A freeway driving risk assessment method based on predicted trajectories

Abstract

The present invention discloses a method for evaluating driving risks on expressways based on predicted trajectories, comprising the following working steps: Step 1: Predict the predicted trajectories of the host vehicle and surrounding vehicles within the prediction interval; Step 2: Regularize the predicted trajectories within the prediction interval obtained in Step 1 into the predicted trajectories of the host vehicle and surrounding vehicles within each look-ahead time t la，m (m = 1, 2,..., T); Step 3: Obtain the quantified perceived risks posed by obstacle vehicles to the host vehicle within each look-ahead time; Step 4: Obtain a Bayesian model for dividing the quantified perceived risk levels through the Bayesian posterior probability formula; Step 5: Realize the real-time evaluation of driving risks; The beneficial effects of the present invention are that it improves the accuracy and practicality of driving risk assessment; it has good technical applicability to intelligent vehicles with different levels of intelligence.

Description

Technical Field

[0001] The present invention relates to the technical fields of traffic safety evaluation and active safety technology of intelligent transportation systems, and particularly to a method for evaluating driving risks on expressways based on predicted trajectories. Background Art

[0002] Driving risk assessment is an important basis for the safe operation of driving assistance systems and intelligent vehicles. With the development of intelligent perception and communication technologies, more abundant information on the real-time motion and environmental characteristics of vehicles can be obtained, and driving risk assessment methods are also more diversified. Currently, they mainly include methods based on collision probability, methods based on driving risk fields, and methods based on machine learning. Among them, methods based on collision probability usually only consider the uncertain relationship between the host vehicle and the preceding vehicle, ignoring the mutual influence of the behaviors or motions of multiple surrounding vehicles. Methods based on driving risk fields mostly calculate based on the relative position of the current vehicle, and during actual high-speed driving, even the position may change greatly within a short period of time, which is not conducive to the subsequent construction of an accurate collision avoidance control scheme. Methods based on machine learning generally do not consider the multi-vehicle collision risk problem in the complex scenarios of multiple surrounding vehicles that need to be faced during actual driving, affecting the practicality of such algorithms in reality. Therefore, it is necessary to study a driving risk estimation method that can adapt to the complexity and uncertainty characteristics of the actual driving scenarios on expressways. Summary of the Invention

[0003] The purpose of the present invention is to propose a design solution to solve the above-mentioned existing problems, specifically a method for evaluating driving risks on expressways based on predicted trajectories.

[0004] To achieve the above purpose, the technical solution of the present invention is a method for evaluating driving risks on expressways based on predicted trajectories, including the following working steps:

[0005] Step 1: In the vehicle networking environment, real-time obtain the motion information of the host vehicle and multiple surrounding vehicles through sensors and V2X interaction, and based on the long short-term memory (LSTM) neural network structure, fuse the environmental interaction information attention module to predict the predicted trajectories of the host vehicle and surrounding vehicles within the prediction interval.

[0006] Step 2: Construct T moving time windows for the prediction interval. Step 2: Construct T moving time windows for the prediction interval As the T-segment look-ahead time of the host vehicle, regularize the predicted trajectories within the prediction interval obtained in Step 1 into the predicted trajectories of the host vehicle and surrounding vehicles for each segment of look-ahead time. within the host vehicle and surrounding vehicles.

[0007] Step 3: Based on the predicted trajectories of the host vehicle and surrounding vehicles within the look-ahead time described in Step 2 respectively, construct the driving risk domain of the host vehicle and the environmental event cost within each look-ahead time in the prediction interval, and obtain the quantified perceived risk posed by the obstacle vehicle to the host vehicle within each look-ahead time;

[0008] Step 4: Construct a conditional distribution probability model of the quantified perceived risk under different driving risk levels, calibrate the thresholds of different risk levels in the model based on the real vehicle trajectory data, and obtain a Bayesian model for dividing the quantified perceived risk level through the Bayesian posterior probability formula;

[0009] Step 5: Input the quantified perceived risk within each look-ahead time in the prediction interval described in Step 3 into the Bayesian model for dividing the risk level described in Step 4 respectively to obtain the driving risk level probability within each look-ahead time. Finally, through probability fusion calculation considering duration attenuation, obtain the fused perceived risk of the entire prediction interval to realize the real-time assessment of driving risk.

[0010] For further supplement to this technical solution, the calculation method of the quantified perceived risk posed by the obstacle vehicle to the host vehicle in Step 3 includes the following steps:

[0011] Step 1: Based on the predicted trajectory of the host vehicle, construct the driving risk domain of the host vehicle within each look-ahead time in the prediction interval;

[0012] Step 2: Based on the predicted trajectories of surrounding vehicles, construct the environmental event cost within each look-ahead time in the prediction interval;

[0013] Step 3: Calculate the quantified perceived risk of the host vehicle for each look-ahead time in the prediction interval. The quantified perceived risk of the host vehicle within the look-ahead time is:

[0014] ;

[0015] where, is the driving risk domain of the host vehicle within the look-ahead time described in Step 1, is the environmental event cost within the corresponding look-ahead time described in Step 2, where the dynamic obstacle is defined as the vehicle ; represents the coordinates of all grid points within the surrounding area of driving.

[0016] For further supplement to this technical solution, Step 1 is specifically implemented through the following sub-steps:

[0017] Step 1-1): Taking the intersection of the road cross-section at the centroid of the host vehicle at the initial moment of each preview time and the median strip as the origin, with the forward direction along the lane as the positive x-axis and the direction of its counterclockwise rotation by 90° as the positive y-axis, a dynamic interval coordinate system is established;

[0018] Step 1-2): The driving range around the host vehicle is meshed, and the coordinates of each grid point in the dynamic interval coordinate system described in Step 1-1) are ( , ). The coordinate of the centroid position of the host vehicle at the initial moment is ( , ), and the driving risk field DRF of the host vehicle changing with the grid position coordinates is calculated.

[0019] For further supplement to this technical solution, assuming that the distribution of the driving risk field DRF of the host vehicle is a surface with a Gaussian cross-section extending along the predicted trajectory, the driving risk field DRF model of the host vehicle within the preview time is constructed through the following steps :

[0020] Step 1-2-1): The height of the Gaussian cross-section of the driving risk field DRF is a parabola with respect to the driving path length L of the predicted trajectory point from the initial position ( , ), and is positively correlated with the speed V of the host vehicle at the initial moment within the preview time:

[0021] ;

[0022] Among them, the parameter p defines the steepness of the parabola, and s is the total path length of the vehicle prediction trajectory within the preview time;

[0023] Step 1-2-2): The width of the Gaussian cross-section of the driving risk field DRF is a linear function of the driving path length L:

[0024] ;

[0025] Among them, w is the vehicle width correction parameter, which can be taken as 1 / 4 of the vehicle width; m defines the slope of the expansion or contraction of the Gaussian cross-section; represents the vehicle steering angle; k represents the edge parameter of the DRF model; except for the vehicle width correction parameter w, other parameters p, m, k can be estimated by the grid search algorithm according to the driver risk perception characteristics;

[0026] Step 1-2-3): Based on the Gaussian cross-section height variable and width variable , the driving risk field DRF model of the vehicle within the preview time is:

[0027] ;

[0028] Among them, ([[]] , ) is the perpendicular point coordinates from the grid point ( , ) to the predicted trajectory of the host vehicle;

[0029] Step 1-2-4): In order to quantify the potential risks that dynamic obstacles within the range behind the vehicle may bring, with the cross-section of the center of gravity position of the host vehicle as the symmetry plane, a risk domain symmetric to the DRF described in Steps 1-2-1) to 1-2-3) is established backward, so as to realize the continuous characterization of the quantified risks within the range around the vehicle.

[0030] For further supplementation of this technical solution, the specific implementation of Step 2 is as follows through the following sub-steps:

[0031] Step 2-1): Define the environmental event cost parameters that may pose potential risks to the driving of the host vehicle as shown in Table 1;

[0032] Table 1 Environmental event cost parameters

[0033] ;

[0034] Among them, , , represent the dynamic obstacle, lane line, and out-of-lane costs respectively; for When the host vehicle forms a following or cutting-in scenario with an adjacent obstacle vehicle, the obstacle vehicle cost is represented by the distance Δd and speed difference Δv between the two vehicles, denoted as ; in other driving scenarios, the obstacle vehicle cost is fixed at 1200, denoted as . For static obstacles, it can be assumed that the lane line cost is 100, and the out-of-lane boundary cost is 500;

[0035] Step 2-2): Use the same method as described in Step 1-2) to grid the driving range around the host vehicle. The determination method of the environmental event cost within the driving range around the host vehicle during the look-ahead time is as follows:

[0036]

[0037] Among them , and are respectively the 0-1 discrimination functions of the environmental event type at the initial moment of the host vehicle's look-ahead time:

[0038] ;

[0039] ;

[0040] 。

[0041] For further supplement to this technical solution, the method for constructing the Bayesian model for partitioning the quantified perceived risk level in step 4 is as follows:

[0042] Step 1): Construct a conditional distribution probability model for quantified perceived risk under different driving risk levels;

[0043] Step 2): Obtain the Bayesian model for partitioning the quantified perceived risk level through the Bayesian posterior probability formula:

[0044] ;

[0045] In the formula, represents the probability that the vehicle is in a certain risk level given the quantified perceived risk C, is the conditional distribution probability of the quantified perceived risk under different driving risk levels described in step 1, is the prior probability of each risk level and satisfies the constraint .

[0046] For further supplement to this technical solution, step 1) is implemented through the following sub-steps:

[0047] Step 1.1): Define the risk level set :

[0048] ;

[0049] where D, A, and S represent the dangerous, medium, and safe levels respectively, and are assigned quantified values 2, 1, and 0 to describe the risk level, that is, the risk level is defined as:

[0050] ;

[0051] Step 1.2): Let the thresholds corresponding to the quantified perceived risk C under the high and low driving risk levels be and respectively, and the expected value of C under the medium risk level be , then the conditional distribution probability of the quantified perceived risk C can be constructed as follows:

[0052] ;

[0053] ;

[0054] ;

[0055] ;

[0056] In the formula, , represents the probability of observing the quantified perceived risk C at different risk levels, and its magnitude is determined by the relative magnitudes of C and the respective risk level thresholds , and ; the parameter is the uncertainty factor, which is mainly used to control the smoothness of the curves of different driving risk levels;

[0057] Step 1.3): The preset parameters , and in Step 1.2) can be determined according to the distribution characteristics of the real vehicle trajectory data.

[0058] For further supplement to this technical solution, the specific implementation method of real-time driving risk assessment in the fifth step is as follows:

[0059] Step (1): Using the obstacle vehicle as a dynamic obstacle, calculate the quantified perceived risk { , , , } within each look-ahead time in the prediction interval according to the method in Step 3;

[0060] Step (2): Input the sequence of quantified perceived risks { , , , } in the prediction interval into the risk level division Bayesian model described in Step 4 to obtain the posterior distribution probabilities that the quantified perceived risk generated by the obstacle vehicle for the host vehicle is at different risk levels (including D, A, S), which are respectively denoted as { , , }, { , , }, { , };

[0061] Step (3): Using the posterior distribution probabilities of the quantified perceived risks as weights and considering the duration decay effect, perform a fusion calculation on the risk levels within T look-ahead times to obtain the obstacle vehicle within the prediction interval Fusion perception risks generated by the host vehicle :

[0062] ;

[0063] wherein , , are the Bayesian posterior probabilities of the host vehicle at different risk levels; , and take 2, 1, and 0 respectively to represent the risk level quantization values; is the attenuation coefficient representing the duration decay effect, with a value range of (0, 1]. When takes 1, it means that the duration decay is not considered in the risk assessment process;

[0064] Step (4): Select the 1st, 2nd,..., ith vehicles around the host vehicle as obstacle vehicles, denoted as , , , ; Calculate the fusion perception risks generated by each obstacle vehicle for the host vehicle in real time according to the method described in steps (1) to (3) within the prediction interval .

[0065] For further supplement to this technical solution, the real-time assessment result of driving risk is determined according to the following steps:

[0066] Step (4-1): Take the maximum value among the fusion perception risks generated by all obstacle vehicles for the host vehicle as the risk quantization value of the host vehicle at the current moment, denoted as ;

[0067] Step (4-2): Compare with of the preset critical warning threshold in real time; If , the risk degree of the interaction behavior between the current host vehicle and the obstacle vehicle is relatively high, and the system should issue a risk warning prompt; otherwise, there is no potential risk in the interaction behavior between the host vehicle and the obstacle vehicle.

[0068] Its beneficial effects are as follows: 1. The present invention constructs a driving risk domain with a gradually changing Gaussian cross-section feature on both sides of the prediction trajectory, which has a relatively high risk value near the vehicle prediction trajectory and decreases with the increase of the lateral and longitudinal distances from the vehicle position at the initial moment, and can characterize the uncertainty of the target vehicle's perception and manipulation behavior, improving the accuracy and practicality of driving risk assessment;

[0069] 2. The present invention constructs a driving risk domain from the perspective of the host vehicle. For the current host vehicle with manual driving, its predicted trajectory can be output in real time through a multi-vehicle trajectory prediction model based on environmental interaction information; for future vehicles with autonomous driving, its predicted trajectory can be set as the vehicle's autonomous planning trajectory, which has good technical applicability to intelligent vehicles with different intelligent levels;

[0070] 3. Based on Bayesian theory, the present invention converts the quantified perception risk value based on predicted trajectory points in each moving time window of the prediction interval into the probability of the risk level, and performs probability fusion calculation with consideration of duration decay on the risk levels of all time windows, reducing the influence that the trajectory prediction error in a single time window may have on the risk assessment result. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 is a flowchart of the freeway driving risk assessment based on the predicted trajectory of the present invention;

[0072] Figure 2 is a schematic diagram of the moving time window of the prediction interval of the present invention;

[0073] Figure 3 is a schematic diagram of the concept of the driving risk domain of the present invention;

[0074] Figure 4 is a schematic diagram of the dynamic event cost of the present invention; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0075] To make the technical solution of the present invention clearer to those skilled in the art, the following will Figures 1-4 describe the technical solution in detail:

[0076] A freeway driving risk assessment method based on the predicted trajectory mainly achieves the above technical objectives through the following technical means, including the following working steps:

[0077] Step 1: In the vehicle networking environment, the motion information of the host vehicle and surrounding multi-vehicles is obtained in real time through sensors and V2X interaction. Based on the long short-term memory (LSTM) neural network structure and integrating the environmental interaction information attention module, the predicted trajectories of the host vehicle and surrounding vehicles in the prediction interval are predicted.

[0078] Step 2: Construct T moving time windows (as shown in Figure 2 ) as the T segments of the look-ahead time of the host vehicle, and regularize the predicted trajectories of the host vehicle and surrounding vehicles in the prediction interval obtained in Step 1 into the predicted trajectories of the host vehicle and surrounding vehicles within each segment of the look-ahead time .

[0079] Step 3: Based on the predicted trajectories of the host vehicle and surrounding vehicles within the preview time described in Step 2, construct the driving risk domain of the host vehicle and the cost of surrounding environment events for each segment of the preview time within the prediction interval, and obtain the quantified perceived risk posed by the obstacle vehicle to the host vehicle for each preview time.

[0080] Step 4: Construct a conditional distribution probability model of the quantified perceived risk under different driving risk levels, calibrate the thresholds of different risk levels in the model based on the real vehicle trajectory data, and obtain a Bayesian model for dividing the quantified perceived risk level through the Bayesian posterior probability formula.

[0081] Step 5: Input the quantified perceived risk for each segment of the preview time within the prediction interval described in Step 3 into the Bayesian model for dividing the risk level described in Step 4 to obtain the probability of the driving risk level for each preview time. Finally, through the probability fusion calculation considering the duration decay, obtain the fusion perceived risk for the entire prediction interval to achieve real-time assessment of the driving risk.

[0082] The calculation method of the quantified perceived risk posed by the obstacle vehicle to the host vehicle in Step 3 is as follows:

[0083] Step 1: Based on the predicted trajectory of the host vehicle, construct the driving risk domain of the host vehicle for each segment of the preview time within the prediction interval. The specific implementation of Step 1 is as follows through the following sub-steps:

[0084] Step 1-1): Take the intersection point of the road cross-section where the centroid of the host vehicle is located at the initial moment of each preview time and the central median as the origin, with the forward direction along the lane as the positive x axis, and establish a dynamic interval coordinate system with the direction of its counterclockwise rotation of 90° as the positive y-axis.

[0085] Step 1-2): Grid the driving range around the host vehicle. The coordinates of each grid point in the dynamic interval coordinate system described in Step 1-1) are ( , ). The coordinate of the initial moment center of gravity position of the host vehicle is ( , ). Calculate the driving risk domain DRF of the host vehicle that changes with the grid position coordinates. Assume that the distribution of the driving risk domain DRF of the host vehicle is a surface that extends along the predicted trajectory and has a Gaussian cross-section (as shown in Figure 3 ). Construct the driving risk domain DRF model of the host vehicle within the preview time through the following steps:

[0086] Step 1-2-1): The height of the Gaussian cross-section of the driving risk domain DRF is a function of the distance of the predicted trajectory point from the initial position ( , ( ), the parabola of the driving path length L, and is positively correlated with the speed V of the vehicle at the initial moment within the look-ahead time:

[0087] ;

[0088] Among them, the parameter p defines the steepness of the parabola, and s is the total path length of the predicted trajectory of the vehicle within the look-ahead time.

[0089] Step 1-2-2): The width of the Gaussian cross-section of the driving risk field DRF is a linear function of the driving path length L:

[0090] ;

[0091] Among them, w is the vehicle width correction parameter, which can be taken as 1 / 4 of the vehicle width; m defines the slope of the expansion or contraction of the Gaussian cross-section; represents the vehicle steering angle; k represents the edge parameter of the DRF model; except for the vehicle width correction parameter w, the other parameters p, m, k can be estimated by the grid search algorithm according to the driver's risk perception characteristics.

[0092] Step 1-2-3): Based on the Gaussian cross-section height variable and width variable , within the look-ahead time , the driving risk field DRF model of the vehicle is:

[0093] ;

[0094] Among them, ( , ) is the vertical point coordinates of the grid point ( , ) to the predicted trajectory of the vehicle itself.

[0095] Step 1-2-4): In order to quantify the risks that may be brought by dynamic obstacles within the range behind the vehicle, a risk field symmetric to the DRF described in Steps 1-2-1) to 1-2-3) is established backward with the cross-section of the vehicle's center of gravity position as the symmetry plane, so as to realize the continuous characterization of the quantified risks within the range around the vehicle.

[0096] Step 2: Based on the predicted trajectories of surrounding vehicles, construct the environmental event costs for each look-ahead time within the prediction interval. The specific implementation of Step 2 is as follows:

[0097] Step 2-1): Define the environmental event cost parameters that may pose potential risks to the vehicle's driving as shown in Table 1.

[0098] Table 1 Environmental event cost parameters

[0099] ;

[0100] Among them, , , respectively represent dynamic obstacles (obstacle vehicles, i.e., surrounding driving vehicles), lane lines, and costs outside the lane; for When a following vehicle scenario or an overtaking vehicle scenario is formed between the host vehicle and an adjacent obstacle vehicle, the cost of the obstacle vehicle is represented by the distance Δd and speed difference Δv between the two vehicles, denoted as ; In other driving scenarios, the cost of the obstacle vehicle is fixed at 1200, denoted as . For static obstacles, it can be assumed that the lane line cost is 100, and the cost outside the lane boundary is 500.

[0101] Step 2-2): The driving range around the host vehicle is meshed in the same way as described in Step 1-2), and the method for determining the environmental event cost within the driving range around the host vehicle during the preview time is as follows:

[0102] ;

[0103] Among them , and are respectively 0-1 discriminant functions for the environmental event type at the initial moment of the host vehicle's preview time:

[0104] ;

[0105] ;

[0106] ;

[0107] Step 3: Calculate the quantified perception risk of the host vehicle for each preview time within the prediction interval. The quantified perception risk within the host vehicle's preview time is:

[0108] ;

[0109] Among them, is the driving risk domain of the host vehicle within the preview time described in Step 1, is the environmental event cost corresponding to the preview time in Step 2, where the dynamic obstacle is defined as the vehicle ; represents the coordinates of all grid points within the driving perimeter.

[0110] The method for constructing the Bayesian model for dividing the quantified perception risk level in Step 4 is as follows:

[0111] Step 1): Construct a conditional distribution probability model for quantifying perceived risk under different driving risk levels. The specific implementation of Step 1 is as follows:

[0112] Step 1.1): Define the set of risk levels :

[0113] ;

[0114] where D, A, and S represent the dangerous, medium, and safe levels respectively, and are assigned the quantization values 2, 1, and 0 to describe the risk level, that is, the risk level is defined as:

[0115] ;

[0116] Step 1.2): Let the thresholds corresponding to the quantified perceived risk C under high and low driving risk levels be and , and the expected value of C under the medium risk level be . Then the conditional distribution probability of the quantified perceived risk C can be constructed as follows:

[0117] ;

[0118] ;

[0119] ;

[0120] ;

[0121] In the formula, , represents the probability of observing the quantified perceived risk C under different risk levels, and its magnitude is determined by the relative magnitudes of C and the thresholds of each risk level , and ; The parameter is the uncertainty factor, which is mainly used to control the smoothness of the curves of different driving risk levels.

[0122] Step 1.3): The preset parameters , and in Step 1.2) can be determined according to the distribution characteristics of the real vehicle trajectory data.

[0123] Taking the NGSIM dataset of real vehicle driving trajectory data in the United States as an example, randomly select N pieces (a total of N’The actual vehicle trajectory data of each data frame is used to calculate the look-ahead time of each frame according to the method described in step 3. The quantified perceived risk of the host vehicle within it (look-ahead time can be taken as 0.2 s). Select the 90% and 10% percentile values sorted from smallest to largest of the quantified perceived risk as the thresholds corresponding to the high and low driving risk levels and , and the 50% percentile value as the expected value of the medium risk level. .

[0124] Step 2): Obtain the Bayesian model for dividing the quantified perceived risk level through the Bayesian posterior probability formula:

[0125] ;

[0126] In the formula, represents the probability that the vehicle is in a certain risk level given the quantified perceived risk C, is the conditional distribution probability of the quantified perceived risk under different driving risk levels described in step 1, is the prior probability of each risk level and satisfies the constraint . In practical applications, it can be assumed that different risk levels have the same prior probability (both 1 / 3).

[0127] The specific implementation method of real-time driving risk assessment in step 5 is as follows:

[0128] Step (1): Taking the obstacle vehicle as a dynamic obstacle, calculate the quantified perceived risk within each look-ahead time in the prediction interval according to the method in step 3 { , , , }}.

[0129] Step (2): Input the sequence of quantified perceived risks { , , , }} in the prediction interval into the Bayesian model for risk level division described in step 4 to obtain the posterior distribution probabilities that the quantified perceived risk generated by the obstacle vehicle for the host vehicle is in different risk levels (including D, A, S), which are respectively denoted as { , , }}, { , , }, { , }.

[0130] Step (3): Using the posterior distribution probability of the quantified perceived risk as the weight and considering the duration decay effect (since the longer the prediction time, the lower the accuracy of the trajectory prediction, the lower the weight should be given to the look-ahead time period farther from the initial moment of the prediction interval), T perform a fusion calculation on the risk levels within the look-ahead time period to obtain the integrated perceived risk of the obstacle vehicle to the host vehicle within the prediction interval (taking values between 0 and 1, where the larger the value, the greater the risk of the interaction behavior between the obstacle vehicle and the host vehicle):

[0131] ;

[0132] where , , is the Bayesian posterior probability of the host vehicle at different risk levels; , and respectively take 2, 1, 0 to represent the quantified values of the risk levels; is the decay coefficient representing the duration decay effect, with a value range of (0, 1]. When takes 1, it means that the duration decay is not considered in the risk assessment process.

[0133] Step (4): Screen out the 1st, 2nd,..., i vehicles around the host vehicle as obstacle vehicles, denoted as , , , . Calculate the integrated perceived risk of each obstacle vehicle to the host vehicle within the prediction interval in real time according to the method described in steps (1) to (3). Determine the real-time assessment result of the driving risk according to the following sub-steps:

[0134] Step (4-1): Take the maximum value among the integrated perceived risks of all obstacle vehicles to the host vehicle as the risk quantification value of the host vehicle at the current moment, denoted as .

[0135] Step (4-2): Compare with the preset critical warning threshold (which can be set as according to the actual application requirements ) in real time; if , then the risk degree of the interaction behavior between the current host vehicle and the obstacle vehicle is relatively high, and the system should issue a risk warning prompt; otherwise, there is no potential risk in the interaction behavior between the host vehicle and the obstacle vehicle. ​

[0136] The above technical solutions only reflect the preferred technical solutions of the technical solutions of the present invention. Some changes that those skilled in the art may make to some parts thereof all reflect the principles of the present invention and fall within the protection scope of the present invention.

Claims

1. A freeway driving risk assessment method based on predicted trajectories, characterized in that, It includes the following working steps: Step 1: In the vehicle networking environment, real-time motion information of the host vehicle and multiple surrounding vehicles is obtained through sensors and V2X interaction. Based on the long short-term memory (LSTM) neural network structure and integrating the environmental interaction information attention module, the predicted trajectories of the host vehicle and surrounding vehicles within the prediction interval are obtained; Step 2: Construct T moving time windows for the prediction interval As the T segments of the look-ahead time of the ego vehicle, the predicted trajectories within the prediction interval obtained in Step 1 are sorted into the predicted trajectories of the ego vehicle and surrounding vehicles for each segment of the look-ahead time inside; Step 3: Based on the predicted trajectories of the host vehicle and surrounding vehicles within the look-ahead time described in Step 2 respectively, the driving risk domain of the host vehicle and the environmental event cost within each look-ahead time within the prediction interval are constructed, and the quantified perceived risk posed by the obstacle vehicle to the host vehicle within each look-ahead time is obtained; Step 4: A conditional distribution probability model of the quantified perceived risk under different driving risk levels is constructed, and the thresholds of different risk levels in the model are calibrated based on the real vehicle trajectory data. The Bayesian model for dividing the quantified perceived risk level is obtained through the Bayesian posterior probability formula; Step 5: The quantified perceived risks within each look-ahead time within the prediction interval described in Step 3 are respectively input into the Bayesian model for dividing the risk level described in Step 4 to obtain the driving risk level probabilities within each look-ahead time. Finally, the fused perceived risk of the entire prediction interval is calculated through probability fusion considering duration attenuation, realizing the real-time assessment of driving risk.

2. The method for evaluating driving risks on expressways based on predicted trajectories according to claim 1, wherein, The calculation method for the quantified perceived risk posed by the obstacle vehicle to the host vehicle in Step 3 includes the following steps: Step 1: Based on the predicted trajectory of the host vehicle, the driving risk domain of the host vehicle within each look-ahead time within the prediction interval is constructed; Step 2: Based on the predicted trajectories of the surrounding vehicles, the environmental event cost within each look-ahead time within the prediction interval is constructed; Step 3: Calculate the quantified perception risk of the host vehicle for each preview time within the prediction interval. The quantified perception risk within the host vehicle preview time is as follows: ； Among them, is the look-ahead time described in step 1 the driving risk area of the host vehicle within it, is the environmental event cost within the corresponding look-ahead time of step 2, where dynamic obstacles are defined as vehicles ; represents the coordinates of all grid points within the surrounding area of driving.

3. The method for evaluating the driving risk on expressways based on predicted trajectories according to claim 2, characterized in that The specific implementation of Step 1 is achieved through the following sub-steps: Step 1-1): Taking the intersection of the road cross-section at the centroid position of the host vehicle at the initial moment of each look-ahead time and the central median as the origin, taking the forward direction along the lane as the positive x-axis, and taking the direction of its counterclockwise rotation by 90° as the positive y-axis, a dynamic interval coordinate system is established; Step 1-2): Perform grid processing on the driving range around the host vehicle. The coordinates of each grid point in the dynamic interval coordinate system described in Step 1-1) are ( , ). The coordinates of the center of gravity position of the host vehicle at the initial moment are ( , ). Calculate the driving risk field DRF of the host vehicle that changes with the grid position coordinates.

4. The method for evaluating driving risks on expressways based on predicted trajectories according to claim 3, characterized in that, Assume that the distribution of the driving risk field (DRF) of the ego vehicle is a surface with a Gaussian cross-section that extends along the predicted trajectory. The DRF model of the ego vehicle within the preview time is constructed through the following steps: for the DRF model of the ego vehicle: Step 1-2-1): Height of the Gaussian cross-section of the Driving Risk Field (DRF) is a parabola with respect to the driving path length L from the predicted trajectory point to the initial position ( , ), and is positively correlated with the speed V of the host vehicle at the initial moment within the preview time: ； Among them, the parameter p defines the steepness of the parabola, and s is the total path length of the vehicle predicted trajectory within the look-ahead time; Step 1-2-2): Width of the Gaussian cross-section of the Driving Risk Field (DRF) is a linear function of the driving path length L: ； Among them, w is the vehicle width correction parameter, which can be taken as 1 / 4 of the vehicle width; m defines the slope of the expansion or contraction of the Gaussian cross-section; represents the vehicle steering angle; k represents the edge parameter of the DRF model; except for the vehicle width correction parameter w, other parameters p, m, and k can be estimated by the grid search algorithm according to the driver's risk perception characteristics; Step 1-2-3): Based on the Gaussian cross-section height variable and width variable , the vehicle driving risk domain DRF model within the preview time is as follows: ； Among them, ( , ) is the perpendicular point coordinate from the grid point ( , ) to the predicted trajectory of the host vehicle; Step 1-2-4): To quantify the possible risks brought by dynamic obstacles within the range behind the vehicle, a risk domain symmetric to the DRF described in Steps 1-2-1) to 1-2-3) is established backward with the cross-section at the center of gravity position of the host vehicle as the symmetry plane, so as to realize the continuous characterization of the quantified risk within the range around the vehicle.

5. The method for evaluating driving risks on expressways based on predicted trajectories according to claim 2, characterized in that, The specific implementation of Step 2 is achieved through the following sub-steps: Step 2-1): Define the environmental event cost parameters that may pose potential risks to the driving of the host vehicle as shown in Table 1; Table 1 Environmental event cost parameters ; Among them, , , respectively represent dynamic obstacles, lane lines, and costs outside the lane; for When the host vehicle forms a following scenario or an overtaking scenario with an adjacent obstacle vehicle, the cost of the obstacle vehicle is represented by the distance Δd and speed difference Δv between the two vehicles, denoted as ; in other driving scenarios, the cost of the obstacle vehicle is fixed at 1200, denoted as ; for static obstacles, it can be assumed that the lane line cost is 100, and the cost outside the lane boundary is 500; Step 2-2): Process the driving range around the host vehicle into a grid in the same way as described in Step 1-2). The method for determining the environmental event cost within the driving range around the host vehicle during the preview time is as follows: ； Among them , and are respectively the 0-1 discrimination functions of the environmental event type at the initial moment of the forward-looking time of the host vehicle: ； ； 。 6. The method for evaluating the driving risk on an expressway based on a predicted trajectory according to claim 1, wherein The construction method of the Bayesian model for dividing the quantified perceived risk level in Step 4 is as follows: Step 1): Construct a conditional distribution probability model of the quantified perceived risk under different driving risk levels; Step 2): Obtain the Bayesian model for dividing the quantified perceived risk level through the Bayesian posterior probability formula: ； wherein, represents the probability that the vehicle is in a certain risk level under a given quantization-aware risk C, is the conditional distribution probability of the quantization-aware risk under different driving risk levels described in step 1, is the prior probability of each risk level and satisfies the constraint .

7. A method for evaluating driving risks on expressways based on predicted trajectories according to claim 6, characterized in that The implementation of Step 1) is achieved through the following sub-steps: Step 1.1): Define the set of risk levels : ； Where D, A, and S represent the dangerous, medium, and safe levels respectively, and are assigned quantification values of 2, 1, and 0 respectively to describe the risk level, that is, the risk level is defined as: ； Step 1.2): Let the thresholds of the quantified perceived risk C under high and low driving risk levels be and respectively. The expected value of C under the medium risk level is . Then, the conditional distribution probability of the quantified perceived risk C can be constructed as follows: ； ； ； ； Wherein, , represents the probability of observing the quantified perceived risk C at different risk levels, and its magnitude is determined by the relative magnitudes of C and the respective risk level thresholds , and ; the parameter is an uncertainty factor, mainly used to control the smoothness of different driving risk level curves; Step 1.3): The preset parameters in Step 1.2) , and can be determined according to the distribution characteristics of the actual vehicle trajectory data.

8. A method for evaluating driving risks on expressways based on predicted trajectories according to claim 1, characterized in that The specific implementation method of the real-time assessment of driving risk in Step 5 is as follows: Step (1): Select the 1st, 2nd, …, i-th vehicles around the host vehicle as obstacle vehicles, denoted as , , , respectively; Taking the obstacle vehicle as a dynamic obstacle, calculate the quantified perception risk within each look-ahead time in the prediction interval according to the method in Step 3, { , , , }; Step (2): For the quantized perceived risk sequence { , , , } within the prediction interval, input it into the Bayesian model for risk level division described in Step 4 to obtain the posterior distribution probabilities that the quantized perceived risk generated by the obstacle vehicle for the host vehicle at different risk levels (including D, A, S), which are respectively denoted as { , , }, { , , }, { , }; Step (3): Using the posterior distribution probability of the quantified perceived risk as the weight and considering the duration decay effect, perform a fusion calculation on the risk levels within the T-section look-ahead time to obtain the merged perceived risk of the obstacle vehicle for the ego vehicle within the prediction interval generated by the ego vehicle : ； where , , are the Bayesian posterior probabilities of the host vehicle at different risk levels; , and take 2, 1, and 0 respectively to represent the risk level quantization values; is the attenuation coefficient of the pointer duration attenuation effect, and its value range is (0, 1]. When takes 1, it means that the duration attenuation is not considered in the risk assessment process; Step (4): Calculate in real time the integrated perception risk to the host vehicle caused by each obstacle vehicle within the prediction interval according to the method described in steps (1) to (3). caused by each obstacle vehicle .

9. The method for evaluating driving risks on expressways based on predicted trajectories according to claim 8, characterized in that Determine the real-time assessment result of driving risk according to the following steps: Step (4-1): Take the maximum value among the fusion perception risks generated by all obstacle vehicles for the host vehicle as the risk quantification value of the host vehicle at the current moment, denoted as ; Step (4-2): Compare in real time with preset critical warning threshold in size; if , the risk level of the interaction behavior between the current host vehicle and the obstacle vehicle is relatively high, and the system should issue a risk warning prompt; On the contrary, there is no potential risk in the interaction behavior between the host vehicle and the obstacle vehicle.

Images