Auxiliary driving vehicle lane changing decision-making method aiming at heterogeneous traffic flow environment and having human-like dynamic decision-making characteristic

Abstract

The invention discloses an auxiliary driving vehicle lane change decision-making method aiming at a heterogeneous traffic flow environment and having a human-like dynamic decision-making characteristic, relates to the field of auxiliary driving vehicle lane change decision-making, and aims to solve the problem that a behavior strategy generated by an algorithm in the prior art is deviated from the actual expectation of a human driver. According to the method, information such as own vehicle and surrounding vehicle environment is jointly used as reference information of a decision-making module, the decision-making module adopts a PPO-DDM fusion module, dynamic weight distribution is performed on a PPO model and a DDM model according to traffic conditions, and an environment self-adaptive parameter adjustment mechanism sub-module and an evidence accumulation process numerical value implementation sub-module are introduced into a DDM decision-making sub-module. According to the invention, the decision-making module can predict future behaviors of surrounding vehicles, the decision-making has a continuous characteristic of progressive adjustment, a human driver can accurately pre-judge subsequent behaviors of the human driver conveniently, and adaptive adjustment can be carried out according to real-time traffic scene changes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of assisting vehicle lane-changing decision, in particular to an assisting vehicle lane-changing decision method with human-like dynamic decision characteristics for heterogeneous traffic flow environment. BACKGROUND

[0002] With the gradual popularization of assisting vehicles in open road scenes, how to realize safe, efficient and human-machine interaction compatible lane-changing behavior of assisting vehicles in complex traffic environment has become a core problem to be solved. Especially in high-density interactive areas such as weaving areas, vehicle lane-changing behavior is frequent and high-risk, and the existing assisting driving system still has deficiencies in understanding human driving behavior and responding to high-risk lane-changing decisions. Although the existing lane-changing decision research based on reinforcement learning can learn the optimal driving strategy through continuous trial and error in a simulated environment, there are still many challenges when facing complex traffic scenes. First, the reinforcement learning method usually lacks dynamic description of the decision-making process in the modeling process, ignoring the continuity characteristics such as decision accumulation and gradual adjustment in human driving behavior, resulting in deviation between the behavior strategy generated by the algorithm and the actual expectation of human drivers. Second, although some research introduces the running state information of surrounding vehicles as environmental input, it generally fails to predict their future behavior, which may lead to excessive aggressiveness or conservatism in the strategy execution process, reducing the safety and efficiency of the system. Finally, the objective function in the current reinforcement learning framework usually uses static weights to balance multiple objectives, lacking the ability to adaptively adjust according to real-time traffic scene changes, thereby limiting the adaptability and robustness of the model in dynamic and variable environments. SUMMARY

[0003] The technical problem to be solved by the present application is to overcome the existing defects and provide an assisting vehicle lane-changing decision method with human-like dynamic decision characteristics for heterogeneous traffic flow environment, which can effectively solve the problems in the background art.

[0004] In order to achieve the above purpose, the present application discloses an assisting vehicle lane-changing decision method with human-like dynamic decision characteristics for heterogeneous traffic flow environment, which adopts the technical solution of including the following steps:

[0005] Step 1, obtaining self-vehicle running state information, surrounding vehicle relative state information, traffic scene information, and environmental context information, and transmitting the above information to a multi-dimensional state space module by a traffic environment perception module;

[0006] Step 2, the multi-dimensional state space module converts the ego vehicle running state information output by the traffic environment perception module into ego vehicle characteristic encoding, converts the relative state information of surrounding vehicles into environment interaction encoding, converts the traffic scene information into traffic scene encoding, converts the environment context information into environment context encoding, and finally realizes multi-dimensional feature fusion through the encoder fusion mode, and outputs the fused result as a state vector to the dynamic fusion decision module;

[0007] Step 3, the dynamic fusion decision module makes a decision on the next driving state of the vehicle according to the state vector output result of the current vehicle, and relies on the evidence accumulation of the state vector output result of the current vehicle when making the decision. When the evidence accumulation reaches a threshold, the lane changing intention is determined, and the output result is output. The output result is the lane changing intention, the total reaction time and the comprehensive confidence.

[0008] As a preferred technical solution of the present application, the dynamic fusion decision module comprises a PPO-DDM decision sub-module and a weight adaptive mechanism sub-module. The PPO-DDM decision sub-module comprises a PPO decision sub-module and a DDM decision sub-module. The PPO decision sub-module outputs a lane changing decision result and an acceleration value based on PPO using an Actor network. The DDM decision sub-module outputs a lane changing decision result based on DDM. The output result of the PPO-DDM decision sub-module is input into the weight adaptive mechanism sub-module. The weight adaptive mechanism sub-module dynamically allocates the lane changing decision weight output by the PPO decision sub-module and the DDM decision sub-module according to the changes of weather, traffic density and risk level factors. In a low-risk, low-density scene, the weight of the PPO decision sub-module is increased. In a high-risk or severe weather, the weight of the DDM decision sub-module is increased. The final weight fusion decision result is obtained as the module output.

[0009] As a preferred technical solution of the present application, the calculation formula of the weight fusion decision result is:

[0010]

[0011] wherein, is the final comprehensive action probability distribution; is the action probability distribution output by the PPO decision sub-module; is the action probability distribution of the DDM decision sub-module; and are the weight coefficients of the PPO decision sub-module and the DDM decision sub-module, respectively;

[0012]

[0013]

[0014] wherein, is a default weight of DDM under basic conditions; represents a comprehensive risk score; is a traffic density normalized value, reflecting the degree of congestion;

[0015] The comprehensive risk score is composed of at least one of collision risk, road environment risk, and traffic rule / road structure risk, and is obtained by normalization processing a risk score in the range of [0, 1];

[0016] The comprehensive risk score is calculated as follows:

[0017]

[0018] wherein, represents clipping a numerical value to the interval [0, 1]; represents a collision risk index; represents a road environment risk index; represents a traffic rule / road structure risk index; is a weight coefficient, and satisfies .

[0019] As a preferred technical solution of the present application, the DDM decision-making sub-module includes a DDM model construction sub-module, an environment adaptive parameter adjustment mechanism sub-module, an evidence accumulation process numerical implementation sub-module, and a confidence and decision quality quantification sub-module. The results of the DDM model construction sub-module and the environment adaptive parameter adjustment mechanism sub-module are input into the evidence accumulation process numerical implementation sub-module to perform evidence accumulation, and finally obtain the lane changing intention and the total reaction time. The evidence accumulation result of the evidence accumulation process numerical implementation sub-module is input into the confidence and decision quality quantification sub-module, and the confidence and decision quality quantification sub-module outputs a comprehensive confidence.

[0020] As a preferred technical solution of the present application, the decision rule of the DDM model construction sub-module is:

[0021]

[0022] wherein, is an evidence accumulation value at time t, is a reaction time of decision-making, is a decision threshold value, representing the amount of evidence required to trigger decision-making;

[0023] The reaction time of decision-making includes the time when the evidence first reaches the threshold value and the perception delay, and the specific calculation formula is as follows: ​

[0024]

[0025] where, is the time when the evidence first reaches the threshold, is the non-decision time including perception delay, cognitive processing and action execution time.

[0026] The DDM model is:

[0027]

[0028] where, is the evidence accumulation value at time t, is the drift rate, indicating how fast and in which direction the evidence is accumulating towards a decision; is the random noise, used to represent the uncertainty of information;

[0029] Drift rate The calculation formula of is:

[0030]

[0031] where, represents the drift rate contribution of the th evidence source at time ; is the weight of the th evidence source; is the total number of evidence sources.

[0032] As a preferred technical solution of the present application, the environmental adaptive parameter adjustment mechanism submodule, on the basis of the DDM model construction submodule, calculates the dynamic threshold related to the model, including the dynamic drift rate and the adaptive decision threshold;

[0033] The dynamic drift rate calculation method is as follows:

[0034]

[0035] where, is the basic drift rate, represents the model basic drift rate bias term, is the weight coefficient, controlling the influence strength of PPO policy output on decision tendency; is the intention preference of PPO network to lane changing under the current state, which can be obtained from the softmax probability output by PPO;

[0036]

[0037] where, is the final drift rate; is a weather correction coefficient; is a lane-changing type correction coefficient;

[0038] The adaptive decision threshold calculation method is as follows:

[0039]

[0040] wherein, is a basic decision threshold; is a weather correction coefficient; is a lane-changing type coefficient.

[0041] As a preferred technical solution of the present application, the evidence includes PPO preference evidence, safety urgency evidence, and traffic efficiency evidence;

[0042] The PPO preference evidence calculation method is as follows:

[0043]

[0044] wherein, represents the probability of the PPO policy selecting the "lane-changing" action under the state . represents the probability of the PPO policy selecting the "lane-keeping" action under the state . is a weight coefficient of the PPO preference evidence;

[0045] The safety urgency evidence calculation method is as follows:

[0046]

[0047] wherein, is calculated by using a situational instantaneous risk assessment method, is a weight coefficient of the safety urgency evidence;

[0048] The traffic efficiency evidence calculation method is as follows:

[0049]

[0050] wherein, is a traffic efficiency index, which can be calculated by the difference of lane traffic flow density, is a weight coefficient of the traffic efficiency evidence;

[0051] The weight coefficient of the PPO preference evidence, the weight coefficient of the safety urgency evidence, and the weight coefficient of the traffic efficiency evidence satisfy .

[0052] As a preferred technical solution of the present application, the evidence accumulation adopts Euler-Markov approximation for discretization, comprising the following steps:

[0053] Step 31, initializing the accumulated evidence value and a time counter;

[0054] Step 32, in each time step, calculating an evidence increment composed of a drift term and a diffusion term;

[0055] The drift term is the product of the drift rate and the time step;

[0056] The diffusion term is the product of a Gaussian random number and a diffusion coefficient;

[0057] Step 33, superimposing the current accumulated evidence value and the evidence increment as a new accumulated evidence value, comparing the new accumulated evidence value with a decision threshold, when the evidence first exceeds the threshold, determining the final decision of lane changing or keeping the lane and the confidence degree thereof; if the threshold is not reached within a maximum time range, defaulting to determining the direction according to the evidence sign, and setting the confidence degree to 0.5.

[0058] As a preferred technical solution of the present application, in step 33, a boundary condition and a reflection boundary mechanism are introduced in the model implementation to ensure the numerical stability of the evidence accumulation process, when the absolute value of the accumulated evidence value approaches the decision threshold, it is considered that the decision accumulation value enters a soft trigger interval, and a soft boundary trigger probability is calculated :

[0059]

[0060] wherein, the absolute value of the accumulated evidence, the decision threshold of lane changing / keeping straight, the reflection boundary trigger sensitivity coefficient, the starting value of the soft trigger interval, .

[0061] According to the soft boundary trigger probability the drift rate is adaptively corrected, that is, an additional drift term in the same direction as is superimposed on the original drift rate , so as to accelerate the evidence to reach the threshold in the current decision direction, reduce the probability of long-time oscillation near the threshold, and improve the decision efficiency and stability;

[0062] wherein, is a sign indication function, used to indicate the direction of the current evidence accumulation value; when is positive, +1 is outputted to indicate the bias of lane changing, and when is negative, −1 is outputted to indicate the bias of keeping the lane.

[0063] As a preferred technical solution of the present application, the calculation method of the comprehensive confidence is as follows:

[0064]

[0065] Wherein, , is a weighting coefficient, satisfying + + =1;

[0066] is the evidence strength, and the calculation method thereof is as follows:

[0067]

[0068] Wherein, is the final accumulated evidence value, is a decision threshold; the closer the evidence strength is to 1, the more sufficient the evidence for lane changing or keeping is, and the higher the reliability is;

[0069] is the decision time consistency, and the calculation method thereof is as follows:

[0070]

[0071] Wherein, is the actual decision time, is an experience expectation value, which can be calculated by the median of historical decision time in the same scene (same weather, same speed interval, same lane changing type), is a time consistency sensitivity parameter; the closer the decision time consistency is to 1, the more the time sequence performance of the model conforms to the expectation;

[0072] is the drift rate stability, and the calculation method thereof is as follows:

[0073]

[0074] Wherein, and respectively represent the variance and mean of the drift rate, is a small constant to prevent the denominator from being zero; the closer the drift rate stability is to 1, the more smooth the drift rate process is, and the more reliable the decision is.

[0075] Compared with the prior art, the application has the beneficial effects that: by inputting the self-vehicle information, the relative situation of the surrounding vehicles, the traffic scene and other information into the decision module, and combining the evidence accumulation in the decision module, the decision module can predict the future behavior of the surrounding vehicles, not only can make the decision have the continuity characteristics of gradual adjustment, so as to be more in line with the behavior logic of human driving behavior, facilitate the accurate prediction of the subsequent behavior of the human driver, but also can adaptively adjust according to the real-time traffic scene change.

[0076] Further, by introducing the PPO-DDM decision sub-module, the PPO-DDM model provided by the PPO-DDM decision sub-module can autonomously switch between the PPO model and the DDM model according to the current traffic situation, and select aggressive driving or conservative driving according to the driving environment, so that the lane changing decision is more flexible and the auxiliary comfort is better.

[0077] Further, the application can effectively reduce the risk of traffic accidents and significantly improve driving safety. At the same time, the test results show that the PPO-DDM model exhibits faster convergence speed and stronger learning ability in the model training stage. In terms of model performance, the average score of decision consistency of the PPO-DDM model is improved from 7.2 of the PPO basic model to 9.1, with an increase of 26.4%. In the front vehicle deceleration and high-speed split scene, the average score is further improved to 9.4, indicating that the model has a significant advantage in dynamic decision consistency. This performance improvement is derived from the continuity modeling of DDM in lane changing behavior. In terms of response performance, the application improves the drift rate amplitude and enhances the evidence direction consistency through multi-source evidence fusion, and introduces a soft trigger reflex boundary mechanism to reduce oscillation and long tail delay near the threshold, so that the evidence accumulation reaches the decision threshold faster, thereby shortening the decision time and reducing the model response delay. The average response time is shortened from 0.37 seconds to 0.22 seconds, with a reduction of 40.5%, significantly enhancing the safety redundancy in emergency traffic situations. In terms of decision success rate, the average level is improved from 85.7% to 94.1%, achieving an effective balance between speed and accuracy, further verifying the advantages of the fusion framework in multi-scene driving decision-making. It can help enterprises save a lot of R&D and testing costs and accelerate the market landing of assisted driving products. BRIEF DESCRIPTION OF DRAWINGS

[0078] Figure 1 Fig. 1 is a schematic diagram of the PPO-DDM dynamic fusion lane changing decision model of the application;

[0079] Figure 2 Fig. 5 is a SUMO traffic simulation scene diagram of the application;

[0080] Figure 3 Fig. 9 is a model learning curve comparison diagram of the application;

[0081] Figure 4 This is a comparative analysis chart of the overall performance of the model of this invention;

[0082] Figure 5 This is a comparative analysis chart of the stability of the model of this invention. Detailed Implementation

[0083] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0084] Example 1

[0085] like Figure 1 As shown, this invention proposes a PPO-DDM dynamic fusion lane-changing decision model, including a traffic environment perception module, a multi-dimensional state space module, and a dynamic fusion decision module. The dynamic fusion decision module includes a PPO-DDM decision submodule and a weight adaptive mechanism submodule. The PPO-DDM decision submodule includes a PPO decision submodule and a DDM decision submodule. The DDM decision submodule includes a DDM model construction submodule, an environmental adaptive parameter adjustment mechanism submodule, a numerical implementation submodule of the evidence accumulation process, and a confidence and decision quality quantification submodule.

[0086] This invention also discloses a lane-changing decision-making method for assisted driving vehicles based on the above-mentioned PPO-DDM dynamic fusion lane-changing decision-making model, which is suitable for heterogeneous traffic flow environments and has human-like dynamic decision-making characteristics. The technical solution adopted includes the following steps:

[0087] Step 1: The traffic environment perception module acquires vehicle operating status information, relative status information of surrounding vehicles, traffic scene information, and environmental context information through onboard sensors and roadside equipment. The information from this module will serve as the input information for the multi-dimensional state space module.

[0088] Step 2: The multi-dimensional state space module converts the vehicle's operating state information output by the traffic environment perception module into vehicle characteristic codes, the relative state information of surrounding vehicles into environmental interaction codes, the traffic scene information into traffic scene codes, and the environmental context information into environmental context codes. Finally, multi-dimensional feature fusion is achieved through encoder fusion, and the fused result is used as a state vector input to the dynamic fusion decision module; specifically as follows:

[0089] The vehicle's operating status information includes: vehicle speed v, acceleration a, position (x, y), heading angle θ, and steering wheel angle. The output vehicle characteristic code is: ;

[0090] The relative state information of the surrounding vehicles includes the relative distance of key vehicles such as the front vehicle, the rear vehicle, the left front vehicle, the left rear vehicle, the right front vehicle, and the right rear vehicle , the relative speed , and the time to collision TTC; the output environment interaction code is ;

[0091] The traffic scene information includes the encoded scene type, the road structure, and the traffic density; the scene type (highway / urban / congestion) is encoded by One-hot, the road features (number of lanes, curvature, speed limit) are numerically encoded, and the statistical features (average speed, density) are encoded, and the output traffic scene code is

[0092] ;

[0093] The environmental context information includes the weather, the time of day, the visibility, and the road state (such as flat, potholes, muddy, icy, etc.); the output environmental context code is

[0094] ;

[0095] The encoder fusion result is as follows:

[0096] .

[0097] Step 3, the dynamic fusion decision module makes a decision on the next driving state of the vehicle according to the current state information of the vehicle, and outputs the result as a lane changing intention, a total reaction time, and a comprehensive confidence level;

[0098] The input of the PPO-DDM decision submodule is the state vector output by the multi-dimensional state space module, the PPO decision submodule in the PPO-DDM decision submodule outputs the lane changing decision result and the acceleration value based on PPO by using the Actor network, which is a prior art, and the specific method of the lane changing decision result and the acceleration value based on PPO will not be described here, and the DDM decision submodule of the PPO-DDM decision submodule outputs the lane changing decision result based on DDM;

[0099] The DDM model construction submodule of the DDM decision submodule abstracts the DDM model into a binary decision problem of determining "whether to change lanes", i.e., changing lanes or keeping straight; the environmental adaptive parameter adjustment mechanism submodule calculates the dynamic threshold values related to the model, including the dynamic drift rate and the adaptive decision threshold, based on the DDM model construction submodule;

[0100] The output of the DDM model construction submodule and the environmental adaptive parameter adjustment mechanism submodule is the input of the evidence accumulation process numerical implementation submodule, and the output of the evidence accumulation process numerical implementation submodule is the lane change decision, total decision time, and decision confidence, which are used as the input of the confidence and decision quality quantification submodule. The confidence and decision quality quantification submodule is used to obtain a comprehensive confidence by comprehensively quantifying the terminal state, time consistency, and drift rate volatility of the evidence accumulation process, so as to ensure the interpretability and actual availability of the model output.

[0101] Specifically, the decision rule of the DDM model construction submodule is defined as:

[0102]

[0103] wherein, is the evidence accumulation value at time t, t is the reaction time of the decision, and the initial value is usually set to ; is the decision threshold, which represents the amount of evidence required to trigger a decision , and is usually set to 1.05; z is the initial bias, which reflects the initial decision inclination, and is usually set to 0.

[0104] The reaction time of the decision is composed of two parts, including the time when the evidence first reaches the threshold and the perception delay, and the specific calculation formula is as follows:

[0105]

[0106] wherein, is the time when the evidence first reaches the threshold, is the non-decision time such as perception delay, cognitive processing, and action execution time.

[0107] In order to adapt to complex traffic decision-making environment, the DDM model is expanded to a multi-source evidence fusion form:

[0108]

[0109] wherein, is the evidence accumulation value at time t, is the drift rate, which represents the average direction and speed of evidence accumulation per unit time, is the random noise, which is used to represent the uncertainty of information.

[0110] The calculation formula of the drift rate is as follows:

[0111]

[0112] wherein, represents the drift rate contribution of the evidence source at time ; is the weight of the evidence source; is the total number of evidence sources;

[0113] In the environmental adaptive parameter adjustment mechanism submodule, the dynamic drift rate calculation method is as follows:

[0114]

[0115] wherein, is the basic drift rate; represents the model basic drift rate bias term, which is usually set to 0; is the weight coefficient, which controls the influence strength of the PPO policy output on the decision tendency, and the value range ; is the intention preference of the PPO network to lane changing under the current state, which can be obtained from the softmax probability output by the PPO network. The method for the PPO network to output the softmax probability is a prior art, which will not be described here.

[0116]

[0117] wherein, is the final drift rate; is the weather correction coefficient, and the value range is shown in Table 1; is the lane changing type correction coefficient, and the value range is shown in Table 2.

[0118] Table 1 Example Value Table

[0119]

[0120] Table 2: Example Value Table

[0121]

[0122] In the environmental adaptive parameter adjustment mechanism submodule, the adaptive decision threshold calculation method is as follows:

[0123]

[0124] wherein, is the basic decision threshold, which is consistent with the foregoing, and the value is 1.05; is the weather correction coefficient; is the lane changing type coefficient.

[0125] In the numerical implementation submodule of the evidence accumulation process, the evidence accumulation involves PPO preference evidence, safety urgency evidence, and traffic efficiency evidence, and the specific calculation methods are as follows:

[0126] The calculation method of PPO preference evidence is as follows:

[0127]

[0128] wherein, represents the probability of the PPO policy selecting the "lane changing" action under the state . represents the probability of the PPO policy selecting the "lane keeping" action under the state , which is obtained from the softmax probability output by the PPO network; is the weight coefficient of the PPO preference evidence for measuring its importance in decision-making, and the specific value of the weight coefficient can be adjusted according to the training scene requirements and training results.

[0129] The calculation method of safety urgency evidence is as follows:

[0130]

[0131] wherein, is calculated by using the scenario instantaneous risk assessment method, is the weight coefficient of the safety urgency evidence, indicating the relative influence of this item in decision-making, and the specific value can be adjusted according to the training scene requirements .

[0132] The calculation method of is a prior art, and for details, please refer to:

[0133] Guo Y J, Chen Y Y, Gu X, Guo J F, Zheng S Y, Zhou Y T. Dynamic traffic graph based risk assessment of multivehicle lane change interaction scenarios [J]. Physica A: Statistical Mechanics and its Applications, 2024, 643: 129791.

[0134] The calculation method of traffic efficiency evidence is as follows:

[0135]

[0136] wherein, Traffic efficiency indicator, which can be calculated by the difference of traffic flow density in the lane, is the prior art, which is not described here, Traffic efficiency evidence weight coefficient, indicating the proportion of this item in the decision trade-off, The specific value can be adjusted according to the training scene demand and the training result; wherein, For example .

[0137] The numerical implementation of the evidence accumulation process uses Euler-Markov approximation for discretization; first, initialize the accumulated evidence value and the time counter; in each time step, calculate the evidence increment composed of the drift term (the product of the drift rate and the time step) and the diffusion term (the product of the Gaussian random number and the diffusion coefficient); after updating the accumulated evidence at each step, compare it with the decision threshold value, when the evidence first exceeds the threshold value (positive or negative), the final decision of lane change or lane keeping and its confidence are determined; if the threshold value is not reached within the maximum time range, the direction is determined according to the sign of the evidence by default, and the confidence is set to 0.5.

[0138] The numerical implementation process of the above evidence accumulation process is as follows:

[0139] Step a, initialization: let ; let the maximum decision time window (preferably 2-4s);

[0140] Step b, at each output: lane change decision, total decision time and initial decision confidence. Time step Update the evidence increment:

[0141]

[0142] Step c, update , and judge:

[0143] · If , output Lane Change;

[0144] · If , output Keep Lane;

[0145] · Otherwise, continue to accumulate until ;

[0146] If the threshold value is not reached within , the direction is given according to the sign of (X output lane change, otherwise keep lane), and the initial confidence is set to 0.5.

[0147] Step d, after obtaining , calculate the total reaction time:

[0148]

[0149] Furthermore, boundary conditions and a reflection boundary mechanism are introduced into the model implementation to ensure the numerical stability of the evidence accumulation process. The boundary conditions define the binary decision output when the accumulated evidence first reaches the threshold, while the reflection boundary, when the evidence approaches the threshold range (soft trigger range, such as 0.8–1.0a), avoids long-term oscillation of the accumulated evidence by increasing the trigger probability or additional drift rate, thereby improving the model's decision efficiency and security. The implementation method is as follows:

[0150] Calculate the soft boundary trigger probability :

[0151]

[0152] in, Represents the absolute value of accumulated evidence. The decision threshold for changing lanes or staying straight. For the sensitivity coefficient of the reflection boundary trigger, the preferred value is... , This is the starting value of the soft-trigger interval. The value is 0.8.

[0153] When the absolute value of the accumulated evidence enters the soft trigger zone When calculating the soft boundary trigger probability. The system adaptively corrects the drift rate based on the stated probability, that is, at the original drift rate... Based on the superposition of Additional drift terms in the same direction accelerate the arrival of evidence towards the current decision direction at the threshold, reduce the probability of prolonged oscillations around the threshold, and improve decision-making efficiency and stability. Among these, This is a sign indicator function used to indicate the direction of the current evidence accumulation value; when When it is positive, output +1 (i.e. ) indicates a tendency to change lanes, when When it is negative, output -1 (i.e. This indicates a bias towards keeping the lane, thereby accelerating the amount of evidence. Exceeding the decision threshold .

[0154] The calculation method for the overall confidence score output by the confidence score and decision quality quantification submodule is as follows:

[0155]

[0156] in, , For the weighting coefficients, satisfying + + =1, e.g. , ;

[0157] is the evidence strength, used to measure the proportion of the final accumulated evidence relative to the decision threshold, representing the degree of "certainty" of the decision output, which is calculated as:

[0158]

[0159] where, is the final accumulated evidence value, is the decision threshold; the closer the evidence strength is to 1, the more sufficient the evidence for the lane change or keep decision is, and the higher the reliability is;

[0160] is the decision time consistency, used to depict the matching degree of the decision time of the model output and the expected value, reflecting the timing stability of the model in multiple simulations, which is calculated as:

[0161]

[0162] where, is the actual decision time, is the empirical expected value, which can be calculated by the median of the historical decision time in the same scenario (same weather, same speed interval, same lane change type), is the sensitivity parameter of the time consistency, the closer the decision time consistency is to 1, the more the timing performance of the model conforms to the expectation, and the recommended value is ;

[0163] is the drift rate stability, reflecting the fluctuation of the drift rate in the decision-making process, which is calculated as:

[0164]

[0165] where, and represent the variance and mean of the drift rate, respectively, is a small constant to prevent the denominator from being zero, and the recommended value is ; the closer the drift rate stability is to 1, the more smooth the drift rate process is, and the more reliable the decision is.

[0166] In step 3, the input of the weight adaptive mechanism submodule is the output result of the PPO-DDM decision submodule, which considers weather, traffic density, risk level factors, and dynamically allocates the lane change decision weight output by the PPO decision submodule and the DDM decision submodule according to the changes of the factors. In low-risk, low-density scenarios, PPO dominates the flexible and efficient decision-making. In high-risk or adverse weather, the DDM weight is increased to make the decision more biased towards the cautious lane change habit of human drivers. The final weight fusion decision result is obtained as the module output. The specific formula of the lane change decision weight allocation is as follows:

[0167]

[0168] wherein, is the fused action probability distribution; is the action probability distribution output by the PPO policy network (obtained by the Actor network output after softmax normalization); is the action probability distribution output by the DDM evidence accumulation mechanism (obtained by statistically analyzing the boundary reaching frequency through multiple random simulations, or obtained by monotonically mapping the normalized ratio of the evidence accumulation value and the threshold value); and are the weight coefficients of PPO and DDM respectively, and satisfy:

[0169]

[0170] To achieve self-adaptation to different traffic risks and congestion levels, the application preferably dynamically calculates according to the risk score and traffic density.

[0171]

[0172] wherein, is the default weight of DDM under basic conditions, preferably set according to weather conditions: 0.4 for sunny days and 0.6 for rainy days; in other weather conditions, it can also be set according to pre-set rules or table lookup. is the comprehensive risk score, with a value range of , used to represent the risk level of multiple factors such as collision risk and road risk. The higher the risk, the more inclined to increase the DDM weight to enhance the cautiousness and interpretability of the decision. is the traffic density normalized value, with a value range of , used to represent the degree of road congestion. The higher the congestion, the more inclined to increase the DDM weight to improve the stability of the decision.

[0173] The comprehensive risk score is composed of at least one of collision risk, road environment risk, and traffic rule / road structure risk, and is obtained by normalization processing within a range. is calculated as follows:

[0174]

[0175] wherein, is clipped to the interval ; represents a collision risk index; represents a road environment risk index; represents a traffic rule / road structure risk index; , , is a weight coefficient, and satisfies , the weight can be taken to reflect the collision risk priority principle.

[0176] 1) Calculation of collision risk index (preferably based on TTC)

[0177] Collision risk is obtained by calculating the minimum collision time margin (TTC):

[0178]

[0179] wherein, is a target set that has a potential conflict with the ego vehicle (may include the front and rear vehicles of the current lane and target lane, etc.); is the relative distance between the ego vehicle and the target vehicle ; is the closing speed; is a small constant to prevent division by zero. Based on the threshold , it is normalized to obtain:

[0180]

[0181] wherein, is clipped to the interval ; is a collision risk threshold, preferably 4s in high-speed scenarios and 3s in urban low-speed scenarios.

[0182] 2) Calculation of road environment risk index

[0183]

[0184] wherein, is clipped to the interval ;​ μ(t) represents the road surface adhesion coefficient of the road segment where the vehicle is currently driving at time t; μdry represents the dry road surface reference adhesion coefficient (e.g. 0.8), μmin represents the low adhesion lower bound (e.g. 0.3).

[0185] 3) Traffic rule / road structure risk indicator Calculation (based on road rules and map constraints)

[0186] The traffic rule / road structure risk indicator is used to characterize the degree of limitation of lane change feasibility, which can be defined based on high-definition map or lane line recognition results:

[0187] ;

[0188] Finally, the lane change intention, total reaction time and comprehensive confidence are output;

[0189] Step 4, the input of the vehicle lane change control system is the lane change intention, total reaction time and comprehensive confidence output by the dynamic fusion decision module, the control system sets up a behavior state machine, when the decision is lane change and the comprehensive confidence exceeds the preset threshold and passes the safety executability gate, it enters the lane change preparation state; in the preparation state, the system maintains the lane and adjusts the longitudinal speed to meet the target lane gap requirement, when the corresponding execution trigger time is reached and the gate condition is still met, the system switches to the lane change execution state and calls the trajectory planning module to generate a lane change trajectory that meets the vehicle dynamics constraints; the bottom steering and longitudinal controller tracks the trajectory to complete the lane change. If the risk indicator deteriorates or the confidence drops below the threshold during execution, the abort logic is triggered to return to the keep lane state, thereby realizing explainable lane change timing control and safe and robust execution. This step is prior art and will not be described in detail.

[0190] Model training

[0191] The PPO-DDM dynamic fusion lane change decision model proposed in Embodiment 1 is trained using the public aerial trajectory dataset UCF-SST,

[0192] (I) Training data preprocessing

[0193] The preprocessing goal is to unify the coordinates, extract the features, construct the labels (lane change events / decision times), and perform normalization and slicing.

[0194] (1) Coordinate and lane representation

[0195] Convert global coordinates to lane coordinates (Frenet or "along the lane center line s / d"):

[0196] : along the road direction

[0197] : lateral offset (0 for lane centerline)

[0198] Lane annotation / assignment: based on Determine current lane, target lane with lane boundaries.

[0199] (2) Lane change event and decision time label

[0200] Lane change start: lateral velocity Crosses threshold (e.g., 0.2 m / s) or lateral offset crosses lane line;

[0201] Lane change end: vehicle center enters target lane and lateral velocity falls back;

[0202] Decision time label (suggested to be defined as the time difference from "first time when a lane change opportunity appears to the start of execution"):

[0203] Lane change opportunity: first time when target lane meets minimum gap / rules allow;

[0204] Start of execution: First time when threshold is crossed.

[0205] (3) State features

[0206] Composition

[0207] Suggested to include at least (and normalize):

[0208] ego: speed etc.

[0209] target lane front / rear: relative distance , relative speed , time to collision TTC / headway THW;

[0210] lane level: current lane and target lane density, average speed;

[0211] rules / map: whether lane change is allowed, distance to exit, whether under construction / lane end;

[0212] weather: visibility, rainfall intensity or weather level (used for ).

[0213] (4) Calculation of risk and density indicators

[0214] : normalized to [0, 1] by or predicted collision probability;

[0215] : with window length : number of vehicles inside the vehicle or speed type congestion index normalized to [0,1].

[0216] : suggest to do for all continuous features:

[0217] : clipping (de-extreme, e.g., P1–P99)

[0218] : z-score or min-max normalization

[0219] : missing value filling (0 or last observation carry)

[0220] : (5) data slicing and partitioning

[0221] : sequence window: e.g., cut each trajectory into pieces of length 5–10 s;

[0222] : train / validation / test partitioned by “scenario / link” (avoid same link leakage).

[0223] : (2) training environment

[0224] : (1) hardware environment

[0225] : CPU: 16–64-core x86 (e.g., AMD EPYC / Intel Xeon)

[0226] : GPU: NVIDIA RTX 4090 (24GB) or A100 (40 / 80GB) ×1–4

[0227] : memory: 64–256GB

[0228] : storage: NVMe SSD ≥ 2TB

[0229] : (2) software environment

[0230] : OS: Ubuntu 22.04 LTS

[0231] : Python: 3.10

[0232] : deep learning: PyTorch 2.x + CUDA 12.x

[0233] : RL library: Stable-Baselines3 / RLlib (either one)

[0234] : simulation: SUMO (with TraCI) or CARLA

[0235] Training interface: Gymnasium / OpenAI Gym style environment wrapper

[0236] Logging: TensorBoard / Weights & Biases (optional)

[0237] (III) Training process

[0238] Training of PPO-DDM dynamic fusion lane-changing decision model (in simulation environment)

[0239] (1) Training objectives and rewards

[0240] Combined reward (example):

[0241]

[0242] Safety : strong penalty for collision, low TTC penalty

[0243] Efficiency : target speed tracking, traffic efficiency

[0244] Comfort : lateral / longitudinal acceleration and jerk penalty

[0245] Rules : solid line / no entry / out-of-boundary penalty

[0246] (2) PPO core hyperparameters

[0247] Optimizer: AdamW

[0248] Initial learning rate (actor / critic): 3e-4

[0249] Weight decay: 1e-4

[0250] Discount factor γ: 0.99

[0251] GAE λ: 0.95

[0252] Clip range: 0.2

[0253] Entropy coef: 0.01

[0254] Value loss coef: 0.5

[0255] Max grad norm: 0.5

[0256] Rollout length (number of sampling steps): 2048

[0257] batch size (total samples per update): 65536

[0258] mini-batch size: 2048

[0259] epochs per update: 10

[0260] gradient accumulation steps: 4 (enabled when GPU is insufficient)

[0261] total training steps (total timesteps): 3e7 - 1e8 (commonly used: 5e7)

[0262] convergence strategy (early stop): validation set average return, stop or reduce learning rate if consecutive N=10 evaluations do not improve

[0263] (3) Learning rate decay strategy

[0264] warmup: linear from 1e-4 -> 3e-4 for the first 1e5 steps

[0265] decay: cosine decay to 3e-5

[0266] decay period: once throughout the training steps (or restart every 1e7 steps).

[0267] Simulation verification

[0268] In order to verify the effectiveness of the hierarchical decision-making architecture and multi-scene adaptive reward function of the proposed PPO-DDM dynamic fusion lane-changing decision-making model, an implementation and verification platform for simulation based on the integration of SUMO (Simulation of Urban Mobility) traffic simulator and reinforcement learning is constructed. This platform can simulate real road traffic environment with high fidelity, and flexibly support modeling and dynamic testing of various driving scenarios, providing a solid technical foundation for the development and evaluation of intelligent driving algorithms.

[0269] (I) Construction of SUMO traffic simulation scene

[0270] In order to verify the performance of the model in a relatively real environment, the data set scenarios used for model training are modeled in SUMO in a geometric progression, and four typical highway driving scenarios are constructed: front vehicle deceleration scenario, lane-changing insertion scenario, merging scenario and splitting scenario. Different types of background vehicles are introduced in each scenario to simulate diversified driving behaviors in real traffic environment and increase the strategic challenges faced by the agent. The simulation scene diagram is shown in Figure 2 .

[0271] (1) Front vehicle deceleration scenario

[0272] This scenario simulates a common traffic situation where a front vehicle suddenly decelerates. From step 5 of the simulation, the front vehicle uniformly decelerates from 16 m / s to 10 m / s within 30 steps. The deceleration uses a smooth, non-linear curve with an approximate deceleration of -0.2 m / s². The assisted driving vehicle follows the front vehicle in the same lane at a slightly lower speed (15 m / s) and needs to adjust lane changing and acceleration flexibly during the front vehicle's deceleration to ensure driving safety and efficiency. Background vehicles are generated in the adjacent lanes every 20 steps, with driving styles randomly selected, including normal (60%), cautious (30%), and aggressive (10%). Different types of vehicles have differences in speed (14-18 m / s), minimum safe distance (1.5-3.0 m), and other aspects, forming a diverse traffic flow.

[0273] (2) Lane-changing insertion scenario

[0274] This scenario is mainly used to test the response of the agent to the lane-changing behavior of vehicles in adjacent lanes and test its perception and decision-making performance in lateral interaction scenarios. The front vehicle randomly starts in lane 1 or 3 (adjacent lanes), and the rear vehicle (assisted driving vehicle) starts in lane 2 (middle lane). The front vehicle performs a lane-changing operation within 5 seconds at step 1 of the simulation if the distance to the rear vehicle is more than 10 meters, inserting into lane 2. This disrupts the agent's longitudinal following rhythm, forcing it to re-evaluate safety and execute the corresponding driving strategy. New vehicles are randomly generated in lanes 1 and 3 every 20 steps, with a speed distribution range of 14-18 m / s, and a vehicle distance of 2-3 seconds to make the simulation scenario closer to the real traffic environment.

[0275] (3) Merging scenario

[0276] The merging scenario is mainly used to test the merging lane-changing strategy of the assisted driving vehicle. The initial speed of the vehicle on the ramp is 10 m / s, and it is positioned at the 70% position of the ramp end, simulating a real scenario of entering the main road. From the "approaching merging point" (distance from the merging area <10 meters) to the "merging in progress" and then to the "merging completed" stage, the strategy adapts to the merging decision in stages. Every step generates a new main road vehicle of different types with a probability of 10%, and the density is controlled at 1200-1800 vehicles / hour / lane to ensure the challenge and diversity during the merging process.

[0277] (4) Splitting scenario

[0278] The exit split scenario examines the agent's ability to prepare for and safely exit the highway. The assisted driving vehicle starts in lane 1 (the middle lane), and the simulation triggers exit guidance at step 10. The goal is to exit in lane 0 (the rightmost lane), simulating the warning signs and split requirements of a real highway exit area. Approximately 70% of vehicles eventually exit the exit, and the assisted driving vehicle needs to change lanes and slow down in advance. Some vehicles travel in exit lane 0 but continue to maintain high speed, creating an "exit lane occupancy" scenario that increases the complexity of the agent's lane changing and deceleration decisions.

[0279] (B) Stage training method

[0280] To ensure that the assisted driving vehicle gradually adapts to different risks and complexities, this embodiment introduces a four-stage curriculum learning mechanism: (1) The first stage only includes front vehicle deceleration and merging scenarios, with low background vehicle density and large safety distance; (2) The second stage adds lane changing scenarios, challenging the assisted driving vehicle with more lateral interaction challenges; (3) The third stage introduces exit split scenarios to test lane selection and safe lane changing capabilities; (4) The fourth stage superimposes multiple mixed scenarios, adverse weather (rain), and high-density traffic to simulate complex real-world environments. The curriculum difficulty is dynamically adjusted through background vehicle behavior characteristics (minimum distance, speed distribution, etc.) and traffic density parameters (0.3-0.8), forming a multi-level and diversified verification system. In addition to traffic scenarios, the platform supports random weather conditions (such as rain and sunshine), with a maximum rain probability of 30%. In rain scenarios, the agent needs to dynamically adjust the following distance, safety threshold, and lane changing intention to further enhance its generalization ability and robustness. To improve learning stability, the platform has built-in state normalization and reward normalization modules to reduce oscillation and numerical bias during training, ensuring model robustness and convergence.

[0281] (Three) Simulation interface

[0282] The simulation platform built in this embodiment strictly follows the Gymnasium standard, and the core interface of the simulation environment includes the following parts:

[0283] (1) State space: contains a 12-dimensional vector covering vehicle speed, surrounding traffic, lane changing safety, risk, current lane position, scenario progress, and weather information;

[0284] (2) Action space: a two-dimensional vector corresponding to acceleration and lane changing decision;

[0285] (3) Reward function: integrates multi-scenario adaptive rewards to respond in real time to the safety, efficiency, comfort, and social perception factors of the vehicle.

[0286] (Four) Comparative analysis

[0287] The embodiment compares and analyzes the PPO-DDM dynamic fusion lane changing decision model (referred to as PPO-DDM model) and the PPO reinforcement learning model (referred to as the basic model) in multiple dimensions. The comparative experiment covers 10,000 training episodes, and comprehensively evaluates the performance of the model in learning speed, performance stability, environmental adaptability and other key indicators.

[0288] (1) Comparison of learning ability and convergence characteristics

[0289] As shown in Figure 3 , the PPO-DDM model shows relatively good initial performance (4.33) at the beginning of training (Episode 100), while the basic model starts from a negative value (-454.91), and the initial performance gap between the two is as high as 459.24 points. This indicates that the PPO-DDM model has a significant advantage in parameter initialization and prior knowledge integration. The PPO-DDM model achieved a 37.44-point improvement from Episode 100 to Episode 1000, while the basic model improved by 282.35 points during the same period, although the absolute value is larger, but from the relative improvement rate, the learning efficiency of the PPO-DDM model is higher. The PPO-DDM model only needs 52 episodes to achieve a continuous 3 positive scores, showing extremely fast positive learning ability, while the basic model needs 957 episodes, with a difference of 18 times. This huge difference indicates that the PPO-DDM model has outstanding advantages in rapid environmental adaptation. The PPO-DDM model reached a maximum performance of 202.89 at Episode 10000, compared with the basic model of 50.53, with an advantage of 152.36 points (about 4 times the performance gap), fully proving the long-term learning advantage of the PPO-DDM model.

[0290] (2) Comprehensive evaluation of model performance

[0291] The embodiment evaluates the comprehensive performance level of the two models through four dimensions of average Score, maximum Score, Score standard deviation and positive Score ratio (as shown in Figure 4The average Score of the PPO-DDM model is 124.56, which is 60.46 points higher than the 64.10 of the base model, with a 94.3% improvement. This significant gap indicates that the PPO-DDM architecture has a significant advantage in overall performance stability. As a key indicator of measuring daily performance, the advantage of the PPO-DDM model means that it can provide more reliable and consistent performance output in actual deployment. Second, the maximum Score of the PPO-DDM model reaches 421.04, surpassing the 302.72 of the base model by 118.32 points (39.0% performance improvement). This result shows that the PPO-DDM fusion architecture not only performs well in average performance, but also has strong potential in peak performance mining. The advantage of the maximum Score proves that the fusion architecture has strong ability to handle complex scenarios and extreme cases. Moreover, the Score standard deviation of the PPO-DDM model is 90.46, which is significantly lower than the 150.55 of the base model, with a difference of 60.09 points (39.9% stability improvement). A lower standard deviation indicates that the PPO-DDM model has smaller performance fluctuations under different test scenarios and higher reliability. Finally, the proportion of positive Scores of the PPO-DDM model reaches 65%, which is 10 percentage points higher than the 55% of the base model. This indicator shows that the PPO-DDM model has stronger consistency in maintaining positive performance, reflecting its superior learning effect and environmental adaptability.

[0292] (3) Training stability and volatility characteristics

[0293] The PPO-DDM model exhibits significant stability and robustness in reinforcement learning training. As shown in Figure 5 , the average value of the PPO-DDM model is -0.15, which is significantly smaller than the -0.84 of the base model, indicating the gentleness of policy updates and stronger convergence. Its standard deviation is reduced from 4.31 to 1.16, with a reduction of 73.1%, effectively avoiding fluctuations in training and improving learning consistency and predictability. In terms of value network learning, the average value of the PPO-DDM model is reduced from 2.07 of the base model to 0.54, with an error reduction of 73.9%, indicating its ability to accurately estimate state values; The standard deviation of the value also decreased from 1.06 to 0.42, a 60.4% reduction in volatility, further enhancing the stability of value learning and providing more reliable gradient signals for the policy network. Overall, the PPO-DDM model achieved a dual reduction in the mean loss and volatility in both policy and value dimensions, forming a stable and efficient Actor-Critic training cycle. This coordinated feature not only accelerates the convergence speed of the model, reduces the time and computational resources required for training, but also improves the reliability and repeatability of model training, significantly optimizing the instability and performance degradation risks of the base model during the training process.

[0294] (4) Comparative analysis of model effectiveness

[0295] The overall performance indicators of the PPO-DDM model obtained after training and the base model were compared and analyzed comprehensively in four typical traffic scenarios. The test scenarios include front vehicle deceleration, lane insertion, merging and splitting, and 1000 simulation experiments were conducted for each scenario. Through multi-dimensional index evaluation, the PPO-DDM model showed significant advantages in safety, response speed, decision quality, etc.

[0296] 1. Comparative analysis of overall performance indicators

[0297] Table 3 shows the performance improvement of the PPO-DDM model compared to the base model from four dimensions. The table uses both absolute and relative improvement indicators to present a more comprehensive comparison perspective. The average decision consistency score improved from 7.2 for the base model to 9.1 for the PPO-DDM model, an absolute improvement of 1.9 points and a relative improvement of 26.4%, highlighting the significant advantage of the PPO-DDM model in decision predictability and consistency. The average response time was shortened from 0.37 seconds to 0.22 seconds, an absolute reduction of 0.15 seconds and an improvement of 40.5%. In assisted driving applications, this 0.15-second difference corresponds to a braking distance of about 2.5 meters at 60 km / h, significantly improving safety in emergency situations. The average success rate improved from 85.7% to 94.1%, an absolute improvement of 8.4%, indicating that while speeding up the response, the PPO-DDM model still guarantees a high level of decision success rate, achieving a balance between fast response and high-quality decision-making.

[0298] Table 3 Comparison of overall performance indicators

[0299]

[0300] Note: The relative improvement is positive, indicating performance improvement. Response time is negative, indicating time reduction (performance improvement). The average decision consistency is based on the arithmetic mean of the four scenarios. All indicators are statistically significant.

[0301] 2. Comparison and analysis of performance indicators of each scenario

[0302] ① Front vehicle deceleration scenario

[0303] In this scenario, the number of conflict scenarios of the PPO-DDM model is only 0.56, which is reduced by 34% compared to the basic model, significantly reducing the probability of rear-end accidents. At the same time, the reduction of the average safety distance of 2.8 meters and the decision delay avoids the influence of overly conservative driving strategies on traffic efficiency under the premise of ensuring safety. The speed adjustment amplitude of 1.2 m / s reflects a more smooth and gentle driving style (as shown in Table 4).

[0304] Table 4 Comparison of performance indicators of front vehicle deceleration scenario

[0305] ②

[0306] ③ Insertion lane change scenario

[0307] In this scenario, the PPO-DDM model achieved a lane change success rate of 93.5%, showing good gap judgment and lane change timing ability. Most importantly, the number of conflicts decreased from 1.8 to 0.7, a reduction of 61%, significantly reducing the risk of side impact. The average lane change time of 3.2 seconds ensures the efficiency of the operation while ensuring the stability of the lane change process. The minimum safety distance of 1.8 meters further illustrates the comprehensiveness and accuracy of the system in safety evaluation (as shown in Table 5).

[0308] Table 5 Comparison of performance indicators of insertion lane change scenario

[0309] ④

[0310] ⑤ Merge scenario

[0311] The success rate of the PPO-DDM model in the merge scenario reached 96.2%, significantly better than the basic model of 92.5%, and the merge time was reduced by 12.7%, improving the traffic efficiency. At the same time, the gap acceptance rate of 87.6% reflects the model's good balance between risk control and efficiency, avoiding both excessive conservatism and overly aggressive merge operations. Notably, the number of conflict scenarios decreased from 3.2 to 2.6, a reduction of 23.1%, indicating that the merge behavior significantly reduces the disturbance to the main road traffic flow, which is conducive to improving the overall traffic capacity and safety of the weaving area (as shown in Table 6).

[0312] Table 6 Comparison of performance indicators of merge lane change scenario

[0313] ⑥

[0314] ⑦Splitting scenarios

[0315] The PPO-DDM model shows good prediction ability in the exit scenario, with an early lane-changing distance of 25 meters, an increase of 21% over the base model, effectively avoiding safety risks caused by temporary and urgent lane changes. The exit deceleration smoothness score is as high as 8.9, showing that the algorithm can plan for deceleration in advance when approaching the exit, further optimizing ride comfort. In addition, the lane occupancy time of 1.8 seconds and the rear vehicle influence index of 1.2 seconds are significantly lower than the base model, reflecting the better performance of the model in terms of operational efficiency and human-machine interaction compatibility (as shown in Table 7).

[0316] Table 7 Comparison of performance indicators in splitting lane-changing scenarios

[0317]

[0318] Through the above verification and test, compared with the base model, the PPO-DDM model can significantly improve the safety and continuity of decision-making in complex scenarios such as multi-vehicle interaction, traffic flow density fluctuation and driving intention uncertainty, and reduce the frequency of high-risk lane-changing behavior.

[0319] Although embodiments of the present application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and alterations can be made without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A lane-changing decision-making method for assisted driving vehicles with human-like dynamic decision-making characteristics, characterized in that, Includes the following steps: Step 1: Obtain the vehicle's operating status information, the relative status information of surrounding vehicles, traffic scene information, and environmental context information. The traffic environment perception module then transmits this information to the multi-dimensional state space module. Step 2: The multi-dimensional state space module converts the vehicle operation status information output by the traffic environment perception module into vehicle characteristic codes, converts the relative status information of surrounding vehicles into environmental interaction codes, converts traffic scene information into traffic scene codes, converts environmental context information into environmental context codes, and finally realizes multi-dimensional feature fusion through encoder fusion, and outputs the fused result as a state vector and passes it to the dynamic fusion decision module. Step 3: The dynamic fusion decision module makes a decision on the next driving state of the vehicle based on the current state vector output result. The decision depends on the accumulation of evidence from the current state vector output result. When the evidence accumulates to a threshold, the lane change intention is determined and the result is output. The output result is the lane change intention, the total reaction time, and the overall confidence level.

2. The lane-changing decision-making method for assisted driving vehicles with human-like dynamic decision-making characteristics in heterogeneous traffic flow environments, as described in claim 1, is characterized in that: The dynamic fusion decision module includes a PPO-DDM decision submodule and a weight adaptive mechanism submodule. The PPO-DDM decision submodule includes a PPO decision submodule and a DDM decision submodule. The PPO decision submodule uses an Actor network to output lane-changing decision results and acceleration values ​​based on PPO. The DDM decision submodule outputs lane-changing decision results based on DDM. The output results of the PPO-DDM decision submodule are input to the weight adaptive mechanism submodule. The weight adaptive mechanism submodule dynamically allocates the lane-changing decision weights output by the PPO decision submodule and the DDM decision submodule according to changes in weather, traffic density, and risk level factors. In low-risk, low-density scenarios, the weight of the PPO decision submodule is increased, and in high-risk or severe weather conditions, the weight of the DDM decision submodule is increased, obtaining the final weighted fusion decision result as the module output.

3. The lane-changing decision-making method for assisted driving vehicles with human-like dynamic decision-making characteristics in heterogeneous traffic flow environments, as described in claim 2, is characterized in that... The formula for calculating the weighted fusion decision result is as follows: ， in, This represents the final, synthesized probability distribution of actions. The action probability distribution output by the PPO decision submodule; The action probability distribution of the DDM decision submodule; and These are the weight coefficients for the PPO decision submodule and the DDM decision submodule, respectively. ， ， in, These are the default weights for DDM under the basic conditions; This represents the overall risk score; This is a normalized value for traffic density, reflecting the degree of congestion.

4. The lane-changing decision-making method for assisted driving vehicles with human-like dynamic decision-making characteristics in heterogeneous traffic flow environments, as described in claim 1, is characterized in that: The DDM decision-making submodule includes a DDM model construction submodule, an environmental adaptive parameter adjustment mechanism submodule, an evidence accumulation process numerical implementation submodule, and a confidence level and decision quality quantification submodule. The results of the DDM model construction submodule and the environmental adaptive parameter adjustment mechanism submodule are input to the evidence accumulation process numerical implementation submodule for evidence accumulation, ultimately obtaining the lane-changing intention and total reaction time. The evidence accumulation results of the evidence accumulation process numerical implementation submodule are input to the confidence level and decision quality quantification submodule, which outputs the comprehensive confidence level.

5. The lane-changing decision-making method for assisted driving vehicles with human-like dynamic decision-making characteristics in heterogeneous traffic flow environments, as described in claim 4, is characterized in that... The decision rules for the DDM model construction submodule are as follows: ， in, The accumulated evidence value at time n. For decision-making reaction time, The decision threshold represents the amount of evidence required to trigger a decision. Decision reaction time This includes the time it takes for evidence to first reach the threshold and the perception delay, calculated using the following formula: ， in, The time when the evidence first reaches the threshold. This is a non-decision-making time; The DDM model is: ， in, The accumulated evidence value at time n. For drift rate, It is random noise; Drift rate The calculation formula is: ， in, Indicates the first One source of evidence at any time The contribution of drift rate; For the first The weight of each source of evidence; This represents the total number of evidence sources.

6. The lane-changing decision-making method for assisted driving vehicles with human-like dynamic decision-making characteristics in heterogeneous traffic flow environments, as described in claim 5, is characterized in that: The environmental adaptive parameter adjustment mechanism submodule, based on the DDM model construction submodule, calculates model-related dynamic thresholds, including dynamic drift rate and adaptive decision threshold. The dynamic drift rate is calculated as follows: ， in, Based on the drift rate, This represents the model's basic drift rate bias term. These are the weighting coefficients; This represents the PPO network's intention preference for lane changing in the current state. ； in, This is the final drift rate; This is a weather correction factor; This is a correction factor for lane change type; The adaptive decision threshold calculation method is as follows: ， in, Basic decision threshold; This is a weather correction factor; This is the lane-changing type coefficient.

7. The lane-changing decision-making method for assisted driving vehicles with human-like dynamic decision-making characteristics in heterogeneous traffic flow environments, as described in any one of claims 6, is characterized in that: The evidence includes evidence of PPO preference, evidence of safety urgency, and evidence of traffic efficiency; The method for calculating the PPO preference evidence is as follows: ， in, Indicates the state The probability of the PPO strategy choosing the "lane change" action; Indicates the state Under these conditions, the probability of the PPO strategy selecting the "lane keeping" action; The weighting coefficients for PPO preference evidence; The method for calculating the evidence of security urgency is as follows: ， in, The scenario-based instantaneous risk assessment method is used for calculation. Weighting coefficients for evidence of security urgency; The method for calculating the traffic efficiency evidence is as follows: ， in, As a traffic efficiency indicator, it can be calculated based on the difference in traffic flow density between lanes. Weighting coefficients for evidence of traffic efficiency; The weighting coefficients of the PPO preference evidence, the weighting coefficients of the safety urgency evidence, and the weighting coefficients of the traffic efficiency evidence satisfy the following: .

8. The lane-changing decision-making method for assisted driving vehicles with human-like dynamic decision-making characteristics in heterogeneous traffic flow environments, as described in claim 7, is characterized in that: The evidence accumulation is discretized using the Euler-Markov approximation, including the following steps: Step 31: Initialize the accumulated evidence value and time counter; Step 32: Within each time step, calculate the evidence increment consisting of the drift term and the diffusion term; The drift term is the product of the drift rate and the time step. The diffusion term is the product of a Gaussian random number and a diffusion coefficient; Step 33: Overlay the current cumulative evidence value and the evidence increment as a new cumulative evidence value. Compare the new cumulative evidence value with the decision threshold. When the evidence exceeds the threshold for the first time, determine the final decision of lane changing or lane keeping and its confidence level. If the threshold is not reached within the maximum time range, the direction is determined by default based on the evidence symbol, and the confidence level is set to 0.

5.

9. The lane-changing decision-making method for assisted driving vehicles with human-like dynamic decision-making characteristics in heterogeneous traffic flow environments, as described in claim 8, is characterized in that: In step 33, when the absolute value of the accumulated evidence value approaches the decision threshold, the accumulated decision value is considered to have entered the soft triggering range, and the soft boundary triggering probability is calculated. : ， in, Represents the absolute value of accumulated evidence. The decision threshold for changing lanes or staying straight. The sensitivity coefficient for triggering reflection boundary. This is the starting value of the soft-trigger interval. ; Based on the soft boundary trigger probability Adaptive correction is applied to the drift rate, i.e., based on the original drift rate... Based on the superposition and Additional drift terms in the same direction accelerate the arrival of evidence toward the current decision direction at the threshold, reduce the probability of prolonged oscillations around the threshold, and improve decision efficiency and stability. in, This is a sign indicator function used to indicate the direction of the current evidence accumulation value; when A timeout of +1 indicates a bias towards lane changing. When the value is negative, the output is −1, indicating bias to maintain lane position.

10. The lane-changing decision-making method for assisted driving vehicles with human-like dynamic decision-making characteristics in heterogeneous traffic flow environments, as described in claim 4, is characterized in that: The method for calculating the overall confidence level is as follows: ， in, , For the weighting coefficients, satisfying + + =1; The strength of evidence is calculated as follows: ， in, For the final accumulated value of evidence, This is the decision threshold; To ensure consistency in decision-making time, the calculation method is as follows: ， in, For the actual decision-making time, For the expected value based on experience, This is a sensitivity parameter for time consistency; For drift rate stability, the calculation method is as follows: ， in, and Let Variance and mean of the drift rate be represented, respectively. To prevent small constants with a denominator of zero.

Images