Ice and snow road vehicle lane-changing risk prediction method based on driving intention recognition

Abstract

The present application relates to the technical field of vehicle driving safety, in particular to an ice and snow road vehicle lane-changing risk prediction method based on driving intention recognition, comprising: collecting ice and snow road multi-source heterogeneous driving data to identify lane-changing intention; in the execution stage, based on real-time kinematics and surrounding interaction information, a modified collision time index is calculated by introducing an ice and snow adhesion coefficient correction factor, and a distance deviation index is calculated in combination with the actual stopping sight distance; the two indexes are mapped into risk exposure and severity quantitative values, a comprehensive index is obtained through fault tree analysis, and high, medium and low risk levels are divided. The method takes multi-dimensional features in the intention stage as input, uses gradient boosting decision tree to train and construct a dynamic prediction model with a graded index as a label, realizes early classification and prediction of ice and snow road lane-changing risk, and improves risk quantification accuracy and scene adaptability.

Description

Technical Field

[0001] This invention relates to the field of vehicle driving safety technology, and in particular to a method for predicting the risk of lane changing on icy and snowy roads based on driving intention recognition. Background Technology

[0002] Existing vehicle lane-changing risk prediction technologies are mostly based on ordinary road conditions. They collect data such as vehicle speed, acceleration, and relative positions of adjacent vehicles, and use indicators such as conventional collision time and minimum safe distance to assess risk. The models often rely on data from the entire lane-changing process or linear algorithms for construction. These technologies have significant limitations when applied to icy and snowy roads: the low adhesion coefficient on icy and snowy roads leads to braking delay and increased stopping sight distance; conventional TTC does not correct for the impact of friction coefficient decay; and the spacing indicator does not take into account actual stopping sight distance requirements, resulting in a large deviation between the calculated results and the actual risk. At the same time, the models mostly use data from the entire process after the lane-changing is initiated, failing to capture the early pre-action characteristics of the intent stage, making it difficult to achieve dynamic risk prediction.

[0003] In lane-changing scenarios on icy and snowy roads, the following issues need to be addressed: The execution phase lacks specific risk indicator calculation methods that consider low-adhesion characteristics, leading to distorted risk quantification; existing prediction models struggle to dynamically output graded risks based on fine-grained features from the intent phase. It is necessary to overcome the limitations of conventional indicators that are not adapted to the physical characteristics of icy and snowy roads, and the problem that model inputs rely on data from all stages and cannot utilize early signals from the intent phase. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for predicting vehicle lane-changing risks on icy and snowy roads based on driver intention recognition.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for predicting vehicle lane-changing risks on icy and snowy roads based on driving intention recognition, comprising: Collect multi-source heterogeneous driving data on icy and snowy roads to identify vehicle lane-changing intentions; When a vehicle enters the lane-changing execution phase, a corrected collision time index and a spacing deviation index based on parking sight distance are calculated based on real-time vehicle kinematic data and interaction information with surrounding vehicles. The modified collision time index and the distance deviation index based on parking sight distance are mapped to a quantitative value of risk exposure level and a quantitative value of risk severity. By combining the quantitative values ​​of the risk exposure level and the quantitative values ​​of the risk severity, a comprehensive lane-changing risk index is calculated using the fault tree analysis method. The comprehensive lane-changing risk index described in historical cases was analyzed to determine the numerical boundaries of high-risk, medium-risk, and low-risk levels. Using a gradient boosting decision tree model, multidimensional features extracted from the lane change intention stage data segment are used as input, and the comprehensive lane change risk index, which has been classified into levels, is used as a label for training to construct a dynamic risk prediction model.

[0006] As a further aspect of the present invention, the step of collecting multi-source heterogeneous driving data under icy and snowy road conditions and identifying vehicle lane-changing intentions includes: The multi-source heterogeneous driving data includes vehicle kinematics time series, driving operation time series, surrounding vehicle interaction information, driver biosignal time series, and road surface adhesion coefficient. The multi-source heterogeneous driving data is timestamped and frequency standardized to form a synchronous data stream with a unified sampling frequency. Based on the driver's eye movement pattern and the starting point of the vehicle's lateral movement, a data segment of the lane change intention stage is extracted from the synchronous data stream; From the lane change intention stage data segment, calculate and extract driving operation behavior features, physiological visual features, vehicle dynamics features and vehicle interaction features; The calculated and extracted multidimensional features are input into a multi-layer bidirectional long short-term memory neural network, which is used to identify left lane change intention, right lane change intention and lane keeping intention; The step of extracting lane-changing intention phase data segments from the synchronized data stream based on the driver's eye-tracking pattern and the vehicle's lateral movement start point specifically includes: The eye movement event of the first gaze at the side rearview mirror is identified from the time sequence of the driver's biosignals, and the timestamp of the eye movement event is recorded as the stage start point. The moment when the lateral displacement or yaw rate first undergoes a sustained change is detected from the vehicle kinematic time series, and the timestamp of the moment is recorded as the end point of the stage; The time interval between the start and end of the stage is used as a candidate lane-changing intention stage; If the duration of the candidate lane change intention phase is within a preset empirical time window, then all the synchronous data streams within the candidate lane change intention phase are extracted as the lane change intention phase data segment.

[0007] As a further aspect of the present invention, the step of calculating and extracting driving operation behavior features, physiological visual features, vehicle dynamics features, and vehicle interaction features from the lane change intention stage data segment includes: Differential and statistical analysis are performed on the driving operation time sequence to obtain the standard deviation of steering wheel angle, peak value of steering wheel angular velocity, average brake pedal opening and accelerator pedal change rate, which constitute the driving operation behavior characteristics. The eye movement signal and skin conductance signal in the time sequence of the driver's biosignal are processed to obtain the discrete entropy of the fixation point, the average speed of saccades, the average heart rate growth rate, and the cumulative area of ​​the skin conductance response, which constitute the physiological visual features. The vehicle kinematic time series is subjected to frequency domain transformation and statistics to obtain the longitudinal acceleration variance, the absolute value integral of lateral acceleration, the power spectral density of yaw rate, and the vehicle speed change trend, which constitute the vehicle dynamic characteristics. Based on the surrounding vehicle interaction information, the absolute value of the relative speed between the vehicle and the vehicle in front in the adjacent lane, the rate of change of the headway between the vehicle and the vehicle behind in the target lane, and the comprehensive intrusion index between the vehicle and multiple surrounding vehicles are calculated to constitute the vehicle interaction features.

[0008] As a further aspect of the present invention, the step of inputting the calculated and extracted multidimensional features into a multilayer bidirectional long short-term memory neural network includes: Construct a long short-term memory network structure with four hidden layers, and deploy forward propagation units and backward propagation units in each layer simultaneously; The driving operation behavior features, physiological visual features, vehicle dynamics features and vehicle interaction features are concatenated in the time dimension to form a multi-dimensional feature time sequence input with equal time steps; The multidimensional feature time-series input is simultaneously fed into the forward propagation unit and the backward propagation unit of the multilayer bidirectional long short-term memory neural network; The feature vector output by the forward propagation unit at the final time step is fused with the feature vector output by the backward propagation unit at the initial time step to obtain a comprehensive intent representation vector. The integrated intent representation vector is input into a fully connected classification layer, which outputs the classification probabilities of left lane change intent, right lane change intent, and lane keeping intent.

[0009] As a further aspect of the present invention, the calculation of the corrected collision time index and the distance deviation index based on parking sight distance includes: Obtain the longitudinal relative speed and longitudinal relative distance between this vehicle and the vehicle in front in the target lane; Calculate the maximum deceleration on icy and snowy roads based on the road surface adhesion coefficient; The parameters in the classical collision time formula are corrected using the maximum deceleration to calculate the corrected collision time index. Obtain the actual lateral distance between this vehicle and vehicles in the adjacent lane; Calculate the minimum safe stopping sight distance on icy and snowy roads based on the vehicle's current speed and the road surface adhesion coefficient. By combining the minimum safe parking sight distance with the actual road width, the required safe lateral clearance is derived. The ratio of the actual lateral spacing to the required safe lateral spacing is used as the spacing deviation index based on parking sight distance.

[0010] As a further aspect of the present invention, the modified collision time index and the distance deviation index based on parking sight distance are mapped to a quantitative value of risk exposure level and a quantitative value of risk severity, including: A critical threshold is set for the modified collision time index. When the modified collision time index value is lower than the critical threshold, it is considered that there is a collision exposure in the time dimension. A continuous quantitative value of the risk exposure level is calculated based on the degree and duration to which the modified collision time index is below the critical threshold. The analysis of the spacing deviation index based on parking sight distance shows that when the spacing deviation index value is less than one, it indicates insufficient lateral space. Based on the degree to which the distance deviation index based on parking sight distance deviates from the safety benchmark, and in conjunction with the current vehicle speed, the quantified value of the risk severity is calculated.

[0011] As a further aspect of the present invention, a comprehensive lane-changing risk index is calculated using a fault tree analysis method by combining the quantitative value of the risk exposure level and the quantitative value of the risk severity, including: The occurrence of a lane-changing collision or loss of control is taken as the top event in the fault tree analysis; The time-dimensional risk and the spatial-dimensional risk are considered as intermediate events leading to the top event; The risk exposure level quantification value exceeding the preset upper limit is regarded as the basic event of time dimension risk, and the risk severity quantification value exceeding the preset upper limit is regarded as the basic event of spatial dimension risk. Based on driving experience data on icy and snowy roads, prior failure probabilities are assigned to the basic events of the time dimension risk and the basic events of the spatial dimension risk. Based on the logical AND and OR gate relationships of the fault tree, the failure probabilities of the basic events are passed up to the upper layer and aggregated to finally calculate the occurrence probability of the top event, which is the comprehensive lane-changing risk index.

[0012] As a further aspect of the present invention, the analysis of the comprehensive lane-changing risk index of historical cases, and the delineation of numerical boundaries between high-risk, medium-risk, and low-risk levels, includes: A large number of samples of the comprehensive lane change risk index generated during lane change processes in icy and snowy road conditions were collected. The comprehensive lane change risk index samples were clustered using an unsupervised clustering algorithm, with the number of clusters set to three. Based on the clustering results, the comprehensive lane change risk index samples are divided into three disjoint sets; Calculate the maximum and minimum values ​​of the comprehensive lane-changing risk index in each set, and determine the range of index values ​​corresponding to each risk level; The set with the smallest value is defined as the low-risk level, the set with the middle value is defined as the medium-risk level, and the set with the largest value is defined as the high-risk level.

[0013] As a further aspect of the present invention, the gradient boosting decision tree model is used, with multidimensional features extracted from the lane-changing intention stage data segment as input, and the comprehensive lane-changing risk index after hierarchical classification as label for training, to construct a dynamic risk prediction model, including: All multi-dimensional features extracted from the data segment of the lane change intention stage are used as the feature set of the model. The comprehensive lane-changing risk index, which is generated at the end of the corresponding lane-changing process and has been classified into risk levels, is used as the training label for the model. A gradient boosting framework is adopted, with decision trees as base learners, to construct a multi-class prediction model; During model training, a histogram-based algorithm is used to accelerate the search for optimal split points for features, and a special processing method for category features is introduced. The multi-class prediction model is iteratively trained using training data until the model loss function converges, thus obtaining the trained dynamic risk prediction model. The construction steps of the gradient boosting decision tree model include: Initialize the multi-class prediction model by assigning an initial prediction value to each sample in the training dataset; Calculate the multi-class cross-entropy loss between the current model prediction and the training labels; Calculate the first and second gradients of each sample based on the multi-class cross-entropy loss; Based on the first and second gradients, the histogram-based algorithm is used to find the split point that maximizes the decrease in the loss function for each feature. Develop a decision tree based on the found optimal split point, and determine the weights of the leaf nodes of the tree. The newly generated decision tree is added to the model, and the predicted values ​​for all samples are updated. Repeat the steps of calculating loss, calculating gradient, finding split points, growing decision trees, and updating predictions until the preset number of iterations is reached or the loss function no longer decreases significantly.

[0014] As a further aspect of the present invention, it also includes an online application step of the risk prediction model: During actual vehicle operation, the multi-source heterogeneous driving data is collected and preprocessed in real time. Continuously monitor driving intentions, and immediately initiate the risk prediction process when a left lane change intention or a right lane change intention is detected; Extract multidimensional features from a complete intent phase window preceding the current moment; The multidimensional features are input into the trained risk dynamic prediction model; The risk dynamic prediction model outputs a predicted risk level, which corresponds to the risk state that will occur during the lane change execution phase a few seconds later.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: To address the characteristics of braking performance degradation and increased stopping sight distance under different adhesion coefficients on icy and snowy roads, risk assessment indicators are specifically modified. This adapts to the quantification needs of lane-changing risks in various icy and snowy road scenarios, including icing and snow melting. Even in icy and snowy driving environments where the road adhesion coefficient continuously changes, it accurately reflects the actual collision probability and safe distance of the vehicle. When the vehicle enters the lane-changing execution phase, based on real-time vehicle kinematic data and interaction information with surrounding vehicles, a modified collision time indicator and a distance deviation indicator based on stopping sight distance are calculated. The modified collision time indicator incorporates an icy and snowy road adhesion coefficient correction factor, adjusting relative speed or safe distance thresholds to reflect braking delay. The distance deviation indicator based on stopping sight distance calculates the difference or ratio between the current distance and the ideal value, combined with the actual stopping sight distance. This technology differs from conventional unmodified indicators, making risk quantification more closely aligned with the immediate collision probability and insufficient distance under low-adhesion conditions on icy and snowy roads, accurately capturing the dynamic risk status during the execution phase.

[0016] This technique employs a gradient boosting decision tree model, using multi-dimensional features extracted from lane-changing intention stage data segments as input and a comprehensive lane-changing risk index, after being graded, as labels for training. This differs from conventional full-stage data input or simple linear models by utilizing fine-grained features from the intention stage to capture early risk warnings. Combined with the non-linear fitting capability of gradient boosting decision trees, it achieves early dynamic grading and prediction of lane-changing risks on icy and snowy roads, improving prediction timeliness and scenario adaptability. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method for predicting vehicle lane-changing risks on icy and snowy roads based on driving intention recognition, as described in this invention. Figure 2 A flowchart for multi-source heterogeneous data acquisition and lane change intention recognition; Figure 3 A flowchart for multi-dimensional feature extraction during the lane change intention stage; Figure 4Map showing the risk exposure level of lane changing under different road surface adhesion coefficients; Figure 5 Heatmap of the performance of the lane change risk prediction model. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and 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, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] See Figure 1 Multi-source heterogeneous driving data from icy and snowy roads is collected to identify vehicle lane-changing intentions. When a vehicle enters the lane-changing execution phase, a corrected collision time index and a distance deviation index based on stopping sight distance are calculated based on real-time vehicle kinematic data and interaction information with surrounding vehicles. These are then mapped to quantified risk exposure levels and risk severity. A comprehensive lane-changing risk index is calculated using fault tree analysis, combining the quantified risk exposure levels and risk severity. The comprehensive lane-changing risk index from historical cases is analyzed to define numerical boundaries for high, medium, and low risk levels. A gradient boosting decision tree model is used, with multi-dimensional features extracted from the lane-changing intention phase data segments as input and the graded comprehensive lane-changing risk index as labels, for training, thereby constructing a dynamic risk prediction model.

[0021] See Figure 2In one embodiment of the present invention, the collected multi-source heterogeneous driving data is acquired by vehicle bus, millimeter-wave radar, camera, driver state monitoring system, and road perception sensor. Vehicle kinematic time series include velocity, longitudinal acceleration, lateral acceleration, and yaw rate; driving operation time series includes steering wheel angle, steering wheel torque, brake pedal opening, and accelerator pedal opening; surrounding vehicle interaction information includes relative distance, relative speed, and azimuth angle sensed by onboard sensors; driver biosignal time series includes eye movement trajectory, pupil diameter, heart rate, and skin conductance signal; the road adhesion coefficient is estimated online through onboard road condition sensors or based on a vehicle dynamics inversion model. During actual driving on icy and snowy roads, the road adhesion coefficient dynamically changes with the degree of icing, snow thickness, and snow melting status of the driving section. The online estimation process captures these changes in real time, providing basic parameters that fit the actual working conditions of the current icy and snowy road surface for subsequent risk indicator calculation and risk level prediction. Simultaneously, to address the sensor detection bias problem that is prone to occur on icy and snowy roads, the collected road condition data undergoes multi-source verification and filtering to ensure the accuracy of the adhesion coefficient data. These data from different sources have different sampling frequencies and transmission delays; for example, vehicle bus data is sampled at 100 Hz, while eye tracker data is sampled at 60 Hz. In some embodiments, data timestamp alignment is synchronized using timestamps based on global positioning system timing, and frequency normalization uses an interpolation algorithm to resample low-frequency data to the same frequency as the highest frequency data source, forming a unified synchronized data stream.

[0022] Extracting lane-change intention phase data segments from the synchronous data stream based on the driver's eye movement pattern and the starting point of the vehicle's lateral movement can be understood as a process of identifying and extracting specific time windows related to the generation of lane-change intentions from continuous driving data. In specific implementation, identifying the first eye movement event of gazing at the side mirror from the driver's biosignal time sequence is achieved by analyzing the gaze point coordinate sequence provided by the eye tracker. When the gaze point is located in a predefined in-vehicle rearview mirror area for several consecutive frames, and no gaze at the same area has occurred within a preset time period prior to this gaze, it is determined as a new side mirror gaze event, and the timestamp of this event is recorded as the starting point of the lane-change intention phase. Detecting the moment when the first sustained change in lateral displacement or yaw rate occurs from the vehicle kinematic time sequence is achieved by applying a change point detection algorithm to the lateral displacement or yaw rate signal. When the first-order difference or fitting residual of the signal exceeds a fixed threshold and remains for more than a minimum duration, it is determined as the start of lateral movement of the vehicle, and the timestamp of this moment is recorded as the end point of the lane-change intention phase. The time interval between the start and end of a lane-change intention phase is used as a candidate phase. If the duration of a candidate lane-change intention phase falls within a preset empirical time window, such as 0.5 to 3 seconds, all synchronous data streams within the candidate phase are extracted as the final lane-change intention phase data segment. Optionally, the empirical time window range can be calibrated based on a large amount of real-vehicle test data to filter out false candidate phases caused by unintentional driver glances or slight vehicle movement. On icy and snowy roads, drivers are more cautious due to the slippery surface, and vehicles are also prone to slight lateral movement due to uneven road surface adhesion. The calibration of this empirical time window fully considers the driver's operating habits and vehicle motion characteristics on icy and snowy roads, avoiding misinterpreting the driver's cautious observation actions as lane-change intentions and filtering out false lateral movement signals caused by road factors.

[0023] The calculation and extraction of driving operation behavior features, physiological visual features, vehicle dynamics features, and vehicle interaction features from the lane-changing intention stage data segment aims to transform the raw time-series data into a multi-dimensional static vector capable of representing the intention pattern. The calculated and extracted multi-dimensional features are input into a multi-layer bidirectional long short-term memory neural network for intention recognition. In specific implementations, this neural network receives a feature sequence formed by the evolution of the aforementioned features over time within the intention stage. In some embodiments, the neural network structure includes an input layer, four hidden layers, and an output layer. Each hidden layer consists of forward long short-term memory units and backward long short-term memory units. For example, for an intention stage data segment of length T time steps, at each time step t, the input vector is formed by concatenating all the features extracted at time t, creating a feature sequence. This sequence is simultaneously fed into the forward propagation unit and the backward propagation unit. The forward propagation unit processes the sequence from t=1 to t=T, and the backward propagation unit processes the sequence from t=T to t=1. Finally, the feature vector output by the forward propagation unit at the final time step T is fused with the feature vector output by the backpropagation unit at the initial time step 1. The fusion method can be vector concatenation or weighted summation to obtain a comprehensive intent representation vector. Optionally, to improve the model's ability to capture dependencies at different time scales, an attention mechanism can be introduced into some hidden layers. This comprehensive intent representation vector can be understood as aggregating information from the entire intent phase context. This comprehensive intent representation vector is input into a fully connected classification layer and normalized to output the classification probabilities of left lane change intent, right lane change intent, and lane keeping intent. The classification process can be described by the following formula:

[0024] Where: symbol This represents a three-dimensional probability vector, whose three components correspond to the predicted probabilities of left lane change, right lane change, and lane keeping, respectively; (symbol) This represents the feature vector output by the forward long short-term memory propagation chain at the final time step; symbol This represents the feature vector output by the backward long short-term memory propagation chain at the initial time step; symbol and These are the trainable weight matrices corresponding to the forward and backward eigenvectors, respectively; (symbol) Represents a trainable bias vector; function Used to transform the results of linear combinations into probability distributions.

[0025] See Figure 3In one embodiment of the present invention, the driving operation time sequence is differentiated and statistically analyzed to obtain driving operation behavior characteristics. For example, the standard deviation of the steering wheel angle is obtained by calculating the sample standard deviation of the steering wheel angle sequence during the intention phase; the peak value of the steering wheel angular velocity is obtained by performing a first-order difference on the steering wheel angle sequence to obtain the steering wheel angular velocity sequence and taking its maximum value. The average brake pedal opening is obtained by calculating the arithmetic mean of the brake pedal opening sequence during the intention phase; the accelerator pedal change rate is obtained by performing a first-order difference on the accelerator pedal opening sequence and calculating the average absolute value of the sequence. The eye movement signals and skin conductance signals in the driver's biosignal time sequence are processed to constitute physiological visual features. The gaze point discrete entropy is obtained by statistically analyzing the distribution of the driver's gaze point on different preset regions of interest during the intention phase and calculating the Shannon entropy; the average saccade speed is obtained by identifying saccade events in the eye movement signal and calculating the average speed of all saccade events. The average heart rate growth rate is calculated by dividing the difference in heart rate between the end and start of the intention phase by the phase duration. The cumulative area of ​​skin conductance response is calculated by subtracting the baseline value from the skin conductance signal during the intention phase and integrating over time. Frequency domain transformation and statistics are performed on the vehicle kinematic time series to construct vehicle dynamics characteristics. The longitudinal acceleration variance is obtained by calculating the sample variance of the longitudinal acceleration sequence during the intention phase. The absolute value integral of lateral acceleration is obtained by numerically integrating the absolute value of the lateral acceleration sequence over time. The power spectral density of yaw rate is obtained by performing a fast Fourier transform on the yaw rate sequence and calculating the square of the spectral amplitude. The vehicle speed change trend is characterized by the slope of the fitted straight line obtained by linearly fitting the vehicle speed sequence during the intention phase. Based on surrounding vehicle interaction information, vehicle interaction characteristics are calculated. The absolute value of the relative speed between the vehicle and the vehicle in front in the adjacent lane is obtained by acquiring the relative speed of the vehicle closest to the vehicle in the target lane and calculating its absolute value. The rate of change of headway between the vehicle and the vehicle behind in the target lane is obtained by calculating the headway sequence between the vehicle and the nearest vehicle behind in the target lane and averaging its first-order difference sequence. The comprehensive intrusion index between the vehicle and surrounding multiple vehicles is obtained by calculating the weighted sum of the intrusion risk posed to the vehicle by all vehicles within a certain range around the vehicle based on their relative positions and speeds. The weighting coefficients are related to the distance and relative speed of the vehicles.

[0026] A Long Short-Term Memory (LSTM) network structure with four hidden layers is constructed, and forward and backward propagation units are deployed simultaneously in each layer. In the specific implementation, the input gate, forget gate, output gate, and activation function of candidate cell states for each LSM unit adopt the sigmoid function or tanh function, and the network weights are initialized using the Xavier initialization method. Driving operation behavior features, physiological visual features, vehicle dynamics features, and vehicle interaction features are concatenated in the time dimension to form a multidimensional feature time sequence input with equal time steps. It can be understood that the dimension of all feature vectors is unified during extraction, so that the feature vector at each time step is composed of sub-vectors of fixed dimensions. In some embodiments, if data from a feature source is missing at a certain time step, linear interpolation or the previous valid value is used to fill in the missing data to ensure the continuity of the feature time sequence. The multidimensional feature time sequence input is simultaneously fed into the forward and backward propagation units of the multilayer bidirectional LSM neural network. The forward propagation unit processes data from the start time step to the end time step, and the backward propagation unit processes data from the end time step to the start time step. For each hidden layer in the network, the output of its forward propagation unit is combined with the output of its backward propagation unit at the corresponding time step.

[0027] The feature vector output by the forward propagation unit at the final time step is fused with the feature vector output by the back propagation unit at the initial time step. In some embodiments, the fusion operation uses vector concatenation. The synthesized intent representation vector is input into a fully connected classification layer, which outputs the classification probabilities of left lane change intent, right lane change intent, and lane keeping intent. Optionally, to prevent overfitting, a dropout layer can be introduced before the fully connected classification layer, or weight decay can be used during network training. The calculation process of the classification probabilities can be described by the following formula:

[0028] Where: symbol Represents the comprehensive intent representation vector after processing by the activation function; symbol Represents the hyperbolic tangent activation function; symbol It is the weight matrix multiplied by the forward final state vector; symbol This indicates that the forward long short-term memory propagation chain reaches its final time step. Output hidden state vector; symbol It is the weight matrix multiplied by the backward initial state vector; symbol This indicates the backward long short-term memory propagation chain at the initial time step. Output hidden state vector; symbol This represents a trainable bias vector. It can be understood that a vector... The data is then fed into a fully connected classification layer for the final intent category probability calculation. The weights and biases of the fully connected classification layer are independent of the matrix in the above formula. ,matrix sum vector Another set of parameters. Optionally, the fully connected classification layer can contain multiple sub-layers to increase the model's non-linear expressiveness.

[0029] In one embodiment of the present invention, the longitudinal relative speed and longitudinal relative distance between the vehicle and the preceding vehicle in the target lane are obtained through an onboard millimeter-wave radar or lidar sensor. The sensor provides a target list at a fixed frequency. The vehicle closest to the vehicle in the target lane is selected from the list, and its relative speed and relative distance to the vehicle are measured. The maximum deceleration under icy and snowy road conditions is calculated based on the road adhesion coefficient. The road adhesion coefficient is obtained by an onboard road surface recognition sensor or by estimating data from the vehicle's anti-lock braking system and electronic stability control system. The maximum deceleration is calculated by multiplying the road adhesion coefficient by the gravitational acceleration and then by a safety factor. The parameters in the classic collision time formula are corrected using the maximum deceleration. The classic collision time formula is the longitudinal relative distance divided by the absolute value of the longitudinal relative speed. The correction method is to replace the longitudinal relative speed in the denominator with a term that considers the maximum braking capacity of the vehicle. The specific formula for calculating the corrected collision time index is as follows:

[0030] Where: symbol Represents the corrected collision time index; symbol Represents the longitudinal relative distance between this vehicle and the vehicle in front in the target lane; symbol This represents the longitudinal relative speed between the vehicle and the vehicle in front in the target lane. The relative speed is positive when the vehicle's speed is greater than the vehicle in front's speed; (symbol) This represents the maximum deceleration on icy and snowy roads calculated based on the road surface adhesion coefficient, and is a positive value; the function Used to select the larger of two parameters; symbol This indicates the square root operation. It can be understood that when the longitudinal relative speed is large, the corrected collision time index is close to the classic collision time; when the longitudinal relative speed is small but the distance is short, the square root term in the denominator of the formula dominates. This correction reduces the unreasonable situation of infinitely large collision time in low-speed, close-range scenarios due to near-zero relative speed, and more accurately reflects the time urgency under limited braking capacity on icy and snowy roads.

[0031] The actual lateral distance between the vehicle and vehicles in adjacent lanes is calculated by combining the position information of adjacent vehicles sensed by onboard sensors with the vehicle's positioning information. The sensors can detect vehicles in adjacent lanes whose longitudinal position overlaps with the vehicle's and measure their lateral distance from the vehicle's side. The minimum safe stopping sight distance on icy and snowy roads is calculated based on the vehicle's current speed and the road surface adhesion coefficient. This minimum safe stopping sight distance includes the distance traveled within the driver's reaction time and the distance required to stop using maximum deceleration. The required safe lateral distance is derived by combining the minimum safe stopping sight distance with the actual road width. The derivation process considers possible lateral deviations of the vehicle within the minimum safe stopping sight distance. The required safe lateral distance is directly proportional to the minimum safe stopping sight distance and inversely proportional to the actual road width. The ratio of the actual lateral distance to the required safe lateral distance is used as a spacing deviation index based on the stopping sight distance. When this ratio is greater than or equal to one, it indicates that the actual lateral space meets the offset margin required for safe parking; when the ratio is less than one, it indicates insufficient lateral space.

[0032] A critical threshold for a modified collision time index is set, referencing the standard collision time threshold on dry roads and amplified to account for the increased braking distance on icy and snowy roads. When the modified collision time index value falls below the critical threshold, collision exposure in the time dimension is considered to exist. A continuous risk exposure level quantification is calculated based on the degree and duration of the modified collision time index falling below the critical threshold. This quantification can be designed as a monotonically increasing function of the magnitude of the index falling below the threshold versus time. A spacing deviation index based on stopping sight distance is analyzed; a spacing deviation index value less than one indicates insufficient lateral space. Based on the degree to which the spacing deviation index deviates from the safety benchmark, combined with the current vehicle speed, a risk severity quantification is calculated. In some embodiments, the risk severity quantification can be expressed as a combination function of the spacing deviation index value and the current vehicle speed; for example, when the spacing deviation index value is fixed, the higher the vehicle speed, the greater the risk severity quantification. It can be understood that the risk exposure level quantification focuses on measuring time urgency, while the risk severity quantification focuses on a comprehensive measurement of insufficient space and energy levels.

[0033] In one embodiment of the present invention, lane-changing collision or loss of control is used as the top event in fault tree analysis. The event where a vehicle makes physical contact with surrounding vehicles or deviates from its predetermined trajectory due to instability during a lane-changing process is defined as the top event. Time-dimensional risk and spatial-dimensional risk are considered intermediate events leading to the top event. Time-dimensional risk refers to events where the risk of collision increases due to time constraints, and spatial-dimensional risk refers to events where the risk of collision or loss of control increases due to insufficient space. A risk exposure level exceeding a preset upper limit is considered a basic event of time-dimensional risk. The preset upper limit is a numerical threshold set based on the statistical distribution of risk exposure level quantification values ​​or expert experience. A risk severity quantification value exceeding a preset upper limit is considered a basic event of spatial-dimensional risk. The preset upper limit is a numerical threshold set based on the statistical distribution of risk severity quantification values ​​or expert experience. Based on driving experience data on icy and snowy roads, prior failure probabilities are assigned to the basic events of time-dimensional and spatial-dimensional risks. Experience data can be derived from historical accident databases, natural driving research data, or driving simulator experimental data. The prior failure probabilities reflect the probability of each basic event occurring individually under icy and snowy road conditions. Based on the logical AND and OR gate relationships of the fault tree, the failure probabilities of basic events are passed up to the next layer and aggregated to finally calculate the occurrence probability of the top event. In some embodiments, the logical relationships between the top event and intermediate events, and between intermediate events and basic events, are represented by Boolean algebra AND and OR gates. For example, the occurrence of the top event is defined as the simultaneous occurrence of a time-dimensional risk event and a spatial-dimensional risk event (logical AND), or the occurrence of any one of these events leads to the occurrence of the top event (logical OR). The specific logical relationship is chosen based on the analysis of the causal mechanism of lane-changing accidents on icy and snowy roads. The final calculated occurrence probability of the top event is the comprehensive lane-changing risk index.

[0034] A large number of comprehensive lane-change risk index samples were collected from lane-changing operations performed on icy and snowy roads. These samples were obtained through actual road tests or high-fidelity simulators, involving numerous lane-change operations with different traffic scenarios and driving styles, and the comprehensive lane-change risk index for each operation was calculated according to the aforementioned steps. An unsupervised clustering algorithm was used to cluster the comprehensive lane-change risk index samples, with a maximum of three clusters. In practice, the unsupervised clustering algorithm could be K-means, spectral clustering, or Gaussian mixture model. Based on the clustering results, the comprehensive lane-change risk index samples were divided into three disjoint sets, with each sample belonging to only one set. The maximum and minimum values ​​of the comprehensive lane-change risk index in each set were calculated to determine the range of index values ​​corresponding to each risk level. The set with the smallest value was defined as low-risk, the set with intermediate values ​​as medium-risk, and the set with the largest value as high-risk. It can be understood that the clustering process is based on the inherent similarity of the samples in terms of risk index values, without relying on pre-set thresholds, making the level classification more data-driven. Table 1 shows an example of the numerical boundary of risk level determined after clustering a set of samples using the K-means algorithm.

[0035] Table 1: Example of Lane Change Risk Level Classification Based on Cluster Analysis

[0036] In practice, the comprehensive lane-changing risk index can be calculated using fault tree logic and prior failure probabilities. Assuming the failure probability of the basic risk event in the time dimension is... The probability of failure of basic events in spatial dimension risk is Furthermore, the occurrence of a top event is defined as the simultaneous occurrence of two basic events (logical AND relationship), thus the comprehensive lane-changing risk index... The calculation formula is:

[0037] Where: symbol Represents the comprehensive lane-changing risk index; symbol The probability of failure of a basic risk event representing the time dimension; symbol The failure probability of a basic risk event in the spatial dimension; symbol This represents a multiplication operation. It can be understood as probability. and Instead of being directly determined by a binary judgment exceeding a threshold, the comprehensive lane-changing risk index obtains a continuous probability value between 0 and 1 by inputting the quantified values ​​of risk exposure level and risk severity into a preset membership function or probability distribution function. It is also a continuous probability value. Optionally, if the top event logic relationship is defined as an "OR" gate, then the calculation formula is: The range of exponent values ​​in Table 1 is based on a large number of calculations. The result obtained after clustering the value samples.

[0038] See Figure 4 This is a map showing the risk exposure level of lane changes under different road surface adhesion coefficients. It illustrates the impact of different road surface adhesion coefficients on the quantitative value of risk exposure during lane changes on icy and snowy roads. The horizontal axis represents the lane change process time (0–10 seconds), and the vertical axis represents the quantitative value of risk exposure. The black dashed line represents the critical threshold of risk exposure (0.5). The lower the adhesion coefficient, the higher the overall risk exposure level, and it remains above the critical threshold of 0.5 for most of the time, indicating that the collision risk in the time dimension is consistently high when changing lanes on icy roads. The higher the adhesion coefficient, the lower the overall risk exposure level, and it remains below the critical threshold, indicating that the collision risk in the time dimension is controllable when changing lanes on snow-melting or dry roads. When the risk exposure level exceeds 0.5, it means that the collision exposure risk in the time dimension has increased significantly, requiring immediate intervention from the driver or assistance systems.

[0039] In one embodiment of the present invention, all multidimensional features extracted from the lane-changing intention stage data segment are used as the feature set of the model. These multidimensional features include, but are not limited to, driving operation behavior features, physiological visual features, vehicle dynamics features, and vehicle interaction features. All features need to be standardized before model training to eliminate the influence of dimensions. The comprehensive lane-changing risk index, which corresponds to the final lane-changing process and has been classified into risk levels, is used as the training label of the model. The training label is a discrete category label, for example, 0 represents low risk level, 1 represents medium risk level, and 2 represents high risk level. A gradient boosting framework is adopted, using decision trees as base learners to construct a multi-class prediction model. The gradient boosting framework corrects the prediction error of the previous tree by iteratively adding decision trees. During model training, a histogram-based algorithm is used to accelerate the search for optimal split points for features. The histogram-based algorithm discretizes continuous feature values ​​into a series of buckets and finds split points by traversing the boundaries of the buckets instead of all feature values, thereby improving training speed. A category-specific processing method is introduced, either by directly splitting category features based on the category or by sequentially encoding category features before training. The multi-class prediction model is iteratively trained using training data until the model's loss function converges. The loss function is usually a multi-class log loss. The trained model can directly output the predicted risk level based on the multi-dimensional features of the input intent stage.

[0040] The construction steps of a gradient boosting decision tree model include initializing the multi-class prediction model by assigning an initial predicted value to each sample in the training dataset. The initial predicted value can be set to the marginal log probability of the labels of all training samples. The multi-class cross-entropy loss is calculated between the current model's prediction and the training labels; this loss measures the difference between the probability distribution predicted by the model and the true label distribution. The first and second gradients of each sample are calculated based on the multi-class cross-entropy loss. The first gradient is the partial derivative of the loss function with respect to the current model's predicted value, and the second gradient is the second partial derivative of the loss function with respect to the current model's predicted value. Based on the first and second gradients, a histogram-based algorithm is used to find the split point that maximizes the decrease in the loss function for each feature. Finding the split point involves traversing all features and their possible split points at each node of the decision tree, calculating the loss gain after each split. The decision tree is then grown based on the found optimal split point, and the weights of the leaf nodes are determined. The leaf node weights are determined to minimize the loss function and are typically calculated from the first and second gradients of the samples falling into that leaf node. The newly generated decision tree is added to the model, and the predicted values ​​for all samples are updated. The update of the predicted values ​​is achieved by multiplying the predicted value of the new decision tree by a learning rate parameter and then adding it to the original predicted value. The steps of calculating loss, calculating gradient, finding split points, growing decision trees, and updating predicted values ​​are repeated until the preset number of iterations is reached or the loss function no longer decreases significantly.

[0041] During actual vehicle operation, multi-source heterogeneous driving data is collected and preprocessed in real time. The collection and preprocessing process is consistent with the training data preparation stage, ensuring consistency between online and offline feature extraction. Driving intentions are continuously detected; upon identifying a left or right lane change intention, the risk prediction process is immediately initiated. Multidimensional features are extracted from a complete intention phase window preceding the current moment. The start and end points of the intention phase window are dynamically determined based on the real-time identified driver eye movement pattern and the vehicle's lateral movement start point. The feature extraction calculation method is identical to that used in the training phase. These multidimensional features are input into the trained dynamic risk prediction model, which operates using forward inference. The dynamic risk prediction model outputs a predicted risk level, corresponding to the risk state that will occur several seconds later during the lane change execution phase. The predicted risk level can provide warning information to the driver through the in-vehicle human-machine interface in the form of visual, auditory, or tactile feedback. Considering the characteristics of driving conditions where visibility is easily affected by wind, snow, and road glare on icy and snowy roads, tactile warnings use steering wheel vibration and seat side bolster pushes; auditory warnings use low-frequency continuous alerts; and visual warnings use large-size, high-brightness color indicators to ensure that the warning information can be quickly perceived by the driver amidst the complex visual and environmental interference of icy and snowy roads, allowing for timely lane changes or abandoning lane changes. For example, a green alert is displayed for low risk, a yellow alert for medium risk, and a red alert accompanied by an audible warning for high risk. In some embodiments, the calculation of the prediction model can be described by the formula:

[0042] Where: symbol Represents the risk level category ultimately predicted by the model; symbol This represents the category that maximizes the subsequent summation terms. As output; symbol Represents the total number of decision trees in a gradient boosting decision tree model; symbol Represents the learning rate parameter; symbol Representing the Each decision tree is used for the input feature vector In category Output value on; symbol This represents a multi-dimensional feature vector extracted from the real-time lane change intent window. It can be understood that the model inputs this feature vector during inference. Each decision tree outputs a score for each possible category. The scores of all trees for the corresponding category are multiplied by the learning rate and then summed. The category with the largest sum is selected as the predicted risk level.

[0043] See Figure 5This is a heatmap of the lane change risk prediction model's performance, used to evaluate the model's performance on icy and snowy roads. The high-risk recall rate is outstanding; out of real high-risk samples, 30 were correctly identified as high-risk, resulting in a recall rate of 76.9%. This is crucial in safety-critical scenarios, effectively warning of high-risk lane change behaviors. Six real medium-risk samples were misclassified as high-risk, constituting "over-warning" and potentially leading to unnecessary driving intervention. Three real high-risk samples were misclassified as medium-risk, constituting "missed warnings" and posing a safety hazard requiring focused optimization. Adding features with high distinguishability between medium and high risks, such as vehicle lateral acceleration and relative speed with surrounding vehicles, is recommended. To address the issue of missed high-risk detections, adjusting the classification threshold or using a weighted loss function can increase the attention given to high-risk samples.

[0044] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for predicting vehicle lane-changing risks on icy and snowy roads based on driver intent recognition, characterized in that, include: Collect multi-source heterogeneous driving data on icy and snowy roads to identify vehicle lane-changing intentions; When a vehicle enters the lane-changing execution phase, a corrected collision time index and a spacing deviation index based on parking sight distance are calculated based on real-time vehicle kinematic data and interaction information with surrounding vehicles. The modified collision time index and the distance deviation index based on parking sight distance are mapped to a quantitative value of risk exposure level and a quantitative value of risk severity. By combining the quantitative values ​​of the risk exposure level and the quantitative values ​​of the risk severity, a comprehensive lane-changing risk index is calculated using the fault tree analysis method. The comprehensive lane-changing risk index described in historical cases was analyzed to determine the numerical boundaries of high-risk, medium-risk, and low-risk levels. Using a gradient boosting decision tree model, the multidimensional features extracted from the lane change intention stage data segment are used as input, and the comprehensive lane change risk index after level division is used as the label to train a risk dynamic prediction model. The calculated and corrected collision time index and the distance deviation index based on parking sight distance include: Obtain the longitudinal relative speed and longitudinal relative distance between this vehicle and the vehicle in front in the target lane; Calculate the maximum deceleration on icy and snowy roads based on the road surface adhesion coefficient; The parameters in the classical collision time formula are corrected using the maximum deceleration to calculate the corrected collision time index. Obtain the actual lateral distance between this vehicle and vehicles in the adjacent lane; Calculate the minimum safe stopping sight distance on icy and snowy roads based on the vehicle's current speed and the road surface adhesion coefficient. By combining the minimum safe parking sight distance with the actual road width, the required safe lateral clearance is derived. The ratio of the actual lateral spacing to the required safe lateral spacing is used as the spacing deviation index based on parking sight distance.

2. The method for predicting vehicle lane-changing risks on icy and snowy roads based on driving intention recognition as described in claim 1, characterized in that, The process of collecting multi-source heterogeneous driving data on icy and snowy roads to identify vehicle lane-changing intentions includes: The multi-source heterogeneous driving data includes vehicle kinematics time series, driving operation time series, surrounding vehicle interaction information, driver biosignal time series, and road surface adhesion coefficient. The multi-source heterogeneous driving data is timestamped and frequency standardized to form a synchronous data stream with a unified sampling frequency. Based on the driver's eye movement pattern and the starting point of the vehicle's lateral movement, a data segment of the lane change intention stage is extracted from the synchronous data stream; From the lane change intention stage data segment, calculate and extract driving operation behavior features, physiological visual features, vehicle dynamics features and vehicle interaction features; The calculated and extracted multidimensional features are input into a multi-layer bidirectional long short-term memory neural network, which is used to identify left lane change intention, right lane change intention and lane keeping intention; The step of extracting lane-changing intention phase data segments from the synchronized data stream based on the driver's eye-tracking pattern and the vehicle's lateral movement start point specifically includes: The eye movement event of the first gaze at the side rearview mirror is identified from the time sequence of the driver's biosignals, and the timestamp of the eye movement event is recorded as the stage start point. The moment when the lateral displacement or yaw rate first undergoes a sustained change is detected from the vehicle kinematic time series, and the timestamp of the moment is recorded as the end point of the stage; The time interval between the start and end of the stage is used as a candidate lane-changing intention stage; If the duration of the candidate lane change intention phase is within a preset empirical time window, then all the synchronous data streams within the candidate lane change intention phase are extracted as the lane change intention phase data segment.

3. The method for predicting vehicle lane-changing risks on icy and snowy roads based on driving intention recognition as described in claim 2, characterized in that, The calculation and extraction of driving operation behavior features, physiological visual features, vehicle dynamics features, and vehicle interaction features from the lane change intention stage data segment includes: Differential and statistical analysis are performed on the driving operation time sequence to obtain the standard deviation of steering wheel angle, peak value of steering wheel angular velocity, average brake pedal opening and accelerator pedal change rate, which constitute the driving operation behavior characteristics. The eye movement signal and skin conductance signal in the time sequence of the driver's biosignal are processed to obtain the discrete entropy of the fixation point, the average speed of saccades, the average heart rate growth rate, and the cumulative area of ​​the skin conductance response, which constitute the physiological visual features. The vehicle kinematic time series is subjected to frequency domain transformation and statistics to obtain the longitudinal acceleration variance, the absolute value integral of lateral acceleration, the power spectral density of yaw rate, and the vehicle speed change trend, which constitute the vehicle dynamic characteristics. Based on the surrounding vehicle interaction information, the absolute value of the relative speed between the vehicle and the vehicle in front in the adjacent lane, the rate of change of the headway between the vehicle and the vehicle behind in the target lane, and the comprehensive intrusion index between the vehicle and multiple surrounding vehicles are calculated to constitute the vehicle interaction features.

4. The method for predicting vehicle lane-changing risks on icy and snowy roads based on driving intention recognition as described in claim 3, characterized in that, The step of inputting the calculated and extracted multidimensional features into a multilayer bidirectional long short-term memory neural network includes: Construct a long short-term memory network structure with four hidden layers, and deploy forward propagation units and backward propagation units in each layer simultaneously; The driving operation behavior features, physiological visual features, vehicle dynamics features and vehicle interaction features are concatenated in the time dimension to form a multi-dimensional feature time sequence input with equal time steps; The multidimensional feature time-series input is simultaneously fed into the forward propagation unit and the backward propagation unit of the multilayer bidirectional long short-term memory neural network; The feature vector output by the forward propagation unit at the final time step is fused with the feature vector output by the backward propagation unit at the initial time step to obtain a comprehensive intent representation vector. The integrated intent representation vector is input into a fully connected classification layer, which outputs the classification probabilities of left lane change intent, right lane change intent, and lane keeping intent.

5. The method for predicting vehicle lane-changing risks on icy and snowy roads based on driving intention recognition as described in claim 4, characterized in that, The modified collision time index and the distance deviation index based on parking sight distance are mapped to a quantitative value of risk exposure level and a quantitative value of risk severity, including: A critical threshold is set for the modified collision time index. When the modified collision time index value is lower than the critical threshold, it is considered that there is a collision exposure in the time dimension. A continuous quantitative value of the risk exposure level is calculated based on the degree and duration to which the modified collision time index is below the critical threshold. The analysis of the spacing deviation index based on parking sight distance shows that when the spacing deviation index value is less than one, it indicates insufficient lateral space. Based on the degree to which the distance deviation index based on parking sight distance deviates from the safety benchmark, and in conjunction with the current vehicle speed, the quantified value of the risk severity is calculated.

6. The method for predicting vehicle lane-changing risks on icy and snowy roads based on driving intention recognition as described in claim 5, characterized in that, Combining the quantitative values ​​of risk exposure level and risk severity, a comprehensive lane-changing risk index is calculated using fault tree analysis, including: The occurrence of a lane-changing collision or loss of control is taken as the top event in the fault tree analysis; The time-dimensional risk and the spatial-dimensional risk are considered as intermediate events leading to the top event; The risk exposure level exceeding a preset upper limit is taken as the basic event of time-dimensional risk, and the risk severity exceeding a preset upper limit is taken as the basic event of spatial-dimensional risk. Based on driving experience data on icy and snowy roads, prior failure probabilities are assigned to the basic events of the time dimension risk and the basic events of the spatial dimension risk. Based on the logical AND and OR gate relationships of the fault tree, the failure probabilities of the basic events are passed up to the upper layer and aggregated to finally calculate the occurrence probability of the top event, which is the comprehensive lane-changing risk index.

7. The method for predicting vehicle lane-changing risks on icy and snowy roads based on driving intention recognition as described in claim 6, characterized in that, The analysis of the comprehensive lane-changing risk index of historical cases delineates the numerical boundaries of high-risk, medium-risk, and low-risk levels, including: A large number of samples of the comprehensive lane change risk index generated during lane change processes in icy and snowy road conditions were collected. The comprehensive lane change risk index samples were clustered using an unsupervised clustering algorithm, with the number of clusters set to three. Based on the clustering results, the comprehensive lane change risk index samples are divided into three disjoint sets; Calculate the maximum and minimum values ​​of the comprehensive lane-changing risk index in each set, and determine the range of index values ​​corresponding to each risk level; The set with the smallest value is defined as the low-risk level, the set with the middle value is defined as the medium-risk level, and the set with the largest value is defined as the high-risk level.

8. The method for predicting vehicle lane-changing risks on icy and snowy roads based on driving intention recognition as described in claim 7, characterized in that, The gradient boosting decision tree model is used, with multi-dimensional features extracted from the lane change intention stage data segment as input, and the comprehensive lane change risk index after hierarchical classification as label for training, to construct a dynamic risk prediction model, including: All multi-dimensional features extracted from the data segment of the lane change intention stage are used as the feature set of the model. The comprehensive lane-changing risk index, which is generated at the end of the corresponding lane-changing process and has been classified into risk levels, is used as the training label for the model. A gradient boosting framework is adopted, with decision trees as base learners, to construct a multi-class prediction model; During model training, a histogram-based algorithm is used to accelerate the search for optimal split points for features, and a special processing method for category features is introduced. The multi-class prediction model is iteratively trained using training data until the model loss function converges, thus obtaining the trained dynamic risk prediction model. The construction steps of the gradient boosting decision tree model include: Initialize the multi-class prediction model by assigning an initial prediction value to each sample in the training dataset; Calculate the multi-class cross-entropy loss between the current model prediction and the training labels; Calculate the first and second gradients of each sample based on the multi-class cross-entropy loss; Based on the first and second gradients, the histogram-based algorithm is used to find the split point that maximizes the decrease in the loss function for each feature. Develop a decision tree based on the found optimal split point, and determine the weights of the leaf nodes of the tree. The newly generated decision tree is added to the model, and the predicted values ​​for all samples are updated. Repeat the steps of calculating loss, calculating gradient, finding split points, growing decision trees, and updating predictions until the preset number of iterations is reached or the loss function no longer decreases significantly.

9. The method for predicting vehicle lane-changing risks on icy and snowy roads based on driving intention recognition as described in claim 8, characterized in that, It also includes online application steps for risk prediction models: During actual vehicle operation, the multi-source heterogeneous driving data is collected and preprocessed in real time. Continuously monitor driving intentions, and immediately initiate the risk prediction process when a left lane change intention or a right lane change intention is detected; Extract multidimensional features from a complete intent phase window preceding the current moment; The multidimensional features are input into the trained risk dynamic prediction model; The risk dynamic prediction model outputs a predicted risk level, which corresponds to the risk state that will occur during the lane change execution phase a few seconds later.

Images