Automatic driving decision-making method and system with dynamic risk perception and attention focusing functions and vehicle

Abstract

The invention belongs to the technical field of automatic driving, and particularly relates to an automatic driving decision-making method and system with dynamic risk perception and attention focusing and a vehicle, and the method comprises the steps: predicting the track of a surrounding vehicle in real time through a multi-feature Gaussian weighted particle filtering algorithm, and improving the prediction precision through combining a vehicle kinematic model and resampling optimization; constructing a comprehensive evaluation model fusing transverse and longitudinal risks, and dynamically quantifying the collision risk of the vehicle and surrounding vehicles; and inputting the risk value as a key state feature into a double-depth Q network based on attention mechanism enhancement, focusing key information through a feature attention distribution mechanism, and generating an optimal driving decision in combination with a multi-target reward function. Compared with the prior art, the method solves the problems of insufficient quantification of uncertainty factors, incomplete risk assessment and low decision-making efficiency of automatic driving in a complex dynamic environment, and significantly improves the risk perception capability and decision-making safety of the automatic driving vehicle.

Description

Technical Field

[0001] The present invention belongs to the field of autonomous driving technology, and specifically relates to an autonomous driving decision-making method, system and vehicle with dynamic risk perception and attention focus. Background Art

[0002] Autonomous driving technology has made significant progress in recent years, with continuous improvements in various perception, decision-making, and control systems, enabling autonomous vehicles to operate autonomously in various scenarios. The decision-making system is one of the core modules of autonomous vehicles, responsible for planning the vehicle's future trajectory and maneuvers based on perceived environmental information.

[0003] Existing autonomous driving decision-making systems are primarily designed based on rules or models. Rule-based decision-making systems use a pre-defined rule base to guide vehicle driving behavior. For example, if the preceding vehicle decelerates, the vehicle will decelerate. Model-based decision-making systems, on the other hand, utilize machine learning and other technologies to train models to learn the driver's driving behavior and make decisions based on the model's predictions.

[0004] However, existing autonomous driving decision-making systems still have several flaws that limit their application and safety. First, rule-based decision-making systems struggle to cope with complex and changing traffic environments. For example, when faced with unexpected situations, the rule base may not be able to provide effective decisions. Second, model-based decision-making systems often require large amounts of training data, and the model's generalization capabilities are limited, making it difficult to adapt to different driving scenarios. Furthermore, existing decision-making systems often lack the ability to dynamically perceive risks and are unable to adjust decision-making strategies in a timely manner based on current environmental conditions and traffic conditions, which can lead to safety accidents. Summary of the Invention

[0005] The purpose of the present invention is to provide an autonomous driving decision-making method, system and vehicle with dynamic risk perception and attention focus, so as to enable autonomous driving vehicles to have accurate risk perception and intelligent decision-making capabilities in complex dynamic traffic environments, so as to overcome the shortcomings of existing methods such as insufficient quantification of uncertainty factors, incomplete risk assessment and low decision-making efficiency.

[0006] The present invention achieves the above-mentioned purpose through the following technical solutions: This application proposes an autonomous driving decision-making method with dynamic risk perception and attention focus, including the following steps: S1. Initialize a particle set based on the currently observed vehicle state, update the particle set state using the vehicle kinematic model, introduce Gaussian noise to simulate uncertainty, calculate particle weights using a multidimensional Gaussian distribution, and output a weighted average predicted trajectory after resampling and optimization; the particle weights include position error, heading angle error, and velocity error; S2. Calculate the longitudinal collision risk value based on the vehicle's current position, predicted position, speed, and safety distance. and lateral collision risk value , and dynamically adjust the weight according to the traffic density, and output the comprehensive driving risk value IDR; S3. The comprehensive driving risk value IDR and the vehicle driving intention are integrated into features and input into a dual-depth Q network enhanced by an attention mechanism to generate a lane-changing decision for the autonomous driving vehicle.

[0007] As a preferred embodiment of the present invention, step S1 includes: S11. Obtain the real-time status data of the observed vehicle at the current moment, including the vehicle's position coordinates, heading angle, and speed; S12. In the three-dimensional state space, N initial particles are generated based on the current observed state data, and the state of each particle obeys a Gaussian distribution with the observed value as the mean; S13. For each particle, calculate the next state prediction value based on the vehicle kinematic model; S14. For each predicted particle, calculate its three-dimensional error with the latest observed data, including position error, heading angle error, and velocity error; S15. Calculate particle weights based on the three-dimensional error, normalize the weights, and perform system resampling to retain high-weight particles; S16. Output the weighted average state of the resampled particles as the final prediction result.

[0008] As a preferred solution of the present invention, in step S13, the particle state update adopts the following kinematic model:

[0009] in They are the predicted position, predicted heading angle and predicted speed of the vehicle; They are the current position, current heading angle and current speed of the vehicle; is the yaw angle, is the prediction time step, L is the vehicle's wheelbase, is Gaussian noise, are the current acceleration of the vehicle and the variance of Gaussian noise, respectively.

[0010] As a preferred embodiment of the present invention, step S2 includes: S21. Obtain the current state and predicted state of the vehicle and surrounding vehicles; S22. Calculate longitudinal risk value , as follows:

[0011] in and Represent the longitudinal positions of the surrounding vehicles and the vehicle, and Represent the predicted longitudinal positions of surrounding vehicles and the ego vehicle respectively; is the longitudinal safety distance between the vehicle and surrounding vehicles, is the longitudinal true distance between the ego vehicle and surrounding vehicles; is the maximum value of the longitudinal risk, is the shape adjustment factor; , are the dynamic convergence coefficient and predicted dynamic convergence coefficient along the longitudinal direction, respectively, where is the reference speed, is the current speed of the vehicle, is the predicted speed of the ego vehicle, is a constant coefficient, is the longitudinal constant convergence coefficient; S23. Calculate the horizontal risk value , as follows:

[0012] in is the maximum value of the horizontal risk, It is the lateral safety distance between the vehicle and surrounding vehicles, which is set based on the road width; the lateral real distance ,in is the lateral distance between the vehicle and surrounding vehicles, represents the distance from the center of mass of the vehicle to the side of the vehicle; and , , are the dynamic convergence coefficient and predicted dynamic convergence coefficient along the transverse direction, respectively, where is a constant coefficient, is the transverse constant convergence coefficient; S24. Calculate the integrated risk value (IDR) as follows:

[0013] in and are the dynamic weight coefficients of longitudinal and lateral risks related to traffic density, , ;in is the current traffic density, which is calculated based on the ratio of the number of vehicles to the number of roads within a certain range. and are the minimum and maximum values ​​of traffic density respectively.

[0014] As a preferred embodiment of the present invention, in step S22, the calculation of the longitudinal safety distance and the true distance includes the following formula:

[0015]

[0016] in and Represent the speed and predicted speed of the vehicle and surrounding vehicles respectively; and Represent the heading angle and predicted heading angle of the vehicle and surrounding vehicles respectively, is the longitudinal distance between the vehicle and surrounding vehicles, is the distance from the center of mass of the vehicle to the front of the vehicle, Represents the distance from the center of mass of the vehicle to the side of the vehicle body, is the safety distance margin, is the driver's reaction time, It is the response time of the vehicle's braking system. is the maximum deceleration of the vehicle.

[0017] As a preferred embodiment of the present invention, step S3 includes: S31. Encode the states of the vehicle and surrounding vehicles into feature vectors; S32. Calculate the similarity weights between the ego vehicle and surrounding vehicles through the self-attention mechanism, and generate weighted fusion to generate environmental features that represent the attention distribution; S33. The features are combined with the original vehicle state and input into the DDQN network to generate an action strategy. The action strategy includes: left lane change, right lane change, acceleration, deceleration, and hold.

[0018] As a preferred embodiment of the present invention, the method further includes reward function design, specifically including: The safety reward function is set to punish the vehicle in a collision in a positive correlation with its own vehicle speed, as shown in the following formula:

[0019] is a safety reward function, where is the forward speed of the vehicle, is the maximum speed allowed on the road, is the safety reward weight coefficient; An efficiency reward function is set to encourage vehicles to drive close to the road speed limit, as follows:

[0020] is the efficiency reward function, where is the efficiency reward weight coefficient, is the minimum speed allowed on the road; Set a scenario-specific reward function for merging into a lane to avoid blocking other vehicles, as shown below:

[0021] is the merging lane speed penalty reward function, is its weight coefficient, is the target speed for the ego vehicle to pass through the merging lane.

[0022] This application proposes an autonomous driving decision-making system, which is applied to implement the autonomous driving decision-making method proposed above. The system includes: Environmental perception module, used to obtain vehicle status information; The trajectory prediction module is used to initialize the particle set based on the currently observed vehicle state, update the particle set state through the vehicle kinematic model, introduce Gaussian noise to simulate uncertainty, calculate the particle weights using a multi-dimensional Gaussian distribution, and output the weighted average predicted trajectory after resampling and optimization; the particle weights include position error, heading angle error, and velocity error; Risk assessment module, used to calculate the longitudinal collision risk value based on the vehicle's current position, predicted position, speed and safety distance and lateral collision risk value , and dynamically adjust the weight according to the traffic density, and output the comprehensive driving risk value IDR; A decision generation module is used to integrate the comprehensive driving risk value IDR and the vehicle's driving intention and input the integrated driving risk value IDR and the vehicle's driving intention into a dual deep Q network enhanced by an attention mechanism to generate a lane change decision for the autonomous driving vehicle; The control execution module is used to convert lane change decisions into control signals.

[0023] The present application proposes a vehicle, comprising an autonomous driving decision-making system as proposed above.

[0024] The beneficial effects of the present invention are: The present invention achieves high-precision prediction of surrounding vehicle trajectories through a multi-feature Gaussian weighted particle filtering algorithm. The algorithm combines vehicle kinematic constraints and multi-dimensional error calculation to significantly improve the accuracy of the prediction and environmental adaptability. The comprehensive driving risk assessment model constructed based on the prediction results establishes a more comprehensive risk perception system by dynamically integrating lateral and longitudinal risk factors, enabling the autonomous driving system to more accurately identify potential dangers. The deep reinforcement learning decision-making algorithm that introduces the attention mechanism effectively focuses on key environmental information and optimizes the information processing efficiency of the decision-making process. While ensuring real-time performance, the overall technical solution greatly improves the decision-making safety and reliability of the autonomous driving system in complex traffic environments, and provides an effective solution to the problems of incomplete risk assessment and low decision-making efficiency in existing autonomous driving technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a flow chart of the autonomous driving decision-making method of the present invention; Figure 2 A flowchart of the automatic driving decision-making method of the present invention; Figure 3 This is a detailed flow chart of the particle filter trajectory prediction module; Figure 4 is a schematic diagram of the integrated risk assessment module; Figure 5 Schematic diagram of the simulation environment for a four-lane highway scenario; Figure 6 Schematic diagram of the simulation environment for the highway merging lane scenario; Figure 7 This is the collision rate comparison curve for a four-lane highway scenario; Figure 8 This is the average speed comparison curve for a four-lane highway scenario; Figure 9 This is the average reward comparison curve for the four-lane highway scenario; Figure 10 This is the collision rate comparison curve for the merging lane scenario; Figure 11 This is the average speed comparison curve in the merging lane scenario; Figure 12 This is the average reward comparison curve for the merging lane scenario; Figure 13 This is a driving behavior analysis diagram for a four-lane highway scenario; Figure 14 This is a driving behavior analysis diagram for the merging lane scenario. DETAILED DESCRIPTION

[0026] The following description provides specific application scenarios and requirements for this specification, with the goal of enabling those skilled in the art to make and use the contents of this specification. Various modifications to the disclosed embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other embodiments and applications without departing from the spirit and scope of this specification. Therefore, this specification is not limited to the embodiments shown, but is intended to be accorded the broadest scope consistent with the claims.

[0027] The terms used herein are for the purpose of describing specific example embodiments only and are not intended to be limiting. For example, as used herein, the singular forms "a," "an," and "the" may also include the plural forms unless the context clearly indicates otherwise. When used in this specification, the terms "comprise," "include," and / or "contain" are intended to refer to the presence of the associated integers, steps, operations, elements, and / or components, but do not preclude the presence of one or more other features, integers, steps, operations, elements, components, and / or groups or the addition of other features, integers, steps, operations, elements, components, and / or groups in the system / method.

[0028] These and other features of this specification, as well as the operation and function of the associated elements of the structure, and the economical assembly and manufacture of the components, can be significantly improved with consideration of the following description. Reference is made to the accompanying drawings, all of which form a part of this specification. However, it should be expressly understood that the drawings are for illustration and description purposes only and are not intended to limit the scope of this specification. It should also be understood that the drawings are not drawn to scale.

[0029] The flowcharts used in this specification illustrate operations implemented by systems according to some embodiments of the present specification. It should be clearly understood that the operations of the flowcharts may not be implemented in sequence. Rather, the operations may be implemented in reverse order or simultaneously. Furthermore, one or more additional operations may be added to the flowcharts. One or more operations may be removed from the flowcharts.

[0030] Example 1

[0031] See also Figure 1 and Figure 2 This embodiment proposes an autonomous driving decision-making method with dynamic risk perception and attention focus. By combining predicted trajectory information, current vehicle state information, and comprehensive risk assessment values, a reasonable lane change decision scheme is designed, thereby improving the safety and real-time performance of lane change decisions made by autonomous vehicles. The method specifically includes the following steps: S1. Initialize a particle set based on the currently observed vehicle state. Update the particle set state through the vehicle kinematic model and introduce Gaussian noise to simulate uncertainty. Calculate particle weights using a multidimensional Gaussian distribution. After resampling and optimization, output a weighted average predicted trajectory; the particle weights include position error, heading angle error, and velocity error.

[0032] S2. Calculate the longitudinal collision risk value based on the vehicle's current position, predicted position, speed, and safety distance. and lateral collision risk value , and dynamically adjust the weight according to the traffic density, and output the comprehensive driving risk value IDR.

[0033] S3. The comprehensive driving risk value IDR and the vehicle driving intention are integrated into features and input into a dual-depth Q network enhanced by an attention mechanism to generate a lane-changing decision for the autonomous driving vehicle.

[0034] It can be understood that the present invention realizes intelligent decision-making of autonomous vehicles through the collaboration of multiple modules. First, trajectory prediction is based on the particle filter algorithm. By initializing the particle set and using the vehicle kinematic model to update the state, the particle weight is calculated in combination with the multi-dimensional Gaussian distribution, and finally the weighted average predicted trajectory is output to provide the ability to predict future traffic conditions. Secondly, risk assessment calculates the longitudinal and lateral collision risk values ​​respectively by real-time analysis of the current state of the vehicle and the predicted trajectory, and dynamically adjusts the weight coefficient based on the traffic density to generate a comprehensive driving risk value IDR, thereby achieving a quantitative assessment of environmental risks. Finally, decision generation integrates the comprehensive risk value with the vehicle state characteristics, and performs feature selection and decision optimization through a dual-depth Q network enhanced by an attention mechanism. The attention mechanism dynamically focuses on key environmental information and suppresses irrelevant interference, while the dual-depth Q network is based on a reinforcement learning framework to output the optimal lane change decision while balancing safety and efficiency.

[0035] Specifically, step S1 includes multi-feature Gaussian weighted particle filter trajectory prediction, more specifically: The particle filter directly approximates the posterior probability distribution through Monte Carlo sampling, so it is suitable for traffic scenarios with uncertainty and nonlinear characteristics. Compared with deep learning methods, particle filtering does not require a large amount of offline training data, can achieve real-time prediction under limited computing resources, and has strong applicability. Considering that the vehicle kinematic model has certain physical constraints, high computational efficiency and high interpretability, the state of the particles is updated according to the vehicle kinematic model. In addition, the weight calculation of the particles is designed based on the multi-dimensional Gaussian distribution, taking into account the position error, heading angle error and speed error to ensure the accuracy and robustness of the weight calculation and improve the reliability of the state estimation. The entire prediction process is as follows Figure 3 shown.

[0036] (a) Particle initialization

[0037] At the beginning of the prediction, a set of particles is randomly generated in the particle state space based on the observed vehicle states (vehicle position, heading angle, and speed) in the highway scenario. The distribution of these particles should cover the range of possible vehicle states as much as possible. The state space boundaries are bounded by the a priori constraints of the highway scenario (such as lane boundaries and speed limits) to ensure the rationality of the particle set.

[0038] (b) Particle update

[0039] This paper implements particle state prediction based on a vehicle kinematic model. While ensuring physical rationality and computational efficiency, it fully describes the core characteristics of vehicle motion (including key state variables such as position, velocity, acceleration, and yaw angle). The particle update formula designed by this invention is as follows:

[0040] in They are the predicted position, predicted heading angle and predicted speed of the vehicle; They are the vehicle's current position, current heading angle and current speed; is the yaw angle, is the prediction time step, L is the vehicle's wheelbase, is Gaussian noise, are the current acceleration of the ego vehicle and the variance of Gaussian noise, respectively. Particle updates based on the above model ensure the interpretability and smoothness of vehicle trajectory prediction.

[0041] (c) Weight calculation

[0042] Each particle in the particle filter represents a state hypothesis, and its weight is calculated by the likelihood function, and the particle weight is proportional to the likelihood function. The likelihood function of the present invention is designed based on position error, heading angle error, and velocity error. This multi-dimensional weight calculation has stronger anti-interference ability than a single error calculation. The likelihood function is shown below:

[0043] in Represent the errors between the observed and predicted values ​​of vehicle position, heading angle and speed respectively, The standard deviations of vehicle position, heading angle, and speed are calculated and assigned to each particle, providing a basis for subsequent particle resampling.

[0044] (d) Resampling

[0045] Resampling is performed by selecting particle indices based on their importance weights, prioritizing particles with higher weights. , ,... , ) parameterized categorical distribution for the index Take samples.

[0046] (e) Prediction output

[0047] The model predicts the vehicle state including the vehicle's lateral and longitudinal positions, navigation angles, and speed. The predicted output result is the weighted sum of all particles, and the predicted output value will be used as part of the input information of the comprehensive risk assessment model, providing the risk assessment model with future trajectory state information to improve the comprehensiveness of the risk assessment model.

[0048] Specifically, step S2 includes a comprehensive risk assessment, more specifically: In order to enable autonomous vehicles to make decisions in real time and accurately, the present invention designs a comprehensive risk assessment function, which combines the vehicle's predicted trajectory to quantify the uncertainty risk in a dynamic environment. Unlike traditional rule-based or heuristic methods, the designed method considers a more comprehensive set of vehicle motion states (including the current vehicle state, the vehicle's predicted trajectory state, and safety distance factors) to quantitatively assess the lateral and longitudinal risks of the vehicle, and dynamically weights and fuses the lateral and longitudinal risks, enabling a more comprehensive and detailed analysis and calculation of potential dangers. The model evaluates the integrated driving risk (IDR) of the four vehicles closest to the vehicle, and uses the quantified risk as the state input of the DDQN (Double Deep Q-Network) decision model, thereby enhancing the vehicle's risk perception of traffic conditions. The main advantage of this method is its continuous risk estimation, which allows for smoother and more adaptive risk assessment under uncertainty. The schematic diagram of the comprehensive risk assessment is shown in the figure below. Figure 4 As shown. The longitudinal risk assessment function is defined as follows:

[0049]

[0050]

[0051] in is the maximum value of the longitudinal risk, is the shape adjustment factor. and Represent the longitudinal positions of the surrounding vehicles and the vehicle, and Represent the predicted longitudinal positions of surrounding vehicles and the ego vehicle, respectively. and Represent the speed and predicted speed of the vehicle and surrounding vehicles respectively. and Represent the heading angle and predicted heading angle of the vehicle and surrounding vehicles respectively. ， , are the dynamic convergence coefficient and predicted dynamic convergence coefficient along the longitudinal direction respectively. is the reference speed, set to 25 m / s, is a constant convergence coefficient set to 1.5, Set to 0.8. If the vehicle speed and the predicted speed of the vehicle If it is too high, the dynamic convergence coefficient will decrease accordingly, which will lead to an increase in the sensitivity of the risk value to the change in the relative distance between the vehicle and the surrounding vehicles, thereby enhancing the model's ability to respond to close-range risks. It is the ratio of the longitudinal safety distance to the actual distance. If the actual distance between the vehicle and the surrounding vehicles is less than the safety distance, the risk value will be greater, indicating that the current following or lane change behavior has a higher collision risk. Conversely, if the actual distance is greater than the safety distance, the risk value will be smaller, indicating that the current driving state is relatively safe. and are the longitudinal and lateral distances between the vehicle and surrounding vehicles, is the distance from the center of mass of the vehicle to the front of the vehicle, Represents the distance from the center of mass of the vehicle to the side of the vehicle. is the safety distance margin, set to 2 m, is the driver's reaction time, set to 0.4 s, is the vehicle braking system response time, set to 0.15 s, is the maximum deceleration of the vehicle, set to 10 m / ; The horizontal risk assessment function has a similar structure to the vertical risk assessment function and is defined as follows:

[0052] in is the maximum value of the horizontal risk, is the lateral safety distance, which is set to 1.5 m based on the road width; the lateral actual distance and , , are the dynamic convergence coefficient and predicted dynamic convergence coefficient along the transverse direction, respectively, where is the transverse constant with a convergence coefficient of 0.8, Set to 0.5; By integrating the longitudinal and lateral risk assessment functions of the vehicle and considering the impact of traffic density on the longitudinal and lateral risk assessment, the final comprehensive risk assessment function is defined as follows:

[0053] In a typical road environment, the higher the traffic density, the smaller the distance between vehicles, the more likely vehicles are to rear-end, and the longitudinal risk accounts for a larger proportion of the total risk. Therefore, a dynamic weight related to traffic density is designed. and , the greater the traffic density, the higher the longitudinal risk. , , is the current traffic density, which is calculated based on the ratio of the number of vehicles to the number of roads within a certain range. and The final calculated comprehensive driving risk value will be used as the state input of the DDQN decision module, so that the autonomous driving vehicle has a certain risk awareness ability.

[0054] Specifically, step S3 includes a DDQN decision algorithm based on the attention mechanism, more specifically: (a) State processing module based on attention mechanism To further improve agent training efficiency, this paper introduces an attention mechanism into the DDQN decision-making algorithm to dynamically focus on key vehicle states. This mechanism enables the autonomous vehicle to autonomously allocate attention resources based on the current driving situation, more efficiently extracting decision-relevant information, thereby improving the effectiveness and generalization of policy learning.

[0055] The attention mechanism offers advantages over traditional feature processing methods such as those based on convolutional neural networks (CNNs) or long short-term memory (LSTM) networks. Compared to traditional models, the self-attention mechanism can adaptively focus on key vehicle states without being constrained by a fixed perception range, thereby improving decision-making precision. Furthermore, it can model global relationships and capture the potential impact of distant vehicles on current driving decisions, making it ideally suited to the complex traffic environment understanding required for autonomous driving. Furthermore, it supports parallel computing and does not rely on time series expansion, offering processing efficiency advantages over recurrent neural networks, making it suitable for real-time decision-making in autonomous driving.

[0056] The present invention uses the self-attention mechanism to weight the vehicle state feature process in the environment: First, the input state tensor is split into features of the ego vehicle and other vehicles, and the corresponding attention masks are generated.

[0057] The features of the ego vehicle and other vehicles are then encoded separately through a multi-layer perceptron to produce embedded representations. The ego vehicle's embedding serves as the query, while the embeddings of the ego vehicle and other vehicles are concatenated to form the key and value. Next, an attention mechanism is used to calculate the similarity weight between the query and the key, and this is combined with the value to produce the attention output.

[0058] Finally, the output is fused with the original ego-vehicle embedding using a residual connection to generate updated ego-vehicle features for subsequent decision tasks based on the DDQN decision algorithm.

[0059] (b) States and Actions

[0060] In deep reinforcement learning, the state space is the set of all possible states that an agent can perceive when interacting with its environment. The proper design of the state space is crucial to the convergence and performance of the algorithm. For the baseline method, the environmental states selected in this paper are: the vehicle's lateral and longitudinal position, lateral and longitudinal position deviations, lateral and longitudinal velocities, and the vehicle's heading angle.

[0061] For the risk-aware DDQN decision framework designed by the present invention, a comprehensive driving risk feature is added to the state space ( State characteristics related to vehicle driving intention. The comprehensive driving risk characteristics are calculated using formula (7), where i = {0, 1, 2, 3}, representing the serial number of the surrounding vehicle closest to the vehicle.

[0062] The present invention uses discrete actions as the action space of the DDQN decision algorithm, which include: left lane change, right lane change, acceleration, holding and deceleration.

[0063] (c) Reward function

[0064] In autonomous driving decision-making based on reinforcement learning, the design of the reward function directly impacts the agent's learning outcomes and decision-making behavior. Safety and efficiency are typically the primary considerations in the decision-making process, so this invention considers these factors when designing the reward function. However, in the autonomous driving decision-making process, complex reward functions, such as those based on maintaining a safe distance between the vehicle and the preceding vehicle to avoid collisions, while intended to ensure driving safety, can reduce the agent's ability to explore. This is because under such complex reward rules, the agent becomes overly focused on collision avoidance and reduces its attempts at other driving strategies, limiting its ability to find the optimal decision-making strategy in different driving scenarios. Therefore, a simple and straightforward reward function is used to address vehicle decision-making.

[0065] The safety reward function is defined as follows:

[0066] is a safety reward function, where is the safety reward weight coefficient, is the forward speed of the vehicle, is the maximum speed allowed on the road, set to 30 m / s. The faster the vehicle moves, the greater the penalty value after a collision.

[0067] The efficiency reward function is designed as follows:

[0068] is the efficiency reward function, where is the efficiency reward weight coefficient, is the minimum permitted speed on the road, set to 20 m / s. The speed reward is designed to encourage autonomous vehicles to drive within the permitted speed range. The closer the speed is to the maximum permitted speed, the higher the reward. This guides the agent to learn a strategy of driving at a high and appropriate speed.

[0069] For the merging lane scenario constructed in this invention, another reward function is designed. In addition to the above reward function, it also includes a merging lane speed penalty reward function, as shown in the following formula:

[0070] is the merging lane speed penalty reward function, where is its weight coefficient, is the target speed for the ego vehicle in the merging lane, set to 30 m / s. The goal of this reward function is to guide the autonomous vehicle through the merging section at high speed or to make appropriate lane changes to avoid obstructing merging vehicles.

[0071] Example 2

[0072] This embodiment proposes an autonomous driving decision-making system, which is applied to implement the autonomous driving decision-making method proposed in Example 1. The system includes: Environmental perception module, used to obtain vehicle status information; The trajectory prediction module is used to initialize the particle set based on the currently observed vehicle state, update the particle set state through the vehicle kinematic model, introduce Gaussian noise to simulate uncertainty, calculate the particle weights using a multi-dimensional Gaussian distribution, and output the weighted average predicted trajectory after resampling and optimization; the particle weights include position error, heading angle error, and velocity error; Risk assessment module, used to calculate the longitudinal collision risk value based on the vehicle's current position, predicted position, speed and safety distance and lateral collision risk value , and dynamically adjust the weight according to the traffic density, and output the comprehensive driving risk value IDR; A decision generation module is used to integrate the comprehensive driving risk value IDR and the vehicle's driving intention and input the integrated driving risk value IDR and the vehicle's driving intention into a dual deep Q network enhanced by an attention mechanism to generate a lane change decision for the autonomous driving vehicle; The control execution module is used to convert lane change decisions into control signals.

[0073] According to the above embodiments, the working principle of the present invention is to achieve intelligent decision-making for autonomous vehicles through the collaboration of multiple modules. First, the environmental perception module collects real-time status information of the vehicle and surrounding vehicles, including key parameters such as position, heading angle, and velocity. The trajectory prediction module, based on a particle filtering algorithm, initializes a particle set and updates the state using a vehicle kinematic model. Gaussian noise is introduced to simulate the uncertainty of the traffic environment. A multi-dimensional Gaussian distribution is used to calculate particle weights and perform resampling optimization. The module ultimately outputs a weighted average predicted trajectory, providing the system with the ability to predict future traffic conditions. The risk assessment module analyzes the vehicle's current state and predicted trajectory in real time, calculating longitudinal and lateral collision risk values, dynamically adjusting the weight coefficients based on traffic density, and generating an integrated driving risk (IDR) to quantitatively assess environmental risk. The decision generation module integrates the integrated risk value with vehicle state features and performs feature selection and decision optimization using a dual-deep Q-network enhanced by an attention mechanism. The attention mechanism dynamically focuses on key environmental information and suppresses irrelevant interference. The dual-deep Q-network, based on a reinforcement learning framework, outputs optimal lane change decisions while balancing safety and efficiency. The control execution module converts decision instructions into vehicle control signals.

[0074] In another embodiment of the present invention, a vehicle is proposed, comprising an automatic driving decision system as in Example 2.

[0075] Specifically, the vehicle described in the present invention can be a passenger car, commercial vehicle, specialized vehicle, or any type of intelligent connected vehicle. The vehicle incorporates the autonomous driving decision-making system of the present invention as its "brain" or "decision-making center," closely connected and coordinated with the vehicle's sensors (e.g., cameras, lidar, millimeter-wave radar), positioning unit, high-precision map, and actuators (e.g., steering, drive, and braking systems).

[0076] The decision-making method proposed in the present invention is further elaborated and illustrated below with reference to specific test cases and some figures.

[0077] To investigate the performance of the proposed decision framework, we first used the intelligent driver model to generate random traffic flows to restore the uncertainty of the real traffic environment. We simulated two common highway traffic flow scenarios, where the road width is 4m, the vehicle length is 5m, and the vehicle width is 1.9m. Figure 5 and Figure 6 As shown. Scenario 1 is a four-lane highway scenario with high-density traffic flow, and scenario 2 is a highway merging lane scenario. In order to reflect the uncertainty of real-world traffic conditions, the vehicle positions and speeds in all scenarios contain a certain degree of randomness. The observable distance of the autonomous driving vehicle is 100 meters in front and behind the vehicle. In order to improve the training speed, this case generated four parallel training environments, and compared the decision-making scheme designed by the present invention with four common decision-making algorithms to illustrate its advantages, including PPO, DQN, DDQN and A2C. DQN (Deep Q-Network) is a value-based deep reinforcement learning algorithm, A2C (Advantage Actor-Critic) is an Actor-Critic framework algorithm based on policy gradient, and PPO (Proximal Policy Optimization) is an improved algorithm based on policy gradient; training results Figure 7-12 shown.

[0078] Figure 7-Figure 9 The training results for Scenario 1 are presented. Compared with traditional methods, the proposed decision-making model, PDMM (Perception-Decision Model with Multi-feature fusion), achieves a lower collision rate. However, this strategy also exhibits a lower average speed compared to other algorithms, resulting in slightly reduced driving efficiency. This may be due to the close distance between vehicles in this scenario. Due to the risk state, PDMM tends to sacrifice driving speed for driving safety. Overall, its average reward remains slightly higher than that of other methods. Meanwhile, PPO also exhibits a relatively conservative driving strategy, while DQN, DDQN, and A2C choose a riskier and more efficient driving strategy. The figure above demonstrates that the proposed decision-making model can dynamically adjust its decision-making strategy based on a comprehensive risk assessment, allowing the agent to perceive the current risk at each decision step and choose a safer action, effectively reducing the likelihood of collision.

[0079] Figure 10-12 The training results for Scenario 2 are presented. In this scenario, the PDMM achieves a lower collision rate than other methods while maintaining a relatively high driving speed. It performs well in terms of both safety and driving efficiency, with an average reward slightly higher than that of other methods. This demonstrates that the autonomous vehicle learns a more efficient and safer decision-making strategy through the integrated risk state assessment and attention mechanism.

[0080] To better understand the impact of environmental uncertainty on autonomous vehicle decision-making, this case uses DDQN and PDMM strategies to analyze decision-making scenarios. Figure 13 This is the driving situation in Scenario 1. Under the DDQN strategy, initially, with vehicles ahead in all four lanes, the ego vehicle, seeking a higher speed reward, attempts a lane change at T = 1 second. However, the timing of the lane change is incorrect, leading to a collision with Car D at T = 2 seconds. Under the PDMM strategy, the ego vehicle possesses a certain level of risk perception and attention allocation, finding the right moment to overtake Car C at T = 2 seconds, demonstrating that this strategy outperforms the DDQN strategy.

[0081] Figure 14 This is the driving situation in Scenario 2. Under the DDQN strategy, with vehicles A and B in front of it in lanes 1 and 2, the ego vehicle, seeking a higher speed reward, attempted to overtake both vehicles by changing lanes continuously at T = 1 second. However, it failed to fully consider the changes in vehicle B's state during the lane change, resulting in a collision with vehicle B between T = 1.5 seconds and 2 seconds. This demonstrates that the strategy only considers traffic efficiency and ignores the uncertain risks in the environment. In contrast, the PDMM uses particle filtering to predict vehicle trajectories and combines a comprehensive risk assessment function with a self-attention mechanism to determine the appropriate time for lane changes. It successfully overtook vehicles A and B at T = 1.5 seconds, ensuring both driving safety and efficiency.

[0082] In summary, the present invention has fully demonstrated the effectiveness and superiority of the proposed autonomous driving decision-making method through systematic experimental verification. In complex scenarios such as simulated high-density traffic flows and merging lanes, this method demonstrates significant technical advantages: through the innovative particle filter trajectory prediction algorithm, accurate prediction of the motion trajectory of surrounding vehicles is achieved; the designed comprehensive risk assessment model can comprehensively quantify the potential dangers in the traffic environment; the reinforcement learning decision algorithm based on the attention mechanism effectively improves the decision-making efficiency and adaptability of the system. The experimental results show that this method significantly improves the safety of autonomous driving vehicles while ensuring driving efficiency, especially showing stronger robustness when dealing with emergencies and complex traffic scenarios. Compared with traditional decision-making methods, the present invention can more accurately identify risks and more reasonably plan behaviors, enabling autonomous driving vehicles to make safer and more intelligent decisions in dynamically changing traffic environments.

[0083] The above description is merely a preferred embodiment of the present disclosure and an illustration of the technical principles employed. Those skilled in the art should understand that the scope of the present disclosure is not limited to technical solutions formed by specific combinations of the aforementioned technical features. It also encompasses other technical solutions formed by any combination of the aforementioned technical features or their equivalents, without departing from the scope of the above disclosure. For example, a technical solution formed by replacing the aforementioned features with (but not limited to) technical features with similar functions disclosed in this disclosure.

[0084] In addition, although each operation is described in a specific order, this should not be understood as requiring these operations to be performed in the specific order shown or in a sequential order. Under certain circumstances, multitasking and parallel processing may be advantageous. Similarly, although some specific implementation details have been included in the above discussion, these should not be interpreted as limiting the scope of the present disclosure. Some features described in the context of a separate embodiment can also be implemented in a single embodiment in combination. On the contrary, the various features described in the context of a single embodiment can also be implemented in multiple embodiments individually or in any suitable sub-combination mode.

[0085] The above-described embodiments merely illustrate several implementations of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that variations and improvements are possible without departing from the scope of the present invention, and all such variations and improvements fall within the scope of protection of the present invention.

Claims

1. An autonomous driving decision-making method with dynamic risk perception and attention focus, characterized by: The steps include: S1. Initialize a particle set based on the currently observed vehicle state, update the particle set state using the vehicle kinematic model, introduce Gaussian noise to simulate uncertainty, calculate particle weights using a multidimensional Gaussian distribution, and output a weighted average predicted trajectory after resampling and optimization; the particle weights include position error, heading angle error, and velocity error; S2. Calculate the longitudinal collision risk value based on the vehicle's current position, predicted position, speed, and safety distance. and lateral collision risk value , and dynamically adjust the weight according to the traffic density, and output the comprehensive driving risk value IDR; S3. The comprehensive driving risk value IDR and the vehicle driving intention are integrated into features and input into a dual-depth Q network enhanced by an attention mechanism to generate a lane-changing decision for the autonomous driving vehicle.

2. The autonomous driving decision-making method with dynamic risk perception and attention focus according to claim 1, characterized in that: Step S1 includes: S11. Obtain the real-time status data of the observed vehicle at the current moment, including the vehicle's position coordinates, heading angle, and speed; S12. In the three-dimensional state space, generate according to the current observation state data Initial particles, the state of each particle obeys a Gaussian distribution with the observed value as the mean; S13. For each particle, calculate the next state prediction value based on the vehicle kinematic model; S14. For each predicted particle, calculate its three-dimensional error with the latest observed data, including position error, heading angle error, and velocity error; S15. Calculate particle weights based on the three-dimensional error, normalize the weights, and perform system resampling to retain high-weight particles; S16. Output the weighted average state of the resampled particles as the final prediction result.

3. The autonomous driving decision-making method with dynamic risk perception and attention focus according to claim 2, characterized in that: In step S13, the particle state is updated using the following kinematic model: ； in They are the predicted position, predicted heading angle and predicted speed of the vehicle; They are the current position, current heading angle and current speed of the vehicle; is the yaw angle, is the prediction time step, L is the vehicle's wheelbase, is Gaussian noise, are the current acceleration of the vehicle and the variance of Gaussian noise, respectively.

4. The autonomous driving decision-making method with dynamic risk perception and attention focus according to claim 2, characterized in that: Step S2 includes: S21. Obtain the current state and predicted state of the vehicle and surrounding vehicles; S22. Calculate longitudinal risk value , as follows: ； in and Represent the longitudinal positions of the surrounding vehicles and the vehicle, and Represent the predicted longitudinal positions of surrounding vehicles and the ego vehicle respectively; is the longitudinal safety distance between the vehicle and surrounding vehicles, is the longitudinal true distance between the ego vehicle and surrounding vehicles; is the maximum value of the longitudinal risk, is the shape adjustment factor; , are the dynamic convergence coefficient and predicted dynamic convergence coefficient along the longitudinal direction, respectively, where is the reference speed, is the current speed of the vehicle, is the predicted speed of the ego vehicle, is a constant coefficient, is the longitudinal constant convergence coefficient; S23. Calculate the horizontal risk value , as follows: ； in is the maximum value of the horizontal risk, It is the lateral safety distance between the vehicle and surrounding vehicles, which is set based on the road width; the lateral real distance ,in is the lateral distance between the vehicle and surrounding vehicles, represents the distance from the center of mass of the vehicle to the side of the vehicle; and , , are the dynamic convergence coefficient and predicted dynamic convergence coefficient along the transverse direction, respectively, where is a constant coefficient, is the transverse constant convergence coefficient; S24. Calculate the integrated risk value (IDR) as follows: ； in and are the dynamic weight coefficients of longitudinal and lateral risks related to traffic density, , ;in is the current traffic density, which is calculated based on the ratio of the number of vehicles to the number of roads within a certain range; and are the minimum and maximum values ​​of traffic density respectively.

5. The autonomous driving decision-making method with dynamic risk perception and attention focus according to claim 4, characterized in that: In step S22, the calculation of the longitudinal safety distance and the true distance includes the following formula: ； ； in and Represent the speed and predicted speed of the vehicle and surrounding vehicles respectively; and Represent the heading angle and predicted heading angle of the vehicle and surrounding vehicles respectively, is the longitudinal distance between the vehicle and surrounding vehicles, is the distance from the center of mass of the vehicle to the front of the vehicle, Represents the distance from the center of mass of the vehicle to the side of the vehicle body, is the safety distance margin, is the driver's reaction time, It is the response time of the vehicle's braking system. is the maximum deceleration of the vehicle.

6. The autonomous driving decision-making method with dynamic risk perception and attention focus according to claim 5, characterized in that: Step S3 includes: S31. Encode the states of the ego vehicle and surrounding vehicles into feature vectors; S32. Calculate the similarity weights between the ego vehicle and surrounding vehicles through the self-attention mechanism, and perform weighted fusion to generate environmental features that represent the attention distribution; S33. The features are combined with the original vehicle state and input into the DDQN network to generate an action strategy. The action strategy includes: left lane change, right lane change, acceleration, deceleration, and hold.

7. The autonomous driving decision-making method with dynamic risk perception and attention focus according to claim 6, characterized in that: The method further includes reward function design, specifically including: The safety reward function is set to punish the vehicle in a collision in a positive correlation with its own vehicle speed, as shown in the following formula: ； is a safety reward function, where is the forward speed of the vehicle, is the maximum speed allowed on the road, is the safety reward weight coefficient; the efficiency reward function is set to encourage vehicles to drive close to the road speed limit, as follows: ； is the efficiency reward function, where is the efficiency reward weight coefficient, is the minimum speed allowed on the road; Set a scenario-specific reward function for merging into a lane to avoid blocking other vehicles, as shown below: ； is the merging lane speed penalty reward function, is its weight coefficient, is the target speed for the ego vehicle to pass through the merging lane.

8. An autonomous driving decision-making system, characterized in that: Applied to implement the autonomous driving decision-making method according to any one of claims 1 to 7, the system comprises: Environmental perception module, used to obtain vehicle status information; The trajectory prediction module is used to initialize the particle set based on the currently observed vehicle state, update the particle set state through the vehicle kinematic model, introduce Gaussian noise to simulate uncertainty, calculate the particle weights using a multi-dimensional Gaussian distribution, and output the weighted average predicted trajectory after resampling and optimization; the particle weights include position error, heading angle error, and velocity error; Risk assessment module, used to calculate the longitudinal collision risk value based on the vehicle's current position, predicted position, speed and safety distance and lateral collision risk value , and dynamically adjust the weight according to the traffic density, and output the comprehensive driving risk value IDR; A decision generation module is used to integrate the comprehensive driving risk value IDR and the vehicle's driving intention and input the integrated driving risk value IDR and the vehicle's driving intention into a dual deep Q network enhanced by an attention mechanism to generate a lane change decision for the autonomous driving vehicle; The control execution module is used to convert lane change decisions into control signals.

9. A vehicle, characterized in that: Including an autonomous driving decision system as described in claim 8.

Images