Intelligent vehicle lane changing decision-making and trajectory planning method and system fusing subjective risk perception and personalized driving style

Abstract

The invention provides an intelligent vehicle lane changing decision and trajectory planning method fusing subjective risk perception and a personalized driving style, and belongs to the technical field of intelligent driving. According to the method, deep fusion of subjective risk perception and personalized styles is realized through multi-modal data acquisition and preprocessing, subjective risk perception quantitative modeling based on physiological signals, personalized driving style modeling based on a Gaussian mixture model and LSTM, game theory fusion decision, trajectory correction under safety field constraint and integrated output. According to the method, the technical problems that subjective risks cannot be quantified, style modeling is rough and decision and planning are separated in the prior art are solved, and the anthropomorphism, the safety and the personalized adaptive capacity of the lane changing behavior are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving, and in particular to a method and system for intelligent vehicle lane-changing decision-making and trajectory planning that integrates subjective risk perception and personalized driving style. Background Technology

[0002] With the rapid development of autonomous driving technology, lane-changing behavior of intelligent vehicles in complex traffic environments has become a research hotspot in both academia and industry. Lane-changing decision-making and trajectory planning, as core functional modules of autonomous driving systems, are directly related to driving safety, traffic efficiency, and passenger comfort. Most current mainstream lane-changing decision-making algorithms are based on objective physical quantities for risk assessment, such as collision time, minimum safe distance, or deterministic safety margin. These methods, through precise mathematical modeling and sensor measurements, can ensure driving safety to a certain extent and have achieved good application results in structured road scenarios.

[0003] However, existing technologies still suffer from several insurmountable technical shortcomings. First, objective physical quantity assessment methods ignore the subjective differences in drivers' perception of risk during actual driving. Numerous studies have shown that different drivers often perceive significant differences in risk in the same traffic scenario. This difference stems from the combined influence of multiple factors, including individual physiological characteristics, driving experience, and psychological state. A lane-changing maneuver that is completely safe in terms of physical quantity measurements may be subjectively perceived as dangerous or uncomfortable by a driver, thereby reducing their trust in and willingness to use the intelligent driving system. In other words, existing technologies lack effective means to quantify and model drivers' subjective risk perception, leading to a significant discrepancy between the lane-changing behavior of intelligent vehicles and the driver's psychological expectations.

[0004] Secondly, while existing lane-change trajectory planning methods have begun to consider differences in driving habits, most remain at the level of style classification or imitation learning. A common approach is to use cluster analysis to simply categorize drivers into discrete types such as aggressive, average, and conservative, and then generate standardized lane-change trajectories based on the classification results. This approach has two fundamental problems: first, discrete classification cannot capture the dynamic and continuous changes in individual style under different traffic scenarios; second, trajectories generated using fixed templates are difficult to adapt to the real-time interaction needs of complex mixed-traffic environments. More importantly, there is an inherent correlation between driving style and risk perception, but existing methods have failed to incorporate both into a unified modeling framework, resulting in generated trajectories that are either overly safe but lack human-like qualities, or highly personalized but lack guaranteed safety.

[0005] Furthermore, existing intelligent driving systems typically process lane-changing decisions and trajectory planning as two separate modules. The decision-making module only outputs the lane-changing intention (yes or no), while the planning module independently generates the execution trajectory. This separate architecture leads to a lack of collaborative optimization: the decision-making process does not consider the feasibility of the subsequent trajectory, and the planning process cannot provide feedback to correct the decision. In a dynamic and interactive traffic environment, this fragmented approach easily leads to contradictory situations, such as conservative decisions with aggressive trajectories, or aggressive decisions with infeasible trajectories. The overall human-likeness and consistency of the behavior are difficult to meet the actual needs of human-machine co-driving.

[0006] Furthermore, most existing risk assessment models are based on threshold judgments of a single physical quantity, lacking consideration for the cumulative effect of risk in multi-vehicle interaction scenarios. During lane changing, the target vehicle needs to simultaneously pay attention to multiple interacting objects, such as the vehicle in front in the original lane, the vehicle in front in the target lane, and the vehicle behind in the target lane. The risk contributions of different objects may overlap or cancel each other out. Existing methods often only consider the risk of the most dangerous single vehicle, or aggregate the risks of multiple vehicles in a simple weighted manner, failing to truly reflect the risk perception patterns of drivers in complex scenarios. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides an intelligent vehicle lane-changing decision-making and trajectory planning method that integrates subjective risk perception and personalized driving style. The aim is to quantify the driver's subjective risk perception, construct a personalized driving style model, and achieve collaborative optimization of decision-making and trajectory under safety field constraints, thereby generating lane-changing behavior that meets both safety requirements and individual psychological expectations.

[0008] The present invention achieves the above-mentioned technical objectives through the following technical means.

[0009] A method for intelligent vehicle lane-changing decision-making and trajectory planning that integrates subjective risk perception and personalized driving style includes the following steps:

[0010] Multimodal lane change data acquisition and preprocessing: Collect driver's physiological signals, eye movement data, vehicle kinematic data and surrounding traffic environment data, calibrate the start and end times of lane changes, extract time-series segments before and after lane changes, preprocess multi-source data, and construct a standardized lane change sample library.

[0011] Quantitative modeling of driver's subjective risk perception: Constructing a collision risk function based on fault tree analysis principles to calculate the real-time collision risk values ​​between the target vehicle and surrounding vehicles. The maximum value among them is taken as the quantitative indicator of subjective risk perception. A two-layer convolutional neural network-long short-term memory network model was constructed, using physiological signals and eye-tracking data as inputs, and subjective risk perception quantification indicators as inputs. The output is used to train a mapping model from physiological reactions to subjective risk perception, and this mapping model is then used to calibrate the risk threshold of individual drivers offline. ;

[0012] Personalized driving style modeling: Based on the lateral speed curves during lane changes, a Gaussian mixture model is used to cluster various typical driving styles. These styles are then fitted to the lateral speed characteristics of individual drivers using a weighted combination method to determine the weight coefficient τ for each style. c A long short-term memory network model is constructed using real-time traffic environment characteristics as input to predict longitudinal acceleration at future moments and generate personalized lane-changing trajectories that match individual styles.

[0013] Game theory-integrated decision-making: The lane-changing decision-making process is modeled as a dynamic game between an intelligent vehicle and the vehicle behind in the target lane. A strategy set for both sides is defined, and a game payoff function incorporating safety, efficiency, and comfort is constructed. The safety indicator incorporates a real-time subjective risk perception quantification index. And using the obtained style weight coefficient τ c By customizing the weights of the payoff function, the game equilibrium is solved to obtain the lane-changing decision.

[0014] Trajectory correction under safety field constraints: Constructing a driving safety field that includes a road potential energy field, a kinetic energy field, and a behavioral field, where the behavioral field is based on individual risk thresholds. With real-time subjective risk perception quantitative indicators Bounded corrections are performed; the generated personalized lane-changing trajectory is placed in a safe field to calculate the cumulative risk value. If the risk exceeds the preset safety threshold, the trajectory is optimized and corrected under kinematic constraints with the goal of minimizing the trajectory deviation.

[0015] The final lane-changing decision and the corrected trajectory are output to the vehicle control system for execution.

[0016] Furthermore, quantitative indicators of the target vehicle's subjective risk perception Defined as:

[0017] ,

[0018] in: This indicates the collision risk between the target vehicle and the i-th interacting vehicle. ;

[0019] Where: TTC is the collision time; τ is the time normalization constant; D is the actual vehicle distance; σ D Here, DDS is the distance-normalized parameter; DDS is the minimum safe distance, expressed as:

[0020]

[0021] Where: tr For driver reaction time; v f v i These are the speeds of the vehicle in front and the vehicle behind, respectively; μ f μ i , where are the adhesion coefficients of the road surface where the front and rear vehicles are located, respectively; g is the acceleration due to gravity.

[0022] Furthermore, offline calibration of individual drivers' risk thresholds The calibration method is as follows:

[0023] The average physical risk of each lane change sample is calculated. Samples marked as high-risk or low-risk are compared with a preset threshold and used as the true labels.

[0024] Average risk predicted by the model Traversing candidate thresholds ,Will The sample was identified as high-risk, compared with the true label, the classification accuracy was calculated, and the threshold with the highest accuracy was taken as the driver's classification accuracy. .

[0025] Furthermore, the personalized driving style modeling specifically includes:

[0026] Gaussian fitting is performed on the lateral velocity curve of each lane change sample to obtain the fitting parameters [μ,σ]. The K-means algorithm is used to cluster [μ,σ] of all samples to obtain the Gaussian model G of the c-th typical driving style. c (t,p c ,q c ), where p c Let q be the mean. c Standard deviation;

[0027] The lateral velocity curve of an individual driver is represented as:

[0028] ,

[0029] In the formula, C represents the total number of typical driving styles; τ c The weighting coefficients for the c-th typical driving style satisfy the following conditions: The solution is obtained by minimizing the squared Euclidean distance after dynamic time warping of the lane-changing trajectory.

[0030] Furthermore, the game payoff function is:

[0031]

[0032] In the formula, Δv is the speed gain that can be obtained after changing lanes; Δv ref Reference velocity difference; jerk represents longitudinal impact force; jerk refFor reference impact level; weighting coefficient w j Determine by the following formula:

[0033]

[0034] in The preset weights are for the c-th style category, where j=1 corresponds to the safety weight, j=2 corresponds to the efficiency weight, and j=3 corresponds to the comfort weight.

[0035] The autonomous vehicle strategy set includes {execute lane change, abandon lane change, delay lane change}, and the implementation mechanism of the delayed lane change strategy is as follows: the initial delay time is set to T. d After the delay ends, the online assessment is re-executed; if another delay decision is received, the delay time is accumulated; when the accumulated delay exceeds the preset limit or the risk of a rear-end collision in the target lane... Exceeding individual threshold At that time, the decision was made to abandon the lane change.

[0036] Furthermore, the driving safety field is defined as follows:

[0037] ,

[0038] Where E R For the road potential energy field, E V For the kinetic energy field, E D This is the behavior field; the kinetic energy field is obtained by summing the contributions of all surrounding vehicles:

[0039] , ,

[0040] Where: Where Δx j Δy j σ represents the longitudinal and lateral distances between the vehicle and the j-th vehicle, respectively; x The vertical influence range; σ y The horizontal influence range;

[0041] The behavioral field is modified according to the following formula:

[0042] ,

[0043] In the formula, β is the individual risk threshold, which is dimensionless; β is the amplification factor; γ is the sensitivity factor; the tanh function maps the input to the (0,1) interval.

[0044] Furthermore, calculate the cumulative risk value I. e Specifically as follows:

[0045] Discretize the personalized trajectory into time series points (x) k ,y k), k=1,…,N e Each point corresponds to a discrete time, denoted by t. k express;

[0046] t k The total risk field between the vehicle and all surrounding vehicles at any given time In the formula: For the j-th surrounding vehicle at time t k The generated kinetic energy field; For time t k The behavioral field;

[0047] Calculate the cumulative risk value ;

[0048] The safety cost is obtained after normalization. T e For lane change duration, Δt k The sampling interval is denoted as .

[0049] Furthermore, if the risk exceeds a preset safety threshold, the trajectory is optimized and corrected under kinematic constraints with the goal of minimizing trajectory deviation, as follows:

[0050] Define safety likelihood: In the formula: α is the security sensitivity coefficient;

[0051] If L e <Safety likelihood threshold L min If the risk exceeds the preset safety threshold, trajectory correction will be initiated.

[0052] The optimization problem for trajectory correction is formulated as follows:

[0053] ,

[0054] The constraints include:

[0055] Safety constraints ;

[0056] Curvature constraint ;

[0057] Acceleration constraints ;

[0058] Impact constraint ;

[0059] in For the original trajectory points, For the trajectory points to be optimized, w k Let the time weight of the k-th trajectory point satisfy the following condition: .

[0060] A planning system, comprising:

[0061] Memory, which stores computer-executable instructions;

[0062] A processor configured to execute computer-executable instructions, which, when executed by the processor, enable an intelligent vehicle lane-changing decision-making and trajectory planning method that integrates subjective risk perception and personalized driving style.

[0063] The beneficial effects of this invention are as follows:

[0064] 1. The intelligent vehicle lane-changing decision-making and trajectory planning method integrating subjective risk perception and personalized driving style described in this invention achieves accurate quantification of real-time risks in multi-vehicle interaction scenarios by constructing a collision risk function based on fault tree analysis and obtaining a comprehensive collision risk index through maximum value aggregation. Based on this, an end-to-end mapping between physiological signals, eye-tracking data, and risk perception is established through a dual-layer convolutional neural network-long short-term memory network model, enabling offline calibration of individual drivers' risk thresholds. This transforms subjective risk perception, a difficult-to-quantify psychological variable, into a calculable and embeddable mathematical parameter. This solves the technical problem of the inability to quantify subjective risk perception in existing technologies and lays a theoretical foundation for achieving human-like decision-making in human-machine co-driving environments.

[0065] 2. The intelligent vehicle lane-changing decision-making and trajectory planning method described in this invention integrates subjective risk perception and personalized driving style. It uses a Gaussian mixture model to perform multi-style clustering of the lane-changing lateral velocity curve and fits the lateral velocity characteristics of individual drivers using a weighted combination method, achieving continuous quantification of driving style rather than discrete classification. Simultaneously, it combines a long short-term memory network model to predict longitudinal behavior in real time, generating a complete lane-changing trajectory that matches the individual's style. This overcomes the shortcomings of existing style modeling methods, such as discrete coarseness and poor adaptability, enabling the generated trajectory to maintain personalized characteristics while dynamically adjusting according to the real-time traffic environment, significantly improving the human-likeness of intelligent vehicle lane-changing behavior.

[0066] 3. The intelligent vehicle lane-changing decision-making and trajectory planning method of this invention, which integrates subjective risk perception and personalized driving style, models the lane-changing decision-making process as a dynamic game between the intelligent vehicle and the vehicle behind in the target lane. It introduces a real-time comprehensive collision risk index into the game's payoff function and uses style weight coefficients to personalize the safety, efficiency, and comfort weights of the payoff function. This achieves a deep integration of subjective risk perception and game-based decision-making, ensuring that lane-changing decisions not only consider physical safety but also reflect the individual driver's risk preferences and style characteristics, thus solving the problem of existing decision-making models being disconnected from driver expectations.

[0067] 4. The intelligent vehicle lane-changing decision-making and trajectory planning method of this invention, which integrates subjective risk perception and personalized driving style, constructs a driving safety field comprising a road potential energy field, a kinetic energy field, and a behavioral field. The behavioral field is bounded and corrected based on individual risk thresholds and real-time comprehensive collision risk indicators, resulting in a larger equivalent field strength corresponding to higher risk perception. By placing the personalized trajectory within the safety field to calculate the cumulative risk value and using safety likelihood as the judgment threshold, a refined assessment of trajectory safety is achieved. When the risk exceeds a preset threshold, optimization correction is performed under kinematic constraints with the goal of minimizing trajectory deviation, thereby ensuring trajectory safety while maintaining driving style. Integrating decision-making and planning into a unified framework for collaborative optimization solves the contradictory behavior problem caused by the separation of the two in existing technologies. Attached Figure Description

[0068] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings described below are some embodiments of the present invention. For those skilled in the art, it is obvious that other drawings can be obtained from these drawings without creative effort.

[0069] Figure 1 This is a flowchart of the intelligent vehicle lane-changing decision-making and trajectory planning method that integrates subjective risk perception and personalized driving style as described in this invention.

[0070] Figure 2 A schematic diagram of the subjective risk perception quantification model structure.

[0071] Figure 3 A flowchart for integrating personalized driving style modeling with game-theoretic decision-making.

[0072] Figure 4 This is a schematic diagram of the original trajectory and the corrected trajectory. Detailed Implementation

[0073] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0074] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "axial," "radial," "vertical," "horizontal," "inner," and "outer," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0075] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0076] like Figure 1 As shown, the intelligent vehicle lane-changing decision-making and trajectory planning method that integrates subjective risk perception and personalized driving style according to the present invention includes the following steps:

[0077] S01: Collect driver's physiological signals, eye movement data, vehicle kinematic data, and surrounding traffic environment data; calibrate lane change start and end times; extract time segments before and after lane changes; perform time alignment, noise reduction, and resampling on multi-source data; and construct a standardized lane change sample library. Specifically:

[0078] The experiment employed a highly realistic driving simulator equipped with a six-degree-of-freedom motion platform, a circular projection screen, and a realistic vehicle cockpit, capable of realistically recreating highway driving scenarios. Forty volunteers of varying ages (25-55 years old) and driving experience (3-30 years) were recruited, including 22 men and 18 women. All participants held valid driver's licenses and had no major traffic accident records. The experimental road was a six-lane, two-way highway with a single lane width of 3.6 meters and a moderate traffic flow density (approximately 20 vehicles / km / lane). Surrounding vehicles were controlled by an intelligent agent built into the simulator, exhibiting randomness and interactivity. Each participant completed approximately 30 lane-changing operations, including free lane changes and forced lane changes (e.g., due to a slower vehicle ahead), resulting in over 1500 valid lane-changing samples.

[0079] Data acquisition equipment includes: a TobiiProGlasses2 eye tracker (sampling frequency 100Hz) for recording eye movement data such as fixation point, blink, saccades, and pupil diameter; a Biopac MP160 physiological signal acquisition unit (ECG 256Hz, skin conductance 256Hz) for acquiring ECG and skin conductance signals via electrode patches; and a CarSim vehicle dynamics simulation platform (output frequency 20Hz) for outputting the vehicle's position, velocity, acceleration, heading angle, and the motion status of surrounding vehicles. All data is hardware synchronized via a time synchronization box to ensure time alignment of multimodal data.

[0080] The calibration rules for lane change events are as follows: Using the lane centerline as a reference, a lane change is considered to begin when the vehicle's lateral displacement exceeds half the lane width (i.e., 1.8 meters) and remains within that lane; a lane change is considered to end when the vehicle's lateral displacement stabilizes again near the new lane centerline (lateral offset less than 0.3 meters). Based on this rule, a 13-second segment is extracted from the continuous data stream, consisting of the 5 seconds before the lane change, the lane change process, and the 3 seconds after the lane change, serving as one lane change sample.

[0081] Data preprocessing includes the following steps:

[0082] Time alignment: Eye-tracking, physiological, and vehicle data are resampled along a unified absolute time axis, and cubic spline interpolation is used to unify the data of each channel to a sampling frequency of 10Hz.

[0083] Filtering and noise reduction: The ECG signal is filtered by a bandpass filter (0.5~45Hz) to remove power frequency interference and baseline drift. The skin conductance signal is filtered by a low-pass filter (cutoff frequency 1Hz) to smooth instantaneous fluctuations. The fixation point coordinates in the eye movement data are subjected to Kalman filtering to reduce noise.

[0084] Outlier removal: For data missing segments caused by equipment obstruction or abnormal operation, if the missing length exceeds 0.5 seconds, the sample is removed; otherwise, linear interpolation of adjacent data is used.

[0085] Standardization: The physiological feature sequence of each sample is normalized by z-score, and the mean and variance are calculated based on all samples.

[0086] The final standardized lane change sample library contains the following data for each lane change sample: 13 seconds of 12-dimensional physiological feature time series (ECG 4D, skin conductance 4D, eye movement 4D), 13 seconds of vehicle kinematic data (vehicle speed, lateral acceleration, longitudinal acceleration, heading angle, yaw rate), and 13 seconds of traffic environment data (surrounding vehicle numbers, relative distance, relative speed).

[0087] S02: Quantitative Modeling of Driver's Subjective Risk Perception: Based on the fault tree analysis principle, a collision risk function is constructed to calculate the real-time collision risk values ​​between the target vehicle and surrounding vehicles, and the maximum value is taken as the quantitative index of subjective risk perception. Simultaneously, a two-layer convolutional neural network-long short-term memory network model is constructed, using the aforementioned physiological signals and eye-tracking data as input and the aforementioned quantitative index of subjective risk perception as output for training, to obtain a mapping model from physiological reaction to subjective risk perception. This model is then used to offline calibrate the risk threshold of individual drivers, and the difference between the individual risk threshold and the overall sample risk threshold is defined as the safety margin, as detailed below:

[0088] S2.1: Based on the fault tree analysis principle, a subjective risk perception quantification index is constructed. Feature parameters are selected from two dimensions: collision probability and collision severity. The collision probability can be expressed as a function of collision time (TTC) and minimum safe distance (DSS). The theoretical value of the minimum safe distance is calculated using the following formula:

[0089] ,

[0090] In the formula, t r The driver's reaction time is taken as an average of 0.8 seconds based on literature statistics; v f v i The speeds (m / s) of the vehicle in front and the vehicle behind are respectively; μ f μ i These are the adhesion coefficients of the road surface where the front and rear vehicles are located, respectively. For example, 0.8 is taken for dry asphalt road surface; g is the acceleration due to gravity, taken as 9.8 m / s².

[0091] The collision risk between the target vehicle and the i-th interacting vehicle is defined as:

[0092] ,

[0093] In the formula, the first term represents time risk, where TTC is the time to collision (seconds), and the risk approaches 1 when TTC approaches 0; τ is a time standardization constant, taken as 1.5 seconds (based on the sum of the driver's average reaction time and operational delay); the second term represents distance risk, where D is the actual distance between vehicles (meters), measured by millimeter-wave radar; σ D The distance standardization parameter is set to 5 meters (approximately 1.4 times the lane width, reflecting the sensitivity of distance perception). The expression for collision risk ensures that the product of the two dimensionless terms is between 0 and 1, and that the values ​​are reasonably distributed.

[0094] In actual driving, drivers typically focus only on the one or two most dangerous vehicles in the interchange, rather than all surrounding vehicles. Therefore, this invention considers only the top n vehicles with the highest risk, taking n=2, i.e., the vehicle in front and the vehicle behind in the target lane. The subjective risk perception quantification index of the target vehicle is defined as:

[0095] ,

[0096] The reason for using the maximum value instead of a product or weighted average is that a driver's subjective risk perception is often dominated by the most dangerous interactive object, and this method can avoid the problem of inflated risk due to an increased number of surrounding vehicles. If the risk of the vehicle in front is 0.8 and the risk of the vehicle behind is 0.2, then the overall risk is taken as 0.8, which aligns with the driver's intuition that their attention is focused on the most dangerous target.

[0097] S2.2: Construct a two-layer convolutional neural network-long short-term memory network model, using the 12-dimensional physiological feature time series collected in S01 as input, and subjective risk perception quantification index. As the output, an end-to-end mapping model from physiological response to subjective risk perception is trained.

[0098] like Figure 2 As shown, the structure of the two-layer convolutional neural network-long short-term memory network model is as follows: The first layer is a one-dimensional convolutional neural network with a kernel size of 3, a stride of 1, 64 output channels, and a ReLU activation function, used to extract temporal local features; then an LSTM layer with 128 hidden units is connected, returning a sequence to capture temporal dependencies; the second layer also adopts a CNN-LSTM structure, but adds an attention mechanism to enhance interpretability. The attention layer calculates the weights at each time step and sums them to obtain the context vector; finally, a fully connected layer outputs the context vector for each time step. Predicted values. The loss function is mean squared error, the optimizer is Adam, the learning rate is 0.001, the batch size is 32, and the iteration count is 100 rounds. During training, 80% of the 1500 samples are used as the training set and 20% as the validation set. An early stopping strategy is used to prevent overfitting.

[0099] The mapping model can be used to calculate the prediction of each driver's physiological characteristics based on all lane-change samples. The sequence, and simultaneously, calculated according to physical formulas. By combining expert annotations or posterior judgments based on collision risk, the average risk of each sample during the entire lane-changing process is calculated. The data is compared with a preset threshold; if it exceeds the threshold, it is marked as a high-risk sample; otherwise, it is marked as a low-risk sample, and this is used as the true label. Then, for the model prediction... The average risk of the sequence is also calculated. By traversing possible thresholds ,Will Samples deemed high-risk are compared with true labels to calculate classification accuracy. The threshold with the highest accuracy is taken as the individual risk threshold for that driver. At the same time, for all samples The values ​​were statistically analyzed, and the 85th percentile was taken as the overall sample lane-changing risk threshold. When making decisions online, the above physical formulas are directly used to calculate real-time... This ensures causal consistency and real-time performance. Offline-calibrated individual risk thresholds. The game payoff function used in step S04 and the behavior field correction in step S05 will be implemented.

[0100] S03: Based on the lateral velocity curves of lane changes, a Gaussian mixture model is used to cluster various typical driving styles. A weighted combination method is then used to fit the lateral velocity characteristics of individual drivers, determining the weight coefficients for each style. Simultaneously, a long short-term memory network model is constructed using real-time traffic environment features as input to predict the longitudinal acceleration at future moments, generating lane-changing trajectories that match individual driving styles, as detailed below:

[0101] S3.1: Based on the lane-changing lateral velocity curve, a Gaussian function is used to fit the driver's lateral velocity characteristics, specifically:

[0102] Gaussian function fitting was performed on the lateral velocity curve of each lane change sample to obtain the sample's fitting parameters [μ, σ], where μ is the mean (corresponding to the midpoint of the lane change) and σ is the standard deviation (reflecting the duration of the lane change). The K-means algorithm was used to cluster all samples based on [μ, σ], with the number of clusters set to 3, corresponding to three typical driving styles: aggressive, normal, and conservative. The parameters of the three cluster centers after clustering are denoted as p. c and q c (c=1,2,3), representing the parameters of the Gaussian model for the three driving styles. Therefore, the Gaussian model for the c-th driving style can be expressed as:

[0103]

[0104] Among them, the radical style corresponds to a smaller p. c and smaller q c A conservative style corresponds to a larger p c and larger q c The general style is moderate.

[0105] S3.2: For new drivers, their lateral velocity curve is represented as a weighted sum of three typical style models:

[0106] ,

[0107] In the formula, τ c These are the corresponding weighting coefficients. The following constraints must be satisfied during the solution process: The weighting coefficients are determined by minimizing the squared Euclidean distance after dynamic time warping of the lane-changing trajectory, and solved using a constrained optimization algorithm (such as sequential least squares programming or projected gradient method). Specifically, for each lane-changing sample of the driver, the DTW distance between its lateral velocity curve and the weighted Gaussian mixture model is calculated, and the τ distance is optimized with the objective of minimizing the average distance of all samples. c Driving style weighting coefficient τ c It remains constant during a single drive and is recalibrated only when the driver is changed.

[0108] S3.3: Construct an LSTM-based longitudinal behavior prediction model to predict the longitudinal acceleration of a vehicle in real time at future moments.

[0109] The input features are 5-dimensional time-series data: the vehicle's speed v and its relative speed Δv with the vehicle in front. f The relative speed Δv with respect to the vehicle behind in the target lane r The time-to-head distance (THW) and time-to-collision (TTC) are considered. The output is a sequence of longitudinal accelerations for the next 3 seconds (30 time steps, 0.1 seconds each). The LSTM model uses a two-layer structure with 128 hidden units per layer, followed by a fully connected layer to output a 30-dimensional model. During training, measured accelerations are used as labels, the loss function is the mean absolute error, the optimizer is Adam, the learning rate is 0.001, and the iterations are 200 epochs.

[0110] By combining lateral and longitudinal models, a complete lane-change trajectory matching the individual's style is generated. Specifically, firstly, based on the lateral velocity curve v obtained in S3.1... x (t), integrate it to obtain the lateral position sequence; then perform a second integration on the longitudinal acceleration sequence output by the LSTM longitudinal behavior prediction model to obtain the longitudinal position sequence; finally, obtain the complete trajectory point sequence (x) by temporally mapping the lateral and longitudinal positions. k ,y k ), k=1,…,N e , where Ne =30, corresponding to lane change duration T e =3 seconds, update frequency 10Hz. The obtained driving style weighting coefficient τ c The game weights will be used for calculation in step S04. The trajectory point sequence will be used for trajectory correction under the safety field constraints in step S05.

[0111] S04: As Figure 3 As shown, the lane-changing decision-making process is modeled as a dynamic game between an intelligent vehicle and the vehicle following in the target lane. A strategy set for both sides is defined, and a game payoff function incorporating safety, efficiency, and comfort is constructed. The safety indicator incorporates the real-time subjective risk perception quantification obtained in step S02, and the style weight coefficients obtained in step S03 are used to personalize the weights of the payoff function. The game equilibrium is then solved to obtain the lane-changing decision, as detailed below:

[0112] The lane-changing decision-making process is modeled as a dynamic game between the intelligent vehicle (the autonomous vehicle) and the vehicle following in the target lane. The autonomous vehicle's strategy set is defined as {execute lane change, abandon lane change, delay lane change}, and the following vehicle's strategy set is defined as {slow down to yield, maintain speed, accelerate to block}. The payoff function comprehensively considers safety, efficiency, and comfort, with the safety index directly incorporating the real-time risk value calculated online in step S02. To ensure uniformity of dimensions and conformity to physical meaning, the game payoff function is modified as follows:

[0113]

[0114] In the formula, Δv is the speed gain (m / s) that can be obtained after lane changing. If the speed decreases after lane changing, then Δv is negative. ref For reference speed difference, take 5 m / s; sigmoid(x) = 1 / (1+e −x Map any real number to the interval (0,1) to ensure the efficiency term is always positive; jerk is the longitudinal impact force (m / s³), calculated using acceleration difference; jerk ref For reference impact intensity, 2 m / s³ is used; the exponential term penalizes large impacts to reflect comfort. Weighting coefficients w1, w2, and w3 are related to driving style:

[0115]

[0116] in The preset weights are for style category c, where j=1 corresponds to safety weight, j=2 corresponds to efficiency weight, and j=3 corresponds to comfort weight. In this embodiment, using experimental data from 40 drivers, the theoretical equilibrium under each scenario is solved using backward induction, and compared with the actual decisions of drivers. The optimized weights for the three styles are as follows:

[0117] Radical c=1: ,

[0118] General form c=2: ,

[0119] Conservative type c=3: ,

[0120] This model assumes a fully information dynamic game, meaning the vehicle can obtain the motion state of the following vehicle through onboard sensors and the vehicle-to-infrastructure (V2I) system, and estimate its style parameters (such as the following vehicle's risk threshold and style weights) using historical behavior data. To address the uncertainty in parameter estimation, a Bayesian game or robust game framework can be introduced in extended research; however, this embodiment does not consider this. The perfect Nash equilibrium of the subgame is solved using backward induction to obtain the lane-changing decision. Lane-changing decisions include executing a lane-changing decision, abandoning a lane-changing decision, and delaying a lane-changing decision; the following explanation uses the delayed lane-changing decision as an example.

[0121] When the game equilibrium calculation results in a delayed lane change, it indicates that the current situation does not present safe or sufficiently profitable conditions for a lane change, but these conditions may become available in the future. Therefore, a wait and reassessment are chosen. The specific implementation mechanism is as follows: the initial delay time is set to 2 seconds. During the delay, the vehicle maintains its original lane and continuously monitors the environment. After the delay ends, an online assessment is re-executed: real-time risk is calculated based on physical formulas. The longitudinal behavior prediction is updated based on the current environment, and a new game decision is made; if a delayed decision is obtained again, the delay time is accumulated; when the accumulated delay exceeds 5 seconds or the collision risk between the vehicle behind and the vehicle in the target lane (denoted as...) is reached... That is, the formula in step two Where i corresponds to the vehicle following in the target lane exceeding the individual threshold. At that time, the decision was made to abandon the lane change.

[0122] S05: Construct a driving safety field that includes a road potential energy field, a kinetic energy field, and a behavioral field. The behavioral field is bounded and corrected based on individual risk thresholds and real-time subjective risk perception quantification indicators. The personalized lane-changing trajectory generated in step S03 is placed in the safety field to calculate the cumulative risk value. If the risk exceeds the preset safety threshold, the trajectory is optimized and corrected under kinematic constraints with the goal of minimizing the trajectory deviation.

[0123] S5.1: Construct a driving safety field that includes a road potential energy field, a kinetic energy field, and a behavioral field. Its expression is:

[0124]

[0125] Among them, E R The road potential energy field characterizes the influence of stationary obstacles. In this embodiment, the road is a straight highway, and stationary obstacles are not considered, therefore E R =0;EV E represents the kinetic energy field, characterizing the influence of the moving vehicle. D The behavioral field characterizes the influence of driver behavior characteristics.

[0126] For the j-th surrounding vehicle, the kinetic energy field component generated at time t is defined as:

[0127]

[0128] Total kinetic energy field E V The sum of contributions to all surrounding vehicles:

[0129]

[0130] Where: Where Δx j Δy j σ represents the longitudinal and lateral distances (in meters) between the vehicle and the j-th vehicle; x For the longitudinal impact range, a typical following distance of 15 meters is taken (corresponding to medium-density traffic flow on highways, 15-25 vehicles / km / lane); σ y For the lateral impact range, take half a lane width, 1.8 meters.

[0131] To reflect the physical intuition that a higher risk perception corresponds to a larger equivalent field strength, while avoiding [the following]... Too small a value might lead to numerical problems with the field strength exploding, affecting the behavior of the field E. D Correction using bounded functions:

[0132]

[0133] In the formula, Let β be the individual risk threshold, dimensionless; β>0 is the amplification factor, taken as 2.0; γ>0 is the sensitivity factor, controlling the response rate of the field strength to excessive risk; the tanh function maps the input to the (0,1) interval, ensuring that the field strength is always bounded. max(0,⋅) guarantees that the field strength is amplified only when the perceived risk exceeds the threshold. The calibration method for γ is: let =2 The field strength is increased by about 50%, i.e., tanh(γ⋅1)=0.5, and we can solve for γ≈0.55. In this embodiment, we take γ=0.55. At this time, when the perceived risk is twice the threshold, the field strength is increased by about 55%.

[0134] S5.2: Discretize the personalized trajectory into time series points (x k ,y k ), k=1,…,N e Each point corresponds to a discrete time, denoted by t. k This is represented by the sampling frequency being fixed at 10Hz, and the interval Δt between adjacent time points. k =0.1 seconds, therefore tk =t0+k⋅0.1 (t0 is the start time of lane change);

[0135] t k The total risk field between the vehicle and all surrounding vehicles at any given time In the formula: For the j-th surrounding vehicle at time t k The generated kinetic energy field; For time t k The behavioral field.

[0136] Calculate the cumulative risk value .

[0137] S5.3: If the risk exceeds the preset safety threshold, the trajectory will be optimized and corrected under kinematic constraints with the goal of minimizing the trajectory deviation, specifically as follows:

[0138] The safety cost is obtained after normalization. T e =N e ⋅0.1 = 3 seconds;

[0139] Define safety likelihood: ,

[0140] In the formula: α is the safety sensitivity coefficient, determined based on high-risk quantiles; calculated for all lane-changing samples. Take the 90th percentile , set L at this time e =0.1, then In this example, statistical analysis was conducted using 1500 sets of samples. Therefore .

[0141] Safety likelihood threshold L min The calibration was performed using a subjective experimental method: 10 drivers were invited to subjectively rate the safety of 100 randomly selected lane-changing routes (including different risk levels) (1-5 points, with 5 points being very safe). Routes with a score ≥4 points (acceptable) were statistically analyzed and their corresponding L... e The distribution is used, and the 5th percentile of this distribution is taken as L. min This ensures that 95% of acceptable trajectories satisfy L. e ≥L min .

[0142] Therefore, when L e ≥L min At that time, the trajectory is considered safe; if L e <L min This indicates that the risk is too high, and the trajectory needs to be optimized and corrected under kinematic constraints, as follows:

[0143]

[0144] The constraints include:

[0145] Safety constraints: ;

[0146] Curvature constraint: (Corresponding to a minimum turning radius of 10 meters)

[0147] Acceleration constraints: ;

[0148] Impact constraint: ;

[0149] in For the original trajectory points, For the trajectory points to be optimized, w k Let the time weight of the k-th trajectory point satisfy the following condition: In this embodiment, weighting is performed by grouping according to time period: N e The 30 points are divided into three segments (10 points each), with total weights of 0.2, 0.5, and 0.3 for each segment. Points within each segment have equal weights, i.e., weights of 0.02, 0.05, and 0.03 respectively. A sequential quadratic programming algorithm is used to solve the problem, with the initial value set to the original trajectory. Iteration continues until the KKT conditions are met or the maximum number of iterations is reached. The corrected trajectory must satisfy all constraints and be as close as possible to the original trajectory to maintain driving style.

[0150] S06: Output the final lane change decision and the corrected trajectory to the vehicle control system for execution.

[0151] The final decision and the revised trajectory are output as lane-change commands to the vehicle's underlying control system. The output includes: lane-change decision (execute / abandon / delay), target lane indicator (left or right), and trajectory point sequence. (Including position, velocity, and acceleration), updated at a frequency of 10Hz, consistent with the sampling frequency of the preprocessed data. If the vehicle detects a safety likelihood L during execution... e Less than 0.5L min or within two consecutive cycles (0.2 seconds) L e The rate of descent exceeds If this occurs, an emergency replanning is triggered, immediately aborting the current lane change and returning to the original lane. Simultaneously, steps two through five are re-executed to generate a new decision and trajectory. A trajectory correction diagram is shown below. Figure 4 As shown.

[0152] Through the above steps, this embodiment realizes intelligent vehicle lane-changing decision-making and trajectory planning that integrates subjective risk perception and personalized driving style. Simulation experiments show that the lane-changing behavior generated by this method is superior to traditional methods based on TTC thresholds or single styles in terms of safety (no collision, low risk), comfort (small impact), and personalization (high driver subjective score), and can adapt to the differences in subjective risk perception among different drivers.

[0153] It should be understood that although this specification is described according to various embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0154] The detailed descriptions listed above are merely specific illustrations of feasible embodiments of the present invention and are not intended to limit the scope of protection of the present invention. All equivalent embodiments or modifications made without departing from the spirit of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent vehicle lane-changing decision-making and trajectory planning that integrates subjective risk perception and personalized driving style, characterized in that, Includes the following steps: Multimodal lane change data acquisition and preprocessing: Collect driver's physiological signals, eye movement data, vehicle kinematic data and surrounding traffic environment data, calibrate the start and end times of lane changes, extract time-series segments before and after lane changes, preprocess multi-source data, and construct a standardized lane change sample library. Quantitative modeling of driver's subjective risk perception: Constructing a collision risk function based on fault tree analysis principles to calculate the real-time collision risk values ​​between the target vehicle and surrounding vehicles. The maximum value among them is taken as the quantitative indicator of subjective risk perception. A two-layer convolutional neural network-long short-term memory network model was constructed, using physiological signals and eye-tracking data as inputs, and subjective risk perception quantification indicators as inputs. The output is used to train a mapping model from physiological reactions to subjective risk perception, and this mapping model is then used to calibrate the risk threshold of individual drivers offline. ; Personalized driving style modeling: Based on the lateral speed curves during lane changes, a Gaussian mixture model is used to cluster various typical driving styles. These styles are then fitted to the lateral speed characteristics of individual drivers using a weighted combination method to determine the weight coefficient τ for each style. c A long short-term memory network model is constructed using real-time traffic environment characteristics as input to predict longitudinal acceleration at future moments and generate personalized lane-changing trajectories that match individual styles. Game theory-integrated decision-making: The lane-changing decision-making process is modeled as a dynamic game between an intelligent vehicle and the vehicle behind in the target lane. A strategy set for both sides is defined, and a game payoff function incorporating safety, efficiency, and comfort is constructed. The safety indicator incorporates a real-time subjective risk perception quantification index. And using the obtained style weight coefficient τ c By customizing the weights of the payoff function, the game equilibrium is solved to obtain the lane-changing decision. Trajectory correction under safety field constraints: Constructing a driving safety field that includes a road potential energy field, a kinetic energy field, and a behavioral field, where the behavioral field is based on individual risk thresholds. With real-time subjective risk perception quantitative indicators Bounded corrections are performed; the generated personalized lane-changing trajectory is placed in a safe field to calculate the cumulative risk value. If the risk exceeds the preset safety threshold, the trajectory is optimized and corrected under kinematic constraints with the goal of minimizing the trajectory deviation. The final lane-changing decision and the corrected trajectory are output to the vehicle control system for execution.

2. The intelligent vehicle lane-changing decision-making and trajectory planning method integrating subjective risk perception and personalized driving style as described in claim 1, characterized in that, Quantitative indicators of subjective risk perception of the target vehicle Defined as: ， in: This indicates the collision risk between the target vehicle and the i-th interacting vehicle. ; Where: TTC is the collision time; τ is the time normalization constant; D is the actual vehicle distance; σ D Here, DDS is the distance-normalized parameter; DDS is the minimum safe distance, expressed as: ， Where: t r For driver reaction time; v f v i These are the speeds of the vehicle in front and the vehicle behind, respectively; μ f μ i , where are the adhesion coefficients of the road surface where the front and rear vehicles are located, respectively; g is the acceleration due to gravity.

3. The intelligent vehicle lane-changing decision-making and trajectory planning method integrating subjective risk perception and personalized driving style as described in claim 1, characterized in that, Offline calibration of individual driver risk thresholds The calibration method is as follows: The average physical risk of each lane change sample is calculated. Samples marked as high-risk or low-risk are compared with a preset threshold and used as the true labels. Average risk predicted by the model Traversing candidate thresholds ,Will The sample was identified as high-risk, compared with the true label, the classification accuracy was calculated, and the threshold with the highest accuracy was taken as the driver's classification accuracy. .

4. The intelligent vehicle lane-changing decision-making and trajectory planning method integrating subjective risk perception and personalized driving style as described in claim 1, characterized in that, The personalized driving style modeling specifically includes: Gaussian fitting is performed on the lateral velocity curve of each lane change sample to obtain the fitting parameters [μ,σ]. The K-means algorithm is used to cluster [μ,σ] of all samples to obtain the Gaussian model G of the c-th typical driving style. c (t,p c ,q c ), where p c Let q be the mean. c Standard deviation; The lateral velocity curve of an individual driver is represented as: ， In the formula, C represents the total number of typical driving styles; τ c The weighting coefficients for the c-th typical driving style satisfy the following conditions: The solution is obtained by minimizing the squared Euclidean distance after dynamic time warping of the lane-changing trajectory.

5. The intelligent vehicle lane-changing decision-making and trajectory planning method integrating subjective risk perception and personalized driving style as described in claim 1, characterized in that, The game payoff function is: ， In the formula, Δv is the speed gain that can be obtained after changing lanes; Δv ref Reference velocity difference; jerk represents longitudinal impact force; jerk ref For reference impact level; weighting coefficient w j Determine by the following formula: ， in The preset weights are for the c-th style category, where j=1 corresponds to the safety weight, j=2 corresponds to the efficiency weight, and j=3 corresponds to the comfort weight. The autonomous vehicle strategy set includes {execute lane change, abandon lane change, delay lane change}, and the implementation mechanism of the delayed lane change strategy is as follows: the initial delay time is set to T. d The online assessment will be re-executed after the delay ends; If another delay decision is received, the delay time is accumulated; when the accumulated delay exceeds the preset limit or the risk of a rear-end collision in the target lane... Exceeding individual threshold At that time, the decision was made to abandon the lane change.

6. The intelligent vehicle lane-changing decision-making and trajectory planning method integrating subjective risk perception and personalized driving style as described in claim 1, characterized in that, The driving safety field is defined as follows: ， Where E R For the road potential energy field, E V For the kinetic energy field, E D This is the behavior field; the kinetic energy field is obtained by summing the contributions of all surrounding vehicles: ， ， Where: Where Δx j Δy j σ represents the longitudinal and lateral distances between the vehicle and the j-th vehicle, respectively; x The vertical influence range; σ y The horizontal influence range; The behavioral field is modified according to the following formula: ， In the formula, β is the individual risk threshold, which is dimensionless; β is the amplification factor; γ is the sensitivity factor; the tanh function maps the input to the (0,1) interval.

7. The intelligent vehicle lane-changing decision-making and trajectory planning method integrating subjective risk perception and personalized driving style as described in claim 1, characterized in that, Calculate the cumulative risk value I e Specifically as follows: Discretize the personalized trajectory into time series points (x) k ,y k ), k=1,…,N e Each point corresponds to a discrete time, denoted by t. k express; t k The total risk field between the vehicle and all surrounding vehicles at any given time In the formula: For the j-th surrounding vehicle at time t k The generated kinetic energy field; For time t k The behavioral field; Calculate the cumulative risk value ; The safety cost is obtained after normalization. T e For lane change duration, Δt k The sampling interval is denoted as .

8. The intelligent vehicle lane-changing decision-making and trajectory planning method integrating subjective risk perception and personalized driving style according to claim 1, characterized in that, If the risk exceeds the preset safety threshold, the trajectory will be optimized and corrected under kinematic constraints with the goal of minimizing the trajectory deviation, as follows: Define safety likelihood: In the formula: α is the security sensitivity coefficient; If L e <Safety likelihood threshold L min If the risk exceeds the preset safety threshold, trajectory correction will be initiated. The optimization problem for trajectory correction is formulated as follows: ， The constraints include: Safety constraints ; Curvature constraint ; Acceleration constraints ; Impact constraint ; in For the original trajectory points, For the trajectory points to be optimized, w k Let the time weight of the k-th trajectory point satisfy the following condition: .

9. A planning system, characterized in that, include: Memory, which stores computer-executable instructions; A processor configured to execute the computer-executable instructions, which, when executed by the processor, implement the intelligent vehicle lane-changing decision and trajectory planning method according to any one of claims 1-8, which integrates subjective risk perception and personalized driving style.

Images