A Regionalized Decision-Making Method for Human-Machine Shared Control Based on Gaussian Hidden Markov Model

Abstract

This invention discloses a human-machine shared control regional decision-making method based on a Gaussian Hidden Markov Model (HMM), comprising the following steps: Step 1, information collection; Step 2, information processing; Step 3, level judgment; Step 4, preliminary allocation; Step 5, final allocation; Step 6, real-time update. In Step 1, the information collection module collects real-time driving information from the control panel, sensing devices, and the driver assistance system module, as well as real-time driving information from the driver assistance system. This invention comprehensively considers the driver's different coping abilities and sensitivities to environmental risks from different directions, fully utilizing the double-chain structure of the HMM to comprehensively consider driver specificity, relative driving ability, vehicle environmental risk level, and the surrounding area, ultimately achieving adaptive automatic adjustment of the driver's control switching strategy, making driving safer and more comfortable.

Description

Technical Field

[0001] This invention relates to the fields of autonomous driving and human-machine co-driving technology, specifically to a regionalized decision-making method for human-machine shared control based on Gaussian Hidden Markov Model. Background Technology

[0002] With the continuous evolution of autonomous driving technology, driver assistance systems have become standard equipment in intelligent vehicles. However, in open road scenarios, shared control driving systems, represented by human-machine co-driving, still have safety and reliability issues. Therefore, in the "dual-driver" environment of human-machine co-driving, the highly reliable and reasonable allocation of driving rights is crucial to the development of intelligent vehicles and driving safety.

[0003] Current mainstream research focuses on comprehensively considering the driver's state and using the deviation between the driver's decisions or intentions and the intelligent model's decisions as the main basis for allocating servo-level shared control. This process only considers the driver's state or intentions and the vehicle's state to allocate driving control, without taking into account the driver's specificity (specificity refers to the ability of drivers with different driving skills to make different corrections to the same driving risk). This driving control allocation strategy, which only assesses the driver's state, is prone to the drawbacks of over-intervention or under-intervention, weakening the driver's own driving ability and causing driver resentment, reducing the comfort of human-machine interaction, and violating the original intention of human-machine co-driving research. Secondly, existing research on servo-level shared control problems considers the surrounding environmental risks too generally and lacks detail. In reality, drivers' awareness and response capabilities to risks ahead and behind are different. Therefore, it is unreasonable to only consider the total environmental risk in a general way without distinguishing between them. Summary of the Invention

[0004] The purpose of this invention is to provide a human-machine shared control regional decision-making method based on Gaussian Hidden Markov Model to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a human-machine shared control regional decision-making method based on a Gaussian hidden Markov model, comprising the following steps: Step 1, information collection; Step 2, information processing; Step 3, level judgment; Step 4, preliminary allocation; Step 5, final allocation; Step 6, real-time update;

[0006] In step one above, the information collection module collects real-time driving information of the driver and the real-time driving information of the driver assistance system from the control panel, sensing devices and driver assistance system module.

[0007] In step two above, the driver's real-time driving information is transmitted to the information processing module through the driver driving information unit of the information collection module. Then, the driver's real-time driving ability and the driving real-time regional risk calculation unit based on the driving risk field are used to calculate the driver's real-time driving ability and driving real-time regional risk, respectively. The calculation process is as follows:

[0008] First, calculate the driver's real-time driving ability:

[0009] First, seven characteristics representing the lateral and longitudinal motion states of a vehicle are selected as indicators for evaluating a driver's driving ability, including the speed standard deviation (Std). v , acceleration standard deviation Std a Standard deviation of following distance (Std) DHW The vehicle's longitudinal state index at head-to-head distance (THW) and the standard deviation of the vehicle's lateral position offset (Std) d Vehicle lateral state indicators, including steering wheel angle entropy (SE) and steering wheel intersection frequency (TF);

[0010] Then, the analytic hierarchy process (AHP) algorithm based on group decision-making is used to subjectively quantify the individual driver's driving ability. Based on the seven selected vehicle driving state characteristics, the evaluation matrix given by the i-th expert according to the "5 / 5-9 / 1" scaling system is... It is 7×7 dimensional, and the consistency test of the evaluation matrix is ​​as follows:

[0011]

[0012] In the formula, RI=0.4007; To evaluate the largest eigenvalue of a matrix, its corresponding eigenvector is: m is the dimension of the evaluation matrix; when CR i When the value is less than or equal to 0.1, the evaluation matrix meets the consistency requirement, i.e., it passes the consistency test; the normalized evaluation weights given by the i-th expert. The calculation is as follows:

[0013]

[0014] Next, the entropy weight method is used to objectively quantify the individual driving ability of drivers, and the weight values ​​δ of each indicator are obtained. b =[δ b1 , …, δ bj ,...,δ bm [e] represents the information entropy of each indicator. j The corresponding difference coefficient w j The ratio of the total coefficient of difference to the total difference, i.e.:

[0015]

[0016] In the formula, w j The difference coefficient, i.e., w j =1-e j ;

[0017] And the information entropy e of the j-th indicator j as follows:

[0018]

[0019] In the formula, x′ kj This represents the data after standardization using the range method, where N represents the sample size; then, the subjective and objective mixed weights are calculated:

[0020]

[0021] In the formula, n expert This represents the number of experts in the ANP algorithm; thus, the driver's individual driving ability value is obtained:

[0022] Dr aby =δ*Veh f T (6)

[0023] In the formula, Veh f =[Std v ,Std a Std DHW THW,Std d [,SE,TF];

[0024] Secondly, calculate the real-time regional risks of driving:

[0025] First, we analyze the risk quantity of a traffic unit. When a traffic unit collides, kinetic energy is released through compression and collision, resulting in deformation of both colliding parties. This deformation characterizes the risk quantity of the traffic unit. Based on the above analysis, the risk quantity R of the traffic unit is:

[0026]

[0027] In the formula, T veh This refers to the vehicle coefficient, which includes the vehicle type coefficient and the vehicle loading coefficient. For example, the truck coefficient is greater than the passenger car coefficient; m veh Let v be the mass of the traffic unit, v be the speed of the traffic unit, and k1 be a correction factor, where k1 = 0.1. limit This represents the current maximum speed limit for the road.

[0028] Then, distance-based risk correction is performed. For the same risk source, the closer the distance, the greater the risk. The distance-based risk correction R between traffic unit j and traffic unit i can be applied. d，ji Represented as:

[0029]

[0030] In the formula, d ji Let be the distance vector from traffic unit j to traffic unit i, and k2 be a correction coefficient, where k2 = 1.

[0031] Next, risk correction based on motion state is performed, considering the relative magnitude and direction of motion between traffic units, and defining the motion state-based risk correction R of traffic unit j to traffic unit i. mot，ji for:

[0032] R mot,ji =exp[|v r |,cos(θ)] (9)

[0033] In the formula, v r =v j -v i θ is the relative velocity vector between traffic unit j and traffic unit i, and θ is the relative velocity vector v. r With distance vector d ji The included angle;

[0034] Then, a traffic rule-based risk correction is performed: considering that the environmental risk of traffic units is constrained by traffic rules, a traffic rule-based risk correction R is defined. lm for:

[0035]

[0036] In the formula, T lm T is the lane line type coefficient. lm When = 1, it represents a dashed line, T lm =0 indicates a solid line, d p w is the distance vector from the traffic unit to the lane line. road The width of the lane is the lane width. If there is a solid line between two traffic units, there will generally be no crossing of the traffic line, and the risk is relatively low. However, if there is a dashed line between two traffic units and the vehicle is close to the dashed line, the probability of it changing lanes increases, which increases the risk to the main vehicle.

[0037] Finally, taking into account the risk quantification model defined in equations (7) to (10) above, the total risk R of traffic unit j to traffic unit i can be calculated. ji * Defined as:

[0038]

[0039] In the formula, E ji =R d，ji ·R mot，ji ·R lm ;R iand R j Let i and j represent the traffic unit risk quantities calculated using equation (1), respectively.

[0040] In step three above, the driver's real-time driving ability is input into the real-time driving ability relative evaluation module. A Gaussian distribution statistical graph is formed by the driving abilities of all safe drivers in the statistical data set, i.e., the driving abilities of the general public. Based on the position of the quantified driving ability value in the driver's real-time ability calculation unit within this Gaussian distribution statistical graph, the driver's real-time relative driving ability is derived, i.e., the probability of the driver's ability value in the overall distribution. Then, the real-time regional driving risk is input into the regionalized quantitative risk field module, with the intelligent vehicle i as the center and radius D. i All unobstructed risk sources within a circle of / m, i.e., other traffic elements, are identified, and the circle is divided into six regions. Within each region, the total risk field generated by all unobstructed risk sources is calculated to obtain the quantified risk field of the surrounding environment for the main traffic unit, resulting in real-time regionalized driving risk. Then, the driver's relative driving ability level and the real-time key response area risk level are determined. The formula for calculating the total risk field is as follows:

[0041]

[0042] In the formula, S k Let n represent the aggregated risk field of the k-th region, k = 1, 2, ..., 6; k Let γ(k) represent the number of risk sources in the k-th region. Then, the coefficients γ(k) for the six regions are [0.8, 0.7, 0.8, 1, 0.9, 1], respectively.

[0043] In step four above, the driver's relative driving ability level and the real-time key response area risk level are input into the control right calculation module. Through the Gaussian distribution relative ability and risk calculation unit, the maximum value P(k) = max(P1, ..., P6) in the total risk field of the six areas is taken as the key response risk faced by the main traffic unit, and k is the area where the key response risk is located. Based on the distribution level of this key risk [π(1), ... π(5)] and the driver's relative driving ability level, the table is looked up, and the Gaussian distribution is used to determine the initial control right allocation strategy of human-machine co-driving.

[0044] In step five above, the initial control allocation strategy is input into the Hidden Markov Computation Unit (HMM). The HMM is then used to refine the initial control allocation strategy, and the driving power transition between the intelligent driving model and the driver is described through an HMM chain, based on the state set S. s ={1, 2} and state transition matrix In this context, state 1 indicates that the driver has acquired the primary driving control, and state 2 indicates that the intelligent driving model has acquired the primary driving control. The resulting observation probability matrix is ​​then calculated as follows:

[0045] B=[1-γ(k)*P(k), γ(k)*P(k)] T (13)

[0046] That is, the probability of preferring to drive using an intelligent driving model is proportional to the current priority risk response of the main traffic unit, and γ(k) is a coefficient related to region k;

[0047] Based on the Viterbi algorithm, the flexible control weight allocation coefficients for human-machine co-driving control at time t are given:

[0048]

[0049] Servo-level shared control actual output:

[0050] u s =(1-λ) t )u c +λ t u h (15)

[0051] In the formula, u s The actual effective control quantity, u h Input control parameters for the driver, u c Input control quantities into the intelligent driving model, and then calculate the final driving control allocation strategy;

[0052] In step six above, the driver's relative driving ability level is input into the absolute driving ability update module through the control right calculation module. Using the state transition matrix in the Hidden Markov Model, the driver's long-term absolute driving ability is evaluated. The state transition matrix A is updated using the following method with a forgetting factor:

[0053] [φ ij (t), j]=max(a i1 ·(1-γ(k)*P(k)), a i2 ·γ(k)*P(k)) (16)

[0054] In the formula, φ ij The subscripts i and j of (t) represent the main states at time t-1 and t, respectively; max(·) means taking the maximum value in the vector and extracting its position j∈S. s As the main state at time t

[0055]

[0056] From the update equations (16) and (17) of the state transition matrix, it can be seen that for a driver with good driving ability, they can have more control over the driving process during long-term driving, so the state transition matrix a 11 and a 21 Compared to a 12 and a 22 The values ​​will gradually increase, and then be transmitted back to the Gaussian distribution relative capability and risk calculation unit of the control calculation module to update the final driving control allocation strategy in real time, thereby achieving a long-term objective evaluation of the driver's absolute driving ability.

[0057] Preferably, in step one, the input terminal of the information acquisition module is connected to the output terminals of the control panel, the sensing device, and the driver assistance system module, respectively.

[0058] Preferably, in step one, the input terminal of the information acquisition module is connected to the output terminals of the control panel, the sensing device, and the driver assistance system module, respectively.

[0059] Preferably, in step two, the input terminal of the driver's driving information unit is connected to the output terminal of the control panel, and the output terminal of the driver's driving information unit is connected to the input terminal of the information processing module.

[0060] Preferably, in step three, the input end of the real-time driving ability relative evaluation module is connected to the output end of the driver's real-time ability calculation unit, and the input end of the regionalized quantitative risk field module is connected to the output end of the regionalized real-time risk calculation unit based on the driving risk field.

[0061] Preferably, in step four, the input of the control right calculation module is connected to the output of the information processing module, the real-time driving capability relative evaluation module, and the regionalized quantitative risk field module, respectively.

[0062] Preferably, in step four, the output terminals of the control authority calculation module, the driver driving information unit, and the driver assistance system module are all connected to the input terminal of the control panel.

[0063] Preferably, in step five, the input of the Hidden Markov Computation Unit is connected to the output of the Gaussian Distribution Relative Capability and Risk Computation Unit.

[0064] Preferably, in step six, the input and output terminals of the absolute driving capability update module are connected to the output and input terminals of the control authority calculation module, respectively.

[0065] Compared with existing technologies, the beneficial effects of this invention are as follows: This regionalized decision-making method for human-machine shared control based on Gaussian Hidden Markov Model (HMM) establishes a relative driving ability model based on a Gaussian distribution function to characterize the driver's real-time relative driving ability. It further utilizes a driving vector risk field to quantify environmental risk, proposing a regionalized Gaussian vector risk field environmental quantification model to characterize the environmental risk values ​​and fuzzy risk levels of different environmental regions. Finally, it proposes a regionalized decision-making algorithm for human-machine shared control based on HMM that comprehensively considers the driver's absolute ability, driving state, and environmental risk, achieving highly reliable and reasonable allocation of driving rights in human-machine shared control and providing more reasonable control authority. A flexible allocation scheme effectively reduces risks within the controllable range of the intelligent driving model. It proposes the concept of relative driving ability, constructs a real-time relative driving ability model for the driver, and continuously learns and strengthens the driver-specific absolute driving ability model represented by the Hidden Markov State Transition Matrix using the driver's current relative driving ability. This enables a long-term, objective evaluation of the driver's absolute driving ability. Taking into account the driver's different coping abilities and sensitivities to environmental risks from different directions, the area around the vehicle is divided into six zones. A regionalized real-time risk quantification model based on the driving risk field is constructed. Different driving control allocation strategies are adopted for different zones according to the driver's general driving attention concentration habits. Attached Figure Description

[0066] Figure 1 This is a system architecture diagram of the present invention;

[0067] Figure 2 This is a diagram of the judgment method in step three of the present invention;

[0068] Figure 3 This is a flowchart of the method of the present invention. Detailed Implementation

[0069] 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.

[0070] Please see Figure 1-3 The present invention provides an embodiment of a human-machine shared control regional decision-making method based on a Gaussian hidden Markov model, comprising the following steps: Step 1, information collection; Step 2, information processing; Step 3, level judgment; Step 4, preliminary allocation; Step 5, final allocation; Step 6, real-time update.

[0071] In step one above, the information collection module collects real-time driving information of the driver and the driver assistance system from the control panel, the sensing device and the driver assistance system module. The input end of the information collection module is connected to the output end of the control panel, the sensing device and the driver assistance system module respectively.

[0072] In step two above, the driver's real-time driving information is transmitted to the information processing module through the driver driving information unit of the information acquisition module. Then, the driver's real-time driving ability calculation unit and the regionalized real-time risk calculation unit based on the driving risk field calculate the driver's real-time driving ability and the real-time driving area risk, respectively. The input end of the driver driving information unit is connected to the output end of the control panel, and the output end of the driver driving information unit is connected to the input end of the information processing module. The calculation process is as follows:

[0073] First, calculate the driver's real-time driving ability:

[0074] First, seven characteristics representing the lateral and longitudinal motion states of a vehicle are selected as indicators for evaluating a driver's driving ability, including the speed standard deviation (Std). v , acceleration standard deviation Std a Standard deviation of following distance (Std) DHW The vehicle's longitudinal state index at head-to-head distance (THW) and the standard deviation of the vehicle's lateral position offset (Std) d Vehicle lateral state indicators, including steering wheel angle entropy (SE) and steering wheel intersection frequency (TF);

[0075] Then, the analytic hierarchy process (AHP) algorithm based on group decision-making is used to subjectively quantify the individual driver's driving ability. Based on the seven selected vehicle driving state characteristics, the evaluation matrix given by the i-th expert according to the "5 / 5-9 / 1" scaling system is... It is 7×7 dimensional, and its "5 / 5-9 / 1" scaling system is shown in the table below:

[0076]

[0077] The consistency check of the evaluation matrix is ​​as follows:

[0078]

[0079] In the formula, RI=0.4007; To evaluate the largest eigenvalue of a matrix, its corresponding eigenvector is: m is the dimension of the evaluation matrix; when CR i When the value is less than or equal to 0.1, the evaluation matrix meets the consistency requirement, i.e., it passes the consistency test; the normalized evaluation weights given by the i-th expert. The calculation is as follows:

[0080]

[0081] Next, the entropy weight method is used to objectively quantify the individual driving ability of drivers, and the weight values ​​δ of each indicator are obtained. b =[δ b1 , …, δ bj ,...,δ bm [e] represents the information entropy of each indicator. j The corresponding difference coefficient w j The ratio of the total coefficient of difference to the total difference, i.e.:

[0082]

[0083] In the formula, w j The difference coefficient, i.e., w j =1-e j ;

[0084] And the information entropy e of the j-th indicator j as follows:

[0085]

[0086] In the formula, x′ kj This represents the data after standardization using the range method, where N represents the sample size; then, the subjective and objective mixed weights are calculated:

[0087]

[0088] In the formula, n expert This represents the number of experts in the ANP algorithm; thus, the driver's individual driving ability value is obtained:

[0089] Dr aby =δ*Veh f T (6)

[0090] In the formula, Veh f =[Std v Std a Std DHW THW,Std d [SE,TF];

[0091] Secondly, calculate the real-time regional risks of driving:

[0092] First, we analyze the risk quantity of a traffic unit. When a traffic unit collides, kinetic energy is released through compression and collision, resulting in deformation of both colliding parties. This deformation characterizes the risk quantity of the traffic unit. Based on the above analysis, the risk quantity R of the traffic unit is:

[0093]

[0094] In the formula, T veh This refers to the vehicle coefficient, which includes the vehicle type coefficient and the vehicle loading coefficient. For example, the truck coefficient is greater than the passenger car coefficient; m veh Let v be the mass of the traffic unit, v be the speed of the traffic unit, and k1 be a correction factor, where k1 = 0.1. limit This represents the current maximum speed limit for the road.

[0095] Then, distance-based risk correction is performed. For the same risk source, the closer the distance, the greater the risk. The distance-based risk correction R between traffic unit j and traffic unit i can be applied. d，ji Represented as:

[0096]

[0097] In the formula, d ji Let be the distance vector from traffic unit j to traffic unit i, and k2 be a correction coefficient, where k2 = 1.

[0098] Next, risk correction based on motion state is performed, considering the relative magnitude and direction of motion between traffic units, and defining the motion state-based risk correction R of traffic unit j to traffic unit i. mot，ji for:

[0099] R mot，ji =exp[|v r |·cos(θ)] (9)

[0100] In the formula, v r =v j -v i θ is the relative velocity vector between traffic unit j and traffic unit i, and θ is the relative velocity vector v. r With distance vector d ji The included angle;

[0101] Then, a traffic rule-based risk correction is performed: considering that the environmental risk of traffic units is constrained by traffic rules, a traffic rule-based risk correction R is defined. lm for:

[0102]

[0103] In the formula, T lm T is the lane line type coefficient. lm When = 1, it represents a dashed line, T lm =0 indicates a solid line, d p w is the distance vector from the traffic unit to the lane line. roadThe width of the lane is the lane width. If there is a solid line between two traffic units, there will generally be no crossing of the traffic line, and the risk is relatively low. However, if there is a dashed line between two traffic units and the vehicle is close to the dashed line, the probability of it changing lanes increases, which increases the risk to the main vehicle.

[0104] Finally, taking into account the risk quantification model defined in equations (7) to (10) above, the total risk R of traffic unit j to traffic unit i can be calculated. ji * Defined as:

[0105]

[0106] In the formula, E ji =R d，ji ·R mot，ji ·R lm ;R i and R j Let i and j represent the traffic unit risk quantities calculated using equation (1), respectively.

[0107] In step three above, the driver's real-time driving ability is input into the real-time driving ability relative evaluation module. The input of the real-time driving ability relative evaluation module is connected to the output of the driver's real-time ability calculation unit. The input of the regionalized quantified risk field module is connected to the output of the regionalized real-time risk calculation unit based on the driving risk field. A Gaussian distribution statistical graph is formed by statistically analyzing the driving abilities of all safe drivers in the data set, i.e., the driving abilities of the general public. Based on the position of the quantified driving ability value in the driver's real-time ability calculation unit within this Gaussian distribution statistical graph, the driver's real-time relative driving ability, i.e., the probability of the driver's ability value in the overall distribution, is obtained. Then, the real-time regional driving risk is input into the regionalized quantified risk field module, with the intelligent vehicle i as the center and radius D. i All unobstructed risk sources (i.e., other traffic elements) within a circle of / m are identified, and the circle is divided into six regions. Within each region, the total quantified risk field generated by all unobstructed risk sources is calculated to obtain the quantified risk field of the surrounding environment for the main traffic unit region. This yields the real-time regionalized driving risk. Subsequently, the driver's relative driving ability level and the real-time key response area risk level are determined. The determination method is detailed in the appendix. Figure 2 The formula for calculating the aggregated risk field is as follows:

[0108]

[0109] In the formula, S k Let n represent the aggregated risk field of the k-th region, k = 1, 2, ..., 6; kLet γ(k) represent the number of risk sources in the k-th region. Then, the coefficients γ(k) for the six regions are [0.8, 0.7, 0.8, 1, 0.9, 1], respectively.

[0110] In step four above, the driver's relative driving ability level and the real-time key response area risk level are input into the control right calculation module. The input end of the control right calculation module is connected to the output end of the information processing module, the real-time driving ability relative evaluation module, and the regionalized quantitative risk field module, respectively. The output ends of the control right calculation module, the driver driving information unit, and the assisted driving system module are all connected to the input end of the control panel. Through the Gaussian distribution relative ability and risk calculation unit, the maximum value P(k) = max(P1, ..., P6) in the totalized risk field of the six regions is taken as the key response risk faced by the main traffic unit, where k is the region where the key response risk is located. Based on the distribution level of this key risk [π(1), ..., π(5)] and the driver's relative driving ability level, the table is looked up as follows:

[0111] P(k)∈π(1) P(k)∈π(2) P(k)∈π(3) P(k)∈π(4) P(k)∈π(5) <![CDATA[P Dr ∈π(1)]]> HMM Decision HMM Decision intelligent model intelligent model intelligent model <![CDATA[P Dr ∈π(2)]]> HMM Decision HMM Decision HMM Decision intelligent model intelligent model <![CDATA[P Dr ∈π(3)]]> driver HMM Decision HMM Decision HMM Decision intelligent model <![CDATA[P Dr ∈π(4)]]> driver driver HMM Decision HMM Decision HMM Decision <![CDATA[P Dr ∈π(5)]]> driver driver driver HMM Decision HMM Decision

[0112] And the initial control allocation strategy for human-machine co-driving was determined using Gaussian distribution;

[0113] In step five above, the initial control allocation strategy is input into the Hidden Markov Model (HMM) computation unit. The input of the HMM computation unit is connected to the output of the Gaussian distributed relative capability and risk calculation unit. The HMM refines the initial control allocation strategy using the HMM model, and describes the driving power transition between the intelligent driving model and the driver through the HMM chain, based on the state set.

[0114] S s ={1, 2} and state transition matrix In this context, state 1 indicates that the driver has acquired the primary driving control, and state 2 indicates that the intelligent driving model has acquired the primary driving control. The resulting observation probability matrix is ​​then calculated as follows:

[0115] B=[1-γ(k)*P(k), γ(k)*P(k)] T (13)

[0116] That is, the probability of preferring to drive using an intelligent driving model is proportional to the current priority risk response of the main traffic unit, and γ(k) is a coefficient related to region k;

[0117] Based on the Viterbi algorithm, the flexible control weight allocation coefficients for human-machine co-driving control at time t are given:

[0118]

[0119] Servo-level shared control actual output:

[0120] u s =(1-λ) t )u c +λ t u h (15)

[0121] In the formula, u s The actual effective control quantity, u h Input control parameters for the driver, u c Input control quantities into the intelligent driving model, and then calculate the final driving control allocation strategy;

[0122] In step six above, the driver's relative driving ability level is input into the absolute driving ability update module through the control right calculation module. The input and output of the absolute driving ability update module are connected to the output and input of the control right calculation module, respectively. Using the state transition matrix in the Hidden Markov Model, the driver's long-term absolute driving ability is evaluated. The state transition matrix A is updated using the following method with a forgetting factor:

[0123] [φ ij (t), j]=max(a i1 ·(1-γ(k)*P(k)), a i2 ·γ(k)*P(k)) (16)

[0124] In the formula, φ ij The subscripts i and j of (t) represent the main states at time t-1 and t, respectively; max(·) means taking the maximum value in the vector and extracting its position j∈S. s As the main state at time t;

[0125]

[0126] From the update equations (16) and (17) of the state transition matrix, it can be seen that for a driver with good driving ability, they can have more control over the driving process during long-term driving, so the state transition matrix a 11 and a 21 Compared to a 12 and a 22 The values ​​will gradually increase, and then be transmitted back to the Gaussian distribution relative capability and risk calculation unit of the control calculation module to update the final driving control allocation strategy in real time, thereby achieving a long-term objective evaluation of the driver's absolute driving ability.

[0127] Based on the above, the advantages of this invention are as follows: It establishes a relative driving ability model based on a Gaussian distribution function to characterize the driver's real-time relative driving ability; further, it utilizes a driving vector risk field to quantify environmental risk, proposing a regionalized Gaussian vector risk field environmental quantification model to characterize the environmental risk values ​​and fuzzy risk levels of different environmental regions; finally, it proposes a regionalized decision-making algorithm for human-machine shared control based on a hidden Markov model, comprehensively considering the driver's absolute ability, driving state, and environmental risk. This achieves highly reliable and reasonable allocation of driving rights in human-machine shared control, providing a more reasonable flexible allocation scheme for control rights, effectively reducing risk within the controllable range of the intelligent driving model, and ultimately realizing adaptive automatic adjustment of the driver's control switching strategy, making driving safer, improving driver comfort, and reducing the risk of driver assistance system delays.

[0128] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A human-machine shared control regional decision-making method based on Gaussian Hidden Markov Model, comprising the following steps: Step 1: Information Collection; Step 2: Information Processing; Step 3: Level Determination; Step 4: Preliminary Allocation; Step 5: Final Allocation; Step 6: Real-time Update; Its key feature is: In step one above, the information collection module collects real-time driving information of the driver and the real-time driving information of the driver assistance system from the control panel, sensing devices and driver assistance system module. In step two above, the driver's real-time driving information is transmitted to the information processing module through the driver driving information unit of the information collection module. Then, the driver's real-time driving ability and the driving real-time regional risk calculation unit based on the driving risk field are used to calculate the driver's real-time driving ability and driving real-time regional risk, respectively. The calculation process is as follows: First, calculate the driver's real-time driving ability: First, seven characteristics representing the lateral and longitudinal motion states of a vehicle are selected as indicators for evaluating a driver's driving ability, including the speed standard deviation. Acceleration standard deviation Standard deviation of following distance Headway The vehicle's longitudinal condition indicators and the standard deviation of the vehicle's lateral position offset. Steering wheel angle entropy Steering wheel turning frequency Vehicle lateral condition indicators; Then, the analytic hierarchy process (AHP) algorithm based on group decision-making is used to subjectively quantify the individual driver's driving ability. Based on the selected seven vehicle driving state features, therefore the... According to experts The evaluation matrix given by the scaling system for The dimension is defined, and the consistency test of the evaluation matrix is ​​as follows: （1）； In the formula, ; , To evaluate the largest eigenvalue of a matrix, its corresponding eigenvector is: , To evaluate the matrix dimension; when When the evaluation matrix meets the consistency requirement, it passes the consistency test; The normalized evaluation weights given by the experts The calculation is as follows: （2）； Next, the entropy weight method is used to objectively quantify the individual driving ability of drivers, and the weight values ​​of each indicator are obtained. Information entropy for each indicator Corresponding difference coefficient The ratio of the total coefficient of difference to the total difference, i.e.: （3）； In the formula, The coefficient of variation is denoted as , ; And the first Information entropy of each indicator as follows: （4）； In the formula, This represents the data after standardization using the range method. Indicates the number of samples; Then, the subjective and objective combined weights are calculated: （5）； In the formula, This represents the number of experts in the ANP algorithm; thus, the driver's individual driving ability value is obtained: （6）； In the formula, ; Secondly, calculate the real-time regional risks of driving: First, we analyze the risk quantity of a traffic unit. When a traffic unit collides, kinetic energy is released through compression and collision, resulting in deformation of both colliding parties. This deformation characterizes the risk quantity of the traffic unit. Based on the above analysis, the risk quantity R of the traffic unit is: （7）； In the formula, The vehicle coefficient includes the vehicle type coefficient and the vehicle loading coefficient; the truck coefficient is greater than the passenger car coefficient. For the quality of the transportation unit, For the speed of the traffic unit, For the correction factor, here , This represents the current maximum speed limit for the road. Then, distance-based risk correction is performed. For the same risk source, the closer the distance, the greater the risk. This can be achieved by adjusting the traffic unit. For transportation units Distance-based risk correction Represented as: （8）； In the formula, for Transportation unit to Distance vector of traffic unit, For the correction factor, here =1; Next, risk correction based on motion state is performed, taking into account the relative magnitude and direction of motion between traffic units, and defining traffic units. For transportation units Risk correction based on motion state for: （9）； In the formula, It is a transportation unit With transportation units The relative velocity vector, Relative velocity vector With distance vector The included angle; Then, risk correction based on traffic rules is performed: considering that the environmental risks of traffic units during travel are constrained by traffic rules, a risk correction based on traffic rules is defined. for: (10)； In the formula, Lane type coefficient, When, it represents a dashed line. The time indicates a solid line. The distance vector from the traffic unit to the lane line. The width of the lane is the lane width. If there is a solid line between two traffic units, there will generally be no crossing of the traffic line, and the risk is relatively low. However, if there is a dashed line between two traffic units and the vehicle is close to the dashed line, the probability of it changing lanes increases, which increases the risk to the main vehicle. Finally, taking into account the risk quantification model defined in equations (7) to (10) above, the traffic unit can be... For transportation units Total risk Defined as: (11)； In the formula, ; as well as They represent the traffic units calculated using equation (7). and transportation units The risk level of the transportation unit; In step three above, the driver's real-time driving ability is input into the real-time driving ability relative evaluation module. This module uses a Gaussian distribution graph formed by the driving abilities of all safe drivers in a statistical data set (i.e., the driving abilities of the general public). Based on the position of the quantified driving ability value in the driver's real-time ability calculation unit within this Gaussian distribution graph, the driver's real-time relative driving ability is derived, i.e., the probability of the driver's ability value within the overall distribution. Then, the real-time regional driving risk is input into the regionalized quantified risk field module, using intelligent vehicles... With center at and radius at, All unobstructed risk sources within a circle of meters, i.e., other traffic elements, are considered. The circle is divided into six regions. Within each region, the total quantified risk field generated by all unobstructed risk sources is calculated to obtain the quantified risk field of the surrounding environment of the main traffic unit, thus obtaining the real-time regionalized driving risk. Then, the driver's relative driving ability level and the real-time key response area risk level are determined. The formula for calculating the total quantified risk field is as follows: (12)； In the formula, Indicates the first The overall risk field in each region ; Indicates the first The number of risk sources in each region, then the coefficients for the six regions. Take respectively ; In step four above, the driver's relative driving ability level and the real-time risk level of the key response area are input into the control authority calculation module. The maximum value among the total risk fields of the six regions is taken through the Gaussian distribution relative ability and risk calculation unit. The key risks that the main transportation units face. To focus on addressing the areas where risks are located, based on the distribution level of these key risks. And the relative driving ability level of the driver is looked up in a table, and the initial control allocation strategy for human-machine co-driving is determined by using Gaussian distribution; In step five above, the initial control allocation strategy is input into the Hidden Markov Model (HMM) unit. The HMM is then used to refine the initial control allocation strategy, and the driving control transition between the intelligent driving model and the driver is described through an HMM chain, based on the state set. and state transition matrix State 1 indicates that the driver has acquired the primary driving position, and state 2 indicates that the intelligent driving model has acquired the primary driving position. The observation probability matrix is ​​then integrated as follows: (13)； In other words, the probability of favoring intelligent driving models is directly proportional to the main transportation unit's current focus on addressing risks. In order to cooperate with the region Relevant coefficients; Based on the Viterbi algorithm, the flexible control weight allocation coefficients for human-machine co-driving control at time t are given: (14)； Servo-level shared control actual output: (15)； In the formula, The actual effective control quantity. Input control parameters for the driver. Input control quantities into the intelligent driving model, and then calculate the final driving control allocation strategy; In step six above, the driver's relative driving ability level is input into the absolute driving ability update module through the control right calculation module. Using the state transition matrix in the Hidden Markov Model, the driver's long-term absolute driving ability is evaluated. The state transition matrix A is updated using the following method with a forgetting factor: (16)； In the formula, subscript They represent as well as The main state at any given moment; This means taking the maximum value in a vector and extracting its position. As the The main state at any given moment, (17)； From the update equations (16) and (17) of the state transition matrix, it can be seen that for a driver with good driving ability, they can have more control over the driving process over a long period of time. Therefore, their state transition matrix will show a higher degree of control over the driving process. and Compared to and The values ​​will gradually increase, and then be transmitted back to the Gaussian distribution relative capability and risk calculation unit of the control calculation module to update the final driving control allocation strategy in real time, thereby achieving a long-term objective evaluation of the driver's absolute driving ability.

2. The human-machine shared control regional decision-making method based on Gaussian Hidden Markov Model according to claim 1, characterized in that: In step one, the input terminal of the information acquisition module is connected to the output terminals of the control panel, the sensing device, and the driver assistance system module, respectively.

3. The human-machine shared control regional decision-making method based on Gaussian Hidden Markov Model according to claim 1, characterized in that: In step two, the input terminal of the driver's driving information unit is connected to the output terminal of the control panel, and the output terminal of the driver's driving information unit is connected to the input terminal of the information processing module.

4. The human-machine shared control regional decision-making method based on Gaussian Hidden Markov Model according to claim 1, characterized in that: In step three, the input of the real-time driving ability relative evaluation module is connected to the output of the driver's real-time ability calculation unit, and the input of the regionalized quantitative risk field module is connected to the output of the regionalized real-time risk calculation unit based on the driving risk field.

5. The human-machine shared control regional decision-making method based on Gaussian Hidden Markov Model according to claim 1, characterized in that: In step four, the input of the control right calculation module is connected to the output of the information processing module, the real-time driving capability relative evaluation module, and the regionalized quantitative risk field module, respectively.

6. The human-machine shared control regional decision-making method based on Gaussian Hidden Markov Model according to claim 1, characterized in that: In step four, the outputs of the control calculation module, the driver driving information unit, and the driver assistance system module are all connected to the input of the control panel.

7. The human-machine shared control regional decision-making method based on Gaussian Hidden Markov Model according to claim 1, characterized in that: In step five, the input of the Hidden Markov Computation Unit is connected to the output of the Gaussian Distribution Relative Capability and Risk Computation Unit.

8. The human-machine shared control regional decision-making method based on Gaussian Hidden Markov Model according to claim 1, characterized in that: In step six, the input and output terminals of the absolute driving capability update module are connected to the output and input terminals of the control authority calculation module, respectively.

Images