] The quadratic programming (QP) formula for control solving transforms the expected action of pure pursuit control. Using these as initial values, the final action satisfying the CBF constraint is solved using QP. Greedy actions The optimization problem is: in, objective function To minimize actual actions Pure chasing reference action Deviation; constraint From a dynamic perspective, the "time derivative term + decay term" of the safety barrier function is always positive, ensuring that the safety distance converges to a reasonable range over time. for The derivative with respect to time, The attenuation coefficient ensures that the safe distance converges to a reasonable range over time.

[0012] (III) Beneficial Effects Compared with existing technologies, this invention provides an autonomous driving lane change decision-making and control method based on risk trust perception, which has the following beneficial effects: This invention introduces CVaR to measure rear-end collision risk and accurately capture extreme risk scenarios; it adopts CBF-QP safety hard constraints to ensure that the dynamic safety boundary is not exceeded during lane changes to avoid collisions; and it has a neighbor vehicle intent trust penalty mechanism to actively avoid non-cooperative vehicles and reduce interaction risks.

[0013] The multi-target reward function of the application considers traffic efficiency (speed matching), ride comfort (jerk penalty), environmental protection (emission proxy), and balances the comprehensive performance of lane changing; the risk-weighted quadratic Q approximator dynamically adjusts the control weight to adapt to different risk scenarios (such as congestion and high speed).

[0014] The layered architecture (decision-adjustment-execution) of the application clearly defines the functional boundaries, the decision layer focuses on "when to change lanes", and the control layer focuses on "how to safely change lanes", reducing system complexity; the application combines deep reinforcement learning (dynamic decision) and classical control (CBF-QP), combining learning adaptability and control stability.

[0015] The state vector of the application supports the expansion of more risk indicators (such as weather and road friction coefficient), which can adapt to complex scenarios; the simulation environment training and testing mechanism facilitates rapid iteration and optimization of algorithm parameters, reducing the cost of real vehicle testing. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The lane change scene example of the automatic lane change behavior based on deep reinforcement learning proposed by the application is shown in the figure; Figure 2 The proposed framework structure of the application is shown in the figure; Figure 3 The structure of the function approximator proposed by the application is shown in the figure; Figure 4 The path tracking based on the pure pursuit model proposed by the application is shown in the figure; Figure 5 The simulation environment simulation environment structure of the application is shown in the figure; Figure 6 The training loss and reward effect evaluation of the application is shown in the figure; Figure 7 The evaluation of the adjustment step proposed by the application is shown in the figure; Figure 8 The lane change yaw acceleration evaluation of the application is shown in the figure; Figure 9 The lane change trajectory of the lane change performance evaluation of the application is shown in the figure. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0018] Embodiment 1: ReferenceFigures 1-9 , the automatic driving lane changing decision and control method based on risk trust perception, comprising the following steps: Step S1, upper layer lane changing command: receiving the upper layer lane changing instruction from the vehicle path planning module, triggering the lane changing opportunity judgment process of the decision layer.

[0019] Step S2, decision layer lane changing opportunity judgment based on advantage function: the decision layer decides when to change lanes, and introduces conditional risk value measure the tail collision risk, combined with the neighbor car intention trust degree Construct the advantage function to realize the risk-averse decision, which is used to determine when to perform the lane changing operation.

[0020] The advantage function formula is: Among them: is the normalized score of the front / rear safety gap of the target lane; is the lane changing waiting cost; is the tail collision risk measure; Among them, Among them, Take 0.95-0.99, is the collision cost Quantile function, Given by the neighbor car speed, distance uncertainty model; is the neighbor car intention trust penalty; Among them, Among them, is the neighbor car set, is the cooperative intention confidence of the first neighbor car; is the intention classification; is the weight of each index; Action selection logic: If and meet the execution layer safety barrier constraint, trigger lane changing; otherwise, keep the current lane and enter the following mode.

[0021] Step S3, adjust the adaptive control risk and constraint safety of the adjustment layer and the execution layer: the adjustment layer is responsible for how to change lanes, which is divided into longitudinal continuous control and reward function optimization; The execution layer introduces the control barrier function CBF-Quadratic Programming QP safety constraint, which replaces the single geometric tracking of traditional pure pursuit control.

[0022] The adjustment layer in step S3 adopts a risk-weighted quadratic Q approximator to realize a risk-aware dynamic weight structure and adaptive control, wherein the risk-weighted quadratic Q approximator is: wherein: is a state value of the action ; is a value function of a state ; is an optimal action mean value under a state ; is a diagonal matrix of longitudinal / lateral control gains; wherein, corresponds to a longitudinal acceleration, corresponds to a lateral yaw angular velocity; is a risk-weighted matrix, is mapped by , is mapped by ; The state vector is extended, and the risk indicator is spliced into the original state to realize joint learning of risk and action.

[0023] The adjustment layer in step S3 adopts a multi-objective reward function, including a following reward function and a lane-changing reward function. The following reward function is used for the multi-objective reward mechanism of the automatic driving adjustment layer, and quantifies the comprehensive performance of the vehicle in the following scenario, and the formula is: wherein: is a reward value of the following scenario; is a driving efficiency term, representing the ratio of the vehicle speed to the speed limit, reflecting the traffic efficiency; is a road speed limit; is a longitudinal comfort term, which punishes the absolute value of the longitudinal jerk , to reduce the sense of riding jerk; is a longitudinal ; ​For environmental emission item, based on the emission proxy function of ego vehicle speed and acceleration, reduce fuel / energy waste; For emission proxy item; For neighbor vehicle intention trust item, punish the trust degree of neighbor vehicle "non-cooperative" intention, and avoid risky vehicles; For tail collision risk item, use conditional risk value Quantify extreme collision risk and improve safety; Trust and risk weight in following scenario; The lane change reward function explicitly includes tail risk, trust punishment and lateral smoothness, including lane change basic reward formula and lane change total reward formula; Lane change basic reward formula, used to quantify the lane change process, including: Lane change total reward formula, on the basis of the basic reward, supplement the longitudinal efficiency gain, i.e. the speed matching degree with the target lane traffic after lane change, including: Wherein: The weight of target lane gap score; The normalized score of target lane gap score, The yaw angle; The lateral acceleration Weight; The lateral ; The time of merging into the target lane; The merging into the target lane time weight; The tail collision risk weight in lane change scenario; The tail collision risk quantified by conditional risk value CVaR; The neighbor vehicle intention trust degree weight in lane change scenario; The trust punishment of neighbor vehicle intention; The longitudinal efficiency gain weight; The longitudinal speed difference with the target lane traffic after lane change; , Risk and trust weight for lane change scenario.

[0024] The execution layer in step S3 adopts a safety filter combining a control barrier function CBF and a quadratic programming QP, and a pursuit control is used for reference trajectory generation, a CBF safety hard constraint is introduced to embed control solving, and it is ensured that the safety boundary is not broken; The control barrier CBF safety function formula is used for safety distance constraint of the ego vehicle and the neighbor vehicle, and it is ensured that the safety boundary is not broken, and it includes: Wherein: The safety barrier function value corresponding to the first neighbor vehicle; The actual distance between the ego vehicle and the first neighbor vehicle; The safety distance threshold value of the ego vehicle and the first neighbor vehicle; Based on the relative speed, the safety distance threshold value is dynamically adjusted, , The minimum safety distance, The reaction time, The longitudinal / lateral relative speed.

[0025] The quadratic programming QP formula of the control solving, the expected action of the pure pursuit control is taken as the initial value, the final action satisfying the CBF constraint is solved by QP, the greedy action , and the optimization problem is: Wherein, The objective function is to minimize the deviation of the actual action from the pure pursuit reference action ; The constraint guarantees that the time derivative term + decay term of the safety barrier function is always positive from the dynamic perspective, and ensures that the safety distance converges to a reasonable range over time; The derivative of with respect to time, The decay coefficient ensures that the safety distance converges to a reasonable range over time.

[0026] Embodiment 2: The automatic driving lane change decision and control method based on risk and trust perception is used in Figure 1In the typical lane-changing scenario shown, ego vehicle must be able to adjust its action to adapt to the dynamic traffic environment, thus, the feasible lane-changing policy π to be learned should be able to avoid collision with surrounding vehicles and make smooth movements.

[0027] The decision-making includes strategic and tactical strategies, the strategic decision means to initiate lane-changing according to long-term goals, such as driving efficiency (e.g. traffic flow condition of multi-lane), remaining time in the current lane (e.g. need to exit from nearby ramp) and other factors related to route planning in a longer time range, the tactical decision means that ego vehicle has already had the intention to change lanes, but still needs to continue to execute.

[0028] For the first layer, the state space includes the relative distance (Δx leader ,Δx target ,Δx follow ) between surrounding vehicles, the relative speed Δv leader between ego vehicle and its leading vehicle, and the relative speed Δv target between ego vehicle and the leading vehicle on the target lane; the speed v ego and acceleration of ego vehicle.

[0029] (1) The action space is defined as follows: evaluate whether the lane-changing action is beneficial.

[0030] (2) Where “0” means “now do not change lanes”, i.e. stay on the autonomous lane; “1” means “now change lanes”.

[0031] The task of the second control layer is to consider how to properly operate, in this task, we divide it into two steps, i.e. adjustment step and execution step, the “adjustment step” is to adjust the longitudinal action to find a suitable opportunity, at the same time, the “execution step” is to execute a smooth and effective action to complete the maneuver. Specifically, if the lane-changing decision is “1”, i.e. change lanes at this time, then the action adjustment module will play a role, ego vehicle will adjust its longitudinal action to adapt to the gap on the target lane, then in the execution step, the planned trajectory will be generated and tracked by the controller based on pure pursuit control, if the lane-changing decision is “0”, i.e. do not change lanes immediately, then the adjustment step and the execution step will be combined together and the car following module will play a role, ego vehicle will stay on the lane and wait for another gap.

[0032] In the action adjustment step, the action space is defined as (3)​​ Reinforcement learning is used to train the decision-making and longitudinal control modules in a hierarchical structure to help drivers learn not only high-level strategies for deciding when to change lanes, but also low-level control strategies for adjusting their state to feasible maneuvering conditions. The overall decision-making framework structure is as follows: Figure 2 As shown.

[0033] A. Decision-making level At the decision-making level, discrete decisions are given by a Q-network. In our study, we designed two fully connected layers. The reward function of the decision-making module is based on three factors: the distance Δx between the car and the car in front. ego The relative speed Δv between the car and the car in front ego And the target gap d on the target lane gap For example, if the distance Δx ego Large enough, relative velocity Δv ego Smaller, gap d gap If the relative distance Δx is small, the ego vehicle will tend to stay in its autonomous lane and continue following the vehicle in front. On the other hand, if the relative distance Δx is small... ego Smaller, gap d gap If it is large enough, the vehicle will tend to steer into the target lane.

[0034] Therefore, we can construct the reward function for the decision problem as follows: (4) Where, d target The desired distance on the target lane can be calculated according to (5), d ego It is the expected distance on the target lane, which can be calculated according to (6), where d t It is a safety threshold. τ is the maximum acceleration, d0 is the human reaction time, t is the minimum distance, and t is the total time for lane changing.

[0035] (5) (6) B. Adjustment Steps In the adjustment process, we need to calculate the continuous actions output by the decision layer. However, traditional Q-learning methods cannot handle the problem of continuous action space.

[0036] Although there are some policy gradient-based algorithms, such as actor-critic algorithms, that can directly learn the policy without the help of the value function, it still requires a lot of effort to design the policy network. In the adjustment layer, we use an improved Q-function approximator to handle continuous action space, which is similar to our previous work. Specifically, we design a quadratic form of Q function, so that the optimal action has a closed-form solution. Our method can also simplify the network design work in the learning algorithm.

[0037] The representation of the Q-function approximator is as follows: (7) where A, B and C are coefficients, the neural network designed with state information as input, as shown in Figure 3 , because any smooth Q function can be Taylor expanded in this form near the greedy action, and if we stick to the greedy action in Q-learning exploration, there will be no much loss of generality.

[0038] As shown in Figure 3 , the design of A has a hidden layer consisting of 150 neurons, and in particular, A uses a softplus activation function on the output layer, which makes its boundary negative. C is also a single-hidden-layer neural network with the same number of neurons as A, and the output can be any scalar. B is a neural network consisting of two hidden layers, each with 200 neurons, and each layer has a ReLU activation function.

[0039] The adjustment module for two discrete decisions (e.g. "immediately change lane" and "immediately do not change lane") is trained using the same quadratic Q network structure, but the reward function and weight are different, which will be discussed in detail later.

[0040] 1) The reward of the car following module In the case of car following, we need to consider the distance and relative speed between the ego vehicle and the leading vehicle. A good following module should maintain an appropriate relative distance and speed, while controlling the ego vehicle's acceleration within a comfortable range. Based on this idea, we designed the following reward function.

[0041] (8) where R dis is the reward related to the relative distance, w dis is the weight of the reward, x leader is the longitudinal position of the leading vehicle, x ego is the longitudinal position of the ego vehicle, and d ego is the desired distance to the leading vehicle.

[0042] (9) R Δv It's a reward related to relative speed, -w Δv It is the weight of the reward, v ego and v leader These are the speeds of the small car and the lead car, respectively.

[0043] (10) R c It is the total reward of the following module, it is R Δv and R dis The sum of .

[0044] 2) Lane change module rewards In lane-changing situations, the "autonomous" vehicle should adjust its longitudinal acceleration or deceleration. To adapt to the target gap while maintaining a safe distance from the leading vehicle in the "autonomous" lane, the reward function is defined as follows: (11) In other words, the ideal position for your vehicle is between the following vehicles and the lead vehicle, and closer to the target vehicle, such as... Figure 1 As shown.

[0045] (12) R ΔV The meaning is the same as in the following modules.

[0046] (13) Total reward R A It is R dis and R Δv The sum of .

[0047] C. Execution Steps The execution steps are separated from the adjustment steps of the lane change decision, but are combined with the adjustment steps of the following decision. Therefore, in this section, we will introduce how an autonomous vehicle executes a lane change action when it receives a lane change signal from the decision-making layer.

[0048] The initial state is generated based on the fifth-order polynomial curve. and endpoint state The reference trajectory is shown in (14) and (15).

[0049] (14) (15) Then, we can define the time parameter matrix: (16) In terms of lane change motion control, we adopt a pure pursuit control method to track the reference trajectory, by generating target waypoints with speed and position, we can control the ego vehicle to track the target trajectory (as shown in Figure 4 ), by calculating the angle a between the vehicle body and the target waypoint, we can control the steering wheel to track the target point.

[0050] According to the geometric relationship, we can get (17) (18) Where, l d is the front distance, a is the actual front steering angle, d(t) is the controlled front steering angle for tracking direction, L is the wheelbase of the vehicle.

[0051] IV. Implementation of lane change scheme A. Simulation scenario construction We established a simulation environment in which the ego vehicle was controlled by the hierarchical RL framework proposed by us, and the surrounding other vehicles always drove on their lanes and were controlled according to the intelligent driver model (IDM).

[0052] B. Implementation of lane change scenario First, we constructed a simulation environment as shown in Figure 5 , the red vehicle is the ego vehicle, the green vehicle is the vehicle related to the selected gap, and the blue vehicle is the ordinary vehicle controlled by IDM.

[0053] In this simulation platform, the length of the road section is 1000 meters, the width of each lane is 3.75 meters, and the initial speed of the vehicle is from 75 km / h to 136 km / h. After a certain time, the leader vehicle and the ego vehicle are generated at the starting point of the lane, and the period t e between the ego vehicle and the leader vehicle is set to U(0.1, 7), which represents a uniform distribution, so the initial distance between the leader vehicle and the ego vehicle is , these settings allow the environment to produce unique situations, in addition, as shown in Figure 3 , the initial values of A and B are set to 0, and the initial value of C is set to -60.

[0054] During the training process, we choose the learning rate to be 0.0005, the replay buffer size to be 5000, the batch size to be 64, and the decay of E from 1 to 0.1 to be 300000 steps. Initially, the agent tends to take random actions to explore the environment and obtain environmental information, and as the training process deepens, the agent tends to use the learned reward mechanism to make reasonable decisions.

[0055] Figure 6 The loss curve along the training steps and the cumulative reward per episode are described, from which we can see that the loss and reward converge well, indicating that the agent has successfully learned to take actions that maximize the reward. In order to better evaluate the performance of the model, we selected a typical initial condition, i.e., the initial speed of the leading vehicle is 35.5 m / s, the initial speed of the ego vehicle is 17 m / s, and the relative distance between the two is 68 m.

[0056] Figure 7 It is shown that after about 100 time steps (10 seconds), the ego vehicle successfully learns to adjust the speed and relative distance to appropriate values, and although there is a speed difference between the ego vehicle and the leading vehicle at the beginning, we can see that the ego vehicle achieves good performance and achieves a small relative speed.

[0057] We first train the longitudinal adjustment module in the case of car following and lane changing, and then train the decision layer with fixed controller layer parameters. From the simulation results, it can be concluded that the proposed structure can solve the lane changing problem in terms of the designed reward function.

[0058] From the experimental results, both the decision module and the adjustment module have good performance, and the ego vehicle can recognize the safety of the gap when executing the lane changing instruction.

[0059] In addition, we also evaluate the lane changing performance, Figure 8 and Figure 9 The curves show that the ego vehicle can successfully and smoothly change lanes to the target lane without sudden lateral acceleration or deceleration, Figure 8 showing the longitudinal speed and yaw acceleration during lane changing, Figure 9 showing the trajectory of the ego vehicle during lane changing, relative to the X and Y coordinates.

[0060] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A risk-trust perception based automatic driving lane changing decision and control method, comprising the following steps: Step S1, upper layer lane changing command: receiving an upper layer lane changing instruction from a vehicle path planning module, triggering a lane changing timing judgment process of the decision layer; Step S2, decision layer based on advantage function lane change opportunity judgment: the decision layer decides when to change lanes, introduces the conditional value at risk Measure the risk of rear-end collision, combined with the intention trust degree of adjacent vehicles Construct the advantage function to realize the risk-averse decision, which is used to determine when to perform lane change operation; Step S3, adjusting the adaptive control risk and constraint safety of the adjustment layer and the execution layer: the adjustment layer is responsible for how to change lanes, which is divided into longitudinal continuous control and reward function optimization; the execution layer introduces a control barrier function CBF- quadratic programming QP safety constraint to replace the single geometric tracking of traditional pure pursuit control. 2.The risk trust perception based automatic driving lane changing decision and control method according to claim 1, characterized in that, The advantage function formula is: Wherein: a normalized score for a forward / backward safety gap for the target lane; lane change waiting cost; is a tail collision risk metric; wherein wherein, taking 0.95-0.99, is the collision cost is the quantile function, is given by the neighboring vehicle speed, distance uncertainty model; To the neighbor's intention to trust punishment; Wherein, wherein, is a neighboring vehicle set, is a confidence in the cooperative intention of the neighboring vehicle. for intent classification; Wi is the weight of each index; Action selection logic: If and the execution layer safety barrier constraint is satisfied, a lane change is triggered; otherwise, the current lane is maintained and the car-following mode is entered. 3.The risk trust perception based autonomous driving lane changing decision and control method of claim 1, wherein, The adjustment layer in the step S3 adopts a risk weighted quadratic Q approximator to realize a risk perception dynamic weight structure and adaptive control, wherein the risk weighted quadratic Q approximator is: Wherein: for the state the action of the value; is the state value function; the optimal action mean for the state under the policy a diagonal matrix of longitudinal / lateral control gains; wherein, corresponding to a longitudinal acceleration, corresponding to a lateral yaw angular velocity; for the risk-weighted matrix, by mapping, by mapping; state vector expansion, risk indicators are incorporated at training time into the original state , enabling joint learning of risk and action.

4. The risk trust perception based autonomous driving lane changing decision and control method of claim 1, wherein, The adjustment layer in the step S3 adopts a multi-objective reward function, which includes a following reward function and a lane changing reward function. The following reward function is used for a multi-objective reward mechanism of the automatic driving adjustment layer, quantifies the comprehensive performance of the vehicle in the following scenario, and the formula is: Wherein: reward value for the car-following scenario; For the travel efficiency item, a ratio of the host vehicle speed to the speed limit is represented, embodying the traffic efficiency; To limit the speed of a road; For longitudinal comfort, penalize the absolute value of the longitudinal jerk to reduce the perception of jolts. For longitudinal comfort, penalize the absolute value of the longitudinal jerk to reduce the perception of jolts. for longitudinal ; For environmental emissions, based on the speed of the vehicle, acceleration emissions proxy function, reduce fuel / energy waste; to discharge the proxy; For the adjacent vehicle intention trust item, the trust degree of the "non-cooperative" intention of the adjacent vehicle is punished, and the risk vehicle is avoided; For the tail collision risk term, use the conditional value at risk Quantify extreme collision risk, improve safety; : trust and risk weight under the following scenarios of car following; The lane changing reward function explicitly includes tail risk, trust penalty and lateral smoothness, including a lane changing basic reward formula and a lane changing total reward formula. The lane changing basic reward formula is used for quantifying the lane changing process, including: The lane changing total reward formula supplements the longitudinal efficiency gain on the basis reward, that is, the speed matching degree with the target lane traffic after lane changing, including: Wherein: a weight for the target lane gap score; a normalized score for the target lane gap score, is the yaw angle; for lateral acceleration weight; for lateral ; time to incorporate the target lane; To incorporate target lane time weight; tail collision risk weight for a lane change scenario; tail collision risks quantified with Conditional Value at Risk, CVaR; For the lane change scene, the intention trust weight of the neighboring vehicle; Trust penalty for adjacent car intentions; L is the longitudinal efficiency gain weight; is the longitudinal speed difference of the target lane traffic after changing lanes; , Risk and trust weights for lane change scenarios.

5. The risk trust perception based autonomous driving lane changing decision and control method of claim 1, wherein, The execution layer in the step S3 adopts a safety filter combined with a control barrier function CBF and quadratic programming QP, and the pursuit control is used for reference trajectory generation, a CBF safety hard constraint is embedded into the control solution to ensure that the safety boundary is not broken; Wherein, the control barrier CBF safety function formula is used for safety distance constraint of the ego vehicle and the neighbor vehicle to ensure that the safety boundary is not broken, including: Wherein: For the first vehicle corresponding to the safety barrier function value; For self-vehicle and the first The actual distance between two adjacent vehicles; to maintain a safe distance from a vehicle in front of the subject vehicle; and a safety distance threshold from a vehicle in front of the subject vehicle; based on relative speed dynamic adjustment, , for minimum safety distance, for reaction time, for longitudinal / lateral relative speed.

6. Control the solved quadratic program (QP) formula to get the desired action of pure pursuit control As an initial value, solve the final action that satisfies the CBF constraint by QP , greedy action , the optimization problem is: Wherein, Objective function To minimize the actual motion deviation from the pure pursuit reference motion ; Constraint The "time derivative term + decay term" of the safety barrier function is always positive from the dynamic perspective, ensuring that the safety distance converges to a reasonable range over time. For The derivative with respect to time, is a decay coefficient that ensures the safety distance converges to a reasonable range over time.

Images