Lane changing intention recognition method based on driving style

Abstract

The invention provides a lane changing intention recognition method based on a driving style. The method comprises the following steps: firstly, acquiring driving behavior data in a car-following scene through a driving simulation experiment, and constructing a multi-dimensional data set comprising a self-car state, a front-car state and time-space car-following characteristics; then, on the basis of an expected safety margin model, identifying reaction characteristics, steady-state risk characteristics and operation characteristics of a driver, and extracting behavior parameters forming a driving style, including indexes such as reaction time, an expected safety margin interval and operation sensitivity; and further dividing the drivers into different style types by using a clustering algorithm, and endowing style labels with the drivers. On the basis, lane changing intention time information subjectively marked by a driver is collected through a non-control observation experiment and is aligned with the trajectory data, and a trajectory data set with a real intention label is constructed. And finally, constructing a dynamic Bayesian network model based on the data set, taking a historical driving state sequence in a fixed time window as input, fusing a driving style, interaction characteristics of a vehicle and surrounding vehicles and traffic flow characteristics, and realizing real-time identification of whether a lane changing intention exists at the current moment. The method has good interpretability, individual difference adaptability and engineering deployment, and is suitable for dynamic identification and response control of driver intention in an intelligent driving assistance system.

Description

Technical Field

[0001] The present invention relates to the fields of traffic safety and intention recognition, and in particular to a method for recognizing lane-changing intentions based on driving style. Background Art

[0002] With the development of intelligent driver assistance systems (ADAS) and autonomous driving technologies, driver intention has become an important research topic for improving traffic safety and system decision-making rationality. Lane changing, one of the most uncertain and dangerous micro-driving behaviors, is crucial for collision avoidance, path planning optimization, and achieving human-machine collaborative control.

[0003] Scholars at home and abroad have conducted extensive research on lane change intention recognition, using input data sources including vehicle kinematic parameters, traffic environment information, and driver behavior signals. Modeling methods include logical rules, support vector machines, hidden Markov models, and time-series neural networks such as LSTM. However, current research still has significant shortcomings in the following areas: First, the actual mechanism of intention generation is poorly modeled. Most methods simplify lane change intention as a state label representing "a few seconds before the lane change occurred," focusing on retrospective identification of behavioral outcomes while neglecting the formation of lane change intention as an independent cognitive state. This approach fails to distinguish the causal relationship between "intention generation" and "action execution." Second, individual driving style factors are neglected. Drivers vary significantly in reaction speed, risk tolerance, and operational sensitivity. However, existing studies often use uniform thresholds or fixed model structures, failing to model individual driving styles and thus failing to generalize to diverse driving behaviors. Third, the acquisition of intention data is limited. Currently, there is no unified standard for lane-changing intention labeling. Mainstream methods still rely on inferring lane-changing intention from lane-changing behavior, which is unable to form accurate intention labels. They also ignore complex situations where the intention exists but the lane change is not made, or where the lane-changing action is not immediately taken, resulting in an unbalanced intention dataset.

[0004] Therefore, there is an urgent need to propose a lane-changing intention recognition method that integrates driving style modeling, subjective intention dot data, and multi-source temporal state features. This method can accurately identify the lane-changing intention state at the current moment based on the historical state sequence, and has structural interpretability and engineering practicality to meet the real-time intention recognition needs in intelligent driving scenarios. Summary of the Invention

[0005] The present invention addresses the problems of existing lane-change intention research by providing a driving style-based lane-change intention recognition method that improves the accuracy, proactiveness, and personalized adaptability of intention recognition. Furthermore, the present invention's recognition method is more closely aligned with the driver's actual lane-changing cognitive process, enabling dynamic modeling and real-time recognition of lane-change intentions. It possesses both good engineering feasibility and theoretical explanatory power, making it suitable for real-time driver intention recognition tasks within intelligent driving assistance systems.

[0006] The technical solution of the present invention is a method for identifying lane change intention based on driving style, characterized by comprising the following steps:

[0007] Step 1: Use a driving simulator to collect kinematic parameters of vehicles in a car-following scenario and create a multidimensional dataset S. The multidimensional dataset S contains eight features: ego vehicle speed, leading vehicle speed, ego vehicle acceleration, leading vehicle acceleration, ego vehicle position, leading vehicle position, following distance, and time.

[0008] S={v,v p ,a,a p ,x,x p ,D,t}

[0009] D=x p -xL

[0010] Where v is the vehicle speed; v p is the speed of the preceding vehicle; a is the acceleration of the vehicle; a p is the acceleration of the preceding vehicle; x is the position of the vehicle; x p is the position of the preceding vehicle; D is the following distance; t is the recording time; L is the length of the vehicle;

[0011] Step 2: Based on the data set S from step 1, the heterogeneous car-following characteristics of the drivers are identified. Based on the expected safety margin model, the driver reaction characteristic quantitative index, steady-state risk quantitative index, and operational characteristic quantitative index are respectively calculated to obtain a parameter vector p that can effectively characterize the heterogeneous car-following characteristics of the drivers. The driver reaction characteristic quantitative index includes the average driver reaction time; the steady-state risk quantitative index includes the lower limit of the expected safety margin and the upper limit of the expected safety margin; and the operational characteristic quantitative index includes the acceleration sensitivity coefficient and the deceleration sensitivity coefficient.

[0012] p=(τ n ,SM nDL ,SM nDH ,ɑ1,α2)T

[0013] Among them, τ n is the average reaction time of the driver; SM nDL is the lower limit of the expected safety margin; SM nDHis the upper limit of the expected safety margin; ɑ1 is the acceleration sensitivity coefficient; ɑ2 is the deceleration sensitivity coefficient;

[0014] Step 3: Using the parameter vector p based on the expected safety margin model obtained in Step 2, a clustering algorithm is used to classify driving styles and obtain corresponding driving style labels, including conservative, general, and aggressive.

[0015] Step 4: Based on representative samples from the public trajectory dataset, a high-fidelity 3D dynamic traffic scenario was created. An uncontrolled observation experiment was conducted using a driving simulator. The driver's subjective judgment of the intention was made and annotated. The driver's lane change intention annotation data was collected to construct a driving sample set containing lane change intentions.

[0016] Step 5: Using the driving style labels obtained in step 3 and the driving sample set constructed in step 4, a dynamic Bayesian network model is constructed and trained to achieve real-time recognition and dynamic modeling of lane change intentions.

[0017] Furthermore, in step 2, the steps of identifying the car-following characteristics of heterogeneous drivers and calculating the parameter vector p representing the car-following characteristics of heterogeneous drivers are as follows:

[0018] a. Calculate the driver's reaction characteristic quantitative indicator, the average driver reaction time τ n According to the Newell car-following model, the spatiotemporal trajectory of the following car is basically the same as that of the leading car during the car-following process, with only a translation in space and time. Therefore, at each moment t n , the time delay Δt required for the following vehicle to reach the same speed as the preceding vehicle is determined and is defined as:

[0019]

[0020] Among them, τ h is the driver’s reaction time at time t; v(t) is the vehicle speed, v p (t n ) is the speed of the preceding vehicle at the current moment, and ε is the speed tolerance;

[0021] At each time step, h After solving and screening, τ h The statistical mean of the following vehicle driver is taken as the average reaction time τ n The final identification result:

[0022]

[0023] Among them, N τ represents the effective τ solved during the car-following process h The number of

[0024] b. Calculate the driver's steady-state risk quantitative index expected safety margin lower limit SM nDL and the expected upper limit of safety margin SM nDH First, the stable car-following data segment is filtered, and then the sliding window filter is introduced to calculate the safety margin SM within the sliding window. n The maximum and minimum values ​​of (t):

[0025]

[0026] in, and Represents the real-time safety margin and the upper and lower limits of the expected safety margin during the sliding window.

[0027] c. Calculate the acceleration sensitivity coefficient α1 and deceleration sensitivity coefficient α2, which are quantitative indicators of the driver's operating characteristics. According to the DSM model, the sensitivity coefficients can be derived as follows:

[0028]

[0029] Considering the instability of the driver's car-following behavior and the influence of data errors, the acceleration / deceleration car-following data segment should meet the requirement that the absolute value of the acceleration is greater than 0.5m / s 2 Finally, the final identification result of the sensitivity coefficient is obtained:

[0030]

[0031] in, and Represent the number of identified α1 and α2 respectively.

[0032] Furthermore, in step 3, the K-Means++ algorithm is used to simulate and analyze the driver's car-following characteristic parameters to obtain a data set divided into three clusters, that is, the driver's driving style G is divided into conservative, general, and aggressive types;

[0033] G∈{conservative, general, radical}

[0034] Furthermore, in step 4, the high-fidelity three-dimensional dynamic traffic scene used in the uncontrolled observation experiment is generated from real vehicle trajectory data, and a unified observation perspective is used on the driving simulator. Multiple subjects independently judge the intention by holding the steering wheel and pressing the corresponding button on the steering wheel when the lane change intention is generated, thereby collecting and annotating the intention label I;

[0035] Furthermore, in step 5, the driving style label obtained in step 3 and the driving sample set constructed in step 4 are used as input, and the driving state sequence of the fixed time window length T in the past is used, wherein the state characteristics at each moment include the state of the vehicle, the relative relationship between the vehicle and surrounding vehicles, the traffic flow information of the road section and the driving style label, and the output is whether there is a lane change intention at the current moment. During the model training process, the constructed labeled sample set is divided into a training set and a test set in an 8:2 ratio, and the conditional dependency and probability distribution between variables are established using structural learning and parameter learning methods. The model performance is evaluated on the test set by accuracy, recall rate and F1 score. When the F1 score reaches the preset threshold, the model is considered to have practical deployment capabilities. The model supports an online reasoning mechanism in a sliding time window manner to achieve real-time identification and dynamic modeling of lane change intentions at the current moment. The model has the following structure:

[0036] a. The temporal input feature sequence is:

[0037]

[0038] Among them, G is the driving style label; F t is the state feature at time t, including the state of the vehicle, the interaction features between the vehicle and the six neighboring vehicles, and the traffic flow characteristics of the road section;

[0039] b. Divide the constructed labeled sample set into a training set and a test set in an 8:2 ratio. The training set is used for model structure and parameter learning, and the test set is used for model performance evaluation.

[0040] c. The training goal is to learn the current intention state node I t The posterior probability Parameter estimation methods include maximum likelihood estimation or expectation maximization algorithms;

[0041] d. The model performance is evaluated using precision, recall, and F1 score. A model that achieves an F1 score of at least 0.80 in the test set is considered satisfactory for practical use.

[0042] e. The model supports real-time acquisition of the latest state sequence in the form of a sliding time window, completes the identification of the lane change intention state at the current moment, and outputs the posterior probability results as a decision-making aid. It is suitable for real-time identification and control response in embedded intelligent driver assistance systems.

[0043] The advantages of the present invention compared with the prior art are:

[0044] The method of the present invention constructs a multi-dimensional time-series driving state feature sequence based on the integration of driving style, vehicle dynamic state, interaction characteristics between the vehicle and surrounding vehicles, and traffic flow information; based on the lane-changing intention labels subjectively labeled by the subjects, combined with the dynamic Bayesian network modeling method, causal modeling and probabilistic reasoning of the lane-changing intention state are realized. By real-time identification of whether there is a lane-changing intention at the current moment, the potential cognitive intention of the driver's lane-changing behavior can be effectively captured, and the system's understanding and response capabilities to complex driving behaviors can be improved. The structure of this method is interpretable and the reasoning can be updated. It is close to the actual lane-changing decision-making mechanism and has good engineering adaptability and model expansion capabilities. It is suitable for integration into intelligent driving assistance systems to perform real-time intention recognition and early warning tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a block diagram of the overall concept of the present invention;

[0046] Figure 2 Construct process maps for typical scenarios;

[0047] Figure 3 Design a flow chart for an uncontrolled observational experiment. DETAILED DESCRIPTION

[0048] The present invention will be described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the examples are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0049] like Figure 1 As shown, a lane change intention recognition method based on driving style of the present invention specifically includes the following steps:

[0050] Step 1: By constructing a typical car-following driving simulation experiment, obtain the driver's longitudinal operational behavior data under real perception conditions and establish a multi-dimensional dataset S.

[0051] (1) A standard two-lane straight road scene was constructed in the UC-win / Road driving simulator. The environment was set to clear weather, daytime traffic, and no other interfering traffic flow.

[0052] (2) The leading vehicle is set as a program-controlled vehicle, and executes a full-cycle driving strategy including starting, accelerating, constant speed, decelerating, stopping, and restarting. The complete driving cycle is completed periodically to simulate all longitudinal motion conditions that occur on the actual road. The initial distance between the leading vehicle and the experimental vehicle is D0 = 1.9 m;

[0053] (3) Invite multiple subjects to enter the experimental system as drivers and require them to operate the self-vehicle in the simulator to continuously follow the vehicle in front without changing lanes or overtaking, and adjust the acceleration and deceleration according to the actual traffic judgment;

[0054] (4) The dynamic states of the vehicle and the preceding vehicle are recorded in real time through the driving simulator platform, and the following eight core characteristic parameters are collected to construct a basic multi-dimensional data set:

[0055] S={v,v p ,a,a p ,s,s p ,D,t}

[0056] Where v is the vehicle speed; v p is the speed of the preceding vehicle; a is the acceleration of the vehicle; a p is the acceleration of the preceding vehicle; s is the position of the vehicle; s p is the position of the preceding vehicle; D is the following distance; t is the recording time; L is the length of the vehicle;

[0057] The features v,a,s,t in the dataset can be directly exported from the platform data, v p ,a p ,s p The following distance D is calculated based on the original program control calculation of the preceding vehicle as follows:

[0058] D=s p -sL

[0059] Among them, L is the length of the vehicle.

[0060] Step 2: Based on the dataset s from Step 1, identify the heterogeneous car-following characteristics of the drivers and calculate the parameter vector p representing these heterogeneous car-following characteristics. Quantitative indices for driver reaction characteristics, steady-state risk, and operational characteristics are calculated based on the expected safety margin model. The driver reaction characteristic indices include the average driver reaction time; the steady-state risk indices include the lower and upper expected safety margin limits; and the operational characteristics indices include the acceleration sensitivity coefficient and the deceleration sensitivity coefficient.

[0061] (1) Establish model constraints. To ensure the accuracy of parameter identification, it is necessary to first select data segments that meet the constraints from the full working condition data, including:

[0062] a. The vehicle and the preceding vehicle remain in the same lane;

[0063] b. Meet the minimum speed constraint and reduce vehicle idling and data errors:

[0064] v n (t)>3m / s,v n-1 (t)>3m / s

[0065] Among them, v n (t) represents the instantaneous speed of the vehicle at the nth moment;

[0066] c. Minimum distance constraints must be met. If the distance between vehicles is too small, the driver may need to frequently apply the brakes and may not be able to demonstrate true car-following characteristics.

[0067] t(t)>5m

[0068] Where D(t) represents the following distance of the ego vehicle at the tth moment;

[0069] (2) Calculate the driver's reaction characteristic quantitative indicator, the average driver reaction time τ n .

[0070] According to the Newell car-following model, the spatiotemporal trajectory of the following car during the car-following process is basically the same as that of the leading car, with only a translation in space and time. Therefore, at each moment t n , the time delay Δt required for the following vehicle to reach the same speed as the preceding vehicle is determined and is defined as:

[0071]

[0072] Among them, τ h is the driver’s reaction time at time t; v(t) is the vehicle speed, v p (t n ) is the speed of the preceding vehicle at the current moment, and ε is the speed tolerance;

[0073] At each time step, h After solving and screening, τ h The statistical mean of the following vehicle driver is taken as the average reaction time τ n The final identification result:

[0074]

[0075] Among them, N τ represents the effective τ solved during the car-following process h The number of

[0076] (3) Calculate the lower limit of the expected safety margin SM of the driver's steady-state risk quantitative index nDL and the expected upper limit of safety margin SM nDH .

[0077] According to the DSM model, if the current risk level is within an acceptable range, the driver tends to maintain the current speed. This means that when the driver is in a stable following state, that is, the vehicle acceleration is small and the speed is almost constant, the current driver's following risk should be within the acceptable risk level. First, calculate the safety margin SM n :

[0078]

[0079] Secondly, filter the stable following data segment and select the vehicle whose acceleration meets -0.2m / s 2 n (t)<0.2m / s 2 , safety margin meets 0.5 <SM n The data segment under the condition of (t) < 1 and duration t > 3s is regarded as the stable following interval;

[0080] Then the sliding window filter is introduced to calculate the safety margin SM within the sliding window n The maximum and minimum values ​​of (t):

[0081]

[0082] in, and Indicates the real-time SM and DSM upper and lower limits during the sliding window period.

[0083] (4) Calculate the acceleration sensitivity coefficient α1 and deceleration sensitivity coefficient α2, which are quantitative indicators of the driver's operating characteristics.

[0084] According to the DSM model, the sensitivity coefficient can be derived as:

[0085]

[0086] Considering the instability of the driver's car-following behavior and the influence of data errors, the acceleration / deceleration car-following data segment should meet the requirement that the absolute value of the acceleration is greater than 0.5m / s 2 The effective range of α1 and α2 is set to (0,30). Finally, the final identification result of the sensitivity coefficient is obtained:

[0087]

[0088] in, and Represent the number of identified α1 and α2 respectively.

[0089] (5) After completing the calculation of the above three types of indicators, the five-dimensional feature vector of each subject's driving behavior is obtained:

[0090] p=(τ n ,SM nDL ,SM nDH ,ɑ1,α2) T

[0091] This parameter set serves as the basis for subsequent driving style clustering and intention modeling.

[0092] ​Step 3: Use the K-Means++ algorithm to simulate and analyze the driver's car-following characteristic parameter vector p, obtaining a data set divided into three clusters. That is, the driver's driving style G is divided into conservative, general, and aggressive types.

[0093] (1) Sample preparation

[0094] The parameter vector of each subject's car-following behavior in the simulation experiment is normalized to construct the parameter matrix P∈R N ×5 , where N is the number of subjects, and each row represents the five-dimensional parameter vector of a subject.

[0095] The standardization method uses Z-score standardization:

[0096]

[0097] Among them, μ j , σ j are the mean and standard deviation of the jth parameter respectively;

[0098] (2) Clustering method

[0099] The K-Means++ clustering algorithm is used to classify the standardized behavioral feature vectors. Compared to traditional K-Means, K-Means++ has a better initial center point selection strategy, significantly improving clustering stability and classification results. Based on driving psychology literature and experimental experience, three driving styles are defined:

[0100]

[0101] (3) Style tag assignment

[0102] After clustering is completed, the driving style label is defined based on the central eigenvalue and relative index distribution of each class.

[0103] G∈{conservative, general, radical}

[0104] It is then bound to the corresponding behavioral data and lane-changing intention samples of each subject, and the style classification results are finally output:

[0105] G n =Cluster(P n )

[0106] Among them G n is the style label of the nth subject, P n is its eigenvector.

[0107] Step 4: To address the difficulties of indirectly inferring lane-changing intentions through trajectories and the lag and subjectivity of traditional labeling methods, this step designs and implements an uncontrolled observation experiment to obtain the time points at which drivers actually subjectively generate lane-changing intentions under uniform observation conditions, thereby constructing a dataset with lane-changing intention labels.

[0108] (1) Typical scenario construction

[0109] Based on public trajectory datasets (such as NGSIM), typical lane-changing and car-following samples are selected and restored into high-fidelity 3D dynamic traffic scenes using RoadRunner and Unity platforms. The typical scene construction process is as follows: Figure 2 The generated scene has the following characteristics:

[0110] a. Use the driver's unified driving perspective to ensure consistent observation;

[0111] b. Set the actual lane, road markings, surrounding vehicles and background traffic;

[0112] c. The scene generation frame rate is strictly aligned with the original trajectory timestamp;

[0113] (2) Uncontrolled observation experiment

[0114] The same subjects as in step 1 sat in the driving simulator cockpit. The experiment required them not to perform actual driving operations, but to act as observers to subjectively judge whether the driving vehicle in the scene had the intention to change lanes: when it was judged that the main vehicle in the scene had the intention to change lanes, they immediately pressed the marked button on the steering wheel; the system simultaneously recorded the time t when the button was pressed. intention , and synchronized with the timestamp in the data trajectory to obtain the lane change intention label I t , the uncontrolled observational experimental design process is as follows Figure 3 As shown;

[0115] (3) Dataset construction

[0116] Finally, a lane-changing intention label dataset F covering the entire trajectory is constructed. Each frame contains the following information:

[0117] a. Vehicle status information S ego

[0118] S ego ={v,a,x,y,θ}

[0119] Where x is the lateral coordinate of the vehicle; y is the longitudinal coordinate of the vehicle; θ is the heading angle of the vehicle;

[0120] b. Interaction information between the vehicle and surrounding vehicles

[0121]

[0122] in, Indicates the vehicle index set in front, behind, left front, right front, left rear, and right rear around the vehicle; Δv i is the relative speed between the vehicle and surrounding vehicles; Δa i is the relative acceleration between the vehicle and surrounding vehicles; Δx i is the relative lateral position of the vehicle and surrounding vehicles; Δy i is the relative longitudinal position of the vehicle and surrounding vehicles;

[0123] c. Road section traffic flow characteristic information S trf

[0124]

[0125] K=n / s

[0126]

[0127] Among them, Q is the traffic volume of the road section; K is the traffic density of the road section; is the time average speed of vehicles on the road section; is the average acceleration of vehicles on the road section; σ v is the standard deviation of road speed; a is the standard deviation of the acceleration of the road section; s is the length of the road section; n is the number of vehicles at the i-th time stamp of the road section; v r is the speed of the rth vehicle on the road section at the i-th timestamp; a r is the acceleration of the rth vehicle on the road segment at the i-th timestamp;

[0128] d. Lane change intention label

[0129]

[0130] Step 5: To achieve real-time recognition and structured modeling of lane change intentions, this step constructs and trains a dynamic Bayesian network model based on the driving style labels obtained in Step 3 and the trajectory data with lane change intention labels constructed in Step 4. This model takes a fixed-length historical time series state feature sequence as input and outputs whether a lane change intention exists at the current moment, enabling online recognition of the driver's lane change intention state.

[0131] (1) Time parameter description and sliding window update mechanism

[0132] To ensure the real-time availability of the model, the present invention adopts a sliding time window mechanism, continuously inputs data per frame and updates it in real time:

[0133] T: model input window length (unit: frame);

[0134] dt: time step of trajectory data (unit: seconds);

[0135] Sliding window method: Updated every dt seconds, the model uses the most recent T frame state as input and outputs whether there is a lane change intention in the current frame.

[0136] The recommended settings are:

[0137] T = 20, that is, the model is input with the state of the past 2 seconds each time;

[0138] dt = 0.1s, corresponding to 10Hz trajectory sampling;

[0139] The output frequency is 10 times / second.

[0140] (2) Model input and output definition

[0141] The model input is a historical driving state sequence of length T:

[0142]

[0143] Each frame feature Including driving style label G and multi-source fusion input feature set F;

[0144] The model output is whether the lane change intention I occurs at the current time t t , the value comes from the subjective marking results in the uncontrolled observation experiment in step 4.

[0145] (3) Training sample construction and data division

[0146] When constructing the sample dataset, a fixed-length sliding window is used to traverse the trajectory data. For each moment t:

[0147] The input is the state feature sequence of T frames before this moment The label is whether there is a lane change intention at the current moment I t .

[0148] By traversing the entire trajectory data, a temporal supervision dataset containing two types of samples, "intentional" and "unintentional", is formed. It is then divided into a training set and a test set in an 8:2 ratio, which are used for structure learning, parameter learning, and performance verification of the dynamic Bayesian network, respectively.

[0149] (4) Dynamic Bayesian network structure construction

[0150] Construct a dynamic Bayesian network structure with temporal dependencies. The network nodes and dependencies are as follows:

[0151] a. Node collection

[0152] X={G,F t-T+1 , Ft-T+2 ,…,F t , I t}

[0153] b. Conditional dependency structures include:

[0154] Time-dependent paths

[0155] F t-T+1 →F t-T+2 →…→F t

[0156] Intent Identification Path

[0157] G→I t , F t →I t

[0158] Optional enhancement path

[0159] F t-k →I t (1≤k≤T)

[0160] c. Joint probability distribution

[0161]

[0162] Among them, F t-T is considered as the initial input or network input prior.

[0163] (5) Model training and real-time inference

[0164] Model training includes three stages: structure learning, parameter learning, and evaluation indicators:

[0165] a. Structural learning: Preset node dependencies based on driving behavior mechanisms, or use greedy search combined with Bayesian Information Criterion (BIC) to automatically search for the optimal structure;

[0166] b. Parameter learning: Use the maximum likelihood estimation (MLE) or expectation maximization (EM) algorithm to learn the conditional probability table (CPT);

[0167] c. Evaluation Metrics: Performance evaluation is performed using accuracy, recall, and F1 score on the test set. If F1 ≥ 0.80, the model performance is considered to meet actual deployment requirements.

[0168] During the model deployment phase, the system collects the state features of the latest T frames in real time as input sequences. Combined with the driving style label G, the current lane change intention state is inferred through a dynamic Bayesian network:

[0169]

[0170] If the probability exceeds the set threshold of 0.7, it can be determined that the current driver has the intention to change lanes, and the risk warning or auxiliary control mechanism can be triggered accordingly.

[0171] Although the specific implementation methods of the present invention are described above, those skilled in the art should understand that these are merely examples and that various changes or modifications may be made to these implementation methods without departing from the principles and implementations of the present invention. Therefore, the scope of protection of the present invention is limited by the appended claims.

Claims

1. A method for identifying lane change intention based on driving style, characterized in that: The following steps are involved: Step 1: Use a driving simulator to collect kinematic parameters of vehicles in a car-following scenario and create a multidimensional dataset S. The multidimensional dataset S contains eight features: ego vehicle speed, leading vehicle speed, ego vehicle acceleration, leading vehicle acceleration, ego vehicle position, leading vehicle position, following distance, and time. S={v,v p ,a,a p ,x,x p ,D,t} D=x p -x-L Where v is the vehicle speed; v p is the speed of the preceding vehicle; a is the acceleration of the vehicle; a p is the acceleration of the preceding vehicle; x is the position of the vehicle; x p is the position of the preceding vehicle; D is the following distance; t is the recording time; L is the length of the vehicle; Step 2: Based on the data set S from step 1, the heterogeneous car-following characteristics of the drivers are identified. Based on the expected safety margin model, the driver reaction characteristic quantitative index, steady-state risk quantitative index, and operational characteristic quantitative index are respectively calculated to obtain a parameter vector p that can effectively characterize the heterogeneous car-following characteristics of the drivers. The driver reaction characteristic quantitative index includes the average driver reaction time; the steady-state risk quantitative index includes the lower limit of the expected safety margin and the upper limit of the expected safety margin; and the operational characteristic quantitative index includes the acceleration sensitivity coefficient and the deceleration sensitivity coefficient. p=(τ n ,SM nDL ,SM nDH ,a1,a2) T Among them, τ n is the average reaction time of the driver; SM nDL is the lower limit of the expected safety margin; SM nDH is the upper limit of the expected safety margin; ɑ1 is the acceleration sensitivity coefficient; α2 is the deceleration sensitivity coefficient; Step 3: Using the parameter vector p based on the expected safety margin model obtained in Step 2, a clustering algorithm is used to classify driving styles and obtain corresponding driving style labels, including conservative, general, and aggressive. Step 4: Based on representative samples from the public trajectory dataset, a high-fidelity 3D dynamic traffic scenario was established. An uncontrolled observation experiment was conducted using a driving simulator. The driver's subjective judgment of the intention was made and annotated. The driver's lane change intention annotation data was collected to construct a driving sample set containing lane change intentions. Step 5: Using the driving style labels obtained in step 3 and the driving sample set constructed in step 4, a dynamic Bayesian network model is constructed and trained to achieve real-time recognition and dynamic modeling of lane change intentions.

2. The method for recognizing lane change intention based on driving style according to claim 1, characterized in that: In step 2, the steps of identifying the car-following characteristics of heterogeneous drivers and calculating the parameter vector p representing the car-following characteristics of heterogeneous drivers are as follows: a. Calculate the driver's reaction characteristic quantitative indicator, the average driver reaction time τ n According to the Newell car-following model, the spatiotemporal trajectory of the following car is basically the same as that of the leading car during the car-following process, with only a translation in space and time. Therefore, at each moment t n , the time delay Δt required for the following vehicle to reach the same speed as the preceding vehicle is determined and is defined as: Among them, τ h is the driver’s reaction time at time t; v(t) is the vehicle speed, v p (t n ) is the speed of the preceding vehicle at the current moment, and ε is the speed tolerance; At each time step, h After solving and screening, τ h The statistical mean of the following vehicle driver is taken as the average reaction time τ n The final identification result: Among them, N τ represents the effective τ solved during the car-following process h the number of b. Calculate the driver's steady-state risk quantitative index expected safety margin lower limit SM nDL and the expected upper limit of safety margin SM nDH First, the stable car-following data segment is filtered, and then the sliding window filter is introduced to calculate the safety margin SM within the sliding window. n The maximum and minimum values ​​of (t): in, and Represents the real-time safety margin SM and the upper and lower limits of the expected safety margin during the sliding window; c. Calculate the acceleration sensitivity coefficient α1 and deceleration sensitivity coefficient α2, which are quantitative indicators of driver operating characteristics. Based on the desired safety margin (DSM) model, the sensitivity coefficients can be derived as follows: Considering the instability of the driver's car-following behavior and the influence of data errors, the acceleration / deceleration car-following data segment should meet the requirement that the absolute value of the acceleration is greater than 0.5m / s 2 Finally, the final identification result of the sensitivity coefficient is obtained: in, and Represent the number of identified α1 and α2 respectively.

3. The method for recognizing lane change intention based on driving style according to claim 1, characterized in that: In step 3, the K-Means++ algorithm is used to simulate and analyze the driver's car-following characteristic parameters to obtain a data set divided into three clusters, that is, the driver's driving style G is divided into conservative, general and aggressive types. G∈{conservative, general, radical}.

4. The method for recognizing lane change intention based on driving style according to claim 1, characterized in that: In step 4, the high-fidelity three-dimensional dynamic traffic scene used in the uncontrolled observation experiment is generated from real vehicle trajectory data and a unified observation perspective is used on the driving simulator. Multiple subjects hold the steering wheel and independently judge the intention. When the intention to change lanes is generated, the corresponding button on the steering wheel is pressed, and the intention label I is collected and annotated.

5. The method for predicting lane change intention based on driving style recognition according to claim 1, characterized in that: In step 5, the driving style label obtained in step 3 and the driving sample set constructed in step 4 are used as input, and the driving state sequence of the fixed time window length T in the past is used, wherein the state characteristics at each moment include the state of the vehicle, the relative relationship between the vehicle and surrounding vehicles, the traffic flow information of the road section and the driving style label, and the output is whether there is a lane change intention at the current moment. During the model training process, the constructed labeled sample set is divided into a training set and a test set in an 8:2 ratio, and the conditional dependency relationship and probability distribution between variables are established using structural learning and parameter learning methods. The model performance is evaluated by accuracy, recall rate and F1 score on the test set. When the F1 score reaches the preset threshold, the model is considered to have actual deployment capabilities. The model supports an online reasoning mechanism in a sliding time window manner to achieve real-time identification and dynamic modeling of lane change intentions at the current moment. The model has the following structure: a. The temporal input feature sequence is: Among them, G is the driving style label; F t is the state feature at time t, including the state of the vehicle, the interaction features between the vehicle and the six neighboring vehicles, and the traffic flow characteristics of the road section; b. Divide the constructed labeled sample set into a training set and a test set in an 8:2 ratio. The training set is used for model structure and parameter learning, and the test set is used for model performance evaluation. c. The training goal is to learn the current intention state node I t The posterior probability Parameter estimation methods include maximum likelihood estimation or expectation maximization algorithms; d. The model performance is evaluated using precision, recall, and F1 score. A model that achieves an F1 score of at least 0.80 in the test set is considered satisfactory for practical use. e. The model supports real-time acquisition of the latest state sequence in the form of a sliding time window, completes the identification of the lane change intention state at the current moment, and outputs the posterior probability results as a decision-making aid. It is suitable for real-time identification and control response in embedded intelligent driver assistance systems.

Images