Multi-feature fusion driver behavior detection model construction method and detection method
By constructing a multi-feature fusion driver behavior detection model and using deep learning networks to detect driver fatigue and distraction, the problem of incomplete detection in existing technologies is solved, and high-precision and real-time detection results are achieved.
Patent Information
- Application Number
- CN202411116256.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-14
- Publication Date
- 2024-11-05
AI Technical Summary
Existing technologies are insufficient to comprehensively and accurately detect driver fatigue and distraction, and their detection accuracy and real-time performance are poor, making them difficult to implement in practice.
A multi-feature fusion driver behavior detection model is constructed. By combining target detection, whole-body pose estimation, and head pose estimation with a deep learning network, multiple fatigue and distraction features are integrated, and driver behavior detection is performed using a deep learning network.
It achieves high-precision and real-time driver behavior detection, can comprehensively detect fatigue and distraction behaviors, reduces costs, and has model scalability to adapt to actual production needs.
Smart Images

Figure CN118898831A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of driver behavior detection technology, and in particular to a method for constructing a driver behavior detection model and a detection method based on multi-feature fusion. Specifically, a method for intelligently monitoring and detecting the driver's behavior during driving is designed based on a target detection algorithm, a whole body posture estimation algorithm, and a head posture estimation algorithm. Background Art
[0002] With the continuous development of my country's economy, the number of motor vehicles has increased year by year. According to relevant data, the occurrence of road traffic accidents in my country is mostly attributed to the human factors of drivers. The research on driver behavior detection belongs to the field of intelligent transportation and is a key technology of intelligent assisted driving. This technology avoids traffic accidents by detecting driver behavior and issuing warnings, which has important application value and social significance.
[0003] A considerable proportion of traffic accidents are caused by improper behaviors of drivers, including fatigue behaviors such as closing eyes for a long time, yawning, nodding, and distraction behaviors such as making phone calls, eating, smoking, and looking around.
[0004] At present, many countries at home and abroad have developed related technologies in this field, but they all have their own limitations to varying degrees. Most methods are either only aimed at detecting driver fatigue driving behavior, or only at detecting driver distracted driving behavior, and few technologies can comprehensively detect driver behavior. In addition, many methods that can simultaneously detect fatigue driving and distracted driving have their own limitations, and have shortcomings such as low detection accuracy and poor real-time performance, which also makes them difficult to implement. Summary of the invention
[0005] In order to address the deficiencies in the prior art, the present application proposes a multi-feature fusion driver behavior detection model construction method and a detection method. Based on a variety of fatigue features, distraction features and timing information, the system detects the driver's behavior in video clips in driving scenes, solving the problems of low accuracy, slow speed and incomplete detection content in the prior art.
[0006] To achieve the above object, the present invention provides the following technical solutions:
[0007] A method for constructing a driver behavior detection model by fusion of multiple features, comprising the following steps;
[0008] S1, dataset preparation and processing;
[0009] S1.1. Object detection uses a self-made COCO format dataset. The original images of the dataset are obtained from the vehicle camera. At least five types of labels such as people, heads, cigarettes, water cups and mobile phones are marked through the dataset annotation platform Roboflow;
[0010] S1.2, whole body key point estimation uses the public COCO-WholeBody dataset;
[0011] S1.3, head pose estimation uses the public 300W-LP dataset;
[0012] S2. Build a deep learning network, including three parts: target detection part, whole body key point estimation part and head posture estimation part, where the whole body posture estimation part network and the head posture estimation part network are connected in parallel after the target detection part network;
[0013] S2.1. Build the object detection network based on YOLOv10. The input is the original RGB image, and the output is the coordinates and confidence of various labels.
[0014] S2.2. Construct the whole body key point estimation network based on RTMPose, input the RGB image of the driver area, and output the coordinates and confidence of the driver's whole body key points;
[0015] S2.3, construct the head posture estimation network based on 6DRepNet, input the RGB image of the driver's head area, and obtain the driver's head Euler angle information;
[0016] S3. Training a deep learning network, including: using the self-made COCO format dataset to train a YOLOv10-based target detection network, using the COCO-WholeBody dataset to train a RTMpose-based full-body pose estimation network, and using the 300W-LP dataset to train a 6DRepNet-based head pose estimation network.
[0017] To achieve the above object, the present invention also provides the following technical solutions:
[0018] A multi-feature fusion driver behavior detection method comprises the following steps:
[0019] S1. Input the video clip to be detected into the above-mentioned multi-feature fusion driver behavior detection model, obtain the detection results of target detection, whole body posture estimation, and head posture estimation, and obtain the required five types of driving behavior features from the detection results: eye features, mouth features, head features, hand features, and distracting object features. Eye features are eye key point information, mouth features are mouth key point information, head features include head position information and head posture information, hand features include hand key point information, and distracting object features include position information of distracting objects such as cigarettes, water cups, and mobile phones;
[0020] S2. Get single-frame actions through the above five types of driving behavior characteristics. The single-frame actions include: eyes open / eyes closed, mouth open / mouth closed, head down / looking straight, sideways / looking straight, smoking / not smoking, drinking water / not drinking water, using mobile phone / not using mobile phone, other distracting behaviors / normal;
[0021] S3. Combine the single-frame action with the timing information to calculate the fatigue index and distraction index, and determine the fatigue / distraction state through the index. Fatigue detection includes four indicators: the percentage of closed eye time P c , continuous closing time P c , Yawning duration T y and nodding frequency F n , distraction detection includes two indicators: the duration of distraction behavior T d and the frequency of distraction behavior F d ;
[0022] S4. Determine the degree of fatigue and distraction based on the above index results;
[0023] S5. Add a preprocessing stage before starting, and calculate the Euler angle value of the driver's head in a normal state through the target detection algorithm and the head posture estimation algorithm.
[0024] Furthermore, the step S2 includes the following sub-steps:
[0025] S2.1. Given the eye closing threshold EAR and the yawning threshold MAR, calculate the driver's eye aspect ratio E according to the whole body posture estimation result. ar and the aspect ratio of the mouth M ar , if E ar ≤EAR, the driver is deemed to have closed his eyes. ar ≤MAR is considered as the driver yawning;
[0026] S2.2. Given a forward-looking threshold YAW and a head-down threshold PITCH, based on the head posture estimation result, if the yaw angle Yaw ≥ YAW, it is considered that the driver is not looking forward; if the pitch angle Pitch ≥ PITCH, it is considered that the driver is looking down;
[0027] S2.3. Given the overlap coefficient CX, according to the target detection results and the whole body key point estimation results, the hand frame is fitted through the left / right hand key points, and the overlap area Co of the driver's hand frame and the distracting object area and the distracting object area Cd are calculated. If Co≥CD×CX, it is considered that the driver has this distracting behavior and skip step S2.4;
[0028] S2.4. Calculate the overlapping area Ch of the driver’s hand frame and head area and the head area Ch based on the key point estimation results. If Ch≥Ch×CX, it is considered that the driver’s left hand / right hand has left the steering wheel and there are other distracting behaviors.
[0029] Furthermore, in step S2.1, the pseudo code for calculating the eye aspect ratio is as follows:
[0030]
[0031] The aspect ratio M of the mouth ar It is expressed as:
[0032]
[0033] Among them, the mouth key points are the inner circle key points of the mouth, y 0 ,y 1 ...etc. represent the ordinates of the 0th, 1st, ...key points, x 0 、x 1 …etc. represent the horizontal coordinates of the key points No. 0, 1,…
[0034] Furthermore, in step S3, the percentage of eye closing time P c It is expressed as:
[0035]
[0036] Among them, n eye Indicates the total number of closed-eye frames in the video frame, N eye Indicates the total number of video frames;
[0037] Yawning duration T y It is expressed as:
[0038] T y =T end2 -T start2
[0039] Among them, T start2 Indicates the time when a yawn starts in the video frame, T end2 Indicates the time when the yawn stops in the video frame;
[0040] Nodding frequency F n It is expressed as:
[0041] F n =N1 / T end3 -T start3
[0042] Among them, N start3 Indicates the sequence number of the first nod in the video frame, N end3 Indicates the sequence number of the last stop nod in the video frame, T 0 Indicates the time interval for collecting each frame of image, N 1 Indicates the number of nods;
[0043] Distracting behavior duration T d It is expressed as:
[0044] T d =T end4 -T start4
[0045] Among them, T start4 Indicates the time when a distracting behavior starts in the video frame, T end4 Indicates the time in the video frame when the distracting behavior ends;
[0046] Frequency of distraction F d It is expressed as:
[0047] F d =N2 / T end5 -T start5
[0048] Among them, N start5 Indicates the time when the distracting behavior first starts in the video frame, N end5 Indicates the time when the distracting behavior last stopped in the video frame, N 2 Indicates the number of distracting behaviors.
[0049] Furthermore, in step S4, the fatigue degree is determined by combining multiple fatigue characteristics:
[0050]
[0051] Among them, P c is the percentage of eye closing time, T c is the single continuous eye closure time, T y is the duration of a single yawn, F n is the nodding frequency;
[0052] Combining multiple distraction characteristics to determine the degree of distraction:
[0053]
[0054] The beneficial effects of the present invention are as follows:
[0055] 1. Compared with traditional methods, the present invention builds a network based on deep learning, which has the advantages of high accuracy, good real-time performance, good robustness, and strong anti-interference ability.
[0056] 2. The driver behavior detection model proposed in this method is built in a modular way. The coupling between the parts is low, the data set is easy to obtain, the cost is reduced, and the iteration and upgrade of each module are convenient. At the same time, the model has strong scalability, and the detection of distracting behaviors such as eating, bending over to pick up objects, etc. can be added to the model, which meets the need for continuous iteration of technology in actual production.
[0057] 3. Compared with traditional detection methods, the driver behavior detection method proposed in the present invention integrates multiple features in fatigue detection and distraction detection. It can detect fatigue behaviors such as closing eyes, yawning, nodding and dozing off, as well as dangerous behaviors such as smoking, making phone calls and other common distracting behaviors. It has a more comprehensive detection field, fewer misjudgments, stronger versatility and higher accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] Figure 1 This is a diagram of the driver behavior detection model with multi-feature fusion according to the present invention.
[0059] Figure 2 This is a driver behavior detection method based on multi-feature fusion of the present invention.
[0060] Figure 3 Schematic diagram of key points of the eye of the present invention.
[0061] Figure 4 Schematic diagram of the key points of the mouth of the present invention, wherein p12-p19 are selected key points.
[0062] Figure 5 It is a schematic diagram of the overall process of the present invention. DETAILED DESCRIPTION
[0063] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0064] Embodiment 1:
[0065] A method for constructing a driver behavior detection model by fusion of multiple features, comprising the following steps;
[0066] S1, dataset preparation and processing;
[0067] S1.1. Object detection uses a self-made COCO format dataset. Specifically:
[0068] The original images of the dataset are obtained from the on-board camera. Videos are taken in groups according to different vehicle conditions, camera angles, lighting conditions, drivers, driving behaviors, etc., and then images are captured from the videos. As many combinations as possible are involved, and the final number of images is more than 5,000. This example does not limit the number of images. Generally speaking, clear images should be selected for the same situation, repeated images should not be too many, and the types of images should be as rich as possible to enhance the robustness of the model.
[0069] At least five types of labels, including people, heads, cigarettes, cups, and mobile phones, are marked through the data set annotation platform Roboflow. When annotating, objects that are blurry, indistinguishable to the naked eye, or mostly obscured can be left unlabeled, while clear objects should be labeled as much as possible.
[0070] The target detection annotation format of COCO refers to the format of the COCO dataset official website.
[0071] S1.2, whole body key point estimation uses the public COCO-WholeBody dataset;
[0072] S1.3, head pose estimation uses the public 300W-LP dataset;
[0073] S2, such as Figure 1 As shown, a deep learning network is constructed, which includes three parts: a target detection part, a whole body key point estimation part and a head posture estimation part, wherein the whole body key point estimation part network and the head posture estimation part network are connected in parallel after the target detection part network;
[0074] S2.1. In this example, the target detection network is constructed based on YOLOv10. The input is the original RGB image, and the output is the coordinates and confidence of each type of label.
[0075] S2.2. In this example, the whole body key point estimation network is constructed based on RTMPose. Based on the original RGB image of S2.1, the person with the largest area is regarded as the driver according to the target detection result, and the RGB image of the driver's part is cropped as input. The output is the coordinates and confidence of the driver's whole body key points;
[0076] S2.3. In this example, the head posture estimation network is constructed based on 6DRepNet. Based on the original RGB image of S2.1, according to the target detection results, the person with the largest area is regarded as the driver, and the head is regarded as the driver's head. The RGB image of the driver's head area is cropped as input. The output is the driver's head Euler angle information;
[0077] S3, training a deep learning network, including: using the self-made COCO format dataset to train a YOLOv10-based target detection network, using the COCO-WholeBody dataset to train a RTMpose-based full-body pose estimation network, and using the 300W-LP dataset to train a 6DRepNet-based head pose estimation network;
[0078] The Adam adaptive learning rate algorithm is used to train each part of the model independently. The whole body key point estimation network is trained for 210 rounds, and the weights are saved every 10 rounds. After the training is completed, the weight with the highest accuracy is selected; the object detection network and the head pose estimation network do not set an upper limit on the number of training rounds. When the network loss tends to be stable, the Adam adaptive learning is stopped and the weight with the highest accuracy is selected;
[0079] Embodiment 2:
[0080] A multi-feature fusion driver behavior detection method comprises the following steps:
[0081] S1. Input the video clip to be detected into the above-mentioned multi-feature fusion driver behavior detection model, obtain the detection results of target detection, whole body posture estimation, and head posture estimation, and obtain the required five types of driving behavior features from the detection results: eye features, mouth features, head features, hand features, and distracting object features. Eye features are eye key point information, mouth features are mouth key point information, head features include head position information and head posture information, hand features include hand key point information, and distracting object features include position information of distracting objects such as cigarettes, water cups, and mobile phones;
[0082] S2. Get single-frame actions through the above five types of driving behavior characteristics. The single-frame actions include: eyes open / eyes closed, mouth open / mouth closed, head down / looking straight, sideways / looking straight, smoking / not smoking, drinking water / not drinking water, using mobile phone / not using mobile phone, other distracting behaviors / normal;
[0083] S2.1. Given the eye closing threshold EAR and the yawning threshold MAR, calculate the driver's eye aspect ratio E according to the whole body posture estimation result. ar and the aspect ratio of the mouth M ar , if E ar ≤EAR, the driver is deemed to have closed his eyes. ar ≤MAR is considered as the driver yawning. In this example, EAR is 0.2 and MAR is 0.5;
[0084] S2.2. Given the forward-looking threshold YAW and the head-down threshold PITCH, according to the head posture estimation result, if the yaw angle Yaw ≥ YAW, it is considered that the driver is not looking forward, and if the pitch angle Pitch ≥ PITCH, it is considered that the driver is looking down. In this example, YAW is 30° and PITCH is 15°;
[0085] S2.3. Given the overlap coefficient CX, according to the target detection results and the whole body key point estimation results, the hand frame is fitted through the left / right hand key points, and the overlap area Co of the driver's hand frame and the distracting object area and the distracting object area Cd are calculated. If Co≥Cd×CX, it is considered that the driver has this distracting behavior and skip step S2.4. In this example, CX is 0.2;
[0086] Taking the right hand as an example, obtain the estimated information of all key points of the right hand and find the maximum and minimum values of the horizontal and vertical coordinates X min , Y min , X max , Y max , we can get the fitted hand frame coordinates (X min , Y min ) and (X max , Y max ).
[0087] S2.4. Calculate the overlapping area Ch of the driver's hand frame and head area and the head area Ch based on the key point estimation results. If Ch≥Ch×CX, it is considered that the driver's left hand / right hand has left the steering wheel and there are other distracting behaviors.
[0088] S3. Combine the single-frame action with the timing information to calculate the fatigue index and distraction index, and determine the fatigue / distraction state through the index. Fatigue detection includes four indicators: the percentage of closed eye time P c , continuous closing time T c , Yawning duration T y and nodding frequency F n , distraction detection includes two indicators: the duration of distraction behavior T d and the frequency of distraction behavior F d ;
[0089] S4. Given thresholds PC, TC, TY, FN, TD, and FD, determine the degree of fatigue and distraction based on the above index results;
[0090]
[0091] Among them, P c is the percentage of eye closing time, T c is the single continuous eye closure time, T yis the duration of a single yawn, F n is the nodding frequency. In this example, PC is 0.5, TC is 2 seconds, TY is 2.5 seconds, FN is 0.1, TD is 2.5 seconds, FD is 0.1, K1 is 2, and K2 is 2;
[0092] In step S2.1, the pseudo code for calculating the eye aspect ratio is as follows:
[0093]
[0094] The aspect ratio M of the mouth ar It is expressed as:
[0095]
[0096] Among them, the mouth key points are the inner circle key points of the mouth, y 0 ,y 1 ...etc. represent the ordinates of the 0th, 1st, ...key points, x 0 、x 1 ...etc. represent the horizontal coordinates of the key points No. 0, 1, ...;
[0097] In step S3, the percentage of eye closing time P c It is expressed as:
[0098]
[0099] Among them, n eye Indicates the total number of closed-eye frames in the video frame, N eye Indicates the total number of video frames;
[0100] Yawning duration T y It is expressed as:
[0101] T y =T end2 -T start2
[0102] Among them, T start2 Indicates the time when a yawn starts in the video frame, T end2 Indicates the time when the yawn stops in the video frame;
[0103] Nodding frequency F n It is expressed as:
[0104] F n =N 1 / T end3 -T start3
[0105] Among them, N start3 Indicates the sequence number of the first nod in the video frame, N end3Indicates the sequence number of the last stop nod in the video frame, T 0 Indicates the time interval for collecting each frame of image, N 1 Indicates the number of nods;
[0106] Distracting behavior duration T d It is expressed as:
[0107] T d =T end4 -T start4
[0108] Among them, T start4 Indicates the time when a distracting behavior starts in the video frame, T end4 Indicates the time in the video frame when the distracting behavior ends;
[0109] Frequency of distraction F d It is expressed as:
[0110] F d =N 2 / T end5 -T start5
[0111] Among them, N start5 Indicates the time when the distracting behavior first starts in the video frame, N end5 Indicates the time when the distracting behavior last stopped in the video frame, N 2 Indicates the number of distracting behaviors;
[0112] The above embodiments are only used to illustrate the design ideas and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design ideas disclosed by the present invention are within the protection scope of the present invention.
Claims
1. A method for constructing a driver behavior detection model by fusion of multiple features, characterized in that: The following steps are involved: S1, dataset preparation and processing; S1.
1. Object detection uses a self-made COCO format dataset. The original images of the dataset are obtained from the vehicle camera. At least five types of labels such as people, heads, cigarettes, water cups and mobile phones are marked through the dataset annotation platform Roboflow; S1.2, whole body key point estimation uses the public COCO-WholeBody dataset; S1.3, head pose estimation uses the public 300W-LP dataset; S2. Build a deep learning network, including three parts: target detection part, whole body key point estimation part and head posture estimation part, where the whole body posture estimation part network and the head posture estimation part network are connected in parallel after the target detection part network; S2.
1. Build the object detection network based on YOLOv10. The input is the original RGB image, and the output is the coordinates and confidence of various labels. S2.
2. Construct the whole body key point estimation network based on RTMPose, input the RGB image of the driver area, and output the coordinates and confidence of the driver's whole body key points; S2.3, construct the head posture estimation network based on 6DRepNet, input the RGB image of the driver's head area, and obtain the driver's head Euler angle information; S3. Training a deep learning network, including: using the self-made COCO format dataset to train a YOLOv10-based target detection network, using the COCO-WholeBody dataset to train a RTMpose-based full-body pose estimation network, and using the 300W-LP dataset to train a 6DRepNet-based head pose estimation network.
2. A driver behavior detection method based on multi-feature fusion, characterized in that: The following steps are involved: S1. Input the video clip to be detected into the above-mentioned multi-feature fusion driver behavior detection model, obtain the detection results of target detection, whole body posture estimation, and head posture estimation, and obtain the required five types of driving behavior features from the detection results: eye features, mouth features, head features, hand features, and distracting object features. Eye features are eye key point information, mouth features are mouth key point information, head features include head position information and head posture information, hand features include hand key point information, and distracting object features include position information of distracting objects such as cigarettes, water cups, and mobile phones; S2. Get single-frame actions through the above five types of driving behavior characteristics. The single-frame actions include: eyes open / eyes closed, mouth open / mouth closed, head down / looking straight, sideways / looking straight, smoking / not smoking, drinking water / not drinking water, using mobile phone / not using mobile phone, other distracting behaviors / normal; S3. Combine the single-frame action with the timing information to calculate the fatigue index and distraction index, and determine the fatigue / distraction state through the index. Fatigue detection includes four indicators: the percentage of closed eye time P c , continuous closing time T c , Yawning duration T y and nodding frequency F n , distraction detection includes two indicators: the duration of distraction behavior T d and the frequency of distraction behavior F d ; S4. Determine the degree of fatigue and distraction based on the above index results; S5. Add a preprocessing stage before starting, and calculate the Euler angle value of the driver's head in a normal state through the target detection algorithm and the head posture estimation algorithm.
3. The driver behavior detection method based on multi-feature fusion according to claim 2 is characterized in that: The step S2 comprises the following sub-steps: S2.
1. Given the eye closing threshold EAR and the yawning threshold MAR, calculate the driver's eye aspect ratio E according to the whole body posture estimation result. ar and the aspect ratio of the mouth M ar , if E ar ≤EAR, the driver is deemed to have closed his eyes. ar ≤MAR is considered as the driver yawning; S2.
2. Given a forward-looking threshold YAW and a head-down threshold PITCH, based on the head posture estimation result, if the yaw angle Yaw ≥ YAW, it is considered that the driver is not looking forward; if the pitch angle Pitch ≥ PITCH, it is considered that the driver is looking down; S2.
3. Given the overlap coefficient CX, according to the target detection results and the whole body key point estimation results, the hand frame is fitted through the left / right hand key points, and the overlap area Co of the driver's hand frame and the distracting object area and the distracting object area Cd are calculated. If Co≥Cd×CX, it is considered that the driver has this distracting behavior and skip step S2.4; S2.
4. Calculate the overlapping area Ch of the driver’s hand frame and head area and the head area Ch based on the key point estimation results. If Ch≥Ch×CX, it is considered that the driver’s left hand / right hand has left the steering wheel and there are other distracting behaviors.
4. The driver behavior detection method based on multi-feature fusion according to claim 3 is characterized in that: In step S2.1, the pseudo code for calculating the eye aspect ratio is as follows: The aspect ratio M of the mouth ar It is expressed as: Among them, the mouth key points are the inner circle key points of the mouth, y0, y1... etc. represent the vertical coordinates of the 0th, 1st... key points, and x0, x1... etc. represent the horizontal coordinates of the 0th, 1st... key points.
5. The driver behavior detection method of multi-feature fusion according to claim 2 is characterized by: In step S3, the percentage of eye closing time P c It is expressed as: Among them, n eye Indicates the total number of closed-eye frames in the video frame, N eye Indicates the total number of video frames; Yawning duration T y It is expressed as: T y =T end2 -T start2 Among them, T start2 Indicates the time when a yawn starts in the video frame, T end2 Indicates the time when the yawn stops in the video frame; Nodding frequency F n It is expressed as: F n =N1 / T end3 -T start3 Among them, N start3 Indicates the sequence number of the first nod in the video frame, N end3 Indicates the sequence number of the last nodding stop in the video frame, T0 indicates the time interval for collecting each frame of image, and N1 indicates the number of nodding; Distracting behavior duration T d It is expressed as: T d =T end4 -T start4 Among them, T start4 Indicates the time when a distracting behavior starts in the video frame, T end4 Indicates the time in the video frame when the distracting behavior ends; Frequency of distraction F d It is expressed as: F d =N2 / T end5 -T start5 Among them, N start5 Indicates the time when the distracting behavior first starts in the video frame, N end5 It indicates the time when the distracting behavior was last stopped in the video frame, and N2 indicates the number of distracting behaviors.
6. The driver behavior detection method of multi-feature fusion according to claim 2 is characterized by: In step S4, the fatigue degree is determined by combining multiple fatigue characteristics: Among them, P c is the percentage of eye closing time, T c is the single continuous eye closure time, T y is the duration of a single yawn, F n is the nodding frequency; Combining multiple distraction characteristics to determine the degree of distraction:
Citation Information
Cited By
User state monitoring method and device
CN121963095A
