Vehicle auxiliary driving method based on driver individualized fatigue detection

Abstract

The application belongs to the technical field of intelligent transportation, and discloses a vehicle auxiliary driving method based on driver individualized fatigue detection. The method fuses multi-modal data and driving behaviors to cluster drivers, establishes an individualized fatigue detection model for specific driving groups, and adopts adaptive intelligent fatigue control means for different driving groups, so as to realize individualized fatigue detection and intelligent auxiliary control of drivers. The method mainly comprises the following steps: data acquisition and data preprocessing; constructing a similarity matrix; driver clustering; constructing a group-specific fatigue detection model for different categories of drivers; determining a new driver category; determining whether the new driver is in a fatigue state; and according to the most critical features affecting driver fatigue detection, adopting adaptive safety control means. The application can fuse multi-modal information to individually determine the driving state of drivers and adaptively adopt intelligent safety auxiliary control means.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of intelligent transportation, and relates to a vehicle auxiliary driving method based on driver individualized fatigue detection. BACKGROUND

[0002] Fatigue driving is one of the main causes of traffic accidents, because the reaction time, judgment and alertness of the driver are affected. Fatigue driving not only poses a safety risk to the driver himself, but also endangers the safety of other road users.

[0003] Current detection methods for fatigue driving face the following problems: 1. Models relying on a single physiological signal source are difficult to fully capture the multi-dimensional complexity of fatigue state, and their cross-individual generalization ability is limited; 2. Current research mainly relies on large physiological monitoring devices such as electroencephalogram (EEG), electrocardiogram (ECG) and eye tracker to collect data, which not only has high cost, but also easily causes discomfort when worn, and fatigue detection technology based on facial features is extremely sensitive to lighting conditions, and drivers wearing masks or sunglasses will make it difficult to effectively capture facial features; 3. Methods based on computer vision to capture driver facial features are extremely susceptible to external environmental factors, and the recognition accuracy is unstable; 4. There is currently a lack of individualized modeling for fatigue driving detection, as a large amount of individual-specific normal driving and fatigue driving data is required, resulting in a heavy data collection burden; 5. The current method also has a key problem, that is, new drivers will encounter safety hazards caused by detection gaps during the initial data accumulation stage.

[0004] In addition, current methods for driver fatigue intervention and control also have limitations: 1. There is a problem of single strategy and invasiveness in the intervention execution level; current mainstream primary intervention is often too simple and rigid, which is easy to make the driver "habitual numbness" or interfere with driving safety; 2. There is a problem of "only warning, not intervening" in the solution level; after the system detects that the driver is in a fatigue state, it will issue an intervention warning, but it does not take corresponding substantive auxiliary measures to help the driver get out of danger; 3. There is a lack of individualized adaptation methods for different drivers; current fatigue intervention methods for drivers all use a fixed and unchanged standard, without considering individual differences among drivers, so that the intervention does not achieve the expected effect. SUMMARY

[0005] The application aims to provide a vehicle auxiliary driving method based on driver personalized fatigue detection, which fuses multi-modal data and driving behaviors to cluster driver groups, and respectively establishes personalized fatigue detection models for specific driving groups, determines the new driver category to be fatigue detected in actual detection, detects whether the new driver is fatigue based on the fatigue detection model under the category, and takes adaptive intelligent fatigue control means for the new driver according to the most critical features affecting the fatigue of different driving groups, so as to realize personalized fatigue detection and intelligent auxiliary control of drivers.

[0006] In order to achieve the above-mentioned purpose, the application adopts the following technical solutions:

[0007] The vehicle auxiliary driving method based on driver personalized fatigue detection comprises the following steps:

[0008] Step 1. Synchronously collect physiological signal data and driving behavior data of the driver, and perform preprocessing and feature extraction;

[0009] Step 2. Construct a similarity matrix based on all driving data of the driver obtained by preprocessing and feature extraction in step 1, and perform group clustering of the driver based on the obtained similarity matrix by using a clustering algorithm;

[0010] Step 3. Construct a specific fatigue detection model corresponding to the driver group of the category for different categories of driver groups, and train the fatigue detection model by using normal and fatigue driving data samples of the corresponding driver group;

[0011] Step 4. Determine the category group to which the new driver to be fatigue detected belongs based on the normal driving data of the new driver;

[0012] Step 5. Put the driving data of the new driver into the fatigue detection model of the corresponding category, and the fatigue detection model predicts whether the new driver is in a fatigue state; if yes, go to step 6; otherwise, do not process;

[0013] Step 6. For the new driver in a fatigue state, perform feature importance sorting according to the category group to which the new driver belongs, to obtain the most critical features affecting the fatigue detection of the new driver, and then take intelligent fatigue auxiliary control means accordingly.

[0014] The application has the following advantages:

[0015] As described above, this invention discloses a vehicle-assisted driving method based on personalized driver fatigue detection. Addressing the issue of significant individual physiological differences among drivers, this invention employs a similarity matrix-driven clustering method based on physiological signal data and driving behavior data of drivers under normal and fatigued driving conditions. This method groups drivers with similar driving characteristics into one category, achieving clustering based on the mapping relationship between "driving characteristics" and "driving state." It innovatively proposes a group segmentation method based on population similarity. Compared to traditional methods that rely solely on driving data without corresponding driving states, this method provides more accurate results during clustering, improving the accuracy of group segmentation. This enables personalized fatigue detection and intelligent assisted control for different driving groups, significantly improving the accuracy of fatigue detection and the effectiveness of assisted control. For different driver groups, this invention uses a Bayesian-optimized weighted search mechanism to construct a group-specific fatigue detection ensemble model for each category, achieving personalized modeling. This overcomes the problem of insufficient model generalization ability, significantly improving the accuracy and effectiveness of fatigue driving detection. To address the high data costs associated with existing personalized fatigue detection models that rely on both normal and fatigued driving data, this invention proposes a method based on a probability-weighted voting XGBoost model to determine the category of a new driver using only their normal driving data, significantly reducing data costs. Furthermore, this invention proposes an adaptive vehicle intelligent safety control assistance strategy targeting the driving data characteristics of each driver category—the most critical features influencing driver fatigue detection. Based on the new driver's category, targeted adaptive intelligent fatigue assistance control is achieved, overcoming the template-based and unspecific problems of traditional intervention methods, improving the effectiveness of fatigue control, and breaking through the technical bottleneck of "only warning, no intervention." Additionally, addressing the reliance on single physiological signal sources and the bulky and expensive physiological signal monitoring equipment in existing technologies, this invention uses a portable, low-cost physiological wristband to collect multi-source physiological signal data from the driver, thus significantly improving the accuracy of driving state discrimination by fusing multimodal information. Attached Figure Description

[0016] Figure 1 This is a flowchart of a vehicle assisted driving method based on personalized driver fatigue detection in an embodiment of the present invention.

[0017] Figure 2 This is a flowchart illustrating the construction of a similarity matrix and the grouping of drivers in an embodiment of the present invention.

[0018] Figure 3 This is a flowchart illustrating the construction and training process of the fatigue detection model in this embodiment of the invention.

[0019] Figure 4This is a flowchart illustrating the specific process for determining the category of a new driver in an embodiment of the present invention. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0021] Example 1

[0022] This embodiment describes a vehicle-assisted driving method based on personalized driver fatigue detection. It integrates multimodal data and driving behavior to perform driver clustering, establishes a personalized fatigue detection model for a specific driver group, and adopts adaptive intelligent fatigue control measures for different driver groups to achieve personalized driver fatigue detection and intelligent assisted control.

[0023] like Figure 1 As shown, this vehicle-assisted driving method based on personalized driver fatigue detection includes the following steps:

[0024] Step 1. Synchronously collect the driver's physiological signal data and driving behavior data, and perform preprocessing and feature extraction.

[0025] Step 1.1. In order to achieve data collection, this invention uses a physiological wristband and a driving simulator to simultaneously collect physiological signal data and driving behavior data of multiple drivers under normal driving and fatigue driving conditions.

[0026] The collected physiological signal data includes the driver's heart rate, skin conductance, and wrist movement signals; driving behavior data includes vehicle speed, vehicle acceleration, lane departure, steering wheel angle, and yaw rate.

[0027] Step 1.2. For the collected physiological signal data and driving behavior data, box plots are used to identify quartile boundaries and remove outliers from the data. At the same time, Kalman filtering is applied to the physiological signal data to reduce high-frequency noise interference.

[0028] Step 1.3. Using a sliding time window, extract multiple features highly correlated with driving fatigue from the collected physiological signal data and driving behavior data, respectively, including heart rate features, skin conductance features, wrist movement features, and driving behavior features.

[0029] The sliding time window length is 30 seconds, and the sliding step size is 5 seconds.

[0030] Heart rate characteristics include average heart rate, standard deviation of heart rate, maximum heart rate, and minimum heart rate.

[0031] The characteristics of skin conductance include the mean skin conductance, the median skin conductance, and the skin conductance skewness.

[0032] Wrist movement characteristics include average X-axis acceleration, average X-axis angular velocity, average Y-axis acceleration, average Y-axis angular velocity, average Z-axis acceleration, and average Z-axis angular velocity.

[0033] Driving behavior characteristics include average speed, average lateral acceleration, average longitudinal acceleration, average lane departure, average steering wheel angle, and average yaw rate.

[0034] Step 2. Construct a similarity matrix based on all driving data of drivers obtained from preprocessing and feature extraction in Step 1, and use a clustering algorithm to cluster the drivers based on the obtained similarity matrix.

[0035] The process of constructing the similarity matrix is ​​as follows:

[0036] For any two drivers, the normal driving data and fatigue driving data of one driver i are used as the training set and fed into the CatBoost model for training, while the normal driving data and fatigue driving data of the other driver j are used as the test set.

[0037] If two drivers have similar driving characteristics, the model trained on the driving data of the i-th driver will perform better on the driving data of the j-th driver, with a higher number of correctly predicted samples on the test set, resulting in a higher prediction accuracy. Conversely, if the two drivers have significantly different driving characteristics, the model trained on the driving data of the i-th driver will perform worse on the driving data of the j-th driver, with a lower number of correctly predicted samples on the test set, resulting in a lower prediction accuracy.

[0038] The accuracy metric is defined as the ratio of the number of samples correctly predicted by the model on the test set to the total number of samples.

[0039] The construction of the similarity matrix here provides a basis for the driver clustering mentioned below. Clustering is based on the mapping relationship between "driving features and driving state", which makes the clustering results more accurate.

[0040] The accuracy of the CatBoost model prediction is used as an indicator to evaluate driver similarity. The higher the indicator, the better the model trained on the training set performs on the test set, indicating that the two drivers are more similar.

[0041] All drivers are trained in pairs sequentially to obtain the similarity among all drivers, thus forming a similarity matrix.

[0042] The value in the i-th row and j-th column of the similarity matrix represents the similarity between the i-th driver and the j-th driver. Here, i = 0, 1, ..., N-1, and j = 0, 1, ..., N-1. Due to the inherent randomness during model training, the similarity obtained by training with the driving data of the i-th driver as the training set and the driving data of the j-th driver as the test set will not be consistent with the similarity obtained by training with the driving data of the j-th driver as the training set and the driving data of the i-th driver as the test set. Therefore, the resulting similarity matrix will be asymmetric.

[0043] For new drivers who need to be tested for fatigue driving (i.e., those to be tested for fatigue), only their normal driving data is known.

[0044] Next, using the similarity matrix, the optimal number of clusters is determined by the Gap statistic algorithm based on statistical inference. .

[0045] The core principle of the Gap statistic algorithm is to determine the optimal number of clusters m by comparing the cluster compactness difference between the actual data and the uniformly distributed reference data. The algorithm steps are as follows:

[0046] I. For each cluster number Based on the similarity feature vector of each driver, the sum of squared intra-cluster errors among these similarity feature vectors is calculated, which is the sum of squared distances from all points in the cluster to the centroid. The formula is expressed as follows:

[0047] .

[0048] In the formula This represents the sum of squared errors within the cluster; Let represent the set of samples in the t-th cluster; Cluster The center of mass; Indicates sample to the center of mass Euclidean distance, .

[0049] II. Generate 5 reference datasets from random data that follow the same distribution as the actual dataset.

[0050] III. For each The value is used to calculate the sum of squared errors within each cluster for each reference dataset.

[0051] IV. Calculate each The gap statistic of the value is expressed by the following formula:

[0052] .

[0053] In the formula For each Gap statistic of value Indicates the actual dataset in Sum of squared intra-cluster errors for each cluster; Indicates the first A reference dataset in Sum of squared intra-cluster errors for each cluster.

[0054] V. Choose the option that maximizes the Gap statistic. The value, as the optimal number of clusters. .

[0055] Set the maximum number of candidate clusters. N represents the number of drivers. The formula is expressed as follows:

[0056] .

[0057] This invention can classify all drivers into groups based on their similarity. These categories form the basis for enabling personalized fatigue detection and intelligent driving assistance controls for different driver groups.

[0058] Optimal number of clusters The number of cluster centers in the K-Means clustering algorithm divides drivers into... The class, the process is as follows:

[0059] Step 2.1. From Randomly selected from drivers These objects serve as the initial cluster centers. .

[0060] Step 2.2. Calculate for each sample separately. To each cluster center Euclidean distance The formula is as follows:

[0061] ;in For the first Cluster centers, .

[0062] Sample Assign it to the nearest cluster, as shown in the following formula:

[0063] .

[0064] in Indicates the first Each cluster contains a set of drivers; Indicates the first Cluster centers.

[0065] Step 2.3. After all samples have been assigned, recalculate the centers of each cluster using the following formula:

[0066] ;

[0067] Step 2.4. Combine the cluster centers obtained in Step 2.3 with those from the previous iteration. Comparison of cluster centers.

[0068] If the cluster centers change, proceed to step 2.2; otherwise, proceed to step 2.5.

[0069] Step 2.5. When the cluster centers no longer change, stop and output the drivers included in each category; at this point, all drivers have been divided into groups based on their similarity. There are 1 category, numbered 0, 1, ..., m-1 respectively.

[0070] The clustering algorithm described above has good clustering accuracy because it defines the clustering metric by using the driver similarity matrix and clusters drivers based on the mapping relationship between driving data and driving status, thus making the clustering results more accurate.

[0071] Step 3. For different categories of driver groups, construct specific fatigue detection models corresponding to each category of driver group, and train the fatigue detection models using normal and fatigued driving data samples of the corresponding driver groups.

[0072] Step 3.1. For each driver group segmented in Step 2, construct an ensemble model consisting of three base classifiers: a logistic regression classifier, an XGBoost classifier, and a random forest classifier.

[0073] The specific weight allocation of the three base classifiers is dynamically determined by the weight search mechanism in step 3.2.

[0074] Step 3.2. For each category of ensemble model, a weight search mechanism based on Bayesian optimization design is used to automatically search and optimize the weights of the three base classifiers.

[0075] For the logistic regression classifier, StandardScaler is used to standardize the input features; the regularization strength parameter is adjusted to C=0.1; and the standardized data is input into the L2 regularized logistic regression classifier.

[0076] In addition, the logistic regression classifier uses the LBFGS optimizer to achieve efficient parameter solving.

[0077] For the XGBoost classifier, feature randomization is used to limit the subset of split features for each tree; the tree depth is set to max_depth=6; the learning rate is set to learning_rate=0.05; and the split threshold is set to gamma=0.1.

[0078] For the random forest classifier, a class balancing strategy is used to automatically adjust the sample weights.

[0079] The class balancing strategy is used to automatically adjust the sample weights; the feature subset for splitting each tree is limited by feature randomization; the tree depth is set to max_depth=12; the minimum number of split samples is set to 5; and the minimum number of leaf nodes is set to 2.

[0080] In this invention, the logistic regression model excels at constructing a linear relationship between feature values ​​and driving state, the XGBoost model has significant advantages in algorithm improvement, and the random forest model has the advantage of capturing complex and nonlinear data.

[0081] This invention leverages a combination of the three base classifiers mentioned above, its core advantage lying in constructing a complementary and comprehensive "model council." This combination cleverly integrates three distinct modeling philosophies: the logistic regression classifier provides a stable and reliable linear baseline, excelling at capturing global trends; the random forest classifier uses "collective averaging decision-making" to uncover complex nonlinear interactions and local patterns, effectively combating overfitting; and the XGBoost classifier, through sequential optimization, focuses on overcoming the prediction residuals that the former two failed to address, thereby achieving precise capture of complex patterns. This fusion essentially combines the stability of linear models, the nonlinear capabilities of ensemble tree models, and the precise optimization of gradient boosting, fundamentally ensuring the diversity of ensemble models. The complementary strengths of different algorithms help to maximize their optimal performance.

[0082] Specifically, step 3.2 is as follows:

[0083] Step 3.2.1. Dynamically explore the three-dimensional weight space using the Optuna framework, and perform normalization processing on the weights generated in each iteration, where each iteration is defined as a complete base classifier weight search process using the Optuna framework.

[0084] The objective function is defined as maximizing the accuracy on the validation set, and the three-dimensional weight space is dynamically explored using the Optuna framework.

[0085] For the weights found in each iteration Perform the following normalization process:

[0086] ;

[0087] in This represents the weight of the i-th base classifier generated by Bayesian optimization; Indicates the effective weight.

[0088] Step 3.2.2. The predicted probabilities of the driver's driving state from the logistic regression classifier, XGBoost classifier, and random forest classifier are weighted and averaged to obtain the ensemble model's predictions for the class. Overall confidence level The calculation formula is as follows:

[0089] ;

[0090] in This indicates the driver's driving status, i.e., the category. The feature vectors representing the normal driving and fatigue driving data of the drivers in the validation set; This indicates that the base classifier predicts the category to belong to. The probability of.

[0091] This indicates that the vehicle is in a normal driving state. This indicates that the driver is in a state of fatigue.

[0092] Step 3.2.3. Divide the drivers in this category into two parts: use all the driving data of one driver as the validation set of the model, and use all the driving data of the other drivers as the training set of the model.

[0093] The training and validation sets are input into an ensemble model constructed using a weight search mechanism based on Bayesian optimization for training, and the prediction accuracy index of the ensemble model constructed under the weight combination of the base classifiers in this iteration is obtained.

[0094] The driving state with the highest overall confidence calculated in step 3.2.2 of the ensemble model built under the weight combination of the base classifiers in this iteration will be used as the driving state prediction result of the model in this iteration.

[0095] In this invention, a Bayesian optimization hyperparameter tuning method is used to automatically find various weight combinations and continuously optimize them to find the weight combination that maximizes the accuracy of the validation set, thereby constructing the optimal fatigue detection model.

[0096] Step 3.3. For each category of ensemble model, after multiple dynamic search base classifier weight iteration experiments, extract the base classifier weight combination that achieves the highest accuracy on the validation set in the experiment, and use it as the final weight combination.

[0097] In this embodiment, the number of iterations is, for example, 50, but this is just an example.

[0098] Step 3.4. Re-input all driving data of drivers in this category into the ensemble model composed of the three base classifiers under the final weight combination for training, and obtain the specific fatigue detection model for this category.

[0099] Step 4. Based on the normal driving data of the new driver to be fatigued, determine the category group to which he belongs.

[0100] An XGBoost multi-class classification model with probability-weighted voting is used to determine the driver category of new drivers based on the collected normal driving data. The specific determination process is as follows:

[0101] Step 4.1. Based on the driver clustering results obtained in Step 2, assign a label corresponding to the clustering result to each normal driving data sample of each known driver, i.e. , This represents the number of categories.

[0102] Step 4.2. Use the normal driving data of drivers of each category as the training set of the model, and the category label corresponding to each sample as the test set, input them into the XGBoost multi-class classification model for training. Multi-class classification model;

[0103] Step 4.3. Input the normal driving data of the new driver into the XGBoost multi-class classification model as the validation set. The XGBoost multi-class classification model outputs a probability vector of each sample belonging to each category through training. The calculation formula is as follows:

[0104] ;

[0105] In the formula This represents the number of samples of new drivers driving normally. Indicates the first The sample belongs to the first The probability of a class.

[0106] Step 4.4. Based on the probability vector Calculate cumulative probability By summing the probability vectors column-wise, we obtain the probability that all normal driving data samples belong to each category, thus obtaining the probability of all new driver samples. The sum; the calculation formula is as follows:

[0107] .

[0108] Step 4.5. Determine the category with the highest cumulative probability for the new driver and output it.

[0109] In this invention, the probability-weighted voting method is more flexible than the hard voting mechanism that directly determines whether a new driver belongs to the category with the highest probability. It allows more data to participate in the decision-making process, thereby improving the accuracy of category determination.

[0110] Step 5. Input the new driver's driving data into the fatigue detection model of its corresponding category. The fatigue detection model predicts whether the new driver is in a state of fatigue. If the driver is in a state of fatigue, proceed to step 6; otherwise, do nothing.

[0111] For new drivers to be fatigued, physiological signal data and driving behavior data under normal driving conditions are collected simultaneously using a physiological wristband and a driving simulator, and then preprocessed and feature extracted.

[0112] The physiological signal data and driving behavior data of the new driver obtained after preprocessing and feature extraction are input into the fatigue detection model corresponding to the category of the new driver that was pre-trained in step 3.

[0113] The fatigue detection model outputs the new driver's driving status to determine whether the new driver is fatigued.

[0114] If the driver is detected to be in a normal driving state, the vehicle's adaptive safety assistance control method will not be activated, and the system will remain in a silent detection state; conversely, if the driver is in a state of fatigued driving, adaptive vehicle safety assistance control measures will be implemented.

[0115] Step 6. For new drivers who are fatigued, sort the importance of features according to their category group to obtain the most critical features affecting fatigue detection of new drivers, and then take targeted intelligent fatigue assistance control measures.

[0116] First, we built a feature importance ranking model based on the XGBoost model.

[0117] During the training phase, all driving data of the remaining drivers belonging to the same category as the new driver are divided into training and testing sets, and input into the feature importance ranking model for training and testing.

[0118] Based on the feature importance ranking chart obtained from model training, the most critical features affecting driver fatigue detection for this category are identified. These most critical features include driving behavior characteristics, heart rate characteristics, skin conductance characteristics, or wrist movement characteristics.

[0119] If the most critical characteristic affecting driver fatigue detection in the new driver's category is driving behavior characteristics, the vehicle will send a request through the visual and auditory channels to whether to activate the lane keeping assist system.

[0120] If the most critical characteristic affecting driver fatigue detection in the new driver's category is heart rate, the vehicle will display a selection interface on the in-vehicle display asking whether to take over or drive to a service area.

[0121] If the most critical characteristic affecting driver fatigue detection in the new driver's category is skin conductance, the vehicle control system issues a command to drive the window lifting mechanism to lower the window by, for example, 5-10 cm, to help the driver regain alertness.

[0122] If the most critical characteristic affecting driver fatigue detection in the new driver's category is wrist movement, the vehicle will send a request through visual and auditory channels to whether to activate Adaptive Cruise Control (ACC).

[0123] If the driver does not respond within a preset time period (e.g., 10 seconds) after the vehicle issues the above request, the vehicle will automatically activate the corresponding auxiliary control measures to achieve intelligent auxiliary control.

[0124] The vehicle assisted driving method based on personalized driver fatigue detection proposed in this invention can integrate multimodal information to make personalized judgments on the driver's driving state and adaptively adopt intelligent safety assistance control measures.

[0125] Example 2

[0126] This embodiment 2 describes a vehicle assisted driving system based on personalized driver fatigue detection. This system is based on the same inventive concept as the vehicle assisted driving method based on personalized driver fatigue detection in embodiment 1 above.

[0127] The vehicle-assisted driving system based on personalized driver fatigue detection in this embodiment includes the following modules:

[0128] The driving data acquisition and preprocessing module is used to simultaneously collect the driver's physiological signal data and driving behavior data, and perform preprocessing and feature extraction.

[0129] The driver group clustering module is used to construct a similarity matrix based on all driving data of drivers obtained through preprocessing and feature extraction, and to perform group clustering of drivers based on the obtained similarity matrix using a clustering algorithm;

[0130] The fatigue detection module is used to build specific fatigue detection models for different types of driver groups, and to train the fatigue detection models using normal and fatigued driving data samples of the corresponding driver groups.

[0131] The new driver category determination module is used to determine the category group to which a new driver belongs based on their normal driving data.

[0132] The new driver fatigue detection module is used to put the driving data of new drivers into the fatigue detection model of their corresponding category. The fatigue detection model predicts whether the new driver is in a state of fatigue.

[0133] It also includes an auxiliary control module, which first targets new drivers in a state of fatigue, sorts them by feature importance to obtain the most critical features affecting the detection of fatigue in new drivers, and then takes targeted intelligent fatigue auxiliary control measures.

[0134] It should be noted that any content not mentioned in the above-described functional modules of the system described in this embodiment can be referred to the step description of the corresponding method in Embodiment 1 above, and will not be repeated in detail here.

[0135] Example 3

[0136] This embodiment 3 describes a computer device including a memory and one or more processors. Executable code is stored in the memory. When the processor executes the executable code, it implements the steps of the vehicle assisted driving method based on personalized driver fatigue detection in embodiment 1 above.

[0137] In this embodiment, the computer device can be any device or apparatus with data processing capabilities, and will not be described in detail here.

[0138] Example 4

[0139] This embodiment 4 describes a computer-readable storage medium storing a program that, when executed by a processor, is used to implement the steps of the vehicle assisted driving method based on personalized driver fatigue detection in embodiment 1 above.

[0140] The computer-readable storage medium can be an internal storage unit of any device or apparatus with data processing capabilities, such as a hard disk or memory, or an external storage device of any device with data processing capabilities, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc.

[0141] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A vehicle-assisted driving method based on personalized driver fatigue detection, characterized in that, Includes the following steps: Step 1. Synchronously collect the driver's physiological signal data and driving behavior data, and perform preprocessing and feature extraction; Step 2. Construct a similarity matrix based on all driving data of drivers obtained from preprocessing and feature extraction in Step 1, and use a clustering algorithm to cluster the drivers based on the obtained similarity matrix; Step 3. For different categories of driver groups, construct specific fatigue detection models for the corresponding categories of driver groups, and train the fatigue detection models using normal and fatigued driving data samples of the corresponding driver groups; Among them, the specific fatigue detection model refers to an integrated model composed of three base classifiers in parallel: logistic regression classifier, XGBoost classifier and random forest classifier. Step 4. Based on the normal driving data of the new driver to be fatigued, determine their category group; Step 5. Input the new driver's driving data into the fatigue detection model corresponding to its category. The fatigue detection model predicts whether the new driver is in a state of fatigue. If the driver is in a state of fatigue, proceed to step 6; otherwise, do nothing. Step 6. For new drivers who are fatigued, sort the importance of features according to their category group to obtain the most critical features affecting fatigue detection of new drivers, and then take targeted intelligent fatigue assistance control measures. Construct a feature importance ranking model based on the XGBoost model; During the training phase, all driving data of other drivers belonging to the same category as the new driver are divided into training and testing sets and input into the feature importance ranking model for training and testing. Based on the feature importance ranking chart obtained from model training, the most critical features affecting driver fatigue detection for this category are identified. Among these, the most critical features are driving behavior features, heart rate features, skin conductance features, or wrist movement features.

2. The vehicle assisted driving method based on personalized driver fatigue detection according to claim 1, characterized in that, Step 1 specifically involves: Step 1.

1. Use a physiological wristband and driving simulator to synchronously collect physiological signal data and driving behavior data of multiple drivers under normal driving and fatigue driving conditions; The physiological signal data includes the driver's heart rate signal, skin conductance signal, and wrist movement signal; the driving behavior data includes vehicle speed, vehicle acceleration, lane departure, steering wheel angle, and yaw rate. Step 1.

2. For the collected physiological signal data and driving behavior data, box plots are used to identify quartile boundaries and remove outliers from the data. At the same time, Kalman filtering is applied to the physiological signal data to reduce high-frequency noise interference. Step 1.

3. Using a sliding time window, extract multiple features highly correlated with driving fatigue from the collected physiological signal data and driving behavior data, respectively, including heart rate signal, skin conductance signal, wrist movement signal, and driving behavior characteristics.

3. The vehicle assisted driving method based on personalized driver fatigue detection according to claim 1, characterized in that, In step 2, the similarity matrix is ​​constructed as follows: For any two drivers, the normal driving data and fatigue driving data of the i-th driver are used as the training set and fed into the CatBoost model for training, and the normal driving data and fatigue driving data of the j-th driver are used as the test set. The accuracy metric is defined as the ratio of the number of correctly predicted samples on the test set to the total number of samples. The accuracy of the CatBoost model prediction is used as an indicator to evaluate driver similarity. The higher the indicator, the better the model trained on the training set data performs on the test set, which means that the two drivers are more similar. All drivers are trained in pairs sequentially to obtain the similarity among all drivers, thus forming a similarity matrix.

4. The vehicle assisted driving method based on personalized driver fatigue detection according to claim 1, characterized in that, In step 2, the optimal number of clusters is determined using the similarity matrix and the Gap statistic algorithm based on statistical inference. , the optimal number of clusters The number of cluster centers in the K-Means clustering algorithm divides drivers into... The class, its steps are as follows: Step 2.

1. From Randomly selected from drivers These objects serve as the initial cluster centers. ; Step 2.

2. Calculate for each sample separately. To each cluster center Euclidean distance The formula is as follows: ； in For the first Cluster centers, ; take the sample Assign it to the nearest cluster, as shown in the following formula: ； in Indicates the first Each cluster contains a set of drivers; Indicates the first Cluster centers; Step 2.

3. After all samples have been assigned, recalculate the centers of each cluster using the following formula: ； Step 2.

4. Combine the cluster centers obtained in Step 2.3 with those from the previous iteration. Comparison of cluster centers; If the cluster centers change, proceed to step 2.2; otherwise, proceed to step 2.

5. Step 2.

5. When the cluster centers no longer change, stop and output the drivers included in each category; at this point, all drivers have been divided into groups based on their similarity. There are 1 category, numbered 0, 1, ..., m-1 respectively.

5. The vehicle assisted driving method based on personalized driver fatigue detection according to claim 1, characterized in that, In step 2, the Gap statistics algorithm determines the optimal number of clusters. The process is as follows: I. For each cluster number Based on the similarity feature vector of each driver, the sum of squared intra-cluster errors among these similarity feature vectors is calculated, which is the sum of squared distances from all points in the cluster to the centroid. The formula is expressed as follows: ； In the formula This represents the sum of squared errors within the cluster; Let represent the set of samples in the t-th cluster; Cluster The center of mass; Indicates sample to the center of mass Euclidean distance, ; II. Generate 5 reference datasets from random data that follow the same distribution as the actual dataset; III. For each The value is used to calculate the sum of squared errors within each cluster for each reference dataset; IV. Calculate each The gap statistic of the value is expressed by the following formula: ； In the formula For each Gap statistic of value Indicates the actual dataset in Sum of squared intra-cluster errors for each cluster; Indicates the first A reference dataset in Sum of squared intra-cluster errors for each cluster; V. Choose the option that maximizes the Gap statistic. The value, as the optimal number of clusters. ; Set the maximum number of candidate clusters. N represents the number of drivers. The formula is expressed as follows: 。 6. The vehicle assisted driving method based on personalized driver fatigue detection according to claim 1, characterized in that, Step 3 specifically involves: Step 3.

1. For each driver group segmented in Step 2, construct an ensemble model consisting of three base classifiers: a logistic regression classifier, an XGBoost classifier, and a random forest classifier. The specific weight allocation of the three base classifiers is dynamically determined by the weight search mechanism in step 3.2; Step 3.

2. For each category of ensemble model, a weight search mechanism based on Bayesian optimization design is used to automatically search and optimize the weights of the logistic regression classifier, XGBoost classifier, and random forest classifier. Step 3.

3. For each category of ensemble model, after multiple dynamic search base classifier weight iteration experiments, extract the base classifier weight combination that achieves the highest accuracy on the validation set in the experiment, and use it as the final weight combination; Step 3.

4. Retrain the ensemble model consisting of the three base classifiers with their final weights by inputting all driving data of the drivers in this category into the ensemble model to obtain the specific fatigue detection model for this category.

7. The vehicle assisted driving method based on personalized driver fatigue detection according to claim 6, characterized in that, Step 3.2 specifically involves: Step 3.2.

1. Dynamically explore the three-dimensional weight space using the Optuna framework, and perform normalization processing on the weights generated in each iteration, where each iteration is defined as a complete base classifier weight search process using the Optuna framework; The objective function is defined as maximizing the accuracy on the validation set, and the three-dimensional weight space is dynamically explored using the Optuna framework. For the weights found in each iteration Perform the following normalization process: ； in This represents the weight of the i-th base classifier generated by Bayesian optimization; Indicates the effective weight; Step 3.2.

2. The predicted probabilities of the driver's driving state from the logistic regression classifier, XGBoost classifier, and random forest classifier are weighted and averaged to obtain the ensemble model's predictions for the class. Overall confidence level The calculation formula is as follows: ； in This indicates the driver's driving status, i.e., the category. The feature vectors representing the normal driving and fatigue driving data of the drivers in the validation set; This indicates that the base classifier predicts the category to belong to. The probability of; This indicates that the vehicle is in a normal driving state. This indicates that the driver is in a state of fatigue. Step 3.2.

3. Divide the drivers in this category into two parts, with all driving data of one driver serving as the validation set for the model and all driving data of the remaining drivers serving as the training set for the model; The training set and validation set are input into the ensemble model constructed using a weight search mechanism based on Bayesian optimization for training, and the prediction accuracy index of the ensemble model constructed under the weight combination of the base classifiers in this iteration is obtained. The driving state with the highest overall confidence calculated in step 3.2.2 of the ensemble model built under the weight combination of the base classifiers in this iteration will be used as the driving state prediction result of the model in this iteration.

8. The vehicle assisted driving method based on personalized driver fatigue detection according to claim 1, characterized in that, In step 4, the XGBoost multi-classification model with probability-weighted voting is used to determine the driver category based on the collected normal driving data of the new drivers. The specific determination process is as follows: Step 4.

1. Based on the driver clustering results obtained in Step 2, assign a label corresponding to the clustering result to each normal driving data sample of each known driver, i.e. , Number of categories; Step 4.

2. Use the normal driving data of drivers of each category as the training set of the model, and the category label corresponding to each sample as the test set, input them into the XGBoost multi-class classification model for training. Multi-class classification model; Step 4.

3. Input the normal driving data of the new driver into the XGBoost multi-class classification model as the validation set. The XGBoost multi-class classification model outputs a probability vector of each sample belonging to each category through training. The calculation formula is as follows: ； In the formula This represents the number of samples of new drivers driving normally. Indicates the first The sample belongs to the first The probability of a class; Step 4.

4. Based on the probability vector Calculate cumulative probability By summing the probability vectors column-wise, we obtain the probability that all normal driving data samples belong to each category, thus obtaining the probability of all new driver samples. The sum; the calculation formula is as follows: ； Step 4.

5. Determine the category with the highest cumulative probability for the new driver and output it.

9. The vehicle assisted driving method based on personalized driver fatigue detection according to claim 1, characterized in that, Step 5 specifically involves: For new drivers to be fatigued, physiological signal data and driving behavior data of the new driver under normal driving conditions are collected simultaneously using a physiological wristband and a driving simulator, and then preprocessed and feature extracted. The physiological signal data and driving behavior data of the new driver obtained after preprocessing and feature extraction are input into the fatigue detection model corresponding to the category of the new driver that was pre-trained in step 3. The fatigue detection model outputs the new driver's driving status to determine whether the new driver is fatigued.

10. The vehicle assisted driving method based on personalized driver fatigue detection according to claim 1, characterized in that, In step 6, if the most critical feature affecting driver fatigue detection in the new driver's category is driving behavior characteristics, the vehicle sends a request through the visual and auditory channels to whether to activate the lane keeping assist system. If the most critical characteristic affecting driver fatigue detection in the new driver's category is heart rate, the vehicle will display a selection interface on the in-vehicle display asking whether to take over or drive to a service area. If the most critical characteristic affecting driver fatigue detection in the new driver's category is skin conductance, the vehicle control system issues a command to drive the window lifting mechanism to lower the window to help the driver regain consciousness. If the most critical characteristic affecting driver fatigue detection in the new driver's category is wrist movement, the vehicle will send a request through visual and auditory channels to whether to activate adaptive cruise control.

Images