Dynamic prediction method of curve driving risk by integrating multi-source data in connected environment

Abstract

The present invention provides a method for dynamically predicting the risk of driving on a curve by integrating multi-source data in a networked environment, specifically comprising the following steps: Step 1: Acquire multi-source data of a vehicle during driving on a curve; Step 2: Match the multi-source data according to timestamps and unify the recording frequency; Step 3: Convert the multi-source data from a time mode to a spatial mode; Divide the multi-source data in the spatial mode into a number of data points with a length of x meters; Step 4: Use a sliding window with L as the sliding window length and U as the sliding step size to traverse the data points in Step 3 and record the critical lateral acceleration mean a e ; Step 5: Take the critical lateral acceleration mean a e As a clustering index, the curve driving risk is divided; Step 6: Based on the driving risk level results, the driving risk level of the vehicle is predicted. The present invention improves the accuracy of the prediction.

Description

Technical Field

[0001] The invention belongs to the technical field of traffic safety. Background Art

[0002] Curved roads have poor alignment and relatively short sight distances, making driving risks much higher than on straight roads. When driving at high speeds on sharp bends, the centripetal force required to turn the vehicle is insufficient, increasing the risk of accidents such as rollovers and skidding. It can also easily cause the vehicle to run into oncoming lanes, guardrail ditches, and other areas, resulting in casualties and property damage. Existing safety measures for curved roads mostly consist of static warning signs or use a constant speed limit, failing to provide drivers with reasonable driving advice based on the actual alignment of the road. The development of intelligent connected technology has further strengthened the connection between vehicles, roads, and drivers, providing technical support for assisted driving and proactive intervention. Therefore, establishing a dynamic prediction method for driving risks on curved roads in a connected environment can reduce driving risks on curved roads and improve driving safety.

[0003] Most current curve risk prediction and warning methods focus on collision avoidance, rollover prevention, and side-slip prevention on curves. They focus on establishing safe curve speed models or using side-slip speed, lateral load transfer rate, side-slip gradient angle, and curve trajectory deviation as indicators of curve driving risk. Many methods employ static assessment and analysis methods, evaluating driving performance throughout the curve to calibrate driving risk. However, such analysis methods struggle to dynamically identify and predict curve driving risk. However, the ability to predict driving risk is essential for improving driving safety on curved roads. In fact, due to the design characteristics of curves, the radius of curvature and road cross-slope vary at different locations along the curve. Furthermore, vehicle driving characteristics vary when entering, driving through, and exiting a curve. Vehicle speed along a curve is not constant, exhibiting an overall trend of deceleration followed by acceleration. Vehicle trajectories and lateral deviations also vary at different locations along curves of varying turning angles and radii. Existing technologies primarily consider the overall characteristics of a curve, resulting in inaccurate driving risk predictions. Summary of the Invention

[0004] Purpose of the invention: In order to solve the problems existing in the above-mentioned prior art, the present invention provides a method for dynamically predicting the risk of curve driving by integrating multi-source data in a networked environment.

[0005] Technical Solution: The present invention provides a method for dynamically predicting the risk of driving on a curved road by integrating multi-source data in a networked environment. The method comprises the following steps:

[0006] Step 1: Acquire multi-source data of the vehicle during its curve driving process, wherein the multi-source data includes vehicle dynamics data, driver's control behavior, driver's physiological data and road geometric characteristics;

[0007] Step 2: Match multi-source data according to timestamps and unify the recording frequency;

[0008] Step 3: Convert the multi-source data matched by timestamps in step 2 from a temporal mode to a spatial mode; divide the multi-source data in the spatial mode into several data points with a length of x meters;

[0009] Step 4: Use the sliding window to traverse the data points in step 3, with L as the sliding window length and U as the sliding step size; and record the critical lateral acceleration mean a e ;

[0010] Step 5: Take the critical lateral acceleration mean a e As the clustering index, the K-means clustering method is used to divide the risk of driving on curved roads;

[0011] Step 6: Based on the driving risk level results obtained by K-means clustering, use the long short-term memory network algorithm to predict the driving risk level of the vehicle.

[0012] Furthermore, the step 4 uses the following formula to calculate the critical lateral acceleration mean a e :

[0013]

[0014] Where n represents the number of data points in the sliding window; a j is the lateral acceleration of the vehicle at the jth data point in the sliding window, β is the adjustment parameter, and a t is the acceleration, a t The expression is as follows:

[0015] a t =(i h +μ)g

[0016] Among them, i h is the road cross slope value, μ is the road lateral force coefficient, and g is the acceleration of gravity.

[0017] Furthermore, in step 5, the loss function of K-means is:

[0018]

[0019] Among them, x i represents the i-th sample, c i is the sample x iThe cluster to which it belongs, is the center point of the cluster, and N is the total number of samples. Further, the structure of the long short-term memory network algorithm in step 6 is as follows:

[0020] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0021] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0022]

[0023]

[0024] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0025] h t =o t tanh(C t )

[0026] Among them, f t ,i t ,o t are the values ​​of the forget gate, input gate, and output gate after passing through the activation function at time t; C t-1 and C t are the unit states at the previous moment and the current moment respectively; C t Candidate value of h t-1 is the output of the hidden layer at the previous moment; x t is the input of the long short-term memory network algorithm at the current moment; W f , W i , W C , W o is the weight matrix; b f ,b i ,b C ,b o is the bias vector; σ is the sigmoid function.

[0027] Furthermore, the vehicle dynamics data includes vehicle tangential velocity, tangential acceleration, lateral velocity, lateral acceleration, longitudinal velocity, longitudinal acceleration, yaw velocity, yaw acceleration, pitch velocity, pitch acceleration, roll velocity, roll acceleration and lane deviation; the driver's control behavior includes the force of the accelerator and brake pedals, the steering wheel angle and speed; the driver's physiological data includes the driver's heart rate and skin electrodermal activity; and the road linear characteristics include road radius and cross slope.

[0028] Beneficial effects: The present invention combines the curve road linear characteristics and driving behavior characteristics to design curve risk assessment indicators, and dynamically segments the driving process into space to make it risk predictable; the method proposed by the present invention determines the optimal model window combination by comparing the model prediction performance of different observation windows and interval window combinations, which can avoid the influence of subjective determination of model window length on model accuracy; compared with driving risk prediction through a single data source such as vehicle dynamics, the model prediction performance of the present invention that integrates the driver's operating behavior, the driver's physiological characteristics and the road linear characteristics is significantly improved. Compared with the use of machine learning algorithms, the driving risk prediction method based on the long short-term memory network algorithm proposed by the present invention uses complete and continuous time series data to predict driving risks, which can retain the volatility of the data and improve the prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a flow chart of the present invention;

[0030] Figure 2 Schematic diagram of the curved road section road alignment used in the present invention;

[0031] Figure 3 It is the model structure feature of the driving risk prediction model based on the long short-term memory network algorithm of the present invention;

[0032] Figure 4 This is the data input format of the driving risk prediction model based on the long short-term memory network algorithm proposed in the present invention;

[0033] Figure 5 It is a comparison of model performance with different observation windows and interval windows. DETAILED DESCRIPTION

[0034] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0035] like Figure 1 As shown, the present invention provides a method for dynamically predicting the risk of driving on a curved road by integrating multi-source data in a networked environment. The method comprises the following steps:

[0036] Step 1: Obtain multi-source data of the vehicle during cornering through on-board equipment, vehicle sensors, smart wearable devices, etc. The multi-source data includes: vehicle dynamics data, driver control behavior, driver physiological data and road geometric linear characteristics.

[0037] This embodiment conducts a curve driving simulation experiment based on a driving simulation platform. The driving simulation sphere is equipped with 6 projectors, which can realize a 360-degree full-scale display of the road scene. The cockpit is converted from an automatic Hyundai Sonata vehicle to achieve a high degree of fidelity in driving simulation. SCANeRstudio software is used to set up the curve driving scene and record and extract data. All experimental participants wore a non-invasive wireless wrist heart rate belt (Empatica E4) to collect physiological data. The wristband sensor can continuously collect spectral (PPG) signals and skin electrodermal signals (EDA).

[0038] The experimental scenario is a mountainous section of a third-class highway with two lanes in both directions and a design speed of 40 km / h. The experiment is set in a free-flow traffic state. The curve radii are 60m, 100m, and 150m respectively. Each radius curve has two turning options: left turn (L) and right turn (R). The experimental curve line design is as follows: Figure 2 As shown in the figure, the geometric linear design of the curves complies with the relevant highway design specifications (Highway Alignment Design Specifications JTG D20-2017).

[0039] The experiment collected data from 55 drivers (41 males and 14 females), aged 21-50 years (mean = 27.3 years, standard deviation = 6.1 years). All drivers held valid driver's licenses, with a driving experience ranging from 1 to 12 years (mean = 3.9 years, standard deviation = 2.6 years). All drivers had good or corrected-to-good vision and were in good physical condition before the experiment, with no history of fatigue or alcohol-induced driving. Drivers took a 10-minute test drive before the actual experiment to familiarize themselves with the simulated driving environment. Participants who experienced symptoms such as dizziness were asked to terminate the experiment.

[0040] Step 2: Match multi-source data according to timestamps and unify the recording frequency.

[0041] First, the vehicle motion data is matched with the driver's physiological data using timestamps, and the frequency is unified to 10Hz. Then, the data is intercepted based on the collected vehicle travel coordinates, extracting the straight section 100 meters before entering the curve and the data within the curve.

[0042] Then, the data recording mode is converted from time to space by calculating the average value and linear interpolation, that is, the curved road section is divided into several segments of x meters (x=2 in this embodiment). The main data and indicator characteristics collected in the experiment are shown in Table 1:

[0043] Table 1

[0044]

[0045] Step 3: Considering the differences in road alignment and vehicle driving characteristics at different locations on the curve, use the method of calculating the mean and linear interpolation to convert the data recording method from time to space. The data length is x meters, and a number of data points are obtained.

[0046] Step 4: Use the sliding window method to dynamically divide the curve into spatial segments, with L (in this embodiment, L = 10m) as the sliding window length and U (in this embodiment, U = 2m) as the sliding step length to perform sliding segmentation on the data points, and calculate the critical lateral acceleration mean a e :

[0047]

[0048] Where n represents the number of data points in the sliding window; a j is the lateral acceleration of the vehicle at the jth data point in the sliding window. The sliding window length L is determined according to the average vehicle speed and the driver's perception reaction time. β is the adjustment parameter. a t is the acceleration, a t The expression is as follows:

[0049] a t =(i h +μ)g

[0050] Among them, i h is the road cross slope value, μ is the road lateral force coefficient, and g is the acceleration of gravity.

[0051] Step 5: Take the critical lateral acceleration mean a e As the clustering index, the K-means clustering method is used to divide the risk of driving on curved roads into three levels: low, medium, and high. The risk levels are shown in Table 2.

[0052] Table 2

[0053] Risk Level <![CDATA[a e Minimum value]]> <![CDATA[a e Maximum value]]> <![CDATA[a e Mean]]> Low 0.00 0.51 0.01 middle 0.51 1.82 1.01 high 1.82 6.58 2.64

[0054] The K-means clustering method of this embodiment uses an iterative method to find K clusters so that the loss function corresponding to the clustering result is minimized. The loss function is defined as the sum of squared errors of each sample from the center point of the cluster to which it belongs, as shown below:

[0055]

[0056] Among them, x i represents the i-th sample, ci is the sample x i The cluster to which it belongs, is the center point of the cluster, and N is the total number of samples.

[0057] Step 6: Use the long short-term memory network algorithm (LSTM) to predict driving risks; the LSTM network structure is as follows Figure 3 As shown, the LSTM unit formula is as follows:

[0058] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0059] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0060]

[0061]

[0062] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0063] h t =o t tanh(C t )

[0064] Where: f t ,i t ,o t are the values ​​of the forget gate, input gate, and output gate after passing through the activation function at time t; C t-1 and C t are the unit states at the previous moment and the current moment respectively; C t Candidate value of h t-1 is the output of the hidden layer at the previous moment; x t is the input at the current moment; W f , W i , W C , W o is the weight matrix; b f ,b i ,b C ,b ois the bias vector; σ is the sigmoid function. At each time step, the LSTM can read, write, or reset the memory cell through these three gates. This strategy allows the LSTM to remember and access information from multiple time steps ago.

[0065] It is necessary to determine the observation window and interval window length of LSTM. The data input format of LSTM algorithm is as follows Figure 4 Considering factors such as the average speed on the curve, the driver's perception reaction time, and the model prediction accuracy, the search ranges for the observation window and the interval window are set, and the models are tuned separately to select the optimal window length combination.

[0066] The average driving speed on the curve in this embodiment is 36km / h. Taking into account the driver's perception reaction time, the search range of the interval window length is set to 10~50m, and the search range of the observation window length is 2~50m (the search step is 2m). A total of 525 (25×21) observation window and interval window combinations are designed. The LSTM model includes an input layer, two LSTM hidden layers, and one output layer. The model input features are the variables in Table 1, a total of 21 variables, and the input data is Z-score standardized. The data training set, validation set, and test set are defined in a ratio of 6:2:2. Accuracy and F1-score are used as evaluation indicators of the model. In the driving risk prediction scenario, the model's ability to predict medium and high driving risk levels is more important. Therefore, the model's F1-score for medium and high driving risks is averaged to compare the model prediction performance of different observation window and interval window combinations, such as Figure 5 As shown in the figure, the combination of the model observation window and interval window is finally determined to be 20m.

[0067] Secondly, to evaluate the effectiveness of the proposed model (i.e., the prediction model based on the long short-term memory network algorithm, LSTM) in predicting driving risk levels, it is compared with the following three algorithms: random forest (RF), extreme gradient boosting tree (XGB), and lightweight gradient boosting machine (LGB). As mentioned above, the input sample of the LSTM model is the complete time series data, which is a 21×20 two-dimensional matrix with the data format as follows Figure 4 As shown, the model observation window and interval window lengths are both 20 meters. For the RF, XGB, and LGB algorithms, the model input variables are vehicle kinematic data within the 20-meter observation window, driver control behavior, and driver physiological indicators (maximum, minimum, average, and standard deviation), as well as the average curve radius and cross slope within the observation window, i.e., a one-dimensional vector of 19 × 4 + 2 = 78. The key parameters of the LSTM, RF, XGB, and LGB algorithms were selected and adjusted within the specified parameter range to determine the optimal parameter combination, as shown in Table 3:

[0068] Table 3

[0069]

[0070]

[0071] Evaluate the model effect. Use confusion matrix as the model evaluation method, and select four indicators: accuracy, precision, recall, and F1 score to evaluate model performance:

[0072]

[0073]

[0074]

[0075]

[0076] Where: TP, FP, FN, and TN are the number of true positives, false positives, false negatives, and true negatives, respectively.

[0077] The model performance of the four algorithms in predicting the risk level of driving on curved roads is shown in Table 4. The results show that the model proposed in this embodiment is effective in predicting the driving risk level.

[0078] Table 4

[0079]

[0080] Table 5 compares the results of models using only vehicle kinematic data and integrating multi-source data for driving risk prediction under the same window length combination. The model using only vehicle kinematic data achieved prediction accuracies of 81.6% and 76.5% for medium and high driving risk, and recall rates of 77.3% and 84.6%, respectively. Integrating the driver's heart rate, electrodermal data, steering behavior, and road alignment improved both accuracy and recall for medium and high driving risk levels, with the F1-score for medium driving risk increasing from 79.4% to 84.8%. Table 5 demonstrates that the multi-source data integration method proposed in this paper can effectively improve model prediction performance.

[0081] Table 5

[0082]

[0083] The vehicle dynamics data described in this embodiment includes the vehicle's tangential velocity, tangential velocity acceleration, lateral velocity, lateral acceleration, longitudinal velocity, longitudinal acceleration, yaw velocity, yaw acceleration, pitch velocity, pitch acceleration, roll velocity, roll acceleration and lane deviation; the driver's control behavior includes the force of the accelerator and brake pedals, the steering wheel angle and speed; the driver's physiological data includes the driver's heart rate and skin electrodermal activity; and the road linear characteristics include the road radius and cross slope.

[0084] When using the LSTM model to predict driving risks, the grid search method is used to determine the optimal combination of observation window and interval window length. The search range comprehensively considers factors such as the average driving speed on the curve, the driver's perception reaction time, and the model prediction accuracy, to avoid the impact of subjective setting of the model window on model accuracy.

[0085] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any appropriate manner without contradiction. To avoid unnecessary repetition, the present invention will not further describe various possible combinations.

Claims

1. A dynamic prediction method for curve driving risk integrating multi-source data in a connected environment, characterized by: The steps include: Step 1: Acquire multi-source data of the vehicle during its curve driving process, wherein the multi-source data includes vehicle dynamics data, driver's control behavior, driver's physiological data and road geometric characteristics; Step 2: Match multi-source data according to timestamps and unify the recording frequency; Step 3: Convert the multi-source data matched by timestamps in step 2 from a temporal mode to a spatial mode; Divide the multi-source data in the spatial mode into several data points with a length of x meters; Step 4: Use the sliding window to traverse the data points in step 3, with L as the sliding window length and U as the sliding step size; and record the critical lateral acceleration mean a e ; Step 5: Take the critical lateral acceleration mean a e As the clustering index, the K-means clustering method is used to divide the risk of driving on curved roads; Step 6: Based on the driving risk level results obtained by K-means clustering, use the long short-term memory network algorithm to predict the driving risk level of the vehicle; The step 4 uses the following formula to calculate the critical lateral acceleration mean a e : Where n represents the number of data points in the sliding window; a j is the lateral acceleration of the vehicle at the jth data point in the sliding window, β is the adjustment parameter, and a t is the acceleration, a t The expression is as follows: a t =(i h +μ)g Among them, i h is the road cross slope value, μ is the road lateral force coefficient, and g is the acceleration of gravity.

2. The method for dynamic prediction of curve driving risk by integrating multi-source data in a networked environment according to claim 1 is characterized in that: In step 5, the loss function of K-means is: Among them, x i represents the i-th sample, c i is the sample x i The cluster to which it belongs, is the center point of the cluster, and N is the total number of samples.

3. The method for dynamic prediction of curve driving risk by integrating multi-source data in a connected environment according to claim 1, characterized in that: The structure of the long short-term memory network algorithm in step 6 is as follows: f t =σ(W f ·[h t-1 ,x t ]+b f ) i t =σ(W i ·[ h t -1 ,x t ]+b i ) the t =σ(W o ·[h t-1 ,x t ]+b o ) h t =o t ·tanh(C t ) Among them, f t ,i t ,o t are the values ​​of the forget gate, input gate, and output gate after passing through the activation function at time t; C t-1 and C t are the unit states at the previous moment and the current moment respectively; C t Candidate value of h t-1 is the output of the hidden layer at the previous moment; x t is the input of the long short-term memory network algorithm at the current moment; W f , W i , W C , W o is the weight matrix; b f ,b i ,b C ,b o is the bias vector; σ is the sigmoid function.

4. The method for dynamic prediction of curve driving risk by integrating multi-source data in a networked environment according to claim 1 is characterized in that: The vehicle dynamics data includes vehicle tangential velocity, tangential acceleration, lateral velocity, lateral acceleration, longitudinal velocity, longitudinal acceleration, yaw velocity, yaw acceleration, pitch velocity, pitch acceleration, roll velocity, roll acceleration and lane deviation; the driver's control behavior includes the force of the accelerator and brake pedals, the steering wheel angle and speed; the driver's physiological data includes the driver's heart rate and skin electrodermal activity; and the road linear characteristics include road radius and cross slope.

Images