Method for recognizing dangerous turning behavior based on boundary search

Abstract

The application provides a dangerous turning behavior recognition method based on boundary search, uses currently available vehicle driving behavior data, analyzes and calculates a large amount of data to obtain a dangerous turning behavior recognition algorithm which is more adaptive and more accurate in recognition, and can accurately recognize dangerous turning behavior. The basic idea of the method is as follows: first, a turning behavior database containing dangerous turning behavior and safe turning behavior is established; each data in the turning behavior database includes complete information of a turning behavior, such as lateral speed, longitudinal speed, lateral acceleration, roll angle, yaw angular velocity, steering wheel rotation angle and the like of the vehicle; then based on the established turning behavior database, the behavior characteristics are extracted by using the boundary search and neural network method, and finally the driving turning behavior recognition algorithm is constructed and realized.

Description

Technical Field

[0001] This invention relates to an identification method, specifically a method for identifying dangerous steering behavior, and belongs to the field of intelligent vehicle control technology. Background Technology

[0002] Accurate identification of driving behavior allows ESC to make more precise controls. In emergency situations, drivers often make actions that differ from those under normal conditions. These actions can be categorized as braking and steering. If these characteristics can be accurately identified, ESC can issue early warnings, determining that the vehicle is about to or is in an unstable state, thus providing a basis for further intervention and control by ESC.

[0003] Currently, research institutions both domestically and internationally have conducted extensive studies on driver emergency braking behavior, but research on dangerous steering behavior is still in its early stages. Van Winsum et al. explored the relationship between normal steering behavior and vehicle speed, arguing that road curvature affects not only steering wheel angle but also vehicle speed; a smaller turning radius not only leads to increased steering wheel speed but also decreased vehicle speed. Hong T proposed a steering behavior recognition algorithm based on vehicle lateral velocity, constructing a vehicle handling feature map of a "central area." The driver's position within the map is determined by the magnitude of steering wheel torque and vehicle lateral velocity, thereby judging whether the driver has engaged in dangerous steering behavior. Zhang Zexing proposed a driver steering behavior identification method based on a multidimensional Gaussian Hidden Markov Model (HMM). Utilizing the time-series characteristics of HMM modeling theory, he used the Baum-Welch algorithm to obtain and optimize a driver behavior identification model that can distinguish between emergency steering, normal steering, and straight-line driving. Wu Bin et al. used steering primitives fitted with Gaussian functions to describe steering behavior under emergency conditions and classified dangerous steering behavior based on time to collision. Martin Liebne summarized a series of driver behaviors and characteristics, deriving corresponding relationships. He ultimately established a model that can infer driver intent based on features such as gaze distance and vehicle speed. By inputting relevant parameters, the model can provide a hypothetical event probability. This event probability can be applied to vehicle accident avoidance and driver assistance. Mostafa Hossny proposed a method for predicting driving behavior using "Change Point Segmentation," which predicts driving behavior by detecting feature variation points. This method improves the flexibility of driving behavior prediction and can make relatively reasonable predictions even when facing complex behaviors.

[0004] The aforementioned studies have all achieved partial functionality in recognizing dangerous driver steering behavior, but most of them added extra sensors to the vehicle to achieve this goal. Some studies only analyzed the data without completing a comprehensive driving behavior recognition algorithm. Furthermore, these studies are largely based on experimental simulations or a limited number of successful collision avoidance cases, lacking analysis and research on real-world accident cases, and insufficient research on the relationship between vehicle speed and steering behavior in emergency situations. The limited number of cases means that the resulting driving behavior recognition methods cannot be fully validated, and the dangerous steering behaviors obtained through simulation often differ from those in real-world cases.

[0005] The development of intelligent vehicles in recent years has provided new materials for research on driving behavior recognition. Many research institutions have conducted related research on intelligent driving, such as Commai.ai, KITT, Cityscapes, and Oxford. While conducting their own research, these institutions have also published their respective databases. These driving databases contain all or part of the information from driving videos, vehicle speed, steering wheel angle, GPS, and IMU, from which a large amount of safe steering behavior data can be extracted, which can be used for the analysis and calculation of driving behavior recognition.

[0006] The EDR (Event Data Recorder) device in an ESC records the vehicle's state before an accident. The NHTSA (National Highway Traffic Safety Administration) has collected and made public EDR files for some vehicles involved in accidents from 1988 to the present. The NHTSA has also conducted a simple analysis of this data, showing commonalities in vehicle states before accidents. EDR files provide data for establishing a database of driving behavior in emergency situations. All of these databases can collectively form a driving behavior database, which can be used for analyzing driving behavior. Summary of the Invention

[0007] In view of this, the present invention provides a dangerous steering behavior recognition method based on boundary search. By utilizing currently available vehicle driving behavior data, a more adaptable and accurate dangerous steering behavior recognition algorithm is obtained through analysis and calculation of a large amount of data, thereby enabling accurate recognition of dangerous steering behaviors.

[0008] A dangerous turning behavior identification method based on boundary search:

[0009] Step 1: Establish a steering behavior database, including: a hazardous steering behavior database and a safe steering behavior database;

[0010] Step 2: Selection of steering behavior characteristic quantities:

[0011] Steering wheel angle, steering wheel speed, and vehicle speed are selected from the steering behavior database as steering behavior feature quantities;

[0012] Step 3: Based on the steering behavior feature quantity selected in Step 2, perform statistical analysis on the data in the steering behavior database in Step 1 using the boundary search method to obtain the boundary between dangerous steering behavior and safe steering behavior, and calculate the danger index of each steering behavior in the steering behavior database accordingly.

[0013] Step 4: Identification of Dangerous Turning Behaviors Based on Neural Networks

[0014] The neural network model is trained based on the steering behavior data in the steering behavior database in step 1 and the corresponding steering behavior hazard index in step 3 to establish a neural network model for identifying dangerous steering behaviors. The neural network model can identify dangerous steering behaviors by outputting a hazard index.

[0015] As a preferred embodiment of the present invention: in step 3, when performing statistical analysis on the data in the steering behavior database from step 1 based on the boundary search method:

[0016] 301: Region division based on steering wheel angle

[0017] On a two-dimensional plane diagram reflecting the relationship between steering wheel angle and vehicle speed, the boundary search method is used to calculate the boundary curves L1(x) and L2(x) according to p1 = 0.9 and p2 = 0.99, respectively.

[0018] 302: Regional division based on steering wheel rotation speed

[0019] On a two-dimensional plane diagram that reflects the relationship between steering wheel rotation speed and vehicle speed, the boundary search method is used to calculate the boundary curve L3(x) according to the set p3 = 0.99.

[0020] As a preferred embodiment of the present invention: in step 301, the area between 17m / s and 22m / s within the safe area corresponding to the boundary curve L1(x) is designated as the risk area.

[0021] As a preferred embodiment of the present invention, the method for calculating the steering behavior hazard index D is as follows:

[0022]

[0023] Where: k1, k2, k3, k4 are constant parameters, δ is the absolute value of the steering wheel angle, δ' is the absolute value of the steering wheel speed, and x is the vehicle speed in m / s.

[0024] As a preferred embodiment of the present invention: In step 3: when performing statistical analysis on the data in the steering behavior database based on the boundary search method, the maximum steering wheel angle, the maximum steering wheel speed, and the maximum vehicle speed within the corresponding time period for each steering case in the steering behavior database are selected as the three feature values ​​of the steering behavior case, and a relationship graph between steering wheel angle and vehicle speed and a relationship graph between steering wheel speed and vehicle speed are drawn based on the feature values.

[0025] As a preferred embodiment of the present invention: in step 1: a dangerous steering behavior database is constructed using dangerous steering behavior data extracted from published accident vehicle EDRs; a safe steering behavior database is constructed using steering behavior data from at least two published driving databases.

[0026] As a preferred embodiment of the present invention: In step 4, when training the neural network model, each steering behavior data in the steering behavior database in step 1 is used as the training sample set, and 15% of the cases are randomly divided into a validation set, 15% of the cases are used as a test set, and 70% of the cases are used as the training set.

[0027] As a preferred embodiment of the present invention: a neural network model is built and trained in Matlab, with a learning rate of 0.01, a maximum number of iteration layers of 1000, a hidden layer activation function of tansig, and an output layer activation function of logsig; one hidden layer with 15 neurons is selected.

[0028] Beneficial effects:

[0029] (1) Based on the existing driving behavior database, this invention selects steering behavior feature quantities and then uses boundary search and neural network methods to extract behavioral features, thereby constructing a driving steering behavior recognition algorithm that can accurately identify dangerous steering behaviors, thus providing a basis for further intervention control of ESC.

[0030] (2) In this invention, the selected steering behavior characteristics are: vehicle steering wheel angle, steering wheel speed and vehicle speed, which are all obtained by common automotive sensors. No additional sensors are needed when intelligently controlling the vehicle.

[0031] (3) In this invention, when dividing the area based on the steering wheel angle, it is considered that the vehicle speed is around 20 m / s, which has the highest steering sensitivity, but also brings higher driving control risks. Based on this, the area around 16 m / s to 22 m / s within the safe zone is designated as the risk zone, thereby improving the accuracy of dangerous steering behavior identification.

[0032] (4) In this invention, when training the neural network model, the established neural network model is tested by randomly selecting a test set from the behavior database to ensure the accuracy of the neural network model. Attached Figure Description

[0033] Figure 1 This is a sample distribution diagram reflecting the relationship between steering wheel angle and vehicle speed in the steering behavior database;

[0034] Figure 2 Based on Figure 1 A diagram illustrating the boundaries of search results;

[0035] Figure 3 Based on Figure 2 A schematic diagram illustrating the division of steering behavior zones;

[0036] Figure 4 This is a map showing the area division based on steering wheel angle after adding risk zones;

[0037] Figure 5 This is a sample distribution diagram reflecting the relationship between steering wheel rotation speed and vehicle speed in the steering behavior database;

[0038] Figure 6 Based on Figure 5 A diagram illustrating the boundaries of search results;

[0039] Figure 7 Based on Figure 6 A schematic diagram illustrating the division of steering behavior zones;

[0040] Figure 8 This diagram illustrates the iterative calculation process when training a neural network model using the backpropagation algorithm in Matlab.

[0041] Figure 9 This is a confusion matrix diagram drawn after the trained neural network model classifies the training set, validation set, and test set. Detailed Implementation

[0042] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0043] This embodiment provides a dangerous turning behavior identification method based on boundary search, which can more accurately identify dangerous turning behaviors.

[0044] The basic idea of ​​this method is as follows: First, a steering behavior database containing dangerous and safe steering behaviors is established. Each piece of data in the steering behavior database includes complete information about a steering behavior, such as the vehicle's lateral speed, longitudinal speed, lateral acceleration, roll angle, yaw rate, and steering wheel angle. Then, based on the constructed steering behavior database, steering behavior feature quantities are selected, and behavioral features are extracted using boundary search and neural network methods. Finally, a driving steering behavior recognition algorithm is constructed and implemented. After inputting steering-related parameters, the neural network model outputs the driver's hazard index.

[0045] Step 1: Establish a steering behavior database

[0046] The steering behavior database includes: a hazardous steering behavior database and a safe steering behavior database.

[0047] As research into intelligent vehicles deepens, more and more research institutions are publicly releasing their driving behavior databases. Commai.ai has built a database containing 7.25 hours of driving data, including 160×320 resolution video footage, 100Hz accelerometer data, and information on vehicle speed, steering wheel angle, GPS, and IMU. KITT's database primarily includes information about the vehicle's surroundings during driving, including video and point cloud maps. Cityscapes has created a database containing driving data from 50 cities, mainly recording point cloud maps acquired by LiDAR and video footage from cameras, while also labeling some objects in the videos. DBNet has built a database containing over 70 segments of driving data, including information such as steering wheel angle and vehicle speed. Oxford's database includes 1000km of driving records, including video data from six cameras and data from LiDAR, GPS, and INS devices. BDDV's database includes driving data from multiple cities and various weather conditions, including information on vehicle speed, driver behavior, and video data. This solution selects Commai.ai and DBNet as the foundation for the safe steering behavior database. Drivers' dangerous steering behaviors often occur before accidents, and the EDR (Electronic Event Data Recording) system in ESC (Electronic Stability Control) can accurately record the driver's behavior and vehicle state at that moment. Therefore, dangerous steering behavior data is extracted from the EDRs of accident vehicles published by the NHTSA to construct a dangerous steering behavior database. Finally, the safe steering behavior database and the dangerous steering behavior database are combined to form a steering behavior database for analysis.

[0048] Dangerous Turning Behavior Database:

[0049] ESC includes a module called EDR, which records the vehicle's status within 5 seconds before a collision and for a period afterward. With the increasing adoption of ESC, more and more EDR information is being recorded. Therefore, this solution establishes a database of dangerous steering behaviors based on this data. EDR can be considered the car's black box; it records information such as steering wheel angle, yaw rate, lateral speed, vehicle speed, throttle opening, brake switch status, wheel speed, transmission position, and airbag deployment data within 5 seconds before an accident. Normally, EDR is built into the vehicle's ESC. After an accident, relevant information can be extracted from the vehicle's ESC.

[0050] Using the Bosch CDR 900 EDR reading tool, after connecting to the vehicle's ESC, the EDR information can be read and generated as a binary data file in .cdr format. Decoding the .cdr file converts it into a .pdf report. Each report contains a wealth of data suitable for accident analysis. This includes the vehicle model code, vehicle sensor configuration, and data recorded by the sensors prior to the accident. After summarizing and extracting information from the EDR files, data on dangerous steering behavior suitable for analysis can be obtained. Since the files originate from different versions of ESC, the content varies slightly, requiring further filtering and processing. Each EDR file contains information on one accident, recording sensor data from the 5 seconds prior to the accident. To complete the analysis of steering behavior, EDR files containing steering wheel angle, yaw rate, and vehicle speed are selected. After screening and extraction, 688 EDR files with complete information (including the required steering wheel angle, yaw rate, and vehicle speed) were obtained, meeting the analysis requirements. This constitutes the established database of dangerous steering behaviors containing 688 EDR files. Each EDR file includes data on the vehicle's lateral velocity, longitudinal velocity, lateral acceleration, roll angle, yaw rate, and steering wheel angle. Before an accident, different vehicles exhibit different steering behaviors, but commonalities between dangerous steering behaviors are not readily apparent. Therefore, different methods need to be considered for analysis.

[0051] Safe steering behavior database:

[0052] Based on the requirements for analyzing steering behavior characteristics, in addition to a database of dangerous steering behaviors recording pre-accident maneuvering actions, a corresponding database of safe steering behaviors also needs to be established. To avoid differences in steering behavior due to driver habits and considering the requirements for database content, this solution selects two different driving databases. Both databases contain information such as vehicle speed and steering wheel angle. The two databases are merged into a single .json data file with the same format. Analogous to the dangerous steering behavior database, each data entry in the data file is stored in dictionary form, containing information such as vehicle speed, acceleration, and steering wheel angle. The merged database contains 80 independent driving processes. To improve the accuracy of further analysis, this solution extracts the steering behaviors and removes abrupt changes caused by sensor errors. Furthermore, to maintain consistency with the data format of the dangerous steering behavior database, the steering process is uniformly divided into 5-second time intervals.

[0053] The resulting safe steering behavior database, after data cleaning, contains 700 steering behavior cases, each lasting 5 seconds. These cases include data on vehicle lateral velocity, longitudinal velocity, lateral acceleration, roll angle, yaw rate, and steering wheel angle. In other words, the established safe steering behavior database contains 700 steering behavior cases, each including data on vehicle lateral velocity, longitudinal velocity, lateral acceleration, roll angle, yaw rate, and steering wheel angle.

[0054] The resulting database of safe steering behaviors, combined with the database of dangerous steering behaviors, can then be used for steering behavior analysis and identification. The final steering behavior database contains 1388 steering behavior cases, including 700 safe steering behaviors and 688 dangerous steering behaviors, with each case lasting 5 seconds.

[0055] Step Two: Selection of Steering Behavior Characteristics

[0056] Each data point in the constructed steering behavior database includes the vehicle's lateral velocity, longitudinal velocity, lateral acceleration, roll angle, yaw rate, and steering wheel angle. To achieve the goal of identifying dangerous steering behaviors, appropriate features need to be selected from this database as the basis for identification.

[0057] In this scheme, three parameters are selected as steering behavior feature quantities: vehicle steering wheel angle, steering wheel speed (obtained by differentiating the steering wheel angle), and vehicle longitudinal speed (hereinafter referred to as vehicle speed). The steering behavior is preliminarily analyzed using the boundary search method.

[0058] After determining the impact of each steering behavior feature on steering behavior recognition, a neural network method is used to determine the final dangerous steering behavior recognition algorithm. In the boundary search method, the temporal features need to be simplified, with the maximum steering wheel angle and maximum steering wheel speed within each steering phase selected as the features of a single steering behavior. Considering the influence of vehicle speed on steering behavior, the maximum steering wheel angle, maximum steering wheel speed, and maximum vehicle speed recorded for each steering case in the database (each case in the database lasts for 5 seconds, and the maximum value of the corresponding feature within 5 seconds) are selected as the three feature values ​​for that case. A three-dimensional scatter plot is then drawn based on these feature values ​​to visually observe the differences between different cases. In this three-dimensional scatter plot, the X-axis represents the maximum steering wheel angle, the Y-axis represents the maximum steering wheel speed, and the Z-axis represents the maximum vehicle speed.

[0059] Step 3: Identification of Dangerous Turning Behaviors

[0060] The steering behavior data is analyzed based on the steering behavior features selected in step two to achieve the goal of identifying dangerous steering behaviors. This scheme employs two methods to analyze the steering behavior database: boundary search and neural networks. Boundary search uses statistical analysis of the data to obtain the boundary that distinguishes dangerous steering behaviors from safe steering behaviors, proving the difference between them. The neural network, on the other hand, trains a neural network model based on existing driving behavior data to obtain a classifier that can be used for behavior recognition.

[0061] 301: Classification of Dangerous Turning Behaviors Based on Boundary Search

[0062] Boundary search is an image-based classification method that summarizes the features of samples by finding the envelope of sample points. In a two-dimensional plane, for a graph where sample points are similarly distributed along the x-axis, the boundary search problem can be simplified to finding a curve L(x) such that as many sample points (x, x) are similarly distributed along the x-axis. i y i ) satisfies L(x i )-y i >0. Assuming there are a total of n2 sample points, of which n1 sample points satisfy this condition, then p = n1 / n2 can be defined to evaluate the degree of inclusion of L(x) with respect to the sample points.

[0063] This solution further simplifies the problem by using a discretization method, that is, by adjusting each sample point in the behavioral database according to x. i Group the data into m sample points (i.e., group them according to their x-coordinate values, with each group containing m sample points). Then, analyze the m y-coordinates of each group. i Sort the data, and finally select the y-value at the set position p in each group. iThese serve as the boundary values ​​for the group. Connecting all boundary values ​​yields a boundary polygonal line, and fitting this line with a polynomial yields the boundary polynomial L(x) that satisfies condition p. By randomly selecting some data points in a two-dimensional plane, the entire calculation process can be simulated.

[0064] Since the boundary search method does not provide good support for temporal features, a simplification method based on maximizing the temporal features is employed. The maximum steering wheel angle, maximum steering wheel speed, and maximum vehicle speed within the corresponding time period for each case are selected as the features for each single steering case. To further simplify the analysis method, the three feature values ​​are projected from a 3D scatter plot in 3D space onto different 2D planes for analysis.

[0065] The projection of the three eigenvalues ​​of the three-dimensional scatter points in three-dimensional space onto the XZ plane is as follows: Figure 1 As shown, Figure 1 This reflects the relationship between steering wheel angle and vehicle speed. Following the boundary search method described above, curves L1(x) and L2(x) are calculated with p1 = 0.9 and p2 = 0.99 respectively, ultimately yielding the following... Figure 2 The L1(x) and L2(x) are shown.

[0066] The curve L1(x) satisfying the condition p1 = 0.9 is fitted by the following fourth-order polynomial:

[0067] L1(x) = a0 + a1x + a2x 2 +a3x 3 +a4x 4 (1)

[0068] The L2(x) satisfying the condition p2 = 0.99 is fitted by the following cubic polynomial:

[0069] L2(x) = b0 + b1x + b2x 2 +b3x 3 (2)

[0070] Where: a0 = 114.67730805; a1 = 5.34617032; a2 = -1.50551759; a3 = 0.07761676; a4 = -0.00116450; b0 = 196.69485727; b1 = -8.25530994; b2 = 0.53625039; b3 = -0.01141963.

[0071] Let curve L1(x) be the safety line and curve L2(x) be the warning line; from Figure 2It can be seen that 90% of safe steering behaviors fall within the safety line L1(x), and 99% fall within the warning line L2(x). Dangerous steering behaviors, on the other hand, occur as follows: 47.55% fall within the safety line L1(x), 23.63% fall between the safety line L1(x) and the warning line L2(x), and 28.82% fall outside the warning line L2(x). This means that in accidents involving steering behaviors, 52% of the cases involve steering wheel angles that differ from 90% of conventional steering operations. Therefore, steering wheel angle can be used as one of the parameters for accident warning, and its safe range can be described by the safety line L1(x) and the warning line L2(x), ultimately leading to... Figure 4 The steering behavior area is divided as follows: the area within the safety line L1(x) is the safe zone, the area between the safety line L1(x) and the warning line L2(x) is the warning zone, and the area outside the warning line L2(x) is the danger zone.

[0072] Observing the safety line L1(x), we can see that it continuously decreases from low speed to high speed, causing the safety zone to shrink. Around 20 m / s, it experiences a low point in steering wheel angle, before rising again. According to the accident case heat map, around 20 m / s is precisely the high-risk area for accidents.

[0073] like Figure 3 As shown, in safe steering behavior, a vehicle speed of around 20 m / s is also the minimum value for the maximum steering wheel angle. That is, at speeds within this range, drivers tend to use smaller steering wheel angles to control the vehicle. In fact, if the vehicle model is simplified to a two-degree-of-freedom model, the steady-state yaw rate gain can be obtained:

[0074]

[0075] In the formula: ω r δ is the vehicle yaw rate, δ is the vehicle front wheel steering angle, u is the vehicle speed, L is the vehicle wheelbase, m is the vehicle mass, a is the distance from the vehicle's center of gravity to the front axle, b is the distance from the vehicle's center of gravity to the rear axle, k1 is the vehicle's front wheel tire lateral stiffness, k2 is the vehicle's rear wheel tire lateral stiffness, and K is the vehicle stability factor.

[0076] Modern cars often exhibit understeer characteristics, with a K-value typically around 0.0024s. 2 / m 2 ~0.0026s 2 / m 2 Between these, it is easy to obtain the appropriate value by differentiating u in equation (3). At this point, the steady-state yaw rate gain reaches its maximum value, and calculations show that the vehicle speed is approximately 20 m / s. This means the vehicle's characteristic speed is mostly around 20 m / s. This characteristic of the vehicle can be verified through relevant simulation experiments in Matlab / Simulink. By conducting a 90-degree steering wheel angle step input test at seven different vehicle speeds, we can obtain graphs showing the vehicle's yaw rate over time and the steady-state yaw rate over vehicle speed.

[0077] Verification analysis shows that, under the same steering wheel angle input, the vehicle reaches its highest peak yaw rate at a speed of 20 m / s, and its highest steady-state yaw rate between 15 m / s and 20 m / s, consistent with the previously mentioned theoretical analysis. Vehicles traveling within this speed range exhibit the highest steering sensitivity, but this also brings higher risks to vehicle control. In the event of a sudden incident, if the driver over-controls, higher steering sensitivity means the vehicle will reach a higher yaw rate. The magnitude of the yaw rate not only affects whether the vehicle loses control but also significantly impacts the driver's vision and driving judgment. The relationship between the maximum steering wheel angle and vehicle speed at other speeds can be explained by the same theory: below the characteristic speed, the maximum steering wheel angle decreases with increasing speed, meaning steering sensitivity increases; above the characteristic speed, the maximum steering wheel angle increases with increasing speed, meaning steering sensitivity decreases. Taking these factors into account, when defining the safe zone for steering wheel angle, the area between 17m / s and 22m / s within the safe zone is designated as the risk zone, ultimately yielding... Figure 4 The diagram shown is a region division based on the steering wheel angle.

[0078] Similarly, the projection of the three eigenvalues ​​onto the yz plane using a 3D scatter plot in three-dimensional space can reflect the relationship between steering wheel rotation speed and vehicle speed, such as... Figure 5 As shown. Following the aforementioned boundary search method, L3(x) is calculated with p3 = 0.99, as follows... Figure 6 As shown.

[0079] Similar to the analysis of the steering wheel angle, the boundary is fitted using a quadratic polynomial L3(x):

[0080] L3(x) = c0 + c1x + c2x 2 (4)

[0081] Where: c0 = 79.08185361; c1 = -1.44657488; c2 = 0.02488945;

[0082] Based on L3(x), a diagram showing the division of steering behavior regions according to steering wheel speed can be drawn, as follows: Figure 7 As shown.

[0083] Calculations show that approximately 78.1% of pre-accident steering behaviors exceeded the safety boundary in terms of maximum steering wheel speed. Steering wheel speeds exceeding the safety boundary can be considered to be within the danger zone.

[0084] Based on the analysis of normal driving data and accident driving data, a driver steering behavior hazard index can be obtained. Therefore, driver steering behavior can be divided into four zones—safe zone, warning zone, danger zone, and risk zone—according to boundary curves L1(x), L2(x), and L3(x). To quantify the magnitude of the risk in driver steering behavior, a driver steering behavior hazard index D is introduced. Generally speaking, the further the driver's steering behavior deviates from the safe zone, the greater the risk. Therefore, let the absolute value of the driver's steering wheel angle be δ, the absolute value of the steering wheel rotation speed be δ', and the vehicle speed be x m / s. The calculation method for the hazard index D is shown in the table below:

[0085]

[0086]

[0087] Right now:

[0088]

[0089] Where k1, k2, k3, and k4 are constant parameters used to balance the weights among the features and prevent D from becoming too large. In this example, their values ​​are: k1 = 0.01, k2 = 0.02, k3 = 0.005, and k4 = 0.1.

[0090] The method was applied to calculate the steering hazard index of all dangerous and safe steering behaviors in the dangerous steering behavior database and the safe steering behavior database established in step 1. The results are shown in the table below:

[0091] Risk index calculation results

[0092]

[0093] Based on the calculation results, it can be seen that the method can effectively delineate the boundary between dangerous and safe steering behaviors of drivers. The accuracy rate for judging accident cases reached 84.44%, the accuracy rate for judging safe cases reached 88.43%, and the overall accuracy rate obtained by dividing the number of correctly judged cases by the total number of cases reached 87.10%.

[0094] This step processes the data in the steering behavior database established in step 1 to distinguish between normal steering data and dangerous steering data, which will then be used for training the neural network.

[0095] 302: A Neural Network-Based Method for Recognizing Dangerous Turning Behavior

[0096] A neural network model for identifying dangerous steering behaviors is established by training the steering behavior data in the steering behavior database from step 1 and the corresponding hazard index calculated in step 301. Therefore, when steering behavior data is input into the trained neural network model, it can output the corresponding steering hazard index to identify dangerous steering behaviors.

[0097] Neural networks are information processing methods that mimic the neural networks of the human brain. They consist of a large number of interconnected nodes and can perform functions such as pattern recognition, automatic control, and predictive control. They exhibit good performance in classifying and predicting time-series data. A classifier with good performance can be obtained by training a neural network model using existing data. Therefore, based on the database established in Step 1, a neural network model for recognizing dangerous turning behaviors is built. Then, the neural network is trained using backpropagation, ultimately yielding an effective method for recognizing turning behaviors.

[0098] A multi-layer neural network consists of an input layer, hidden layers, and an output layer, where there can be multiple hidden layers. If the activation value of the l-th layer is set to a... (l) Then the activation value a of the (l+1)th layer (l+1) Equations (5), (6), and (7) can be used for calculation via forward propagation:

[0099] f(x) = max(0,x) (5)

[0100] z (l+1) =W (l) a (l) +b (l) (6)

[0101] a (l+1) =f(z) (l+1) (7)

[0102] In equations (5) to (7), f(x) is the activation function, which can be selected from functions such as ReLU, tansig, and logsig. The activation value a of the output layer is the final calculation result, which can be denoted as h. W,b (x).

[0103] Neural network models can be solved using the gradient descent method. For a single sample (x, y), the cost function is:

[0104]

[0105] For a dataset with m samples, the overall loss function is:

[0106]

[0107] In the formula, λ is the weight decay coefficient, which can reduce the magnitude of the weights and prevent overfitting.

[0108] In the gradient descent method, W and b can be iteratively updated using equations (10) and (11):

[0109]

[0110]

[0111] In equations (10) to (11), α is the learning rate.

[0112] The partial derivatives in equations (10) and (11) can be calculated using the backpropagation algorithm, which mainly consists of three steps.

[0113] Step 1: Calculate L2, L3... up to the output layer L using the forward propagation formula. nl The activation value.

[0114] Step 2: Calculate the residuals between each node of the output layer and the hidden layer. For the output layer n... l For each output unit i, the residual is calculated according to the following formula:

[0115]

[0116] For the hidden layer, the residual of the i-th node in the l-th layer is:

[0117]

[0118] Step 3: Calculate the required partial derivatives as follows:

[0119]

[0120]

[0121] After training a neural network using gradient descent, a classification neural network can be obtained. Based on this method, a neural network model can be built and trained in Matlab to obtain a suitable classification model. Since neural networks provide better support for multiple inputs and temporal features, more features can be selected and their temporal characteristics can be utilized. Here, the concept of a "window" can be introduced: a sliding window is constructed over a continuous set of temporal features, and the neural network can use the data within the window to analyze and calculate the current temporal state, ultimately obtaining the state of the monitored target.

[0122] Depending on the required accuracy, precision, and real-time performance, an appropriate window size can be freely selected. The trained neural network model should be able to analyze the data in the window to determine whether a dangerous steering behavior has occurred. In this example, a window length of 5 is selected, and vehicle speed, vehicle acceleration, steering wheel angle, and steering wheel speed are chosen as features to construct a 4×5 window. To facilitate model training, the 4×5 matrix is ​​transformed into a 1×20 vector by concatenating the first and last rows of the matrix into a single row, as shown below:

[0123]

[0124] The database constructed in Step 1 contains 700 safe turning cases and 688 emergency turning cases. Twenty feature points are extracted from each case as a window for that turning case. After processing, 1388 samples are obtained. These 1388 samples are used as the training sample set, from which 15% of the cases are randomly divided into a validation set, 15% into a test set, and 70% into a training set. A neural network model is built in Matlab for training, with a learning rate of 0.01, a maximum number of iteration layers of 1000, a tansig activation function for the hidden layer, and a logsig activation function for the output layer. Experiments show that selecting one hidden layer with 15 neurons yields the best discrimination performance.

[0125] After setup, you can use Matlab to train the neural network model using the backpropagation algorithm, which minimizes the loss function through iterative calculations. Iterative calculation process. Figure 8 As shown:

[0126] From the training process diagram Figure 8As can be seen, with the increase of the number of iterations, the loss function values ​​of the training set, validation set, and test set all decrease continuously until, after 105 iterations, the loss function values ​​of the training set, validation set, and test set remain basically unchanged, and training stops. In the first 105 iterations, the loss function reaches its minimum at the 99th iteration; therefore, the result of the 99th iteration is used as the final neural network model. The trained neural network model is then used to classify the training set, validation set, and test set, and the result can be plotted as a confusion matrix. The confusion matrix, also known as the error matrix, can be used to evaluate the accuracy of the classifier. Therefore, the quality of the obtained neural network can be intuitively judged based on the plotted confusion matrix.

[0127] In the confusion matrix, each row represents the predicted output of the neural network, and each column represents the actual value of the sample. F indicates a prediction error, and T indicates a correct prediction. Each output value has a corresponding accuracy rate. The prediction accuracy rates under different conditions are shown in the table below:

[0128] Prediction accuracy under different conditions

[0129]

[0130] For ESC, which requires avoiding over-control, a higher accuracy rate in identifying dangerous steering behaviors means that the algorithm's conclusion that the driver is engaging in dangerous steering is more reliable. A steering hazard index can be calculated using a neural network, normalized to the range [0,1], and then a threshold can be set to determine the characteristics of the current driving behavior. In this example, the threshold is set to 0.5. After obtaining new data, the trained network can directly determine whether the data represents an emergency steering or a normal steering.

[0131] The performance of the established neural network was tested using a test set. The results show that the neural network model can basically detect the occurrence of abnormal driving behavior. Therefore, it can be used as a driving behavior recognition module for adaptive vehicle stability control. It can also be combined with the vehicle state recognition module to form a stability control algorithm.

[0132] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A method for identifying dangerous turning behaviors based on boundary search, characterized in that: Step 1: Establish a steering behavior database, including: a hazardous steering behavior database and a safe steering behavior database; Step 2: Selection of steering behavior characteristic quantities: Steering wheel angle, steering wheel speed, and vehicle speed are selected from the steering behavior database as steering behavior feature quantities; Step 3: Based on the steering behavior feature quantity selected in Step 2, perform statistical analysis on the data in the steering behavior database in Step 1 using the boundary search method to obtain the boundary between dangerous steering behavior and safe steering behavior, and calculate the danger index of each steering behavior in the steering behavior database accordingly. Step 4: Identification of Dangerous Turning Behaviors Based on Neural Networks The neural network model is trained based on the steering behavior data in the steering behavior database in step 1 and the corresponding hazard index calculated in step 3 to establish a neural network model for identifying dangerous steering behavior. The neural network model can identify dangerous steering behavior by outputting a hazard index. In step 3, when performing statistical analysis on the data in the steering behavior database from step 1 based on the boundary search method: 301: Region division based on steering wheel angle On a two-dimensional plane diagram reflecting the relationship between steering wheel angle and vehicle speed, the boundary search method is used to calculate the boundary curves L1(x) and L2(x) according to p1 = 0.9 and p2 = 0.99, respectively. 302: Regional division based on steering wheel rotation speed On a two-dimensional planar graph that reflects the relationship between steering wheel rotation speed and vehicle speed, the boundary search method is used to calculate the boundary curve L3(x) according to the set p3 = 0.99; The method for calculating the steering behavior hazard index D is as follows: Where: k1, k2, k3, k4 are constant parameters, δ is the absolute value of the steering wheel angle, δ' is the absolute value of the steering wheel speed, and x is the vehicle speed, in m / s.

2. The dangerous turning behavior recognition method based on boundary search as described in claim 1, characterized in that: In step 301, within the safe area corresponding to the boundary curve L1(x), the area between 17m / s and 22m / s is designated as the risk area.

3. The dangerous turning behavior recognition method based on boundary search as described in any one of claims 1-2, characterized in that: In step 3: When performing statistical analysis on the data in the steering behavior database based on the boundary search method, the maximum steering wheel angle, maximum steering wheel speed, and maximum vehicle speed within the corresponding time period for each steering case in the steering behavior database are selected as the three characteristic values ​​of that steering behavior case. Based on these characteristic values, a graph showing the relationship between steering wheel angle and vehicle speed, as well as a graph showing the relationship between steering wheel speed and vehicle speed, are drawn.

4. The dangerous turning behavior recognition method based on boundary search as described in any one of claims 1-2, characterized in that: In step 1: a dangerous steering behavior database is constructed using dangerous steering behavior data extracted from the published EDR of accident vehicles; A safe steering behavior database will be constructed using steering behavior data from at least two published driving databases.

5. The dangerous turning behavior recognition method based on boundary search as described in any one of claims 1-2, characterized in that: In step 4: when training the neural network model, each steering behavior data in the steering behavior database in step 1 is used as the training sample set, and 15% of the cases are randomly divided into a validation set, 15% of the cases are used as a test set, and 70% of the cases are used as the training set.

6. The dangerous turning behavior recognition method based on boundary search as described in claim 5, characterized in that: In step 4: a neural network model is built and trained in Matlab, with the learning rate set to 0.01, the maximum number of iteration layers set to 1000, the activation function of the hidden layer set to tansig, and the activation function of the output layer set to logsig. Choose one hidden layer with 15 neurons.

Images