SVM-based fatigue driving state detection method and system

Abstract

The invention relates to an SVM-based fatigue driving state detection method and system, and the method comprises the following steps: collecting behavior feature data of a driver, and constructing a multi-modal behavior feature vector; classifying the multi-modal behavior feature vectors through a mixed kernel function SVM model to obtain behavior state classification, wherein the mixed kernel function SVM model comprises a dual-kernel function pool of a radial basis kernel and a polynomial kernel and an online incremental learning framework; the behavior state classification is judged according to a judgment threshold value, a driver fatigue driving state detection result is obtained, and the judgment threshold value is obtained through dynamic adjustment of a dynamic threshold value adjustment module according to the multi-mode behavior characteristic data. Compared with the prior art, the fatigue driving state detection method has the advantages that the conditions of misjudgment and missed judgment are effectively reduced, and the accuracy and reliability of fatigue driving state detection are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic safety monitoring, and in particular, to a fatigue driving state detection method and system based on SVM. Background Art

[0002] With the rapid development of modern transportation industry, the number of motor vehicles in possession continues to grow, and road traffic safety issues have attracted increasing attention. Fatigue driving, as one of the main causes of traffic accidents, seriously threatens people's lives and property safety. According to relevant statistical data, traffic accidents caused by fatigue driving account for a relatively large proportion in various traffic accidents. Therefore, how to accurately and timely detect the fatigue state of drivers has become a research hotspot in the field of traffic safety.

[0003] Currently, the detection methods for drivers' fatigue driving state are mainly divided into two categories: subjective detection and objective detection. Subjective detection methods mainly rely on drivers' self-report or questionnaires. This method has disadvantages such as strong subjectivity and poor real-time performance, and cannot reflect the true fatigue state of drivers in a timely and accurate manner. Objective detection methods judge the fatigue state by collecting and analyzing drivers' physiological signals or behavioral characteristics. For example, Chinese Patent CN116844137A discloses an in-vehicle detection method based on detecting facial features. This method includes: inputting the collected driver's face image for feature extraction, extracting and calculating facial features related to fatigue detection. Judging eye opening and closing through the eye opening degree, and judging whether the driver yawns through the mouth opening degree. Using the eye opening degree, blink frequency, and mouth opening degree as the input of the fuzzy control system, and obtaining the fatigue state according to the classification algorithm. Judging the physiological fatigue degree of the output fatigue degree value, and the output value is the fatigue state of the driver, thus realizing fatigue detection. This method can only judge fatigue driving based on the currently obtained driver's face image, and cannot detect fatigue driving according to the driver's real-time driving state. Traditional SVM-based fatigue detection methods have three limitations: Firstly, the static kernel function selection strategy is difficult to adapt to the dynamic changes of feature distributions in different driving scenarios; Secondly, the offline training mode cannot dynamically adjust the model parameters according to the individual physiological characteristics of drivers; Thirdly, the fixed threshold decision mechanism lacks robustness to various environmental interferences. These defects lead to a relatively high false alarm rate and insufficient personalized adaptation ability in the actual in-vehicle scenario of the existing system. Summary of the Invention

[0004] The purpose of the present invention is to overcome the above-mentioned defects existing in the prior art and provide a fatigue driving state detection method and system based on SVM, which improves the accuracy of fatigue driving state detection.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A fatigue driving state detection method based on SVM, comprising the following steps:

[0007] Collect the behavioral feature data of the driver and construct a multi-modal behavioral feature vector;

[0008] Classify the multi-modal behavioral feature vector through a hybrid kernel function SVM model to obtain a behavioral state classification. The hybrid kernel function SVM model includes a dual kernel function pool of a radial basis kernel and a polynomial kernel and an online incremental learning framework. The dual kernel function pool of the radial basis kernel and the polynomial kernel is used to switch kernel functions and adjust parameters according to the classification accuracy of each kernel function, dynamically optimize the kernel function, and the online incremental learning framework is used to dynamically update the model parameters according to the multi-modal behavioral feature data. The hybrid kernel function SVM model obtains a behavioral state classification based on the optimized kernel function and the updated model parameters;

[0009] Judge the behavioral state classification according to a judgment threshold to obtain a detection result of the driver's fatigue driving state. The judgment threshold is dynamically adjusted by a dynamic threshold adjustment module according to the multi-modal behavioral feature data.

[0010] Further, the behavioral features include the driver's eye behavioral features, mouth behavioral features, and head behavioral features. The eye behavioral feature is blinking, the mouth behavioral feature is yawning, and the head behavioral feature is sleepy nodding.

[0011] Further, the function expression of the dual kernel function pool of the radial basis kernel and the polynomial kernel is:

[0012] K 双核 =λK RBF +(1-λ)K 多项式

[0013] In the formula, K 双核 is the dual kernel pool function of the radial basis kernel and the polynomial kernel, λ is the mixing coefficient, K RBF is the radial basis kernel function, and K 多项式 is the polynomial kernel function.

[0014] Further, the online incremental learning framework includes a short-term adaptive layer and a long-term memory layer. The short-term adaptive layer is used to receive in real time the multi-modal behavioral feature vector extracted from the video stream, update the support vector set to adapt to the instantaneous behavior changes of the driver, and at the same time use the KKT condition to screen the support vectors to optimize the decision hyperplane; the long-term memory layer uses the random Fourier feature mapping to linearize the kernel function and updates the weight matrix of the decision hyperplane by the recursive least squares method.

[0015] Further, the decision hyperplane is:

[0016] WT X + b = 0

[0017] Wherein, W is the weight matrix, X is the multi-modal behavior feature vector, and b is the bias term.

[0018] Furthermore, the update formula of the weight matrix is:

[0019] W 新 = W 旧 + η(X T X + ∈I) -1 X T (y - f(X))

[0020] Wherein, W 新 is the updated weight matrix, W 旧 is the weight matrix before update, η is the learning rate, X is the multi-modal behavior feature vector, ∈ is the regularization parameter, I is the identity matrix, y is the true label vector, and f(X) is the predicted value of the multi-modal behavior feature vector.

[0021] Furthermore, the dynamic threshold adjustment module adjusts the determination threshold in real time by collecting the reference features of the driver in the natural state.

[0022] Furthermore, the reference features include the reference blink frequency, the reference yawn count, and the reference nod amplitude.

[0023] Furthermore, the determination thresholds include an eye determination threshold, a mouth determination threshold, and a head determination threshold. The eye determination threshold is:

[0024] T 眼部 = 0.25μ 眼部 + 0.75σ 眼部

[0025] Wherein, T 眼部 is the eye determination threshold, μ 眼部 is the average value of the eye aspect ratio during calibration, and σ 眼部 is the standard deviation of the eye aspect ratio;

[0026] The mouth determination threshold is:

[0027] T 嘴部 = μ 嘴部 + 1.25σ 嘴部

[0028] Wherein, T 嘴部 is the mouth determination threshold, μ 嘴部 is the average value of the mouth aspect ratio during calibration, and σ 嘴部 is the standard deviation of the mouth aspect ratio;

[0029] The head determination threshold is:

[0030] T 头部 = μ 头部 + 1.25σ 头部

[0031] Wherein, T 头部 is the head determination threshold, μ 头部 is the average value of the head aspect ratio during calibration, and σ 头部 is the standard deviation of the head aspect ratio.

[0032] According to another aspect of the present invention, there is provided a fatigue driving state detection system based on SVM, including:

[0033] A behavioral feature data acquisition module, configured to acquire the behavioral feature data of the driver and construct a multi-modal behavioral feature vector;

[0034] A behavioral feature vector classification module, configured to classify the multi-modal behavioral feature vector through a hybrid kernel function SVM model to obtain a behavioral state classification. The hybrid kernel function SVM model includes a dual kernel function pool of a radial basis kernel and a polynomial kernel and an online incremental learning framework. The dual kernel function pool of the radial basis kernel and the polynomial kernel is used to switch the kernel function and adjust the parameters according to the classification accuracy of each kernel function, dynamically optimize the kernel function, and the online incremental learning framework is used to dynamically update the model parameters according to the multi-modal behavioral feature data. The hybrid kernel function SVM model obtains the behavioral state classification based on the optimized kernel function and the updated model parameters;

[0035] A behavioral state classification determination module, configured to determine the behavioral state classification according to a determination threshold to obtain a detection result of the driver's fatigue driving state. The determination threshold is dynamically adjusted by a dynamic threshold adjustment module according to the multi-modal behavioral feature data.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] 1. The present invention improves the accuracy of detecting the driver's fatigue driving state by acquiring the behavioral feature data of the driver, constructing a multi-modal behavioral feature vector, classifying the multi-modal behavioral feature vector through a hybrid kernel function SVM model to obtain a behavioral state classification, and determining the behavioral state classification according to a determination threshold to obtain a detection result of the driver's fatigue driving state.

[0038] 2. The present invention establishes a dual kernel function pool including a radial basis kernel and a polynomial kernel, evaluates the classification accuracy of each kernel function in real time through cross-validation, switches the kernel function and adjusts the parameters according to the classification accuracy of each kernel function, and dynamically optimizes the kernel function, significantly improving the accuracy and robustness of the SVM model.

[0039] 3. In view of the individual physiological differences of drivers, the present invention dynamically updates the decision hyperplane of the SVM model through an online incremental learning framework, and continuously conducts data analysis and threshold adjustment through a dynamic threshold adjustment module to adapt to the changes in the behavior characteristics of drivers in real time, improving the reliability of detecting the fatigue driving state of drivers. Description of the Drawings

[0040] Figure 1 is a schematic flowchart of a fatigue driving state detection method based on SVM proposed by the present invention;

[0041] Figure 2 is a schematic diagram of the eye aspect ratio;

[0042] Figure 3 is a schematic diagram of the mouth aspect ratio;

[0043] Figure 4 is a schematic diagram of the head key point aspect ratio. Detailed Embodiment

[0044] The present invention will be described in detail below with reference to the drawings and specific embodiments. This embodiment is implemented on the premise of the technical solution of the present invention, and gives the detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0045] English abbreviations involved:

[0046] Support Vector Machine: SVM

[0047] Radial Basis Function: RBF

[0048] Random Fourier Features: RFF

[0049] Karush-Kuhn-Tucker: KKT

[0050] Eye Aspect Ratio: EAR

[0051] Mouth Aspect Ratio: MAR

[0052] Head Aspect Ratio: HAR

[0053] Embodiment 1

[0054] This embodiment provides a fatigue driving state detection method based on SVM, as Figure 1 shown, including the following steps:

[0055] S1. Collect the behavioral characteristic data of the driver and construct a multi-modal behavioral feature vector.

[0056] Collect the behavioral characteristic data of the driver at a frame rate of 30fps. The behavioral characteristics of the driver include the eye behavioral characteristics, mouth behavioral characteristics, and head behavioral characteristics. The eye behavioral characteristic is blinking, the mouth behavioral characteristic is yawning, and the head behavioral characteristic is sleepy nodding.

[0057] By calculating the Eye Aspect Ratio (EAR), the blinking action of the driver can be accurately detected. As Figure 2 shown, the EAR value is calculated by the distance ratio of the eye feature points in the vertical and horizontal directions. The calculation formula of EAR is:

[0058]

[0059] In the formula, P1 is the inner eye feature point, P4 is the outer eye feature point, P2 and P3 are the upper eye feature points, and P5 and P6 are the lower eye feature points.

[0060] When the eyes are closed, the EAR value drops rapidly to near zero; when the eyes are open, the EAR value remains relatively stable. A multi-modal feature vector is constructed through the EAR values of consecutive frames.

[0061] Yawning is one of the typical manifestations of fatigue. The opening and closing degree change of the mouth is monitored by calculating the Mouth Aspect Ratio (MAR). As Figure 3 shown, the MAR value is calculated by the distance ratio of the mouth key points in the vertical and horizontal directions. The calculation formula of MAR is:

[0062]

[0063] In the formula, P 49 and P 55 are the outer mouth feature points, P 62 and P 64 are the upper mouth feature points, and P 66 and P 68 are the lower mouth feature points.

[0064] When the driver yawns frequently, the MAR value will increase significantly. A multi-modal feature vector is constructed through the MAR values of consecutive frames.

[0065] Sleepy nodding is another important feature of fatigue driving. The drooping and nodding actions of the head are detected by calculating the Head Aspect Ratio (HAR). As Figure 4As shown, the HAR value is calculated by the ratio of the vertical and horizontal distances of the head key points. The calculation formula of HAR is as follows:

[0066]

[0067] In the formula, P 28 is the feature point at the uppermost part of the nose bridge, P 31 is the feature point at the lowermost part of the nose bridge, and P4 and P 14 are the feature points at the lowermost parts on both sides of the face.

[0068] When the driver nods frequently, the HAR value will change, and a multi-modal feature vector is constructed through the HAR values of consecutive frames.

[0069] The multi-modal behavior feature vector is input into the SVM model for classification to obtain the behavior state classification.

[0070] S2. The multi-modal behavior feature vector is classified by the SVM model with a hybrid kernel function to obtain the behavior state classification.

[0071] The SVM model with a hybrid kernel function includes a dual kernel function pool of a radial basis kernel and a polynomial kernel, an online incremental learning framework, and a dynamic threshold adjustment module.

[0072] The dual kernel function pool of the radial basis kernel and the polynomial kernel is used to switch the kernel function and adjust the parameters according to the classification accuracy of each kernel function, and dynamically optimize the kernel function.

[0073] The dual kernel function pool of the radial basis kernel and the polynomial kernel evaluates the classification accuracy of each kernel function in real time through cross-validation. When the accuracy of the RBF kernel continuously decreases, the system automatically switches to the polynomial kernel function and recalculates the kernel parameters to ensure the classification accuracy.

[0074] The function expression of the dual kernel function pool of the radial basis kernel and the polynomial kernel is as follows:

[0075] K 双核 = λK RBF +(1 - λ)K 多项式

[0076] In the formula, K 双核 is the dual kernel pool function of the radial basis kernel and the polynomial kernel, λ is the mixing coefficient, and the mixing coefficient is dynamically adjusted by the classification confidence within a sliding window (window size = 50 frames). When the confidence is lower than 85%, the kernel function switch is triggered. K RBF is the radial basis kernel function, and K 多项式 is the polynomial kernel function.

[0077] The online incremental learning framework is used to dynamically update model parameters according to behavioral feature data. The online incremental learning framework includes a short-term adaptive layer and a long-term memory layer. The short-term adaptive layer is used to receive in real time the multi-modal behavioral feature vectors extracted from the video stream, and update the support vector set to adapt to the instantaneous behavioral changes of the driver. At the same time, the KKT condition is used to screen the support vectors to optimize the decision hyperplane. The long-term memory layer uses the random Fourier feature mapping to linearize the kernel function, and updates the weight matrix of the decision hyperplane by the recursive least squares method.

[0078] The decision hyperplane is:

[0079] W T X + b = 0

[0080] Wherein, W is the weight matrix, X is the multi-modal behavioral feature vector, and b is the bias term.

[0081] The update formula of the weight matrix is:

[0082] W 新 = W 旧 + η(X T X + ∈I) -1 X T (y - f(X))

[0083] Wherein, W 新 is the updated weight matrix, W 旧 is the weight matrix before update, η is the learning rate, X is the multi-modal behavioral feature vector (X is an n×d matrix, where n is the number of samples and d is the feature dimension. Each row represents the feature vector of a sample), ∈ is the regularization parameter (set to 1×10 -5 , used to prevent matrix singularity and ensure matrix invertibility), I is the identity matrix, y is the true label vector, f(X) is the predicted value of the multi-modal behavioral feature vector, y - f(X) is the prediction error, and X T (y - f(X) represents the projection of the error in the feature space.

[0084] S3. Determine the classification of the behavioral state according to the determination threshold to obtain the detection result of the driver's fatigue driving state.

[0085] The dynamic threshold adjustment module is used to dynamically adjust the determination threshold according to the behavioral feature data. For individual physiological differences, through the baseline feature self-learning mechanism, after the driver's identity verification is passed, the baseline features of the driver in the natural state are collected. The baseline features include the baseline blink frequency, the baseline yawn count, and the baseline nod amplitude. The dynamic threshold adjustment module adjusts the determination threshold in real time by collecting the baseline features of the driver in the natural state to improve the accuracy and robustness of the detection.

[0086] The determination thresholds include an eye determination threshold, a mouth determination threshold, and a head determination threshold.

[0087] The eye determination threshold is:

[0088] T 眼部 = 0.25μ 眼部 + 0.75σ 眼部

[0089] Wherein, T 眼部 is the eye determination threshold, μ 眼部 is the average value of the eye aspect ratio during calibration, and σ 眼部 is the standard deviation of the eye aspect ratio;

[0090] The mouth determination threshold is:

[0091] T 嘴部 = μ 嘴部 + 1.25σ 嘴部

[0092] Wherein, T 嘴部 is the mouth determination threshold, μ 嘴部 is the average value of the mouth aspect ratio during calibration, and σ 嘴部 is the standard deviation of the mouth aspect ratio;

[0093] The head determination threshold is:

[0094] T 头部 = μ 头部 + 1.25σ 头部

[0095] Wherein, T 头部 is the head determination threshold, μ 头部 is the average value of the head aspect ratio during calibration, and σ 头部 is the standard deviation of the head aspect ratio.

[0096] When the EAR is greater than the eye determination threshold, it is determined that the eyes are open; when the EAR is less than the eye determination threshold, it is determined that the eyes are closed. In this embodiment, the eye determination threshold is 0.3, and the number of consecutive frames of closed eyes is set to 3, that is, when the EAR in 3 consecutive frames is less than the eye determination threshold, it indicates that a fatigued eye-closing operation has occurred. That is to say, when the EAR of 3 consecutive frames is less than 0.3 and there are still blinking actions in the next consecutive frames, it can be judged as fatigue and a fatigue driving reminder is given.

[0097] If the MAR is greater than the mouth determination threshold, it is determined that the mouth is open; if the MAR is less than the mouth determination threshold, it is determined that the mouth is closed. To ensure accuracy, in this embodiment, the continuous number of frames with the mouth open is set to 3. If the open-mouth action appears three times consecutively, it is recognized as one yawn. This is to avoid misjudging a single open-mouth as a fatigue state and record the number of yawns.

[0098] When the HAR of three consecutive frames is less than the head determination threshold, it indicates that the driver has had a nodding-off nod. Record the number of nodding-off nods. If the number of consecutive nodding-off nods exceeds ten times, it can be determined that the driver is in a fatigued state, and a warning message is sent to the driver.

[0099] By comprehensively utilizing the three key behavioral characteristics of blinking, yawning, and nodding-off, various physiological manifestations of the driver in a fatigued state are comprehensively captured. As a powerful classifier, the SVM model can effectively learn and analyze the internal relationships and patterns among these complex behavioral characteristics. Verified by a large number of experiments, compared with the traditional single-characteristic judgment method, the accuracy of the judgment of the driver's fatigue driving state in the present invention has been significantly improved. In the test of simulating the actual driving environment, through the online incremental learning framework and the dynamic threshold adjustment module, the system can adapt to the individual differences of different drivers, improve the accuracy and reliability of detection. The judgment accuracy rate of the present invention has increased by 5.2% compared with the traditional method, supporting the adaptation of new driver characteristics within 30 seconds. The individual adaptation has increased the fatigue detection F1-score from 0.82 to 0.93. The area under the ROC curve (AUC) of the system under vibration interference reaches 0.978, significantly superior to the traditional weighted fusion method (AUC = 0.921), effectively reducing the situations of misjudgment and missed judgment.

[0100] Embodiment 2

[0101] This embodiment provides a fatigue driving state detection system based on SVM, including:

[0102] A behavioral characteristic data acquisition module for acquiring the behavioral characteristic data of the driver and constructing a multi-modal behavioral characteristic vector;

[0103] A behavioral characteristic vector classification module for classifying the multi-modal behavioral characteristic vector through a hybrid kernel function SVM model to obtain a behavioral state classification. The hybrid kernel function SVM model includes a dual kernel function pool of a radial basis kernel and a polynomial kernel and an online incremental learning framework. The dual kernel function pool of the radial basis kernel and the polynomial kernel is used to switch the kernel function and adjust the parameters according to the classification accuracy rate of each kernel function, dynamically optimize the kernel function, and the online incremental learning framework is used to dynamically update the model parameters according to the multi-modal behavioral characteristic data. The behavioral state classification is obtained based on the optimized kernel function and the updated model parameters;

[0104] The behavior state classification determination module is used to determine the behavior state classification according to the determination threshold to obtain the detection result of the driver's fatigue driving state, and the determination threshold is dynamically adjusted by the dynamic threshold adjustment module according to the multi-modal behavior feature data.

[0105] The rest is the same as in Embodiment 1.

[0106] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field of the present invention through logical analysis, reasoning, or limited experiments based on the concept of the present invention on the basis of the prior art should be within the protection scope determined by the claims.

Claims

1. A fatigue driving state detection method based on SVM, characterized in that, It includes the following steps: Collect the behavioral feature data of the driver and construct a multi-modal behavioral feature vector; Classify the multi-modal behavioral feature vector through a hybrid kernel function SVM model to obtain a behavioral state classification. The hybrid kernel function SVM model includes a dual kernel function pool of a radial basis kernel and a polynomial kernel and an online incremental learning framework. The dual kernel function pool of the radial basis kernel and the polynomial kernel is used to switch kernel functions and adjust parameters according to the classification accuracy of each kernel function, dynamically optimize the kernel function, and the online incremental learning framework is used to dynamically update the model parameters according to the multi-modal behavioral feature data. The hybrid kernel function SVM model obtains a behavioral state classification based on the optimized kernel function and the updated model parameters; Judge the behavioral state classification according to a judgment threshold to obtain a detection result of the driver's fatigue driving state. The judgment threshold is dynamically adjusted by a dynamic threshold adjustment module according to the multi-modal behavioral feature data.

2. The fatigue driving state detection method based on SVM according to claim 1, wherein The behavioral features include the eye behavioral features, mouth behavioral features, and head behavioral features of the driver. The eye behavioral feature is blinking, the mouth behavioral feature is yawning, and the head behavioral feature is nodding off.

3. The fatigue driving state detection method based on SVM according to claim 1, characterized in that, The function expression of the dual kernel function pool of the radial basis kernel and the polynomial kernel is: K 双核 = λK RBF +(1 - λ)K 多项式 Where, K 双核 is the dual-core pooling function of the radial basis kernel and the polynomial kernel, λ is the mixing coefficient, K RBF is the radial basis kernel function, and K 多项式 is the polynomial kernel function.

4. The fatigue driving state detection method based on SVM according to claim 1, wherein, The online incremental learning framework includes a short-term adaptive layer and a long-term memory layer. The short-term adaptive layer is used to receive the multi-modal behavioral feature vector extracted from the video stream in real time, update the support vector set to adapt to the instantaneous behavior changes of the driver, and at the same time use the KKT condition to screen the support vectors to optimize the decision hyperplane; the long-term memory layer uses the random Fourier feature mapping to linearize the kernel function and updates the weight matrix of the decision hyperplane through the recursive least squares method.

5. The fatigue driving state detection method based on SVM according to claim 4, characterized in that, The decision hyperplane is: W T X + b = 0 In the formula, W is the weight matrix, X is the multi-modal behavioral feature vector, and b is the bias term.

6. The fatigue driving state detection method based on SVM according to claim 4, wherein The update formula of the weight matrix is: W 新 = W 旧 + η(X T X + ∈I) -1 X T (y - f(X)) where, W 新 is the updated weight matrix, W 旧 is the weight matrix before update, η is the learning rate, X is the multi-modal behavior feature vector, ∈ is the regularization parameter, I is the identity matrix, y is the true label vector, and f(X) is the predicted value of the multi-modal behavior feature vector.

7. The fatigue driving state detection method based on SVM according to claim 1, characterized in that, The dynamic threshold adjustment module adjusts the judgment threshold in real time by collecting the reference features of the driver in the natural state.

8. The fatigue driving state detection method based on SVM according to claim 7, characterized in that, The reference features include the reference blinking frequency, the reference yawning times, and the reference nodding amplitude.

9. The fatigue driving state detection method based on SVM according to claim 7, characterized in that, The judgment threshold includes an eye judgment threshold, a mouth judgment threshold, and a head judgment threshold. The eye judgment threshold is: T 眼部 = 0.25 μ 眼部 + 0.75 σ 眼部 where T 眼部 is the eye determination threshold, μ 眼部 is the average of the eye aspect ratio during calibration, and σ 眼部 is the standard deviation of the eye aspect ratio; The mouth judgment threshold is: T 嘴部 = μ 嘴部 + 1.25σ 嘴部 where T 嘴部 is the mouth determination threshold, and μ 嘴部 is the average value of the mouth aspect ratio during calibration, and σ 嘴部 is the standard deviation of the mouth aspect ratio; The head judgment threshold is: T 头部 = μ 头部 + 1.25σ 头部 where T 头部 is the head determination threshold, μ 头部 is the average of the head aspect ratios during calibration, and σ 头部 is the standard deviation of the head aspect ratios.

10. A fatigue driving state detection system based on SVM, characterized in that, It includes: A behavioral feature data acquisition module, which is used to collect the behavioral feature data of the driver and construct a multi-modal behavioral feature vector; A behavioral feature vector classification module, which is used to classify the multi-modal behavioral feature vector through a hybrid kernel function SVM model to obtain a behavioral state classification. The hybrid kernel function SVM model includes a dual kernel function pool of a radial basis kernel and a polynomial kernel and an online incremental learning framework. The dual kernel function pool of the radial basis kernel and the polynomial kernel is used to switch kernel functions and adjust parameters according to the classification accuracy of each kernel function, dynamically optimize the kernel function, and the online incremental learning framework is used to dynamically update the model parameters according to the multi-modal behavioral feature data. The hybrid kernel function SVM model obtains a behavioral state classification based on the optimized kernel function and the updated model parameters; The behavior state classification determination module is used to determine the behavior state classification according to the determination threshold to obtain the detection result of the driver's fatigue driving state, and the determination threshold is dynamically adjusted by the dynamic threshold adjustment module according to the multi-modal behavior feature data.

Images