A non-conservative decision-making method for autonomous vehicles with defensive behavior

Abstract

The present invention discloses a non-conservative decision-making and control method for an autonomous driving vehicle with defensive behavior. The method of the present invention includes combining information about the vehicle and the surrounding traffic environment to predict the probability distribution of the future trajectories of traffic participants, and building a spatiotemporal risk situation assessment model for the traffic environment in combination with spatial position relationships. At the same time, environmental uncertainty is integrated into an integrated decision-making and control architecture to construct a model predictive optimization control problem. By optimizing and solving the front wheel steering angle and longitudinal acceleration control instructions, the vehicle is controlled to achieve autonomous driving tasks. The present invention can ensure that the vehicle does not overreact when facing potential dangers, while not reducing driving quality due to excessive conservatism, thereby achieving safe and non-conservative driving of autonomous driving vehicles in uncertain environments.

Description

Technical Field

[0001] The present invention relates to the field of autonomous vehicle decision-making and control technology, and in particular to a non-conservative decision-making and control method for an autonomous vehicle with defensive behavior. Background Art

[0002] Autonomous driving has become a strategic direction for the development of the global automotive industry. Countries are accelerating the innovative development of autonomous vehicles to ensure they remain at the forefront of the new round of technological revolution and industrial transformation. As technology continues to improve, automakers are focusing on developing key technologies for Level 3-5 autonomous driving, aiming to gradually achieve higher levels of autonomous driving capabilities.

[0003] Environmental perception, decision-making and motion control are three key technologies for autonomous vehicles. Both autonomous decision-making and motion control can be considered as optimization problems of a certain type, and they are strongly coupled. Therefore, they can be integrated into a single functional module, the decision-making and control system. This system is the primary embodiment of driving intelligence. However, in urban road scenarios, traffic participants (motor vehicles, non-motor vehicles, and pedestrians) have varying directions, speeds, and patterns of movement, making the situation faced by autonomous vehicles ever-changing. Furthermore, the large number and variety of traffic participants in the environment, their mutual influences, and their diverse behaviors make it difficult to accurately identify and predict the intentions and behaviors of each participant. Furthermore, the decision-making and control system for autonomous vehicles must comprehensively consider numerous factors, including traffic regulations, road constraints, road conditions, and the status of other traffic participants. In summary, the highly dynamic, random, and complex nature of urban road traffic environments places higher demands on autonomous driving decision-making and control systems, posing significant challenges to their design. Currently demonstrated decision-making and control strategies for autonomous vehicles are often overly conservative and cautious, sometimes exhibiting inhuman driving behavior. This conservative behavior can compromise driving quality and pose safety risks when other road users cannot predict it. Summary of the Invention

[0004] The present invention aims to solve one of the technical problems in the related art at least to a certain extent.

[0005] To this end, the present invention proposes a non-conservative decision-making and control method for autonomous driving vehicles with defensive behavior, which integrates environmental uncertainty information into an integrated decision-making and control architecture to ensure that the autonomous driving vehicle does not overreact when facing potential dangers, while at the same time not reducing the driving quality due to conservative operation, thereby achieving safe and non-conservative driving of autonomous driving vehicles in uncertain environments.

[0006] To achieve the above objectives, the present invention provides, on one hand, a non-conservative decision-making method for an autonomous driving vehicle with defensive behavior, comprising:

[0007] Predict the movement trajectories of traffic participants based on information about the vehicle and surrounding traffic environment, modeling environmental uncertainty information as a multimodal probability distribution, and constructing a spatiotemporal risk situation assessment model based on spatial position relationships;

[0008] Selecting the optimal control strategy for each candidate static path, based on the spatiotemporal risk situation assessment model, from the two strategies of expected travel speed and expected stopping speed; and determining the optimal path for autonomous driving by constructing a cost function based on the evaluation of the candidate ego-vehicle paths; and constructing a constrained optimal control problem based on model predictive control in conjunction with the ego-vehicle dynamics model.

[0009] Based on the driving behavior and decision-making process of human drivers in interactive scenarios, environmental uncertainty information is integrated into the integrated decision-making and control architecture to optimize the front wheel angle and longitudinal acceleration control commands through the constrained optimal control problem, thereby controlling the vehicle to achieve autonomous driving tasks.

[0010] The non-conservative decision-making method for an autonomous driving vehicle with defensive behavior according to an embodiment of the present invention may also have the following additional technical features:

[0011] In one embodiment of the present invention, the environmental uncertainty information includes: movement direction, speed, acceleration of surrounding vehicles and / or pedestrians, and intentions of surrounding vehicles and / or pedestrians.

[0012] In one embodiment of the present invention, predicting the motion trajectory of traffic participants based on the vehicle and surrounding traffic environment information to model the environmental uncertainty information as a multimodal probability distribution includes:

[0013] Defining first network input information; wherein the first network input information includes historical trajectory information and lane centerline information of traffic participants;

[0014] Inputting the second network input information into the fully connected layer, and constructing a long short-term memory (LSTM) network input tensor based on the first network input information and the second network input information to predict and output the multimodal distribution of the trajectories of traffic participants in a preset future time domain; wherein the second network input information includes the vehicle state, static trajectory, road constraints, traffic participants, and traffic light information;

[0015] A trajectory prediction model based on the vehicle's historical trajectory, driving intention, and vehicle-to-vehicle motion interaction is established to output the multimodal probability distribution of the vehicle's trajectory in the future preset time domain according to the multimodal distribution and through an LSTM decoder.

[0016] In one embodiment of the present invention, the historical trajectory information of the participants And lane centerline information L = {l j},in, is the trajectory information of the traffic participant numbered i in the historical time domain [tn,t]; l j =[x,y,j next ,I signal ,I inter ] represents the lane centerline information numbered j in the map, where x and y are the longitudinal and lateral coordinate vectors of the lane centerline respectively, and j next is the number of the subsequent lane of the center line of the jth lane, I signal is the traffic light information of the lane, I inter Indicates whether the vehicle is in an intersection.

[0017] In one embodiment of the present invention, the optimal control strategy of the two strategies of expected passing speed and expected stopping speed set for each candidate static path is selected based on the spatiotemporal risk situation assessment model; and the optimal path for autonomous driving is determined by constructing a cost function by evaluating the indicators of the candidate paths of the ego vehicle, including:

[0018] Based on static information including at least road structure, speed limit, road markings, and traffic regulations, and in combination with human driving experience, a set of candidate paths and expected vehicle speeds without time information is designed. Two strategies, one for expected passing speed and one for expected stopping speed, are set for each candidate static path. These two strategies are selected and switched based on the spatiotemporal risk situation assessment model.

[0019] In each control cycle, the optimal path for autonomous driving is determined by combining the multimodal distribution of traffic participants' trajectories in a preset future time domain and evaluating a cost function constructed by three indicators: potential collision risk, traffic efficiency, and driving compliance.

[0020] In one embodiment of the present invention, combining the vehicle dynamics model and constructing a constrained optimal control problem based on model predictive control includes:

[0021] Combined with the vehicle dynamics model, with traffic participants as the main constraints, and based on road edge constraints and traffic light restrictions, a constrained optimal control problem is constructed based on model predictive control (MPC). This includes:

[0022] The mathematical expression of the optimal control problem is described by the finite-horizon predictive control that tracks the i-th path:

[0023]

[0024] satisfy:

[0025]

[0026] And subject to the constraints of traffic participants:

[0027] x self (τ)-x surr,i (τ)≥d self ,τ∈[t,t+T]

[0028] Road edge constraints:

[0029] x self (τ)-x road (τ)≥δ road ,τ∈[t,t+T]

[0030] Traffic light constraints:

[0031] x self (τ)∈free,ifgreen,τ∈[t,t+T]

[0032] x self (τ)≤L stop ,if red,τ∈[t,t+T]

[0033] Among them, t is the current time, τ is the virtual time of the prediction domain, T is the length of the prediction domain, is the desired tracking position, is the desired driving speed, x self is the state of the vehicle, u self is the vehicle control quantity, x surr (τ) is the state of the surrounding traffic participants, including the constraint relationship between C traffic participants and the vehicle, x road Refers to the drivable area enclosed by the edge of the road, d self is the safe distance of the surrounding vehicle, δ road is the safe distance from the edge of the road, L stop is the distance from the stop line at the intersection, R represents a red light and G represents a green light.

[0034] In one embodiment of the present invention, based on the driving behavior and decision-making process of a human driver in an interactive scenario, environmental uncertainty information is integrated into an integrated decision-making and control architecture to optimize the front wheel angle and longitudinal acceleration control commands through the constrained optimal control problem, thereby controlling the vehicle to achieve an autonomous driving task, including:

[0035] By characterizing the contribution of different future probability distributions to the cost function, the uncertainty problem is structured as a model prediction optimization problem.

[0036] Through scene data, fully connected layers and LSTM networks, the associated vehicle r is selected based on the multimodal distribution of the trajectories of traffic participants in the future preset time domain and the static path set of the autonomous driving vehicle. i and potential collision area ci , for each collision area c i , set a safety line k and stop line e k ;

[0037] The time from the vehicle's current position to the potential collision zone is defined as TTC, and a constraint-modified TTC criterion is introduced. When assessing the risk situation of the traffic environment, a Bayesian network is used to evaluate the potential risks in the traffic scene.

[0038] A shared safety set and a preview safety set are designed within the MPC prediction domain. In the shared safety set, the uncertainty of the traffic environment is not considered; in the preview safety set, the probability distribution of potential threats in the traffic environment is combined, and the probability constraints are integrated into the MPC optimization problem.

[0039] Based on real-time environmental information, the probability model is updated and the MPC optimization problem is re-solved to dynamically handle environmental uncertainties. The MPC prediction domain is divided into the shared safety set and the preview safety set, and a constrained multi-objective optimization control problem is constructed to integrate environmental uncertainty factors into the integrated decision-making and control architecture.

[0040] The shared action sequence and bifurcated action sequence of the front wheel steering angle and longitudinal acceleration are optimized and determined, and the optimized control instructions are input to the vehicle to realize the vehicle's autonomous driving task.

[0041] In one embodiment of the present invention, the time from the current position of the vehicle to the potential collision area is defined as TTC, and a constraint-modified TTC criterion is introduced, and the formula is as follows:

[0042]

[0043] in, The revised TTC, is the self-car and the surrounding car r i distance, is the k-th step surrounding vehicle r in the MPC prediction domain at time t i The driving speed is ξ, and the set safety distance is ξ.

[0044] In one embodiment of the present invention, when assessing the risk situation of a traffic environment, a Bayesian network is used to assess potential risks in a traffic scenario, including:

[0045] For the associated vehicle r i, define a random variable z to represent the three risk levels associated with TTC: danger, warning, and safety; comprehensively consider driving safety and vehicle tracking performance, define threshold values ​​associated with potential risk levels, and construct a likelihood function for TTC to parameterize scenario risks; assuming that the prior probability density function P(z) of the potential threat level of the associated vehicle follows a uniform distribution, the probability distribution of the potential threat is determined using the Bayesian theorem, as follows:

[0046]

[0047] in, The kth step of the MPC prediction domain at time t and the surrounding vehicles r i The risk level, represent The risk level under the condition is The probability of The collision time under the risk level z condition is The probability, N Z Represents the possible risk level, namely danger, warning and safety, The collision time for each risk condition probability.

[0048] In one embodiment of the present invention, in the preview safety set, the probability distribution of potential threats in the traffic environment is combined to construct a probability constraint ε is the set probability constraint threshold, and the probability constraint is finally integrated into the MPC optimization problem.

[0049] The non-conservative decision-making and control method for an autonomous vehicle with defensive behavior in an embodiment of the present invention combines information about the vehicle and the surrounding traffic environment, uses a probability distribution to represent environmental uncertainty information; predicts the multimodal distribution of future motion trajectories of traffic participants, and constructs a spatiotemporal risk situation assessment model for the traffic environment based on spatial position relationships. At the same time, environmental uncertainty is integrated into an integrated decision-making and control architecture, a model predicts and optimizes the control problem, and the vehicle is controlled to achieve autonomous driving tasks by optimizing and solving the front wheel angle and longitudinal acceleration control commands. This ensures that the vehicle does not overreact when faced with potential dangers, while also preventing driving quality from being compromised due to over-conservativeness, thereby achieving safe, non-conservative driving of autonomous vehicles in uncertain environments.

[0050] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0052] Figure 1 is a flow chart of a non-conservative decision-making method for an autonomous driving vehicle with defensive behavior according to an embodiment of the present invention;

[0053] Figure 2 2 is a schematic diagram of a non-conservative decision-making method with defensive behavior according to an embodiment of the present invention. DETAILED DESCRIPTION

[0054] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0055] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0056] The following describes a non-conservative decision-making method for an autonomous driving vehicle with defensive behavior proposed in accordance with an embodiment of the present invention with reference to the accompanying drawings.

[0057] Figure 1 is a flow chart of a non-conservative decision-making method for an autonomous driving vehicle with defensive behavior according to an embodiment of the present invention. Figure 1 As shown, including:

[0058] S1 predicts the movement trajectory of traffic participants based on the information of the vehicle and the surrounding traffic environment, models the environmental uncertainty information as a multimodal probability distribution, and constructs a spatiotemporal risk situation assessment model based on the spatial position relationship.

[0059] It is understood that this step S1 is designed to consider the spatiotemporal risk situation assessment method of environmental uncertainty, modeling the uncertainty in the environment as a probability distribution to better understand possible situations and events. The environmental uncertainty information may include: the movement direction, speed, acceleration of surrounding vehicles and / or pedestrians, and the intentions of surrounding vehicles and / or pedestrians. This step S1 may specifically include:

[0060] S11 comprehensively considers the historical trajectory information and possible future path information of traffic participants, combines the map topology, and predicts the multimodal distribution of future movement trajectories of traffic participants, modeling the uncertainty in the environment as a probability distribution;

[0061] S12, by quantifying uncertainty and combining spatial location relationships to build a spatiotemporal risk situation prediction model, evaluate the risks associated with each decision, select the optimal control strategy, and improve the decision-making and control performance and reliability in complex environments.

[0062] Specifically, the first network input information is defined, including the historical trajectory information of traffic participants And lane centerline information L = {l j},in is the trajectory information of the traffic participant numbered i in the historical time domain [tn,t]; l j =[x,y,j next ,I signal ,I inter ] represents the lane centerline information numbered j in the map, where x and y are the longitudinal and lateral coordinate vectors of the lane centerline respectively, and j next is the number of the subsequent lane of the center line of the jth lane, I signal is the traffic light information of the lane (0: non-red light, 1: red light), I inter Indicates whether the vehicle is in an intersection (0: not in an intersection, 1: in an intersection); at the same time, considering the acceleration and speed information of the historical trajectory and static road facilities, an environmental uncertainty function is constructed.

[0063] The second network input information: vehicle status, static trajectory, road constraints, traffic participants and traffic light information are input to the fully connected layer, and based on the above information, the Long Short-Term Memory (LSTM) network input tensor is constructed to predict the multimodal distribution of the trajectories of traffic participants in the future time domain [t, t+T]. inter By integrating information into the interaction layer of the LSTM network, the network structure of the LSTM can be simplified in simple scenarios. When the vehicle is in an intersection, considering the diversity and uncertainty of vehicle-to-vehicle interactions, all feasible lane centerlines and subsequent lane centerlines are screened from the lane centerline information on the map, and lane centerlines with topological relationships are spliced ​​to form a set of feasible paths. For each feasible path, its longitudinal and lateral uncertainty parameters are calculated separately. When predicting trajectories, high-precision digital maps and static candidate path sets of autonomous driving vehicles are combined to screen threatening vehicles. Using natural driving data and "long tail" data, combined with vehicle motion status and historical trajectory information, a driving intention recognition and vehicle trajectory prediction model is constructed based on the LSTM network. Secondly, a pooling strategy is designed to integrate vehicle-to-vehicle interaction information into the LSTM network to learn and characterize the interaction influence relationship between vehicles. Finally, a trajectory prediction model is established that comprehensively considers the vehicle's historical trajectory, driving intention, and vehicle-to-vehicle motion interaction. The LSTM decoder outputs the multimodal probability distribution of the vehicle's future trajectory, that is, the uncertainty set of the predicted trajectory;

[0064] Furthermore, when assessing the spatiotemporal risk situation of the traffic environment, the autonomous vehicle is represented by three circles, surrounding vehicles by two circles, and pedestrians and cyclists by one circle. A traffic risk penalty function is constructed based on the center distance between the autonomous vehicle and other participants and incorporated into the safety constraints of the MPC to ensure safe driving in various traffic scenarios.

[0065] S2, based on the spatiotemporal risk situation assessment model, selects the optimal control strategy between the expected travel speed and expected stopping speed strategies for each candidate static path; and determines the optimal path for autonomous driving by constructing a cost function through evaluating the indicators of the candidate paths of the ego vehicle; and constructs a constrained optimal control problem based on the ego vehicle dynamics model and model predictive control.

[0066] Step S2 of this embodiment of the present invention designs a non-conservative decision-making method for autonomous vehicles with defensive behavior, integrating autonomous decision-making and motion control into a unified processing framework, including two key submodules: static path planning and dynamic optimal tracking. Specifically, step S2 includes:

[0067] S21. The static path planning module utilizes static information such as road structure, speed limits, road markings, and traffic regulations, combined with human driving experience, to design a set of candidate paths and expected speeds without time information, thereby improving the real-time performance of local path planning. This module sets an expected passing speed and an expected stopping speed for each candidate static path, selecting and switching between these two strategies during strategy application. The static path planning module does not select the optimal trajectory and does not consider any information about dynamic traffic participants, making it capable of pre-storing in a map.

[0068] The S22 and dynamic optimal tracking modules are the core of the integrated decision-making and control framework. They reshape the autonomous driving decision-making and control tasks, integrating them into an optimal control problem that only includes one performance indicator and one dynamic system. They solve a decision-making and control strategy and improve the decision-making and control level. In each control cycle, the optimal path for autonomous driving is determined by combining the future trajectory distribution of traffic participants and evaluating the cost function constructed by three indicators: potential collision risk, traffic efficiency, and driving compliance of the candidate paths of the vehicle.

[0069] S23. Based on the above, combined with the ego-vehicle dynamics model, with traffic participants as the primary constraints, and considering road edge constraints and traffic light restrictions, a constrained optimal control problem is constructed based on MPC. The following uses the finite-horizon predictive control of tracking the i-th path as an example to describe the mathematical expression of the optimal control problem:

[0070]

[0071] satisfy:

[0072]

[0073] And obey the constraints of traffic participants (no collision with C surrounding vehicles within the prediction time domain):

[0074] x self (τ)-x surr,i (τ)≥d self ,τ∈[t,t+T]

[0075] Road edge constraints:

[0076] x self (τ)-x road (τ)≥δ road ,τ∈[t,t+T]

[0077] Traffic light constraints:

[0078] x self (τ)∈free,if green,τ∈[t,t+T]

[0079] x self (τ)≤L stop ,if red,τ∈[t,t+T]

[0080] Among them, t is the current time, τ is the virtual time of the prediction domain, T is the length of the prediction domain, is the desired tracking position, is the desired driving speed, x self is the state of the vehicle, u self is the vehicle control quantity, x surr (τ) is the state of the surrounding traffic participants, including the constraint relationship between C traffic participants and the vehicle, x road Refers to the drivable area enclosed by the edge of the road, d self is the safe distance of the surrounding vehicle, δ road is the safe distance from the edge of the road, L stop is the distance to the stop line at the intersection, R represents the red light, and G represents the green light. When the light is green, the driving state is not restricted by the traffic light; when the light is red, the vehicle position within the predicted time domain cannot exceed the stop line.

[0081] S3, based on the driving behavior and decision-making process of human drivers in interactive scenarios, integrates environmental uncertainty information into an integrated decision-making and control architecture to optimize the front wheel angle and longitudinal acceleration control commands through the constrained optimal control problem, thereby controlling the vehicle to achieve autonomous driving tasks.

[0082] In the embodiment of the present invention, step S3 is used to integrate environmental uncertainty into the integrated decision-making architecture by combining the driving behavior and decision-making process of human drivers in interactive scenarios. Figure 2 As shown, it includes the following sub-steps:

[0083] S31. By characterizing the contribution of different future probability distributions to the cost function, the uncertainty problem is constructed as a model prediction optimization problem, and a non-conservative decision-making method for defensive behavior is designed to deal with various uncertainties in the urban road environment.

[0084] S32, through scene data, fully connected layer and LSTM network, considering the multimodal distribution of future trajectories of traffic participants and the static path set of autonomous driving vehicles, select the associated vehicle r i and potential collision area c i , for each collision area c i , set a safety line k and stop line e k ,These safety lines and stop lines form a set, and the ,integrated decision and control system uses these sets to adopt corresponding ,maneuvering strategies to ensure driving safety and avoid being ,overly conservative.

[0085] S33. Define the time from the current position of the vehicle to the potential collision area as TTC, introduce the constraint-corrected TTC criterion, and the specific formula is as follows:

[0086]

[0087] in, The revised TTC, is the self-car and the surrounding car r i distance, is the k-th step surrounding vehicle r in the MPC prediction domain at time t i The driving speed is ξ, and the set safety distance is ξ.

[0088] In the assessment of traffic environment risk situation, in order to describe the uncertainty of the environment, the Bayesian network is used to evaluate the potential risks in the traffic scene. i , define a random variable z to represent the three risk levels associated with TTC: danger, warning, and safety; integrate driving safety and vehicle tracking performance, define threshold values ​​related to potential risk levels, and construct a likelihood function for TTC to parameterize scenario risks; assuming that the prior probability density function P(z) of the potential threat level of the associated vehicle follows a uniform distribution, the probability distribution of the potential threat can be determined using the Bayesian theorem, as follows:

[0089]

[0090] in, The kth step of the MPC prediction domain at time t and the surrounding vehicles r i The risk level, represent The risk level under the condition is The probability of The collision time under the risk level z condition is The probability, N Z Represents the possible risk level, namely danger, warning and safety, The collision time for each risk condition probability.

[0091] On this basis, probabilistic constraints are constructed to incorporate potential risks into the constraints of MPC to deal with the uncertainty of the system. In order to cope with various changes that may occur in the future, autonomous vehicles need to make decisions and control their driving behaviors under different events. At this time, if MPC optimizes and solves an open-loop control sequence that meets all constraints, the decision-making behavior may be too conservative. To solve this problem, based on the multimodal distribution of the predicted trajectories of traffic participants, the corresponding front wheel angle and longitudinal acceleration control action sequence will be optimized and decided. However, in the rolling horizon optimization framework, autonomous vehicles usually execute the first action in the optimization sequence. Therefore, the short-term control instructions of the vehicle in the prediction domain should be consistent, because the action that the vehicle needs to perform at the next moment must be certain.

[0092] S34. When designing an MPC control system, a shared safety set and a preview safety set are designed in the prediction time domain. In the shared safety set, the uncertainty of the traffic environment is not considered to avoid the potential risks in the future directly affecting the current decision-making results; in the preview safety set, the probability distribution of potential threats in the traffic environment is combined to construct a probability constraint. ε is the set probability constraint threshold, and the probability constraint is finally integrated into the MPC optimization problem.

[0093] S35. Based on real-time environmental information, the probabilistic constraint model is updated and the MPC optimization problem is re-solved to dynamically handle environmental uncertainty. This ensures the stability of current decision-making and control while also accounting for future uncertainties. Specifically, the MPC prediction horizon is divided into two sets: a shared safety set (the first t1 steps within the prediction horizon, which only considers safety factors and ignores environmental uncertainty) and a preview safety set (which considers the multimodal distribution of the predicted trajectory). This constructs a constrained multi-objective optimization control problem, integrating environmental uncertainty into an integrated decision-making and control architecture.

[0094] S36. Optimize and determine the shared and bifurcated action sequences for the front wheel steering angle and longitudinal acceleration, and input the first action of the control sequence into the vehicle verification module to complete the vehicle's autonomous driving mission. Finally, determine whether the vehicle has completed the autonomous driving mission. If so, the algorithm terminates. If not, reacquire information about the vehicle and surrounding traffic environment, and repeat the above steps until the autonomous driving mission is completed. This design avoids overreacting to low-probability potential collision risks. The final decision-making strategy is not immediately formulated, but only when necessary, resulting in the autonomous vehicle exhibiting non-conservative driving characteristics with defensive behavior. This design aims to enable the autonomous vehicle to calmly respond to potential threats and ensure driving safety, while not being overly conservative and compromising driving quality.

[0095] According to the non-conservative decision-making and control method for an autonomous driving vehicle with defensive behavior in an embodiment of the present invention, the method fully utilizes intelligent network information, comprehensively considers the uncertainty in the traffic environment, quantifies the uncertainty in the environment as a probability distribution, and designs a non-conservative decision-making and control method for an autonomous driving vehicle with defensive behavior. This method prevents the vehicle from overreacting when facing potential dangers, while at the same time preventing the driving quality from being reduced due to excessive conservatism, thereby achieving safe and non-conservative driving of autonomous driving vehicles in uncertain environments.

[0096] In the description of this specification, the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0097] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

Claims

1. A non-conservative decision-making method for an autonomous vehicle with defensive behavior, characterized in that: include: Predict the movement trajectories of traffic participants based on information about the vehicle and surrounding traffic environment, modeling environmental uncertainty information as a multimodal probability distribution, and constructing a spatiotemporal risk situation assessment model based on spatial position relationships; Selecting the optimal control strategy of the two strategies of expected passing speed and expected stopping speed set for each candidate static path based on the spatiotemporal risk situation assessment model; The optimal path for autonomous driving is determined by constructing a cost function based on the evaluation of the indicators of the candidate paths of the ego vehicle; the constrained optimal control problem is constructed based on the ego vehicle dynamics model and model predictive control; Based on the human driver's driving behavior and decision-making process in interactive scenarios, environmental uncertainty information is integrated into the integrated decision-making and control architecture to optimize the front wheel angle and longitudinal acceleration control commands through the constrained optimal control problem described above, thereby controlling the vehicle to achieve autonomous driving tasks; The environmental uncertainty information includes: the movement direction, speed, acceleration of surrounding vehicles and / or pedestrians, and the intentions of surrounding vehicles and / or pedestrians; The method of predicting the movement trajectory of traffic participants based on the vehicle and surrounding traffic environment information to model the environmental uncertainty information as a multimodal probability distribution includes: Defining first network input information; wherein the first network input information includes historical trajectory information and lane centerline information of traffic participants; Inputting the second network input information into the fully connected layer, and constructing a long short-term memory (LSTM) network input tensor based on the first network input information and the second network input information to predict and output the multimodal distribution of the trajectories of traffic participants in a preset future time domain; wherein the second network input information includes the vehicle state, static trajectory, road constraints, traffic participants, and traffic light information; Establishing a trajectory prediction model based on the vehicle's historical trajectory, driving intention, and vehicle-to-vehicle motion interaction to output a multimodal probability distribution of the vehicle's trajectory in a preset future time domain based on the multimodal distribution through an LSTM decoder; Historical trajectory information of traffic participants and lane centerline information ,in, For the number i Traffic participants in the historical time domain[ tn, t ] within the trajectory information; The number in the representation map is j Lane centerline information, x , y are the longitudinal and transverse coordinate vectors of the lane centerline, For the j The number of the subsequent lanes of the lane center line, For lane signal light information, Indicates whether the vehicle is in an intersection; The optimal control strategy of the two strategies of expected passing speed and expected stopping speed set for each candidate static path is selected based on the spatiotemporal risk situation assessment model; and the optimal path for autonomous driving is determined by constructing a cost function by evaluating the indicators of the candidate paths of the ego vehicle, including: Based on static information including at least road structure, speed limit, road markings, and traffic regulations, and in combination with human driving experience, a set of candidate paths and expected vehicle speeds without time information is designed. Two strategies, one for expected passing speed and one for expected stopping speed, are set for each candidate static path. These two strategies are selected and switched based on the spatiotemporal risk situation assessment model. In each control cycle, the optimal path for autonomous driving is determined by combining the multimodal distribution of traffic participants' trajectories in a preset future time domain and evaluating a cost function constructed from three indicators: potential collision risk, traffic efficiency, and driving compliance. The constrained optimal control problem constructed by combining the vehicle dynamics model and model predictive control includes: Combined with the vehicle dynamics model, with traffic participants as the main constraints, and based on road edge constraints and traffic light restrictions, a constrained optimal control problem is constructed based on model predictive control (MPC). This includes: To track the i The mathematical expression of the optimal control problem is described as follows: （1） satisfy: And subject to the constraints of traffic participants: Road edge constraints: Traffic light constraints: in, t It is the present moment, is the virtual time of the prediction domain, T is the length of the prediction time domain, is the desired tracking position, is the expected driving speed, It is the state of the car. is the vehicle control quantity, is the state of the surrounding traffic participants, including the constraint relationship formed between C traffic participants and the vehicle, Refers to the drivable area enclosed by the edge of the road. is the safe distance of the surrounding vehicle, A safe distance from the edge of the road. is the distance to the stop line at the road intersection, R Represents red light, G Represents green light; Based on the human driver's driving behavior and decision-making process in interactive scenarios, environmental uncertainty information is integrated into an integrated decision-making and control architecture to optimize the front wheel angle and longitudinal acceleration control commands through the constrained optimal control problem described above, thereby controlling the vehicle to achieve autonomous driving tasks, including: By characterizing the contribution of different future probability distributions to the cost function, the uncertainty problem is structured as a model prediction optimization problem. Through scene data, fully connected layers, and LSTM networks, the relevant vehicles are selected based on the multimodal distribution of the trajectories of traffic participants in the future preset time domain and the static path set of the autonomous driving vehicle. and potential collision areas , for each collision area , set a safety line and stop lines ; The time from the vehicle's current position to the potential collision zone is defined as TTC, and a constraint-modified TTC criterion is introduced. When assessing the risk situation of the traffic environment, a Bayesian network is used to evaluate the potential risks in the traffic scene. A shared safety set and a preview safety set are designed within the MPC prediction domain. In the shared safety set, the uncertainty of the traffic environment is not considered; in the preview safety set, the probability distribution of potential threats in the traffic environment is combined, and the probability constraints are integrated into the MPC optimization problem. Based on real-time environmental information, the probability model is updated and the MPC optimization problem is re-solved to dynamically handle environmental uncertainties. The MPC prediction domain is divided into the shared safety set and the preview safety set, and a constrained multi-objective optimization control problem is constructed to integrate environmental uncertainty factors into the integrated decision-making and control architecture. The shared action sequence and bifurcated action sequence of the front wheel steering angle and longitudinal acceleration are optimized and determined, and the optimized control instructions are input to the vehicle to realize the vehicle's autonomous driving task.

2. The method according to claim 1, characterized in that The time from the current position of the vehicle to the potential collision area is defined as TTC. The constraint-modified TTC criterion is introduced, and the formula is as follows: （2） in, The revised TTC, For the self-car and the surrounding car r i distance, for t The first time in the MPC prediction domain k Walk around the vehicle r i The driving speed, The set safety distance.

3. The method according to claim 2, characterized in that When assessing the risk situation of the traffic environment, a Bayesian network is used to evaluate the potential risks in the traffic scenario, including: For associated vehicles , define the random variable z , used to characterize the three risk levels of danger, warning, and safety associated with TTC; comprehensively consider driving safety and vehicle tracking performance, define threshold values ​​related to potential risk levels, and construct TTC likelihood functions to parameterize scenario risks; assume a priori probability density function of the potential threat level of the associated vehicle P ( z ) follows a uniform distribution, then the probability distribution of potential threats is determined by Bayes’ theorem, as follows: （3） in, for t MPC prediction time domain k Walk with surrounding vehicles r i The risk level, represent The risk level under the condition is The probability of Represents risk level z The collision time under the condition is The probability of Represents the possible risk level, namely danger, warning and safety, The collision time for each risk condition probability.

4. The method according to claim 3, characterized in that In the preview safety set, combined with the probability distribution of potential threats in the traffic environment, a probability constraint is constructed. , The probability constraint threshold is set, and the probability constraint is finally integrated into the MPC optimization problem.

Images