A driver stress load real-time evaluation method based on urban street view images

Abstract

This invention discloses a real-time driver stress load assessment method based on urban street view images, comprising: constructing a dataset containing U samples; determining the stress level of the samples using heart rate variability; subsequently extracting street view image element feature variables and dynamic parameter feature variables from the sample video data and forming a sample feature vector; using correlation analysis and random forest importance analysis to reduce the dimensionality of the sample feature vector, transforming or removing strongly correlated feature variables; using stress level as a label, forming a feature-label matrix with the sample feature vector, and using it to train a fusion model containing m machine learning classifiers; and using the best-performing fusion model for real-time driver stress load assessment. This invention has high assessment accuracy, and the real-time assessment process has almost no impact on the driver's driving experience, ensuring the feasibility of widespread application.

Description

Technical Field

[0001] This invention belongs to the field of urban road traffic safety technology, specifically relating to a method for real-time assessment of driver stress load based on urban street view images. Background Technology

[0002] Road traffic accidents are a major obstacle to the development of the transportation industry. Current research indicates that driver factors are the core cause of traffic accidents, especially poor driving habits, which are the primary culprits. Among these factors, driver stress load has the most significant impact on safety. Therefore, real-time, efficient, and accurate assessment of driver stress load is crucial for reducing accident risk. Current research on driver stress load both domestically and internationally mainly falls into two categories:

[0003] (1) The first type assesses drivers' stress levels in different driving scenarios based on their subjective will. This method is mainly based on questionnaires, collecting drivers' personal information in the questionnaire items, and designing typical driving scenario items. Drivers score the stress load of driving scenarios based on their personal subjective feelings, and finally, the relationship between driving scenarios and stress load is derived from the questionnaire analysis results. The drawback of this method is that it can only cover a limited number of scenarios and is completely dominated by personal subjective will, resulting in insufficient reliability and reference value of its results.

[0004] (2) The second type is based on driver biopsychological parameters and collects data using driving simulators or real-vehicle driving experiments. This method determines the driver's stress level by having them verbally answer stress questionnaires during driving, and then uses advanced models or algorithms to assess driving stress load through driver biopsychological parameters. This method can provide a more objective and accurate evaluation of driving stress load and fully demonstrates the close relationship between biopsychological parameters and driving stress load. However, relying on driver biopsychological parameters, if applied to actual driving, cannot avoid intrusion and interference with the driver, affecting driving safety.

[0005] The invention disclosed in CN114852088A is a driver assistance system and method for identifying and warning of dangerous and fatigued driving behaviors. It eliminates the need to collect normal driving characteristics and known dangerous driving characteristics, and also eliminates the need to train a detection model, thus reducing implementation costs and increasing versatility. It combines road conditions, driver state, and vehicle driving characteristics, and has an adaptive adjustment mechanism to suit different driver styles, making the judgment of driving behavior more comprehensive and valuable, and the detection more accurate and effective. It balances timely warnings of dangerous behaviors with the detection of difficult-to-detect fatigued driving behaviors, effectively assisting drivers in safe driving. It causes minimal interference to the driver, effectively avoiding traffic accidents caused by dangerous and fatigued driving, and helps drivers detect slowly changing driving behaviors. When a driver's driving and decision-making abilities decline, it enables the driver to promptly recognize danger, thus effectively assisting drivers in safe driving. However, this invention mainly collects various factors that cause dangerous driving and fuses these factors using fuzzy inference algorithms to determine whether dangerous driving has occurred. It does not involve the quantitative processing of driver stress load, and the simple fusion of multiple factors is insufficient to draw objective conclusions about driver stress load.

[0006] The invention disclosed in CN107145835A is an image recognition-based vehicle-mounted device for detecting driver load. This device comprises a data acquisition unit, an image processing unit, a road alignment load level calculation unit, a visual load level calculation unit, a display unit, and a storage unit. The display unit receives data from the road alignment load level calculation unit and the visual load level calculation unit, quantifies and fuses the data, and displays the comprehensive driving load of the road environment in real time, as well as the system's operating status. This invention integrates the driver's visual load and the mental load induced by road alignment, demonstrating the driver's comprehensive driving load for analysis by management and design departments. However, this invention divides roads into two categories based on road alignment, evaluates driving load based on heart rate variability, and then fuses the driver's visual load and the mental load induced by road alignment. The article does not provide effective solutions for how to analyze or introduce other environmental factors. Furthermore, this invention only uses one time-domain indicator, namely the standard deviation of the heart rate sampling interval, when evaluating driving load. Quantifying driving load using a single indicator will lead to biased results; therefore, it is necessary to explore quantification methods using multiple indicators. Furthermore, the level of visual load is obtained by calculating the amount of information in the elements, which is a static classification method that ignores the dynamic characteristics of vision.

[0007] The invention disclosed in CN115035687A is a driver fatigue monitoring system based on seat pressure analysis. This system consists of a seat pressure sensing and monitoring unit, a sensing and monitoring data analysis unit, a vehicle control unit, and an alarm prompting unit. The system primarily determines the driver's fatigue state by analyzing the driver's body movement frequency and the multi-point pressure distribution on the seat, and controls the vehicle's operating status based on the determination results, while simultaneously sending alarm voice prompts. This invention integrates fatigue detection, vehicle control, and voice alarms, and can be deployed on physical media through program code to function as a device, possessing high practicality and potential for large-scale application in commercial vehicle safety assurance. However, this invention only controls the vehicle based on the driver's fatigue state, lacking consideration of actual road conditions and failing to account for the coupling effect of pressure load and fatigue state. Summary of the Invention

[0008] Technical problem solved: This invention discloses a real-time driver stress load assessment method based on urban street view images. It fully considers the influence of street view image elements and dynamic parameters on driver stress load, has high assessment accuracy, and the real-time assessment process has almost no impact on the driver's driving experience, ensuring the feasibility of widespread application.

[0009] Technical solution:

[0010] A method for real-time assessment of driver stress load based on urban street view images, the method comprising the following steps:

[0011] S10. Conduct a real-vehicle driving experiment on the target city roads. Match the three types of data in the time series according to the collection frequency of driver's heart rate, vehicle dynamics and street view images. Extract one sample of length t2 seconds every t1 seconds to construct a dataset containing U samples.

[0012] S20. Perform time-domain and frequency-domain analysis on the heart rate data of each sample, extract heart rate variability index based on the interval between two adjacent heartbeats, and use the heart rate variability index to determine the stress level of the sample; then extract the street scene image element feature variables and dynamic parameter feature variables from the sample video data, and form a sample feature vector.

[0013] S30 employs correlation analysis and random forest importance analysis to reduce the dimensionality of the sample feature vectors, transforming or removing strongly correlated feature variables, and ultimately retaining those with an importance not lower than the importance threshold I. o r characteristic variables;

[0014] S40 uses stress level as a label, and forms a feature-label matrix with sample feature vectors. This matrix is ​​used to train a fusion model containing m types of machine learning classifiers. The best-performing fusion model is then used for real-time assessment of driver stress load.

[0015] Furthermore, in step S10, the target urban roads include urban arterial or secondary arterial roads with high traffic volume, urban secondary road sections with severe traffic mixing, and intersections with frequent vehicle interactions.

[0016] Further, in step S20, the heart rate variability indicators include: mean RR interval mRR; RR interval range eRR; RR interval standard deviation SDRR; root mean square value (RMSSD) of the difference between adjacent RR intervals; and the difference between adjacent RR intervals below T. d The ratio of milliseconds pNNT d Low-frequency energy LF, 0.04≤LF<0.15Hz; High-frequency energy HF, 0.15≤HF<0.40Hz; ③ Low-frequency energy ratio LFnu, High-frequency energy ratio HFnu, ⑤ Low-frequency energy ratio (LF / HF).

[0017] Furthermore, the process of determining the stress level of a sample using heart rate variability indices includes:

[0018] Factor analysis was performed on the heart rate variability index, and the N principal factors with eigenvalues ​​greater than 1 were retained to obtain the score F for each principal factor n of sample i. in The pressure index SI of sample i is calculated using the following formula, along with the variance contribution rate pn. i :

[0019]

[0020] Where n = 1, 2, ..., N, SI i A higher stress level indicates higher stress. The stress index SI is analyzed using the K-means clustering algorithm. i Unsupervised clustering was performed to determine the stress levels of the samples as high, medium, or low.

[0021] Furthermore, the process of extracting feature variables of street view image elements from sample video data includes the following steps:

[0022] One image is extracted from the video data at time intervals Δt, resulting in K images for each sample. Image analysis models are used to perform semantic segmentation of elements in the image, and the proportion of each element is calculated:

[0023]

[0024] propis Represents the percentage of the s-th image element in the j-th image of sample i, pixel. is Represents the number of pixels in the s-th image element of the j-th image of sample i. total This represents the total number of pixels in the j-th image of sample i;

[0025] Calculate the rate of change of elements between two adjacent images:

[0026]

[0027] Where df_prop ijs It is the rate of change of the s-th image element between the j-th image and the (j+1)-th image of sample i;

[0028] The corresponding street scene element feature variables were calculated:

[0029]

[0030]

[0031]

[0032]

[0033] Among them Mi s SD represents the mean proportion of element s in sample i. is MD represents the standard deviation of the proportion of element s in sample i. is SDD represents the mean of the rates of change of element s in sample i. is The standard deviation represents the rate of change of element s in sample i.

[0034] Furthermore, the process of extracting dynamic parameter feature variables from sample video data includes the following steps:

[0035] The mean and standard deviation of each variable—vehicle speed, triaxial acceleration, triaxial angular velocity, pitch angle, roll angle, and yaw angle—were calculated as microdynamic characteristic variables.

[0036] The formula for calculating the change in yaw angle is as follows:

[0037] Y = Y t+1 -Y t

[0038] In the formula, Y represents the change in yaw angle, Y t+1 and Y t These represent the yaw angles of the vehicle at times t+1 and t, respectively.

[0039] Furthermore, in step S30, Spearman rank correlation test is performed on the feature variables of street scene elements and the feature variables of dynamic parameters, and feature dimensionality reduction is performed by using the method of calculating the importance of out-of-bag errors of random forests.

[0040] The formula for calculating the rank correlation coefficient is:

[0041]

[0042] Where ρ is the rank correlation coefficient between variables, U is the total number of samples, and d θ Indicates the rank difference between variables;

[0043] The importance of the feature variable is calculated using the following formula:

[0044]

[0045] Where I f Let f represent the importance of feature f, and dt represent the number of decision trees. This represents the error when predicting using all feature variables. This represents the error when using all feature variables except variable f for prediction.

[0046] Further, in step S40, the process of using stress level as a label and combining it with sample feature vectors to form a feature-label matrix, and using this matrix to train a fusion model containing m types of machine learning classifiers, and then using the best-performing fusion model for real-time assessment of driver stress load, includes the following steps:

[0047] The dimensionality-reduced sample dataset is rebalanced using synthetic minority class oversampling. The rebalanced dataset is then used to train decision trees, multivariate logit regression, gradient boosting trees, extreme gradient boosting trees, categorical gradient boosting trees, and lightweight gradient boosting trees.

[0048] Based on the training results of individual classifiers, a stacking method is used to treat each classifier as a meta-classifier and other classifiers as base classifiers to design the structure of the fusion model one by one, and the fusion model with G-means and F1 score greater than 80% is retained.

[0049] Adaboost's dataset resampling technique was used to further optimize the performance of the retained fusion model. The final performance evaluation of the fusion model was then conducted, and the best-performing fusion model was used for real-time assessment of driver stress load.

[0050] Beneficial effects:

[0051] First, the real-time driver stress load assessment method based on urban street view images of the present invention avoids excessive influence of subjective factors, calibrates the stress level through objective indicators, and comprehensively considers the influence of street view image elements and dynamic parameters on driver stress load. The fusion model adopted has high assessment accuracy and efficiency.

[0052] Second, the real-time driver stress load assessment method based on urban street view images of the present invention accurately extracts environmental element information in street view images through image semantic segmentation and numerical analysis, providing a new approach to assess driving stress from the perspective of driver visual perception.

[0053] The present invention provides a real-time driver stress load assessment method based on urban street view images. By using a fusion model, it fully demonstrates the significant correlation between street view images and dynamic parameters and driver stress load, and achieves real-time driver stress load assessment without intruding on the driving process, thus ensuring the feasibility of the method for widespread application. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of the road segment in the data collection area of ​​the example;

[0055] Figure 2 Data acquisition equipment for real vehicle testing and a schematic diagram of the vehicle's three axes;

[0056] Figure 3 This is a flowchart illustrating the use of the HRV index to calibrate driver stress levels according to an embodiment of the present invention.

[0057] Figure 4 This is a schematic diagram of the result after feature dimensionality reduction;

[0058] Figure 5 A flowchart for modeling the fusion model;

[0059] Figure 6 This is a flowchart of a method for real-time assessment of driver stress load based on urban street view images, according to an embodiment of the present invention. Detailed Implementation

[0060] The following embodiments are provided to enable those skilled in the art to more fully understand the present invention, but do not limit the invention in any way.

[0061] This invention discloses a method for real-time assessment of driver stress load based on urban street view images. The method includes the following steps:

[0062] S10, Conduct a real-vehicle driving experiment on the target city roads, match the three types of data according to the time series based on the collection frequency, and extract one sample of length t2 seconds every t1 seconds to construct a dataset containing U samples;

[0063] S20. Perform time-domain and frequency-domain analysis on the heart rate data of each sample, extract the heart rate variability (HRV) index based on the interval between two adjacent heartbeats (RR interval), and use the HRV index to determine the stress level of the sample; extract the feature variables of street scene image elements and dynamic parameter feature variables from the sample video data, and construct feature vectors;

[0064] S30 employs correlation analysis and random forest importance analysis to reduce the dimensionality of the feature vectors, combining or removing strongly correlated feature variables, ultimately retaining those with an importance of at least 1. o r characteristic variables;

[0065] S40 uses stress level as a label, forms a feature-label matrix with feature vectors, and is used to train a fusion model containing m types of machine learning classifiers to obtain the best-performing fusion model framework for real-time assessment of driver stress load.

[0066] Specifically, the HRV metrics extracted for each sample include:

[0067] Time-domain metrics: ① Mean Restriction Count (mRR, in ms); ② Restriction Count Range (eRR, in ms); ③ Restriction Count Standard Deviation (SDRR, in ms); ④ Root Mean Square Value (RMSSD, in ms) of the difference between adjacent Restriction Counts; ⑤ The difference between adjacent Restriction Counts is less than T. d The proportion of milliseconds (pNNT) d ,unit%);

[0068] Frequency domain specifications: ① Low-frequency energy (LF, 0.04≤LF<0.15Hz, unit: ms²); ② High-frequency energy (HF, 0.15≤HF<0.40Hz, unit: ms²); ③ Low-frequency energy ratio (LFnu, Unit %); ④ High-frequency energy ratio (HFnu, (Unit: %); ⑤ Low / High Frequency Energy Ratio (LF / HF).

[0069] Regarding the determination of sample stress levels based on the HRV index, firstly, confirmatory factor analysis is used to extract n principal factors with eigenvalues ​​greater than 1 from the HRV index, and then the factor n score F of sample i is obtained. in The variance contribution rate p of factor n n Then, the pressure index SI of sample i is calculated. i :

[0070]

[0071] SI iA higher stress level indicates higher stress. The stress index SI is analyzed using the K-means clustering algorithm. i Unsupervised clustering was performed to label the stress levels of the samples as high, medium, or low.

[0072] To extract driving environment element information from street view images, the method for extracting street view image element feature variables proposed in this embodiment of the invention is as follows:

[0073] One image is extracted from the video data at time intervals Δt, resulting in K images for each sample. Image analysis models are used to perform semantic segmentation of G-class elements in the image, and the proportion of each element is calculated:

[0074]

[0075] prop is Represents the percentage of the s-th image element in the j-th image of sample i, pixel. is Represents the number of pixels in the s-th image element of the j-th image of sample i. total This represents the total number of pixels in the j-th image of sample i.

[0076] Then, the rate of change of elements between two adjacent images is calculated:

[0077]

[0078] Where df_propi js This represents the rate of change of the s-th image element between the j-th image and the (j+1)-th image of sample i.

[0079] The street scene element feature variables that can be calculated are as follows:

[0080]

[0081]

[0082]

[0083]

[0084] Where M is SD is the mean proportion of element s in sample i. is MD represents the standard deviation of the proportion of element s in sample i. is SDD represents the mean of the rates of change of element s in sample i. is The standard deviation represents the rate of change of element s in sample i.

[0085] The method for extracting characteristic variables of dynamic parameters is as follows.

[0086] To ensure that the vehicle yaw angle accurately reflects the vehicle's steering, the following difference is applied:

[0087] Y = Y t+1 -Y t

[0088] Where Y is the change in yaw angle, Y t+1 and Y t These represent the yaw angles of the vehicle at times t+1 and t, respectively.

[0089] Subsequently, the mean and standard deviation of each variable—the changes in vehicle speed, triaxial acceleration, triaxial angular velocity, pitch angle, roll angle, and yaw angle—were calculated as microdynamic characteristic variables.

[0090] On the other hand, this embodiment also discloses a method for assessing driver stress load using the feature variables extracted above, including:

[0091] Dataset rebalancing and dataset diversification techniques are employed to improve the evaluation performance and generalization ability of the fusion model.

[0092] The fusion model is trained based on the feature-label matrix. The optimal fusion model framework is obtained by comparing the performance metrics (G-mean and F1 score) of different classifier combinations. The optimal fusion model framework is then used to conduct real-time assessment of driver stress load.

[0093] The following will use a section of road in Jiangning District, Nanjing City as an example to explain this method in detail. Figure 1 As shown, the selected road section falls under the jurisdiction of Jiangning District, Nanjing City, and covers three typical urban road driving scenarios:

[0094] (1) Urban main roads or secondary roads with high traffic volume (e.g.) Figure 1 Points D1 and D2);

[0095] (2) Urban secondary roads with severe traffic mixing (such as...) Figure 1 Points B1 and B2);

[0096] (3) Intersections with frequent vehicle interactions (such as...) Figure 1 Points C1 and C2).

[0097] Figure 1 The place names and other textual content in the image are not the focus of this application for protection, as they are only intended to show the target road section.

[0098] In this example, a real-vehicle driving experiment was conducted in the study area. The experimental instruments and vehicle three-axis indicators were as follows: Figure 2 As shown, the main data collected included driver heart rate data, vehicle micro-dynamic parameters, and street view video data in front of the vehicle. (See attached image.) Figure 6The publicly available method for real-time assessment of driver stress load includes the following main steps:

[0099] Step S10: Match the three types of data according to the time series order. Considering that the shortest meaningful period of the HRV index is 100 seconds, extract one 100-second sample every 10 seconds for the instance dataset, and collect a total of 5274 samples.

[0100] Step S20, the process is as follows Figure 3 As shown, the HRV index is extracted according to

[0012] -

[0014] , where T d We selected values ​​of 20 and 50, i.e., pNN20 and pNN50, respectively, and then performed confirmatory factor analysis on the indicators. The results are shown in Table 1.

[0101] Table 1. Results of confirmatory factor analysis

[0102]

[0103] Based on the results of confirmatory factor analysis, the stress index (SI) of 5274 samples was calculated. i for:

[0104]

[0105] Where F in p represents the factor n score of the i-th sample. n This represents the variance contribution rate of factor n.

[0106] Using K-means clustering algorithm to analyze the stress index SI i Unsupervised clustering was performed, resulting in 168 high-stress samples, 4162 medium-stress samples, and 944 low-stress samples.

[0107] Step S30: Capture one street scene image every second from the street scene video in front of the vehicle. In this example, the image analysis model selected is the Deeplabv3 model. The open-source street scene image dataset Cityscape is used to train the Deeplabv3 model. Extract the number of pixels of G=19 elements (road, sidewalk, building, wall, guardrail, pole, traffic light, sign, vegetation, terrain, sky, pedestrian, chess player, car, truck, bus, train, motorcycle, bicycle) from the sample street scene image data, and calculate the proportion of each element:

[0108]

[0109] prop is Represents the percentage of the s-th image element in the j-th image of sample i, pixel. isRepresents the number of pixels in the s-th image element of the j-th image of sample i. total This represents the total number of pixels in the j-th image of sample i.

[0110] Then, the rate of change of elements between two adjacent images is calculated:

[0111]

[0112] Where df_prop ijs This represents the rate of change of the s-th image element between the j-th image and the (j+1)-th image of sample i.

[0113] The street scene element feature variables that can be calculated are as follows:

[0114]

[0115]

[0116]

[0117]

[0118] Where M is SD represents the mean proportion of element s in sample i. is MD represents the standard deviation of the proportion of element s in sample i. is SDD represents the mean of the rates of change of element s in sample i. is Let S represent the standard deviation of the rate of change of element s in sample i, and K represent the total number of images in sample i. In this example, K = 100.

[0119] Step S40: Extract the characteristic variables of the dynamic parameters of the sample vehicles. First, perform difference processing on the yaw angle:

[0120] Y = Y t+1 -Y t

[0121] Where Y is the change in yaw angle, Y t+1 and Y t These represent the yaw angles of the vehicle at times t+1 and t, respectively.

[0122] Subsequently, the mean and standard deviation of each variable, including the changes in vehicle speed, three-axis acceleration, three-axis angular velocity, pitch angle, roll angle, and yaw angle, were calculated as micro-dynamic characteristic variables. The characteristic variables of the dynamic parameters are shown in Table 2.

[0123] Table 2 Characteristic variables of vehicle dynamic parameters

[0124]

[0125] Step S50 involves performing Spearman's rank correlation test on the street scene element feature variables and dynamic parameter feature variables, and then using the random forest out-of-bag error importance calculation method for feature dimensionality reduction. The rank correlation coefficient is calculated as follows:

[0126]

[0127] Where ρ is the rank correlation coefficient between variables, n is the total number of samples, i.e., n = 5274, d i Represents the rank difference between variables

[0128] The importance of feature variables is calculated as follows:

[0129]

[0130] Where I f The importance of feature f is represented by dt, and the number of decision trees is represented by errOOB. i1 ErrOOB represents the error when predicting using all feature variables. i2 This represents the error when using all feature variables except variable f for prediction.

[0131] The importance of feature variables drops sharply in the range less than 0.04. Therefore, in this example, I0 is set to 0.04. The feature variables retained after feature dimensionality reduction are as follows: Figure 4 As shown.

[0132] Step S60, the construction process of the fusion model is as follows: Figure 5 As shown, the dimensionality-reduced sample dataset is rebalanced using the Synthetic Minority Oversampling Technique (SMOTE). The rebalanced dataset is then used to train decision trees (DT), multivariate logit regression (MLR), gradient boosting trees (GBDT), extreme gradient boosting trees (XGBoost), categorical gradient boosting trees (Catboost), and lightweight gradient boosting trees (LightGBM).

[0133] Based on the training results of individual classifiers, a stacking method is used to design the structure of the fusion model by treating each classifier as a meta-classifier and the other classifiers as base classifiers, and retaining the fusion models with G-means and F1 scores greater than 80%.

[0134] In step S70, the dataset resampling technique of Adaboost is used to further optimize the performance of the retained fusion model. The final performance evaluation results of the fusion model are shown in Table 3.

[0135] Table 3 Performance Comparison Results of Fusion Models

[0136]

[0137]

[0138] Therefore, the fusion model obtained in this embodiment, which uses GBDT, XGBoost, Catboost, and LightGBM as base classifiers and DT as a meta classifier, can be used to perform real-time driver stress load assessment.

[0139] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention are within the protection scope of the present invention.

Claims

1. A method for real-time assessment of driver stress load based on urban street view images, characterized in that, The real-time driver stress load assessment method includes the following steps: S10 involves conducting real-vehicle driving experiments on target city roads. Based on the collection frequencies of three types of data—driver's heart rate, vehicle dynamics, and street view images—these three types of data are matched in a time series, and at each interval... Extract 1 per second Samples of length in seconds, constructing a collection A dataset of samples; S20. Perform time-domain and frequency-domain analysis on the heart rate data of each sample, extract heart rate variability index based on the interval between two adjacent heartbeats, and use the heart rate variability index to determine the stress level of the sample; then extract the street scene image element feature variables and dynamic parameter feature variables from the sample video data, and form a sample feature vector. S30 employs correlation analysis and random forest importance analysis to reduce the dimensionality of the sample feature vectors, transforming or removing strongly correlated feature variables, and ultimately retaining those with importance not lower than the importance threshold. of One characteristic variable; S40, using stress level as the label, forms a feature-label matrix with the sample feature vectors, and is used to train features containing... A fusion model of several machine learning classifiers is used to evaluate the driver's stress load in real time. Step S40 includes the following steps: The dimensionality-reduced sample dataset is rebalanced using synthetic minority class oversampling. The rebalanced dataset is then used to train decision trees, multivariate logit regression, gradient boosting trees, extreme gradient boosting trees, categorical gradient boosting trees, and lightweight gradient boosting trees. Based on the training results of individual classifiers, a stacking method is used to treat each classifier as a meta-classifier and other classifiers as base classifiers to design the structure of the fusion model one by one, and the fusion model with G-means and F1 score greater than 80% is retained. Adaboost's dataset resampling technique was used to further optimize the performance of the retained fusion model. The final performance evaluation of the fusion model was then conducted, and the best-performing fusion model was used for real-time assessment of driver stress load.

2. The method for real-time assessment of driver stress load based on urban street view images according to claim 1, characterized in that, In step S10, the target urban roads include urban arterial or secondary arterial roads with high traffic volume, urban secondary road sections with severe traffic mixing, and intersections with frequent vehicle interactions.

3. The method for real-time assessment of driver stress load based on urban street view images according to claim 1, characterized in that, In step S20, the heart rate variability indicators include: mean RR interval mRR; RR interval range eRR; RR interval standard deviation SDRR; root mean square value (RMSSD) of the difference between adjacent RR intervals; and the difference between adjacent RR intervals below a certain threshold. millisecond ratio pNN Low-frequency energy LF, 0.04≤LF<0.15Hz; High-frequency energy HF, 0.15≤HF<0.40Hz; ③ Low-frequency energy ratio LFnu, High-frequency energy ratio HFnu, ⑤ Low-to-high frequency energy ratio (LF / HF).

4. The method for real-time assessment of driver stress load based on urban street view images according to claim 1, characterized in that, The process of determining the stress level of a sample using heart rate variability indicators includes: Factor analysis was performed on the heart rate variability index, and eigenvalues ​​greater than 1 were retained. One principal factor, to obtain the sample Each principal factor Score and variance contribution rate The sample is calculated using the following formula. Stress index : ； in, , A higher stress level indicates higher stress. The stress index is analyzed using the K-means clustering algorithm. Unsupervised clustering was performed to determine the stress levels of the samples as high, medium, or low.

5. The method for real-time assessment of driver stress load based on urban street view images according to claim 1, characterized in that, The process of extracting feature variables of street view image elements from sample video data includes the following steps: For video data at each interval Take a snapshot of the time period and get the result for each sample. One picture, The image analysis model is used to perform semantic segmentation of the image elements, and the proportion of each element is calculated: ； in Indicates sample No. The first picture The proportion of each image element Indicates sample No. Image number 1 The number of pixels in each image element. Indicates sample No. The total number of pixels in the image; Calculate the rate of change of elements between two adjacent images: ； in It is a sample No. The picture and the first The first picture Rate of change of each image element; The corresponding street scene element feature variables were calculated: ； ； ； ； in Indicates sample medium elements The average percentage Indicates sample The standard deviation of the proportion of element s in the middle. Indicates sample element The mean of the rate of change, Indicates sample element Standard deviation of the rate of change.

6. The method for real-time assessment of driver stress load based on urban street view images according to claim 1, characterized in that, The process of extracting dynamic parameter feature variables from sample video data includes the following steps: The mean and standard deviation of each variable—vehicle speed, triaxial acceleration, triaxial angular velocity, pitch angle, roll angle, and yaw angle—were calculated as microdynamic characteristic variables. The formula for calculating the change in yaw angle is as follows: ； In the formula, This indicates the change in yaw angle. and These respectively indicate that the vehicle is in and Yaw angle at any given moment.

7. The method for real-time assessment of driver stress load based on urban street view images according to claim 1, characterized in that, In step S30, Spearman rank correlation test is performed on the feature variables of street scene elements and the feature variables of dynamic parameters, and feature dimensionality reduction is performed by using the method of calculating the importance of out-of-bag errors of random forests. The formula for calculating the rank correlation coefficient is: ； in The rank correlation coefficient between variables. It is the total number of samples. Indicates the rank difference between variables; The importance of the feature variable is calculated using the following formula: ； in Representation of features Importance Indicates the number of decision trees. This represents the error when predicting using all feature variables. Indicates using the division variable Errors in predicting all external feature variables.

Images