Vehicle driver dangerous driving behavior early warning method based on deep learning

Abstract

The invention provides a vehicle driver dangerous driving behavior early warning method based on deep learning. The vehicle driver dangerous driving behavior early warning method comprises the steps that vehicle lamp brightness and angle data collected by a vehicle-mounted sensor in the vehicle driving process are acquired; preprocessing the vehicle lamp brightness and angle data to obtain standardized vehicle lamp brightness and angle data; according to the standardized vehicle lamp brightness and angle data, the current driver visibility is output through a deep learning model; obtaining a dangerous driving behavior prediction result through a dangerous driving behavior prediction model in combination with the visibility of the driver, the vehicle driving state data and the road environment condition data; according to the dangerous driving behavior prediction result and the vehicle lamp brightness and angle data, whether driving is dangerous driving is judged; if dangerous driving is judged, matching an early warning strategy from a preset early warning strategy library, generating an early warning measure, simulating execution and evaluating an early warning effect, and if necessary, adjusting the early warning strategy until a safe driving requirement is met.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method for warning dangerous driving behavior of a vehicle driver based on deep learning. Background Art

[0002] In the deep learning-based warning of dangerous driving behavior of vehicle drivers, the brightness and angle of the headlights are important factors affecting the driver's visibility. Different headlight brightness and angles will lead to differences in driver visibility, which in turn affects the driver's perception and judgment of the road environment. However, the correlation analysis between the brightness and angle of the headlights and the driver's visibility faces the challenges of data collection and quantification. In actual driving scenarios, the brightness and angle of the headlights change dynamically and are affected by multiple factors such as vehicle model, headlight type, and weather conditions. How to accurately collect the headlight brightness and angle data in a complex and changing driving environment and correlate it with the driver's visibility is a technical problem that needs to be solved urgently. In addition, the evaluation of driver visibility also faces the challenges of subjectivity and individual differences. Different drivers have different light sensitivity and visual adaptability, resulting in different actual visibility of the driver under the same headlight brightness and angle. How to design an objective and quantitative driver visibility evaluation method and apply it to the construction of a dangerous driving behavior prediction model is a technical problem that needs to be overcome urgently. Summary of the invention

[0003] The present invention provides a method for warning dangerous driving behavior of vehicle drivers based on deep learning, which mainly includes:

[0004] Obtain the headlight brightness and angle data collected by the vehicle-mounted sensor during vehicle driving;

[0005] Preprocessing the headlight brightness and angle data to obtain standardized headlight brightness and angle data;

[0006] Outputting current driver visibility through a deep learning model according to the standardized headlight brightness and angle data;

[0007] Combining the driver visibility, the vehicle driving status data and the road environment condition data, obtaining a dangerous driving behavior prediction result through a dangerous driving behavior prediction model;

[0008] Determine whether the driving is dangerous based on the predicted dangerous driving behavior and the headlight brightness and angle data;

[0009] If it is judged as dangerous driving, the warning strategy is matched from the preset warning strategy library, the warning measures are generated and simulated, the warning effect is evaluated, and the warning strategy is adjusted if necessary until it meets the safe driving requirements.

[0010] Furthermore, the preprocessing of the headlight brightness and angle data includes:

[0011] According to the headlight brightness and angle data, a preprocessing instruction is sent to the vehicle-mounted computing unit, and the vehicle-mounted computing unit uses a wavelet transform denoising algorithm to preprocess the received headlight brightness and angle data; the preprocessed headlight brightness and angle data are obtained, and the preprocessed headlight brightness and angle data are standardized.

[0012] Furthermore, outputting the current driver visibility through the deep learning model includes:

[0013] The standardized headlight brightness and angle data are input into a convolutional neural network, which includes three convolutional layers and two pooling layers; the current driver visibility value is output through the fully connected layer of the convolutional neural network; the sliding window method is used to perform time series analysis on the driver visibility value to calculate the mean, variance and change trend of the visibility.

[0014] Furthermore, the combining the driver visibility, the vehicle driving status data and the road environment condition data to obtain the dangerous driving behavior prediction result through the dangerous driving behavior prediction model includes:

[0015] Acquire vehicle driving status data and road environment condition data collected by on-board sensors, the vehicle driving status data including vehicle speed, acceleration and steering angle, and the road environment condition data including road type information and weather condition information; perform preprocessing on the vehicle driving status data and road environment condition data, and generate a standardized data set through denoising, normalization and time synchronization; use a sixty-second sliding window, slide once every ten seconds to extract time series features of the standardized data set, calculate statistical features, and construct feature vectors as inputs of the deep learning model; use a pre-trained long short-term memory network model to process the feature vectors, and calculate the predicted probability distribution of dangerous driving behavior through forward propagation.

[0016] Furthermore, judging whether it is dangerous driving according to the dangerous driving behavior prediction result and the headlight brightness and angle data includes:

[0017] The standardized headlight brightness and angle data are subjected to time series analysis using a sliding window, and the mean, variance and rate of change within the window are calculated; the mean, variance and rate of change are compared with the preset headlight brightness and angle danger thresholds to determine whether there is an over-threshold situation; the dangerous driving behavior assessment results are processed using a weighted average method to obtain a comprehensive assessment score for the dangerous driving behavior; a binary decision tree model is constructed using the CART algorithm, and the over-threshold situation of the headlight brightness and angle data and the comprehensive assessment score of the dangerous driving behavior are used as input features; the final dangerous driving probability score is calculated based on the output result of the binary decision tree model and the preset danger judgment weight; it is determined whether the dangerous driving probability score exceeds the preset threshold, and if so, it is determined to be dangerous driving.

[0018] Furthermore, matching the warning strategy from the preset warning strategy library, generating warning measures and simulating their execution, evaluating the warning effect, and adjusting the warning strategy when necessary until it meets the safe driving requirements, include:

[0019] The prediction results, evaluation results and standardized headlight data of dangerous driving behaviors are obtained to construct a multi-dimensional feature vector; according to the multi-dimensional feature vector, a decision tree-based strategy matching algorithm is used to find the most suitable warning strategy in a preset warning strategy library to obtain an initial optimal warning strategy; for the initial optimal warning strategy, specific warning measures are generated, and the warning measures include a percentage of headlight brightness adjustment and an angle adjustment degree; a simulation execution algorithm based on a vehicle dynamics model is used to calculate the new brightness and new angle of the headlights according to the warning measures; the new brightness and new angle data of the headlights are input into a driver visibility evaluation model to obtain the driver visibility evaluation result after simulation execution; it is determined whether the driver visibility evaluation result meets the preset safety range requirements. If it does not meet the preset safety range requirements, a genetic algorithm is used to optimize the warning strategy, and a new warning strategy and warning measures are generated through iterative calculation, and the simulation and evaluation are repeated until the preset safety range requirements are met or the maximum number of iterations is reached.

[0020] Furthermore, the method further comprises:

[0021] The driver operation data, vehicle driving status data and headlight parameter data collected by the vehicle-mounted sensor network are obtained; a feature vector is generated according to the collected data, and the feature vector includes the driver reaction time, the vehicle speed change rate and the steering wheel angle change rate; the feature vector is processed by a multivariate linear regression model to obtain a driver response degree score and a vehicle status improvement degree score; according to the comprehensive effect score, the driver visibility data and the current driving environment parameters, a fuzzy control algorithm is used to adjust the warning trigger threshold, the warning information presentation method and the headlight control parameters; the adjusted warning trigger threshold, the warning information presentation method and the headlight control parameters are applied to the next warning process.

[0022] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0023] The present invention discloses a method for warning dangerous driving behaviors of vehicle drivers based on deep learning. The method collects the brightness and angle data of the headlights in real time, combines the deep learning model to evaluate the driver's visibility, and analyzes its relationship with the safe driving of the vehicle. At the same time, the present invention integrates the vehicle driving status and road environmental conditions to predict dangerous driving behaviors. When the headlight data exceeds the threshold and there is dangerous driving behavior, the present invention matches the optimal warning strategy from the preset strategy library, generates warning measures and performs simulation evaluation. If the evaluation result does not meet the safety range, the warning measures are regenerated. Finally, the present invention issues a warning to the driver through the human-computer interaction system and automatically adjusts the headlights. The present invention also continuously tracks the driver's reaction and vehicle status, and dynamically adjusts the warning strategy, thereby effectively improving the safety of night driving and reducing the risk of traffic accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a flow chart of a method for warning dangerous driving behavior of a vehicle driver based on deep learning of the present invention.

[0025] Figure 2 This is a schematic diagram of a method for warning dangerous driving behavior of a vehicle driver based on deep learning according to the present invention.

[0026] Figure 3 This is another schematic diagram of a method for warning dangerous driving behavior of a vehicle driver based on deep learning of the present invention. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] like Figure 1-3 In this embodiment, a method for warning dangerous driving behavior of a vehicle driver based on deep learning may specifically include:

[0029] S101, collecting headlight brightness and angle data of the vehicle in real time during driving through the vehicle-mounted sensor, transmitting the collected data to the vehicle-mounted computing unit for preprocessing, denoising and normalizing the data, and obtaining standardized headlight brightness and angle data.

[0030] Receive the headlight brightness and angle data collected by the vehicle-mounted sensor during the vehicle's driving process, the headlight brightness and angle data are obtained by the vehicle-mounted sensor by dynamically adjusting the sampling frequency according to the vehicle speed change range; send a preprocessing instruction to the vehicle-mounted computing unit according to the headlight brightness and angle data, the vehicle-mounted computing unit uses a wavelet transform denoising algorithm to preprocess the received headlight brightness and angle data; obtain the preprocessed headlight brightness and angle data, and standardize the preprocessed headlight brightness and angle data; use a sliding window method to perform time series analysis on the standardized headlight brightness and angle data; if the time series analysis result meets the preset conditions, construct a time series prediction model, and the time series prediction model uses an autoregressive integral sliding average model; generate a headlight control instruction sequence according to the prediction result of the time series prediction model, send the headlight control instruction sequence to the vehicle-mounted control system, and the vehicle-mounted control system adjusts the headlight brightness and angle according to the headlight control instruction sequence.

[0031] Specifically, the vehicle-mounted sensors are used to collect the headlight brightness and angle data in real time during the driving process of the vehicle. The sampling frequency of the sensors is dynamically adjusted according to the speed range. The collected raw data is transmitted to the vehicle-mounted computing unit using high-bandwidth wireless communication technology. The received raw data is preprocessed in the vehicle-mounted computing unit, and the wavelet transform denoising algorithm is used to eliminate the noise introduced by the ambient light intensity and vehicle vibration. The denoised data is normalized, where the headlight brightness data uses the maximum and minimum value normalization method, and the headlight angle data uses a linear mapping based on a predefined angle range. According to the mean and standard deviation obtained from a large amount of historical data in advance, a standardized model is constructed to convert the normalized headlight brightness and angle data into standardized data. Time series analysis is performed on the standardized headlight brightness and angle data, and the sliding window method is used to segment the data. Statistical features are calculated in each sliding window, including mean, variance, kurtosis, etc. The most relevant features for headlight control are screened out using the correlation analysis method, and the screened feature data are stored in the vehicle-mounted database. The stored feature data is read from the vehicle-mounted database to construct a time series prediction model. According to the characteristics of the headlight brightness and angle data, the autoregressive integral moving average model is selected for prediction. The model parameters are obtained by training the historical data, and the trained model is used to predict the trend of headlight brightness and angle changes in the future. Based on the prediction results, the headlight control instruction sequence is generated and sent to the vehicle control system. After receiving the instruction sequence, the vehicle control system adjusts the brightness and angle of the headlight according to the instruction to achieve real-time dynamic control of the headlight.

[0032] The on-board sensor uses a high-precision photoelectric sensor to collect headlight brightness and angle data at a frequency of 100Hz. When the vehicle speed is lower than 30km / h, the sampling frequency is reduced to 50Hz, and when it is higher than 80km / h, it is increased to 200Hz to achieve dynamic adjustment. The collected raw data is transmitted to the on-board computing unit via the 5G network, and the transmission bandwidth can reach 1Gbps. After receiving the data, the on-board computing unit uses the db4 wavelet transform to perform a 4-layer decomposition to remove high-frequency noise. The denoised brightness data is normalized using the (x-min) / (max-min) formula to convert 0-10000cd / m 2The range is mapped to 0-1. The angle data linearly maps the range from -45° to +45° to 0-1. Based on the mean μ=0.5 and standard deviation σ=0.2 obtained from historical data statistics, a standardized model (x-μ) / σ is constructed to convert normalized data into standardized data. A sliding window with a width of 60 seconds is used for the standardized data to calculate 10 statistical features such as mean, variance, kurtosis, and skewness in each window. Through Pearson correlation coefficient analysis, features with an absolute value of correlation coefficient greater than 0.7 are selected as key features and stored in the vehicle database. The key feature data of the past 24 hours are read from the database to construct an ARIMA(2,1,2) model. The model parameters are estimated using the maximum likelihood estimation method to predict the trend of headlight brightness and angle changes in the next 10 minutes. The prediction results are converted into headlight control instructions, such as "increase the brightness of the high beam by 20% and adjust it to the left by 5° after 5 seconds." After receiving the command, the on-board control system adjusts the LED drive circuit through the PWM signal to achieve precise control of the brightness and angle of the car lights, thereby completing the automatic adjustment of the entire process.

[0033] S102. Construct a deep learning model for driver visibility, use the standardized headlight brightness and angle data as input to the deep learning model, output the current driver visibility, and establish a driver visibility evaluation system based on the driver visibility and headlight brightness and angle data, and analyze and obtain the driver visibility evaluation results.

[0034] According to the headlight brightness and angle data, the input features are normalized to obtain normalized feature data; the normalized feature data are input into a convolutional neural network, the convolutional neural network includes three convolutional layers and two pooling layers; the current driver visibility value is output through the fully connected layer of the convolutional neural network; the driver visibility value is analyzed in time series by a sliding window method to calculate the mean, variance and change trend of visibility; a multidimensional feature vector is constructed in combination with the headlight brightness and angle data; according to the multidimensional feature vector, the visibility level is divided by a preset fixed threshold to obtain the driver visibility level; based on the driver visibility level, a comprehensive evaluation score is calculated in combination with evaluation indicators, the evaluation indicators include visibility stability, headlight adjustment frequency and environmental adaptability; if the comprehensive evaluation score is lower than the preset threshold, the headlight brightness and angle parameters are automatically adjusted; the adjusted headlight brightness and angle data are received and re-input into the convolutional neural network for processing.

[0035] Specifically, according to the standardized headlight brightness and angle data, the input features are normalized through the data preprocessing module, and the processed data are input into the convolutional neural network. The network is trained by using a large amount of labeled headlight brightness, angle and corresponding visibility data, and the optimal model parameters are selected by the five-fold cross-validation method. The network structure contains three convolutional layers and two pooling layers, and finally the current driver visibility value is output through the fully connected layer. The sliding window method is used to perform time series analysis on the output driver visibility values, and the mean, variance and change trend of visibility are calculated. Combined with the headlight brightness and angle data, a multi-dimensional feature vector is constructed, and the visibility level is divided by a preset fixed threshold: less than 50 meters is low visibility, 50 to 200 meters is medium visibility, and more than 200 meters is high visibility. Based on the divided visibility levels, the comprehensive evaluation score is calculated in combination with the evaluation indicators. The evaluation indicators include visibility stability, headlight adjustment frequency and environmental adaptability, among which visibility stability is expressed by variance, headlight adjustment frequency is expressed by the number of adjustments per unit time, and environmental adaptability is expressed by the visibility change rate. The weights of each indicator are 0.4, 0.3 and 0.3 respectively. The final driver visibility evaluation result is obtained by weighted summation, with a full score of 100 points. A feedback control mechanism is established for the evaluation results. If the evaluation result is lower than 60 points, the headlight brightness and angle parameters are automatically adjusted: increase the headlight brightness by 10% and adjust the angle downward by 2 degrees. Subsequently, the data is re-collected and input into the convolutional neural network to form a closed-loop optimization process, which continuously improves the accuracy and real-time performance of the driver visibility evaluation. The vehicle control system receives the standardized headlight brightness and angle data, with values ​​ranging from 0 to 1. The data preprocessing module uses the Z-score standardization method to convert the input features into a distribution with a mean of 0 and a standard deviation of 1. The preprocessed data is input into the convolutional neural network, which is trained with 50,000 sets of labeled data, each of which contains 10-second time series data of headlight brightness and angle, as well as the corresponding real visibility value. The network structure contains three convolutional layers (with kernel sizes of 3×3, 5×5, and 7×7) and two maximum pooling layers (with pooling window size of 2×2), and finally outputs the current driver visibility value through two fully connected layers (with 128 and 64 nodes, respectively). A sliding window with a width of 60 seconds is used, sliding every 10 seconds, and the mean, variance, and linear regression slope of the visibility in the window are calculated. Combined with the latest headlight brightness and angle data, a 7-dimensional feature vector is constructed. Visibility levels are divided by preset thresholds: less than 50 meters is low visibility, 50 to 200 meters is medium visibility, and more than 200 meters is high visibility. The calculation method of the evaluation indicators is as follows: visibility stability S = 1 / (1+variance), headlight adjustment frequency F = number of adjustments within 10 minutes / 60, environmental adaptability A = |current visibility - visibility 10 minutes ago| / visibility 10 minutes ago.The weights of each indicator are 0.4, 0.3 and 0.3 respectively. The final evaluation score is obtained by weighted summation: Score = 40S + 30 (1-F) + 30 (1-A). If the evaluation score is less than 60 points, the feedback control mechanism is triggered: the brightness of the headlights is increased by 10%, and the angle of the headlights is adjusted downward by 2 degrees. The on-board control system collects and evaluates data every 1 second. Through this closed-loop optimization process, real-time evaluation and dynamic adjustment of the driver's visibility are achieved.

[0036] The input and output results of the driver visibility deep learning model are correlated and analyzed to obtain the correlation between driver visibility and headlight brightness and angle. The safety of different driving changes is evaluated by monitoring the different driving changes of the vehicle under different visibilities. The correlation between driver visibility and headlight brightness and angle is combined with the driving safety of the vehicle under different visibilities to obtain the impact of driver visibility and headlight brightness and angle data on vehicle safe driving.

[0037] The input data and output results of the driver visibility deep learning model are obtained, and the correlation coefficient matrix between the brightness and angle of the headlights and the driver visibility is calculated using the Pearson correlation coefficient. The correlation strength between the driver visibility and the brightness and angle of the headlights is determined according to the correlation coefficient matrix. The vehicle-mounted sensor collects data once a second to obtain the driving change parameters of the vehicle under different visibility, including speed change, steering angle, braking distance and driving trajectory. The driving change parameters are quantitatively evaluated using the fuzzy comprehensive evaluation method, and an evaluation index system is established, including vehicle speed stability, direction control accuracy, braking response time and trajectory smoothness. The trapezoidal membership function is used to describe the fuzzy set of each index, and the index weight is determined by the hierarchical analysis method. If the correlation analysis results and safety evaluation scores have been obtained, a support vector regression model is constructed, with the driver visibility, headlight brightness and angle as input features and the safety score as the output target. The kernel function parameters and regularization coefficient of the support vector regression model are optimized by the grid search method, and the model performance is evaluated by five-fold cross validation. If the support vector regression model training is completed, the influence weights of the driver's visibility and the headlight brightness and angle data on the safe driving of the vehicle are obtained.

[0038] Specifically, the Pearson correlation coefficient is used to analyze the correlation between the input and output results of the driver visibility deep learning model, and the correlation coefficient matrix between the headlight brightness, angle and driver visibility is calculated. The correlation strength is divided into strong, medium and weak according to the absolute value of the correlation coefficient, and the correlation strength between the driver visibility and the headlight brightness and angle is determined. The vehicle-mounted sensor collects data once a second to monitor the driving changes of the vehicle under different visibility, including speed changes, steering angles, braking distances and driving trajectories. The statistical characteristics of these parameters, such as mean, standard deviation and rate of change, are calculated using a 10-second sliding window. According to the statistical characteristics, the driving changes are preliminarily classified to identify typical driving patterns, such as reduced speed and increased steering frequency under low visibility. The fuzzy comprehensive evaluation method is used to quantitatively evaluate the safety of different driving changes, and an evaluation index system is established, including vehicle speed stability, direction control accuracy, braking response time and trajectory smoothness. The trapezoidal membership function is used to describe the fuzzy set of each indicator, and the indicator weight is determined by the hierarchical analysis method. The membership of each indicator is obtained through fuzzy matrix operation, and the comprehensive safety score is calculated. Based on the correlation analysis results and safety assessment scores, a support vector regression model was constructed, with driver visibility, headlight brightness and angle as input features and safety score as output target. The kernel function parameters and regularization coefficients were optimized using the grid search method, and the model performance was evaluated using a five-fold cross validation. Through the trained model, the influence weights of driver visibility and headlight brightness and angle data on vehicle safe driving were obtained. First, a Pearson correlation analysis was performed on the input (headlight brightness and angle) and output (visibility) of the driver visibility deep learning model. The correlation coefficient between headlight brightness and visibility was calculated to be 0.82, and the correlation coefficient between headlight angle and visibility was -0.65, indicating that headlight brightness and visibility are strongly positively correlated, and headlight angle and visibility are moderately negatively correlated. The on-board sensor collects data at a frequency of 1Hz, and the statistical features are calculated within a sliding window of 10 seconds. For example, when the visibility is 50 meters, the mean speed drops to 40km / h, the standard deviation increases to 5km / h; the steering angle change rate increases to 2 degrees per second. These features were input into a pre-trained decision tree to classify driving changes into "low-speed caution" mode. Subsequently, the fuzzy comprehensive evaluation method used a trapezoidal membership function to evaluate safety. The vehicle speed stability index has a membership of 1 between 35-45km / h and a membership of 0 when it is below 30km / h or above 50km / h. The weights were determined by the hierarchical analysis method, with a weight of 0.4 for vehicle speed stability, 0.3 for directional control accuracy, 0.2 for braking response time, and 0.1 for trajectory smoothness. The fuzzy matrix operation resulted in a comprehensive safety score of 78 points out of 100. Finally, the support vector regression model used a radial basis function kernel and determined the optimal parameters C=10 and γ=0.1 through grid search. The average root mean square error of the five-fold cross validation was 0.15, indicating that the model has good generalization ability.The trained model shows that under current conditions, the safety score increases by 2.5 points for every 10-meter increase in visibility, 1.8 points for every 10% increase in headlight brightness, and 0.5 points for every 1-degree downward tilt of the headlight. These values ​​quantify the impact of driver visibility, headlight brightness, and angle on vehicle safety.

[0039] S103. By collecting vehicle driving status data and road environment condition data, including vehicle speed, acceleration, steering angle, road type, and weather conditions, the vehicle driving status and road environment condition data and the driver visibility assessment result are used as inputs of a preset dangerous driving behavior prediction model based on deep learning to obtain dangerous driving behavior prediction results, a dangerous driving behavior assessment system is constructed, the prediction results are evaluated, and dangerous driving behavior assessment results are obtained.

[0040] The vehicle driving state data and road environment condition data collected by the vehicle-mounted sensor are obtained, wherein the vehicle driving state data includes vehicle speed, acceleration and steering angle, and the road environment condition data includes road type information and weather condition information; pre-processing is performed according to the vehicle driving state data and road environment condition data, and a standardized data set is generated through denoising, normalization and time synchronization; a sixty-second sliding window is used, and the standardized data set is extracted from the time series features by sliding once every ten seconds, and statistical features are calculated to construct a feature vector as the input of the deep learning model; the feature vector is processed by using a pre-trained long short-term memory network model, and the long short-term memory network model includes three layers of long short-term memory units and two layers of fully connected layers, and the predicted probability distribution of dangerous driving behavior is obtained by forward propagation calculation; if the predicted probability distribution meets the preset conditions, the membership of each indicator is calculated by a fuzzy comprehensive evaluation method according to the predicted probability distribution and the preset dangerous driving behavior evaluation index system, and the final dangerous driving behavior evaluation score is obtained by weighted summation.

[0041] Specifically, vehicle driving status data, including vehicle speed, acceleration and steering angle, are collected in real time through on-board sensors. At the same time, road type information is obtained from on-board cameras, and weather condition information is updated every 5 minutes through a real-time data interface with the National Meteorological Administration. The collected data is preprocessed, including denoising, normalization and time synchronization, to generate a standardized vehicle driving status and road environment condition data set. The standardized vehicle driving status and road environment condition data are fused with the driver visibility assessment results. A 60-second sliding window is used to extract time series features from the fused data every 10 seconds. Statistical features such as mean, variance and kurtosis are calculated, and feature vectors are constructed as inputs to the deep learning model. The input feature vectors are processed using a pre-trained long short-term memory network model, which contains three layers of long short-term memory units, each with 128 neurons, followed by two fully connected layers with 64 and 32 neurons, respectively. The predicted probability distribution of dangerous driving behavior is obtained through forward propagation calculation, and the category with the highest probability is selected as the prediction result, and the prediction results are preliminarily classified into three levels: low risk, medium risk and high risk. According to the prediction results and risk level, combined with the preset dangerous driving behavior evaluation index system, including speeding degree, number of sudden accelerations, number of sharp turns and dangerous weather adaptability, the fuzzy comprehensive evaluation method is used to calculate the membership of each index. The trapezoidal membership function is used to describe the fuzzy set of each index, the index weight is determined by the hierarchical analysis method, and finally the final dangerous driving behavior evaluation score is obtained by weighted summation. The vehicle-mounted system collects vehicle speed, acceleration and steering angle data at a frequency of 10Hz through high-precision sensors. At the same time, the vehicle-mounted camera identifies the road type once a second and divides it into three categories: highway, urban road and rural road. The meteorological data interface is updated every 5 minutes, providing weather conditions including sunny, cloudy, rainy, snowy and foggy. The collected raw data is denoised by Kalman filtering, and then normalized to the range of 0-1 using the maximum and minimum values. Data synchronization uses timestamp alignment to ensure the time consistency of all data sources. The standardized data is fused with the driver visibility evaluation results, and a 60-second sliding window is used, sliding every 10 seconds for feature extraction. The calculated features include 20 statistics such as vehicle speed mean and variance, acceleration kurtosis, steering angle change rate, etc., which constitute the input feature vector. The long short-term memory network model receives these feature vectors and processes them through 3 layers of LSTM units (128 neurons in each layer) and 2 layers of fully connected layers (64 and 32 neurons). The model outputs the probabilities of 5 types of dangerous driving behaviors: normal driving, fatigue driving, distracted driving, aggressive driving and drunk driving. If the highest probability exceeds 0.7, the category is directly adopted; otherwise, the results are preliminarily classified as low risk, medium risk or high risk according to the probability distribution. In the evaluation index system, the degree of speeding is expressed by the ratio of the actual speed to the speed limit, and rapid acceleration is defined as an acceleration exceeding 3m / s 2, and a sharp turn means the steering angle speed exceeds 20° / s. A trapezoidal membership function is used to describe these indicators, such as speeding less than 1.1 times is safe, 1.1-1.3 times is relatively dangerous, and more than 1.3 times is very dangerous. The weights are determined by the hierarchical analysis method, and finally an evaluation score of 0-100 is obtained, of which 0-60 points are safe driving, 61-80 points are slightly dangerous, and 81-100 points are seriously dangerous.

[0042] The vehicle speed, acceleration and steering angle of the vehicle status data are standardized, and the time series characteristics of the standardized data are extracted to obtain the changing trends and instantaneous values ​​of the vehicle speed, acceleration and steering angle. The road type and weather conditions are used as additional features, combined with the changing trends and instantaneous values ​​of the vehicle status data, as evaluation indicators of the dangerous driving behavior assessment system.

[0043] The vehicle speed, acceleration and steering angle data are obtained, and the maximum and minimum value normalization method is used to map the data to the range of 0-1 to obtain the standardized vehicle status data. According to the standardized vehicle status data, the moving average and difference method are used to calculate the 5-second short-term and 30-second long-term moving average and the first-order difference value to determine the time series characteristics of the vehicle status data. According to the time series characteristics, the trend prediction is performed by the triple exponential smoothing method to determine the change trend of the vehicle speed, acceleration and steering angle. The 3σ principle is used to perform threshold detection on the vehicle status data, and the data points that exceed the preset times of the mean standard deviation are identified as mutation points to obtain the instantaneous value of the vehicle status data. If the vehicle is traveling on a highway, urban road or rural road, the road type information is processed by one-hot encoding and converted into a 3-dimensional numerical feature vector; if the vehicle is traveling in sunny, cloudy, rainy, snowy or foggy weather, the weather condition information is processed by one-hot encoding and converted into a 5-dimensional numerical feature vector.

[0044] Specifically, the maximum and minimum normalization method is used to standardize the vehicle speed, acceleration and steering angle data, and the original data is mapped to the range of 0-1. The mean and standard deviation in the 30-second sliding window are calculated, and the standardized vehicle status data is obtained by sliding once per second. The time series analysis of the standardized vehicle speed, acceleration and steering angle data is performed using the moving average and difference method. The moving average of the short-term 5 seconds and the long-term 30 seconds, as well as the first-order difference value, are calculated to obtain the time series characteristics of the vehicle status data. According to the extracted time series characteristics, the trend prediction is performed by the triple exponential smoothing method. The smoothing parameters α are set to 0.3, β is set to 0.2, and γ is set to 0.1 to predict the change trend of vehicle speed, acceleration and steering angle. At the same time, the 3σ principle is used for threshold detection, and the data points exceeding the mean ±3 times the standard deviation are identified as mutation points to determine the instantaneous value of the vehicle status data. The road type and weather condition information are processed by one-hot encoding and converted into numerical feature vectors. The road type is encoded as a 3D vector, and the weather condition is encoded as a 5D vector. These encoded features are integrated with the changing trends and instantaneous values ​​of the vehicle status data to construct a multi-dimensional evaluation index for the dangerous driving behavior evaluation system. Based on the integrated evaluation index, the dangerous threshold is set: the vehicle speed exceeds 1.2 times the road speed limit, and the absolute value of acceleration exceeds 3m / s 2 , steering angle change rate exceeding 20° / s is considered dangerous driving behavior. These thresholds are adjusted under different weather conditions, such as reducing the speed threshold by 20% in rainy and snowy weather. In the original data collected by the vehicle system, the vehicle speed range is 0-200km / h and the acceleration range is -10 to 10m / s 2 , the steering angle range is -180° to 180°. Maximum and minimum value normalization maps these data to the 0-1 interval, such as 80km / h after normalization is 0.4. Use a 30-second sliding window and slide once per second to calculate the mean and standard deviation. For example, the normalized mean of the speed in a 30-second window is 0.45 and the standard deviation is 0.05. Perform time series analysis on the standardized data, calculate the 5-second short-term moving average and the 30-second long-term moving average, as well as the first-order difference value. Use triple exponential smoothing method for trend prediction, with smoothing parameters α=0.3, β=0.2, γ=0.1, and predict that the speed at the next moment is 0.47. The 3σ principle detects that the speed suddenly jumps from 0.45 to 0.6 and identifies it as an instantaneous mutation point. Road type coding examples: highway [1,0,0], urban road [0,1,0], rural road [0,0,1]. Weather condition code: sunny [1,0,0,0,0], cloudy

[0045] [0,1,0,0,0], Rain [0,0,1,0,0], Snow [0,0,0,1,0], Fog [0,0,0,0,1]. Set the danger threshold: On urban roads [0,1,0], sunny days [1,0,0,0,0], if the predicted speed exceeds 60km / h (normalized value 0.3), the absolute value of acceleration exceeds 3m / s 2 (normalized value 0.65), or the steering angle change rate exceeds 20° / s (normalized value 0.556), it is judged as dangerous driving. Under rainy conditions [0,0,1,0,0], the speed threshold is automatically adjusted to 48km / h (normalized value 0.24). Based on these thresholds, driving behavior is evaluated in real time to provide a basis for subsequent warnings and interventions.

[0046] S104. Analyze and judge the standardized headlight brightness and angle data. If the headlight brightness and angle data exceed the preset headlight brightness and angle danger thresholds, further combine the dangerous driving behavior assessment results. If the dangerous driving behavior assessment results and the headlight brightness and angle data both exceed the corresponding danger thresholds, it is judged as dangerous driving.

[0047] A sliding window is used to perform time series analysis on the standardized headlight brightness and angle data, the sliding window slides once per second, and the mean, variance and change rate within the window are calculated; the mean, variance and change rate are compared with the preset headlight brightness and angle danger thresholds to determine whether there is an over-threshold situation; the weighted average method is used to process the dangerous driving behavior assessment results to obtain a comprehensive assessment score of the dangerous driving behavior; a binary decision tree model is constructed using the CART algorithm, and the over-threshold situation of the headlight brightness and angle data and the comprehensive assessment score of the dangerous driving behavior are used as input features; if the binary decision tree model outputs a dangerous driving judgment result, the headlight data and the dangerous driving behavior assessment result are interacted and complemented; the final dangerous driving probability score is calculated based on the output result of the binary decision tree model and the preset danger judgment weight; it is determined whether the dangerous driving probability score exceeds the preset threshold, and if it exceeds the preset threshold, it is determined to be dangerous driving.

[0048] Specifically, a 10-second sliding window is used, which slides once per second to perform time series analysis on the standardized headlight brightness and angle data, calculate the mean, variance and change rate within the window, and compare with the preset headlight brightness and angle danger thresholds to determine whether there is an over-threshold situation. The weighted average method is used to process the dangerous driving behavior assessment results, and the vehicle speed, acceleration, and steering angle are given 50%, 30%, and 20% weights respectively. The comprehensive score is calculated and compared with the preset dangerous driving behavior threshold to determine whether it exceeds the dangerous threshold. A decision tree model is constructed, and a binary decision tree is constructed using the CART algorithm. The maximum depth is set to 5, and the minimum number of leaf node samples is set to 100. The over-threshold situation of the headlight brightness and angle data and the dangerous driving behavior assessment results are used as input features, and the dangerous driving judgment rules under different feature combinations are learned through training data. In the decision-making process, the headlight data and the driving behavior assessment results interact and complement each other, with the headlight data accounting for 40% of the weight and the driving behavior assessment results accounting for 60% of the weight. According to the output results of the decision tree model, combined with the preset danger judgment weight, the final dangerous driving probability score is calculated. When the score exceeds the preset threshold, it is judged as dangerous driving, triggering the corresponding warning or intervention mechanism. The formula for calculating the probability score of dangerous driving is: score = 0.4 × (the degree of headlight brightness exceeding the threshold + the degree of headlight angle exceeding the threshold) / 2 + 0.6 × driving behavior assessment score. The vehicle-mounted system collects headlight brightness and angle data in real time and records them at a frequency of 10Hz. A 10-second sliding window is used, updated once per second, and the most recent 100 data points are analyzed. Assume that the standardized headlight brightness threshold is 0.8 and the angle threshold is ±30°. At a certain point in time, the calculated headlight brightness mean is 0.85, the variance is 0.02, and the rate of change is 0.01 / s; the angle mean is 25°, the variance is 5°, and the rate of change is 2° / s. Because the brightness mean exceeds the threshold, it is marked as potentially dangerous. At the same time, the dangerous driving behavior assessment module outputs a vehicle speed of 85km / h (weight 0.5) and an acceleration of 2.5m / s 2 (weight 0.3), steering angle change rate 15° / s (weight 0.2). The weighted average calculation results in a comprehensive score of 0.75 (full score 1). The preset danger threshold is 0.7, so this score is also marked as exceeding the threshold. The decision tree model receives these inputs: the headlight brightness exceeds the threshold (1), the headlight angle does not exceed the threshold (0), and the driving behavior exceeds the threshold (1). The trained decision tree (depth 5, minimum leaf node sample 100) processes these features and outputs a danger probability of 0.8. Finally, the comprehensive score is calculated: 0.4×(1+0) / 2+0.6×0.75=0.65. Assuming that the preset dangerous driving judgment threshold is 0.6, 0.65 exceeds the threshold at this time, it is judged as dangerous driving, and an audio and visual warning is immediately triggered, and the driver is advised to slow down and adjust the lights. At the same time, the on-board system records this event for subsequent model optimization and driving behavior analysis.

[0049] S105. If it is determined to be dangerous driving, the optimal warning strategy is matched from the preset warning strategy library in combination with the dangerous driving behavior prediction results, the dangerous driving behavior evaluation results and the standardized rear headlight data, and warning measures are generated. The brightness and angle of the rear headlights after the warning measures are simulated to evaluate the driver's visibility, and the evaluation results are compared with the preset visibility safety range. If it does not meet the safety range, the warning strategy is re-matched and the warning measures are re-generated.

[0050] The prediction results, evaluation results and standardized headlight data of dangerous driving behaviors are obtained to construct a multidimensional feature vector, wherein the multidimensional feature vector includes numerical features and categorical features; according to the multidimensional feature vector, a strategy matching algorithm based on a decision tree is used to find the most suitable warning strategy in a preset warning strategy library to obtain an initial optimal warning strategy; for the initial optimal warning strategy, specific warning measures are generated, wherein the warning measures include a percentage of headlight brightness adjustment and an angle adjustment degree; a simulation execution algorithm based on a vehicle dynamics model is used to calculate a new headlight brightness and a new angle according to the warning measures, wherein the simulation execution algorithm combines the effects of the vehicle driving state and environmental factors on visibility. Impact; input the new brightness and new angle data of the headlights into a driver visibility assessment model, the driver visibility assessment model is obtained by training with historical data using a multivariate linear regression algorithm, and the driver visibility assessment result after the simulation is executed is obtained; determine whether the driver visibility assessment result meets the preset safety range requirements, if not, optimize the early warning strategy using a genetic algorithm, the parameters of the genetic algorithm include population size, crossover rate, mutation rate and maximum number of iterations; generate new early warning strategies and early warning measures through iterative calculation, and repeat the simulation and evaluation until the preset safety range requirements are met or the maximum number of iterations is reached.

[0051] Specifically, a multidimensional feature vector is constructed based on the prediction results, evaluation results and standardized headlight data of dangerous driving behaviors, and the most suitable warning strategy is found in the preset warning strategy library using a decision tree-based strategy matching algorithm. The algorithm considers mixed-type features, including numerical (such as vehicle speed, acceleration) and categorical (such as road type, weather conditions). The strategy with the highest matching degree is selected as the initial optimal warning strategy, and the specific parameters of the strategy, such as the headlight brightness adjustment range and angle adjustment amplitude, are recorded. Based on the selected optimal warning strategy, specific warning measures are generated, including the headlight brightness adjustment percentage and angle adjustment degree. A simulation execution algorithm based on a vehicle dynamics model is used to calculate the new brightness and new angle of the headlights after the warning measures are executed, while considering the impact of vehicle driving status and environmental factors on visibility. The new brightness and new angle data of the headlights after the simulation execution are input into the driver visibility evaluation model, which uses a multivariate linear regression algorithm and is obtained through historical data training to output the driver visibility evaluation results after the simulation execution. According to the road type and weather conditions, the preset visibility safety range is defined as 100-300 meters. Compare the driver visibility assessment results with the preset safety range. If it does not meet the safety range requirements, use a genetic algorithm to optimize the warning strategy. The genetic algorithm parameters are set as: population size 50, crossover rate 0.8, mutation rate 0.1, and maximum number of iterations 100. Generate new warning strategies and warning measures through iterative calculations, and repeat the simulation and evaluation steps until the safety range requirements are met or the maximum number of iterations is reached. During the optimization process, consider the situation where multiple safety strategies run in parallel, and balance safety and driving experience by setting weight coefficients. After the on-board system detects dangerous driving behavior, it immediately starts the warning strategy matching process. A vector containing 10 features was constructed: vehicle speed 85km / h, acceleration 2.5m / s 2 , steering angle 15°, headlight brightness 0.9, headlight angle 20°, road type (urban road), weather conditions (rainy day), time (night), driving behavior score 0.75 and predicted danger level 0.8. The decision tree algorithm matches the library containing 1,000 warning strategies and selects the best strategy: reduce the speed by 20%, increase the headlight brightness by 15%, and adjust the headlight angle downward by 5°. Based on the vehicle dynamics model, the strategy is simulated and executed, and the new speed is calculated to be 68km / h, the new headlight brightness is 1.035, and the new headlight angle is 15°. These data are input into the trained multivariate linear regression model, which is trained using 10,000 sets of historical data, R 2The value is 0.85. The estimated visibility output by the model is 180 meters. This result is compared with the preset safety range (150-250 meters for urban roads on rainy days) and is found to meet the requirements. If it does not meet the requirements, the genetic algorithm will generate 50 candidate strategies and optimize them after up to 100 iterations. In each iteration, the algorithm selects the 20 best strategies for crossover and mutation to produce a new generation of strategies. The crossover operation uses a single-point crossover, and the mutation operation randomly adjusts the strategy parameters by ±10%. Finally, the strategy with visibility closest to the center value of the safety range is selected for execution, while taking into account the driving experience, such as limiting the speed adjustment range to within 30%. The entire process is completed within 0.5 seconds, ensuring real-time response to dangerous driving conditions.

[0052] S106, converting the generated warning measures into control instructions, issuing voice and visual warning prompts to the driver through the vehicle-mounted human-computer interaction system, and applying the control instructions on the headlight adjustment data in the warning measures to the preset headlight adjustment system to adjust the headlight brightness and angle.

[0053] Receive a preset command mapping table, the command mapping table contains the correspondence between warning measures and standardized control commands; convert the warning measures into standardized control commands according to the command mapping table, the standardized control commands include voice prompt commands, visual prompt commands and headlight adjustment commands; use the speech synthesis module of the vehicle-mounted human-computer interaction system to generate a warning voice prompt, the warning voice prompt is composed of pre-recorded audio clips; obtain the visual warning icon and text information on the vehicle-mounted display screen, the visual warning icon is displayed in a red flashing manner; parse the headlight adjustment command into a brightness adjustment percentage and an angle adjustment degree, and send the parsed command to the control unit of the headlight adjustment system through the vehicle-mounted communication bus; if the headlight adjustment system receives the control command, adjust the headlight angle through a stepper motor, and adjust the headlight brightness by PWM, LIN or direct digital control according to the vehicle model; receive a warning confirmation or rejection signal issued by the driver through a steering wheel button or voice command, and adjust the subsequent warning strategy according to the signal.

[0054] Specifically, a preset command mapping table is used to convert the warning measures into standardized control commands. The mapping table uses key-value pairs to store the warning measures and the corresponding standardized commands, and a hash table is used to achieve fast search. The commands include voice prompt commands, visual prompt commands, and headlight adjustment commands. Each command is assigned an execution priority according to the urgency of the warning measures. The voice prompt of the warning is generated by the speech synthesis module of the vehicle-mounted human-computer interaction system. The voice prompt is spliced ​​with pre-recorded audio clips. At the same time, the visual warning icon and text information are rendered on the vehicle display screen, using red flashing icons and text descriptions. The voice prompts and visual warnings are played in sequence according to the execution priority order. The headlight adjustment command is parsed into specific brightness adjustment percentages and angle adjustment degrees, and the parsed commands are sent to the control unit of the headlight adjustment system through the vehicle communication bus. The CAN bus protocol is used to transmit the headlight adjustment command in a standard frame format. After receiving the control command, the headlight adjustment system adjusts the headlight angle through a stepper motor. According to different vehicle models, the headlight brightness is adjusted by PWM, LIN or direct digital control. The feedback data during the adjustment process is monitored in real time. When the target value is reached, the adjustment is stopped and the execution result is fed back to the central control unit. The driver's confirmation or rejection of the warning is received through the steering wheel button or voice command, and the subsequent warning strategy is adjusted according to the feedback. In the case of multiple warning measures at the same time, they are processed in sequence according to the pre-set priority order. According to different road and weather conditions, the warning threshold is dynamically adjusted to ensure the adaptability and effectiveness of the warning measures. After the vehicle system detects dangerous driving behavior, it immediately searches for the corresponding warning measures from the preset 1000 instruction mapping table. For example, if the vehicle speed is detected to be too fast (120km / h, 40% overspeed), the corresponding instructions are matched through hash search within 0.01 seconds: voice prompt "please slow down immediately", visual warning "red flashing speed limit sign", and headlight adjustment "increase brightness by 20%". The speech synthesis module uses the three pre-recorded audio clips of "please", "immediately" and "slow down" to generate a complete speech at a sampling rate of 44.1kHz. At the same time, the display renders a red speed limit icon with a diameter of 10cm at a refresh rate of 60Hz, flashing once every 0.5 seconds. The headlight adjustment command is transmitted via the CAN bus at a rate of 500kbps, with a frame ID of 0x123, and the data field contains a command to increase the brightness by 20%. After receiving the command, the headlight control unit selects the appropriate dimming method according to the vehicle model. For LED headlights, a 20kHz PWM signal is used to increase the duty cycle from 50% to 60%. The headlight brightness is monitored in real time through a photometric sensor. When the brightness is detected to increase by 20% (±2% error range), the adjustment is stopped and feedback is given that the execution is completed. The entire process is completed within 1 second.If the driver responds to the warning by confirming the button on the steering wheel (double-click within 100ms) or by voice command ("Accept warning", confidence > 0.95), the interaction is recorded and the trigger threshold for the same type of warning is increased by 10% in the following 30 minutes. When multiple warnings are triggered at the same time, they are processed in sequence according to the preset priority (safety > efficiency > comfort). In addition, road and weather information is received in real time, and the speed warning threshold is reduced by 20% on rainy days to ensure the dynamic adaptability of the warning measures.

[0055] S107. Continuously track the driver's reaction, changes in the vehicle's driving status, and changes in the brightness and angle of the headlights, evaluate the effectiveness of the warning measures and the driver's visibility after the headlights change, and dynamically adjust the warning strategy and threshold according to the evaluation results.

[0056] The driver operation data, vehicle driving status data and headlight parameter data collected by the vehicle sensor network are obtained; a feature vector is generated according to the collected data, and the feature vector includes the driver reaction time, the vehicle speed change rate and the steering wheel angle change rate; the feature vector is processed by a multivariate linear regression model to obtain a driver response degree score and a vehicle status improvement degree score; if the driver response degree score is lower than a preset threshold, the current driving environment parameters and the driver visibility data are obtained; according to the comprehensive effect score, the driver visibility data and the current driving environment parameters, a fuzzy control algorithm is used to adjust the warning trigger threshold, the warning information presentation method and the headlight control parameters; the adjusted warning trigger threshold, the warning information presentation method and the headlight control parameters are applied to the next warning process; it is determined whether the driver identity exists in a pre-established driver portrait database; if the driver identity exists in the driver portrait database, a warning strategy benchmark corresponding to the driver identity is obtained from the driver portrait database; the sensitivity of the warning trigger threshold is dynamically adjusted according to the warning strategy benchmark.

[0057] Specifically, the vehicle sensor network collects driver operation data, vehicle driving status data and headlight parameter data in real time at a frequency of 10Hz. A 30-second sliding window is used to preprocess and extract features of the collected data once per second to generate a feature vector that reflects the driver's reaction and vehicle state changes. The feature vector is processed using a multivariate linear regression model, with driver reaction time, vehicle speed change rate, steering wheel angle change rate and other variables as independent variables, and the warning effect score as the dependent variable. The driver's response to the warning measures and the vehicle driving status improvement score are output. Combined with the headlight parameter change data, the comprehensive effect score of the warning measures is calculated. The driver's visibility is evaluated through on-board camera image analysis and driver eye tracking. Based on the comprehensive effect score, driver visibility and current driving environment parameters, the fuzzy control algorithm is used to dynamically adjust the warning strategy. The input variables of the fuzzy control algorithm include effect score, visibility and environmental parameters, and the output variables are the adjustment amount of the warning trigger threshold, the selection of warning information presentation mode and the adjustment amount of the headlight control parameters. The optimized warning strategy is generated by reasoning through the preset fuzzy rule library, including the new warning trigger threshold, the updated warning information presentation method and the adjusted headlight control parameters. The optimized warning strategy is applied to the next warning process, and the optimization results are stored in the warning strategy library. The model parameters are updated through the stochastic gradient descent algorithm to achieve incremental learning and continuous optimization of the warning system. According to the individual differences of different drivers, a driver portrait database is established, and the corresponding warning strategy benchmark is selected according to the driver's identity. By dynamically adjusting the sensitivity of the warning trigger threshold, the response speed and false alarm rate of the system are balanced to ensure personalized warning services under the premise of ensuring safety. The vehicle system collects data at a frequency of 10Hz through multiple sensors: the steering wheel angle sensor records steering changes, the vehicle speed sensor monitors speed fluctuations, and the camera captures the driver's eye movements and road environment. Using a 30-second sliding window, the data is updated once a second. For example, at a certain point in time, the vehicle system records that the driver's reaction time is 1.2 seconds, the vehicle speed drops from 80km / h to 60km / h, and the steering wheel angle changes by 20 degrees. The multivariate linear regression model receives these data and calculates the warning effect score as 85 points (out of 100). At the same time, the on-board camera detects that the visibility ahead is 150 meters through the image processing algorithm, and the driver's eye tracking shows that the concentration is 0.8 (out of 1). The fuzzy control algorithm receives these inputs and makes inferences based on preset rules such as "If the effect score is high and visibility improves, then slightly increase the warning threshold." The output results are: the speed warning threshold is increased from 20% to 25% overspeed, the warning information presentation method is changed from voice plus vision to visual prompts only, and the increase in headlight brightness is reduced from 20% to 15%. The on-board system applies this new strategy to the next warning and stores the data in the strategy library. The stochastic gradient descent algorithm updates the model parameters with a learning rate of 0.01.For newly identified drivers, the vehicle system found that they preferred mild warnings, so it relaxed all thresholds by 5% based on the baseline strategy. In this way, the vehicle system accumulated 100 warning experiences in one day, and the warning accuracy rate increased from the initial 80% to 95%, and the false alarm rate was reduced to less than 5%, achieving a balance between safety and user experience.

[0058] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A method for warning dangerous driving behavior of vehicle drivers based on deep learning, characterized in that: include: Obtain the headlight brightness and angle data collected by the vehicle-mounted sensor during vehicle driving; Preprocessing the headlight brightness and angle data to obtain standardized headlight brightness and angle data; Outputting current driver visibility through a deep learning model according to the standardized headlight brightness and angle data; Combining the driver visibility, the vehicle driving status data and the road environment condition data, obtaining a dangerous driving behavior prediction result through a dangerous driving behavior prediction model; Determine whether the driving is dangerous based on the predicted dangerous driving behavior and the headlight brightness and angle data; If it is judged as dangerous driving, the warning strategy is matched from the preset warning strategy library, the warning measures are generated and simulated, the warning effect is evaluated, and the warning strategy is adjusted if necessary until it meets the safe driving requirements.

2. The method according to claim 1, characterized in that The preprocessing of the vehicle light brightness and angle data includes: Sending a preprocessing instruction to the vehicle-mounted computing unit according to the vehicle-mounted lighting brightness and angle data, wherein the vehicle-mounted computing unit preprocesses the received vehicle-mounted lighting brightness and angle data using a wavelet transform denoising algorithm; The preprocessed vehicle headlight brightness and angle data are obtained, and the preprocessed vehicle headlight brightness and angle data are standardized.

3. The method according to claim 1, characterized in that Outputting the current driver visibility through the deep learning model includes: Inputting the standardized headlight brightness and angle data into a convolutional neural network, wherein the convolutional neural network comprises three convolutional layers and two pooling layers; Outputting the current driver visibility value through the fully connected layer of the convolutional neural network; The sliding window method is used to perform time series analysis on the driver visibility values, and the mean, variance and change trend of the visibility are calculated.

4. The method according to claim 1, characterized in that The combining the driver visibility, the vehicle driving state data and the road environment condition data to obtain the dangerous driving behavior prediction result through the dangerous driving behavior prediction model includes: Acquire vehicle driving state data and road environment condition data collected by vehicle-mounted sensors, wherein the vehicle driving state data includes vehicle speed, acceleration and steering angle, and the road environment condition data includes road type information and weather condition information; Preprocessing the vehicle driving state data and the road environment condition data to generate a standardized data set through denoising, normalization and time synchronization; A sixty-second sliding window is used, and the time series features of the standardized data set are extracted by sliding once every ten seconds, statistical features are calculated, and a feature vector is constructed as an input of the deep learning model; The feature vector is processed using a pre-trained long short-term memory network model, and the predicted probability distribution of dangerous driving behavior is obtained through forward propagation calculation.

5. The method according to claim 1, characterized in that The determining whether the driving behavior is dangerous based on the dangerous driving behavior prediction result and the headlight brightness and angle data includes: The standardized headlight brightness and angle data are analyzed using a sliding window to calculate the mean, variance, and rate of change within the window. Comparing the mean, variance and change rate with preset vehicle headlight brightness and angle hazard thresholds to determine whether there is an over-threshold situation; The weighted average method is used to process the dangerous driving behavior assessment results to obtain a comprehensive assessment score of the dangerous driving behavior; A binary decision tree model is constructed using the CART algorithm, and the exceeding threshold of the headlight brightness and angle data and the comprehensive evaluation score of the dangerous driving behavior are used as input features; Calculate the final dangerous driving probability score according to the output result of the binary decision tree model and the preset risk judgment weight; It is determined whether the dangerous driving probability score exceeds a preset threshold value. If it exceeds the preset threshold value, it is determined to be dangerous driving.

6. The method according to claim 1, characterized in that The method of matching the warning strategy from the preset warning strategy library, generating warning measures and simulating their execution, evaluating the warning effect, and adjusting the warning strategy when necessary until it meets the safe driving requirements, includes: Obtain dangerous driving behavior prediction results, evaluation results and standardized headlight data, and construct a multi-dimensional feature vector; According to the multidimensional feature vector, a decision tree-based strategy matching algorithm is used to search for the most suitable warning strategy in a preset warning strategy library to obtain an initial optimal warning strategy; Generate specific warning measures for the initial optimal warning strategy, the warning measures including the percentage of headlight brightness adjustment and the degree of angle adjustment; Using a simulation execution algorithm based on a vehicle dynamics model, a new brightness and a new angle of the headlights are calculated according to the warning measures; Inputting the new brightness and angle data of the headlights into the driver visibility assessment model to obtain the driver visibility assessment result after the simulation execution; Determine whether the driver visibility assessment result meets the preset safety range requirements. If it does not meet the preset safety range requirements, use a genetic algorithm to optimize the warning strategy, generate new warning strategies and warning measures through iterative calculations, and repeat simulation and evaluation until the preset safety range requirements are met or the maximum number of iterations is reached.

7. The method according to claim 1, characterized in that The method further comprises: Obtain driver operation data, vehicle driving status data and headlight parameter data collected by the vehicle sensor network; Generate a feature vector based on the collected data, the feature vector including driver reaction time, vehicle speed change rate, and steering wheel angle change rate; Processing the feature vector using a multivariate linear regression model to obtain a driver response degree score and a vehicle status improvement degree score; According to the comprehensive effect score, driver visibility data and current driving environment parameters, the fuzzy control algorithm is used to adjust the warning trigger threshold, warning information presentation method and headlight control parameters; The adjusted warning trigger threshold, warning information presentation method and headlight control parameters will be applied to the next warning process.

Images