CN-121982935-A - Road changing risk prediction method for ice and snow road surface vehicle based on driving intention recognition

CN121982935ACN 121982935 ACN121982935 ACN 121982935ACN-121982935-A

Abstract

The invention relates to the technical field of vehicle driving safety, in particular to an ice and snow road surface vehicle lane changing risk prediction method based on driving intention recognition, which comprises the steps of collecting ice and snow road surface multi-source heterogeneous driving data to recognize lane changing intention; the execution stage is based on real-time kinematics and peripheral interaction information, calculates a corrected collision time index for introducing an ice and snow attachment coefficient correction factor and a distance deviation index combined with an actual parking sight distance, maps the two indexes into risk exposure and severity quantification values, analyzes the risk exposure and severity quantification values through a fault tree to obtain a comprehensive index, and divides high, medium and low risk grades. According to the method, the intention stage multidimensional feature is used as an input, the grading index is used as a label, a dynamic prediction model is built by training a gradient lifting decision tree, early grading prediction of the lane changing risk of the ice and snow road surface is realized, and the risk quantification accuracy and the scene suitability are improved.

Inventors

GUO NIANCHENG
ZHAO WEI
GAO YANG
CHENG HAO

Assignees

山东大学

Dates

Publication Date: 20260505
Application Date: 20260407

Claims (10)

1. The method for predicting the lane change risk of the ice and snow road surface vehicle based on driving intention recognition is characterized by comprising the following steps of: collecting multi-source heterogeneous driving data under an ice-snow road surface, and identifying the lane changing intention of a vehicle; when the vehicle enters a lane change execution stage, calculating a corrected collision time index and a distance deviation index based on a parking sight distance based on real-time vehicle kinematic data and surrounding vehicle interaction information; Mapping the corrected collision time index and the interval deviation index based on the parking sight distance into a risk exposure level quantized value and a risk severity quantized value; Combining the risk exposure level quantized value and the risk severity quantized value, and calculating by using a fault tree analysis method to obtain a comprehensive lane change risk index; analyzing the comprehensive lane change risk index of the historical case, and dividing numerical boundaries of a high risk level, a medium risk level and a low risk level; and training by using the gradient lifting decision tree model and taking the multi-dimensional features extracted from the lane change intention stage data fragments as input and the comprehensive lane change risk indexes subjected to grading as labels, and constructing a risk dynamic prediction model.
2. The method for predicting the risk of vehicle lane changing on an icy or snowy road surface based on driving intention recognition according to claim 1, wherein the collecting multi-source heterogeneous driving data under the icy or snowy road surface, the recognizing the vehicle lane changing intention, comprises: the multi-source heterogeneous driving data comprises a vehicle kinematics time sequence, a driving operation time sequence, surrounding vehicle interaction information, a driver biological signal time sequence and a road surface adhesion coefficient; performing time stamp alignment and frequency standardization processing on the multi-source heterogeneous driving data to form synchronous data streams with uniform sampling frequency; Based on a driving human eye movement mode and a vehicle transverse movement starting point, intercepting a lane change intention stage data fragment from the synchronous data stream; calculating and extracting driving operation behavior characteristics, physiological visual characteristics, vehicle dynamics characteristics and vehicle interaction characteristics from the lane change intention stage data segments; inputting the calculated and extracted multidimensional features into a multi-layer two-way long-short-term memory neural network, wherein the multi-layer two-way long-short-term memory neural network is used for identifying left lane change intention, right lane change intention and lane keeping intention; The method for intercepting the lane change intention stage data fragment from the synchronous data stream based on the driving human eye movement mode and the vehicle transverse movement starting point specifically comprises the following steps: Identifying an eye movement event of first gazing at the side rearview mirror from the driver biological signal time sequence, and recording a time stamp of the eye movement event as a stage starting point; Detecting the moment when the transverse displacement or the yaw rate continuously changes for the first time from the vehicle kinematics time sequence, and recording the time stamp of the moment as a stage end point; Taking a time interval between the stage starting point and the stage ending point as a candidate lane change intention stage; And if the duration of the candidate lane change intention stage is within a preset experience time window range, intercepting all the synchronous data streams in the candidate lane change intention stage as the lane change intention stage data fragments.
3. The method for predicting the lane change risk of the ice and snow road surface vehicle based on the driving intention recognition according to claim 2, wherein the calculating and extracting the driving operation behavior feature, the physiological visual feature, the vehicle dynamics feature and the vehicle interaction feature from the lane change intention stage data segment comprises: performing differentiation and statistical analysis on the driving operation time sequence to obtain steering wheel angle standard deviation, steering wheel angular velocity peak value, average brake pedal opening and accelerator pedal change rate, and forming driving operation behavior characteristics; Processing an eye movement signal and a skin electric signal in the biological signal time sequence of the driver to obtain a point of regard discrete entropy, a glance average speed, an average heart rate growth rate and a skin electric reaction accumulation area, so as to form the physiological visual characteristic; Performing frequency domain transformation and statistics on the vehicle kinematic time sequence to obtain longitudinal acceleration variance, lateral acceleration absolute value integration, power spectral density of yaw rate and vehicle speed change trend, so as to form the vehicle dynamics characteristic; Based on the peripheral vehicle interaction information, calculating the absolute value of the relative speed of the host vehicle and the front vehicle of the adjacent lane, the change rate of the time interval between the host vehicle and the rear vehicle of the target lane and the comprehensive invasion index of the host vehicle and the peripheral multiple vehicles to form the vehicle interaction characteristics.
4. The method for predicting the risk of vehicle lane change on icy and snowy road surface based on driving intention recognition according to claim 3, wherein the inputting the calculated and extracted multidimensional features into a multi-layer two-way long-short term memory neural network comprises: constructing a long-term and short-term memory network structure comprising four hidden layers, and simultaneously deploying a forward propagation unit and a backward propagation unit on each layer; Splicing the driving operation behavior feature, the physiological visual feature, the vehicle dynamics feature and the vehicle interaction feature in a time dimension to form multidimensional feature time sequence input with equal time steps; The multi-dimensional characteristic time sequence input is simultaneously fed into the forward propagation unit and the backward propagation unit of the multi-layer two-way long-short-term memory neural network; fusing the characteristic vector output by the forward propagation unit in the final time step with the characteristic vector output by the backward propagation unit in the initial time step to obtain a comprehensive intention characterization vector; And inputting the comprehensive intention representation vector into a fully-connected classification layer, and outputting classification probabilities of the left lane change intention, the right lane change intention and the lane keeping intention.
5. The method for predicting the risk of vehicle lane change on icy and snowy road surface based on driving intention recognition according to claim 4, wherein calculating the deviation of the corrected collision time index from the distance between the vehicles based on the stopping visual distance comprises: Acquiring the longitudinal relative speed and the longitudinal relative distance between the vehicle and a front vehicle on a target lane; calculating the maximum deceleration under the ice and snow road surface according to the road surface adhesion coefficient; correcting parameters in a classical collision time formula by using the maximum deceleration, and calculating to obtain a corrected collision time index; Acquiring the actual transverse distance between the vehicle and the side vehicle of the adjacent lane; Calculating the minimum safe parking sight distance under the ice and snow road according to the current speed of the vehicle and the road adhesion coefficient; combining the minimum safe parking sight distance with the actual road width to deduce the required safe transverse distance; And taking the ratio of the actual transverse distance to the required safe transverse distance as the distance deviation index based on the parking sight distance.
6. The method for predicting the risk of a vehicle lane change on an icy or snowy road based on driving intention recognition according to claim 5, wherein mapping the corrected collision time index and the distance deviation index based on the stopping sight distance into a risk exposure level quantization value and a risk severity quantization value comprises: Setting a critical threshold of the corrected collision time index, and when the corrected collision time index is lower than the critical threshold, considering that collision exposure in the time dimension exists; calculating a continuous quantitative value of the risk exposure level according to the degree and the duration of the corrected collision time index being lower than a critical threshold value; Analyzing the interval deviation index based on the parking sight distance, and when the interval deviation index is smaller than a value indicating insufficient transverse space; And calculating the risk severity quantification value according to the degree of deviation of the distance deviation index based on the parking sight distance from the safety standard and the current vehicle speed.
7. The method for predicting the risk of vehicle lane change on icy and snowy road surface based on driving intention recognition according to claim 6, wherein the step of calculating the comprehensive lane change risk index by using a fault tree analysis method in combination with the risk exposure level quantization value and the risk severity quantization value comprises the steps of: Taking the collision or out of control of the lane change as a top event for fault tree analysis; taking the time dimension risk and the space dimension risk as intermediate events which lead to the top event; taking the risk exposure level quantized value exceeding a preset upper limit as a basic event of the time dimension risk, and taking the risk severity quantized value exceeding the preset upper limit as a basic event of the space dimension risk; distributing prior failure probability for the basic event of the time dimension risk and the basic event of the space dimension risk according to the driving experience data of the ice and snow road surface; and transmitting and aggregating the failure probability of the basic event to an upper layer according to the logical AND gate or OR gate relation of the failure tree, and finally calculating to obtain the occurrence probability of the top event, wherein the occurrence probability is the comprehensive lane change risk index.
8. The method for predicting the risk of a vehicle lane change on an icy or snowy road based on driving intention recognition according to claim 7, wherein the analyzing the comprehensive lane change risk index of the historical case to classify a numerical boundary of a high risk level, a medium risk level and a low risk level comprises: Collecting a large number of comprehensive lane change risk index samples generated in the lane change process under the ice and snow road environment; Clustering the comprehensive lane change risk index samples by using an unsupervised clustering algorithm, and setting the clustering number to be three; Dividing the comprehensive lane change risk index sample into three mutually disjoint sets according to a clustering result; calculating the maximum value and the minimum value of the comprehensive lane change risk indexes in each set, and determining an index numerical range corresponding to each risk level; The smallest set of values is defined as the low risk level, the central set of values is defined as the medium risk level, and the largest set of values is defined as the high risk level.
9. The method for predicting the lane change risk of the ice and snow road surface vehicle based on driving intention recognition according to claim 8, wherein the step of constructing a risk dynamic prediction model by using a gradient lifting decision tree model, taking multi-dimensional features extracted from data segments of the lane change intention stage as input, and taking the comprehensive lane change risk index subjected to grading as a label for training comprises the steps of: taking all the multidimensional features extracted from the lane change intention stage data fragment as a feature set of a model; Taking the comprehensive lane change risk index which is finally generated in the lane change corresponding process and is divided into risk grades as a training label of a model; a gradient lifting frame is adopted, a decision tree is used as a base learner, and a multi-classification prediction model is constructed; in the model training process, the histogram-based algorithm is adopted to accelerate feature searching of optimal splitting points, and a processing mode exclusive to category features is introduced; performing iterative training on the multi-classification prediction model by using training data until a model loss function converges, so as to obtain the risk dynamic prediction model after training; the construction step of the gradient lifting decision tree model comprises the following steps: Initializing the multi-classification prediction model, and distributing an initial prediction value for each sample of the training data set; calculating multi-classification cross entropy loss between the current model prediction result and the training label; Calculating a first-order gradient and a second-order gradient of each sample according to the multi-classification cross entropy loss; Searching a split point which makes the loss function most drop for each feature by using the histogram-based algorithm based on the first-order gradient and the second-order gradient; growing a decision tree according to the found optimal splitting point, and determining the weight of leaf nodes of the tree; adding the newly generated decision tree into the model, and updating the predicted values of all samples; repeating the steps of calculating the loss, calculating the gradient, searching the splitting point, growing the decision tree and updating the predicted value until the preset iteration times are reached or the loss function is not obviously reduced.
10. The method for predicting the risk of a vehicle lane change on an icy or snowy road surface based on driving intention recognition according to claim 9, further comprising the step of on-line application of a risk prediction model: in the actual running process of the vehicle, acquiring and preprocessing the multi-source heterogeneous driving data in real time; Continuously detecting driving intention, and immediately starting a risk prediction flow when the left lane change intention or the right lane change intention is identified; extracting multidimensional features in a complete intention phase window before the current moment; inputting the multidimensional features into a trained risk dynamic prediction model; the risk dynamic prediction model outputs a predicted risk level, and the risk level corresponds to a risk state in a channel change execution stage after a few seconds.

Description

Road changing risk prediction method for ice and snow road surface vehicle based on driving intention recognition Technical Field The invention relates to the technical field of vehicle driving safety, in particular to a road changing risk prediction method for vehicles on ice and snow roads based on driving intention recognition. Background The existing vehicle lane change risk prediction technology is mostly based on common road conditions, and adopts indexes such as conventional collision time, minimum safety distance and the like to evaluate risks by collecting data such as vehicle speed, acceleration, relative positions of adjacent vehicles and the like, and a model is often built by depending on lane change full-stage data or a linear algorithm. The technology has the obvious limitations when being applied to ice and snow roads, namely, the low attachment coefficient of the ice and snow roads causes braking delay and parking sight length extension, the conventional TTC does not correct friction coefficient attenuation influence, the interval index is not combined with the actual parking sight length requirement, the deviation between a calculation result and an actual risk is large, meanwhile, the model mostly adopts full-stage data after channel changing starting, early pre-action characteristics of an intention stage cannot be captured, and dynamic risk pre-judgment is difficult to realize. The problem that risk quantization distortion is caused by lack of a specific risk index calculation mode considering low adhesion characteristics in an execution stage is solved under the road exchange scene of an ice and snow road surface, and the existing prediction model is difficult to dynamically output grading risks based on fine granularity characteristics in an intention stage. The calculation logic of the physical characteristics of the ice and snow pavement, which is not adapted to the conventional index, is broken through, and the problem that model input depends on full-stage data and early signals of the intention stage cannot be utilized is solved. Disclosure of Invention The invention aims to solve the defects in the prior art, and provides a road changing risk prediction method for vehicles on ice and snow roads based on driving intention recognition. In order to achieve the purpose, the invention adopts the following technical scheme that the road changing risk prediction method for the ice and snow road surface vehicle based on driving intention identification comprises the following steps: collecting multi-source heterogeneous driving data under an ice-snow road surface, and identifying the lane changing intention of a vehicle; when the vehicle enters a lane change execution stage, calculating a corrected collision time index and a distance deviation index based on a parking sight distance based on real-time vehicle kinematic data and surrounding vehicle interaction information; Mapping the corrected collision time index and the interval deviation index based on the parking sight distance into a risk exposure level quantized value and a risk severity quantized value; Combining the risk exposure level quantized value and the risk severity quantized value, and calculating by using a fault tree analysis method to obtain a comprehensive lane change risk index; analyzing the comprehensive lane change risk index of the historical case, and dividing numerical boundaries of a high risk level, a medium risk level and a low risk level; and training by using the gradient lifting decision tree model and taking the multi-dimensional features extracted from the lane change intention stage data fragments as input and the comprehensive lane change risk indexes subjected to grading as labels, and constructing a risk dynamic prediction model. As a further aspect of the present invention, the collecting multi-source heterogeneous driving data under an icy or snowy road surface, identifying a lane change intention of a vehicle, includes: the multi-source heterogeneous driving data comprises a vehicle kinematics time sequence, a driving operation time sequence, surrounding vehicle interaction information, a driver biological signal time sequence and a road surface adhesion coefficient; performing time stamp alignment and frequency standardization processing on the multi-source heterogeneous driving data to form synchronous data streams with uniform sampling frequency; Based on a driving human eye movement mode and a vehicle transverse movement starting point, intercepting a lane change intention stage data fragment from the synchronous data stream; calculating and extracting driving operation behavior characteristics, physiological visual characteristics, vehicle dynamics characteristics and vehicle interaction characteristics from the lane change intention stage data segments; inputting the calculated and extracted multidimensional features into a multi-layer two-way long-short-term memory neural network, wherein