Layered dynamic feedback mechanism-based abnormal driving behavior monitoring method for automobile data recorder

Abstract

The invention discloses an abnormal driving behavior monitoring method for an automobile data recorder based on a hierarchical dynamic feedback mechanism. The abnormal driving behavior monitoring method comprises the following steps: S1, acquiring a facial image, an automobile speed and acceleration data of a driver; s2, identifying a driving abnormity type, and analyzing facial expressions, eye movement features and vehicle dynamic data; s3, selecting a task specific network architecture based on the exception type; s4, the hybrid neural architecture searches and optimizes the network architecture, and meta-learning training is carried out; s5, performing layered dynamic feedback according to the severity of the driver behavior; s6, dynamically adjusting the feedback intensity and mode according to the response of the driver; and S7, performing continuous optimization and real-time adjustment, and providing personalized driving monitoring. Through the intelligent monitoring system based on a layered dynamic feedback mechanism, abnormal behaviors of a driver are recognized and intervened in real time, the driving safety is improved, and personalized driving behavior monitoring is provided.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and particularly to a method for monitoring abnormal driving behaviors of a driving recorder based on a hierarchical dynamic feedback mechanism. Background Art

[0002] With the rapid development of technology, intelligent transportation systems and in-vehicle monitoring technologies have become important tools for enhancing road safety and reducing traffic accidents. In the field of driver behavior monitoring, existing technologies mainly rely on vehicle sensors, cameras, and some physiological and behavioral data of drivers to detect and evaluate driver behaviors, and then provide safety reminders through warning and pre-warning systems. The popularization of in-vehicle recorders and intelligent driving assistance systems has enabled the wide application of this technology, especially in the monitoring of behaviors such as fatigue driving, sudden braking, and distracted driving. However, existing technologies still face many challenges and deficiencies, affecting their effectiveness and accuracy in practical applications.

[0003] In existing technologies, many driver behavior monitoring systems mainly rely on hardware devices such as in-vehicle cameras, sensors, and vehicle accelerometers to obtain driver behavior data. However, the capabilities and adaptabilities of these devices have certain limitations. First, traditional cameras and sensors often have difficulty in coping with complex driving environments and changing lighting conditions when obtaining driver facial features, body movements, and driving behaviors. For example, changes in light inside the vehicle or weather conditions affect the capture effect of the camera, resulting in a significant reduction in the accuracy of monitoring results. In addition, existing hardware devices can usually only detect some obvious abnormal driving behaviors such as sudden braking or excessive vehicle speed, but it is difficult to accurately judge some subtle signs of fatigue or minor changes in distracted driving.

[0004] In addition, the monitoring algorithms in many existing systems rely on traditional computer vision methods and rule engines for data processing. These algorithms often have difficulty in adapting to complex driving behaviors and diverse driving environments, resulting in poor accuracy and robustness in processing driver behavior data. For example, traditional image processing algorithms can only extract some relatively simple facial features and are difficult to judge the fatigue state of the driver. Even in the monitoring of fatigue driving, existing technologies mostly rely on simple rules or threshold judgments based on single features, ignoring the comprehensive influence of various factors such as driver emotions, physical conditions, and road environments.

[0005] Another drawback in the prior art lies in the singularity of the feedback mechanism. Most existing driving behavior monitoring systems can only provide simple warnings and reminders, such as voice prompts or vibration alerts, lacking a multi-level feedback mechanism for different driving anomalies. Especially when the driver fails to respond to the warning in a timely manner, the existing systems often lack sufficient mechanisms to increase the intensity of the feedback or change the feedback method. This single feedback method reduces the driver's attention to the warning to a certain extent, especially in the face of serious driving anomalies, and cannot effectively enhance the warning effect.

[0006] In addition, most of the feedback mechanisms in the prior art are static, that is, they provide feedback according to a certain fixed rule or threshold, lacking the ability to dynamically adjust according to the driver's personalized behavior and historical driving data. Drivers' behaviors have obvious individual differences, and there are significant differences in reaction speed, fatigue perception, and driving habits among different drivers. Therefore, without personalized adaptation, the feedback of traditional systems often cannot achieve the same effect among different drivers. This "one-size-fits-all" feedback method not only reduces the adaptability of the system but also increases interference to the driver and cannot provide truly effective safety protection.

[0007] The deficiencies in monitoring accuracy and feedback mechanism lead to poor application effects of the prior art in various complex scenarios. For example, when the driver's behavior gradually transitions from slightly abnormal to severely abnormal, the feedback mechanism of the existing system cannot identify and gradually increase the intensity of the feedback in a timely manner, which results in some severe driving anomalies not being intervened in a timely and effective manner, creating potential safety hazards.

[0008] Therefore, how to provide a method for monitoring driving anomalies of a driving recorder based on a hierarchical dynamic feedback mechanism is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0009] An object of the present invention is to propose a method for monitoring driving anomalies of a driving recorder based on a hierarchical dynamic feedback mechanism. The present invention makes full use of technologies such as deep learning, vision transformers, adaptive meta-learning, and graph convolutional temporal networks, and details the algorithms for intelligently implementing the monitoring of driver fatigue driving, sudden braking, and distracted driving behaviors. This method dynamically adjusts the feedback intensity and feedback method according to the driver's behavior score to ensure that appropriate feedback can be given in a timely manner when the driver exhibits abnormal behaviors, and has the advantages of high precision, high adaptability, and personalized feedback.

[0010] The method for monitoring driving anomalies of a driving recorder based on a hierarchical dynamic feedback mechanism according to an embodiment of the present invention includes the following steps:

[0011] S1. Collect the driver's facial images and vehicle dynamic data through the driving recorder;

[0012] S2. Identify the types of abnormal driving, analyze the driver's facial images and vehicle dynamic data, and predict the driver's distracted driving behavior;

[0013] S3. Based on the identified types of abnormal driving, select a task-specific network architecture;

[0014] S4. According to the current task, optimize the network architecture through hybrid neural architecture search and conduct meta-learning training;

[0015] S5. According to the severity of the driver's behavior, conduct hierarchical dynamic feedback. The first layer of feedback is voice prompt, the second layer is vibration prompt, and the third layer is strong sound warning and visual warning;

[0016] S6. According to the driver's reaction to the feedback, dynamically adjust the feedback intensity and method, monitor the driver's behavior in real time, and automatically switch tasks and adjust the feedback strategy when an abnormality occurs;

[0017] S7. Through continuous optimization and real-time adjustment, the system provides personalized monitoring of driving behavior according to different drivers and driving environments.

[0018] Optionally, the specific steps of S2 include:

[0019] S21. Use the facial images and vehicle dynamic data collected by the driving recorder to perform preliminary feature extraction;

[0020] S22. Based on the features of the facial images, use a vision transformer to analyze the driver's facial expressions and eye movement states, and calculate the fatigue index FI:

[0021]

[0022] where E i is the eye closure degree of the i-th frame image, E mean is the average value of the closure degree, E std is the standard deviation of the closure degree, A i is the facial movement amplitude of the i-th frame, A mean is the average value of the movement amplitude, γ1 is the non-linear weighting coefficient of the movement amplitude, α is the exponential weight of the movement amplitude, δ i is the interaction influence coefficient of the closure degree and the movement amplitude, X is the number of frames of the collected images, exp() is the exponential function, and when the fatigue index FI exceeds the preset threshold, it is determined as fatigue driving;

[0023] S23. According to the vehicle's dynamic data, including vehicle speed change and acceleration change, use a temporal convolutional neural network to process the temporal data of acceleration and vehicle speed, and calculate the emergency braking probability EBS:

[0024]

[0025] Among them, Δv j represents the vehicle speed change at the j-th moment, and Δt j is the time interval of the vehicle speed change, γ2 is the weighted index of the vehicle speed change, β2 is the weighted index of the time interval, and a k is the acceleration at the k-th moment, α2 is the weighted index of the acceleration, γ3 is the non-linear activation function of the acceleration, ω1 is the interaction coefficient between the acceleration and the vehicle speed change, N is the number of samples, and Δv l represents the vehicle speed change at the l-th moment, and a l is the acceleration at the l-th moment. When the emergency braking probability EBS exceeds the set threshold, it is determined as an emergency braking behavior;

[0026] S24. Combine the driver's historical behavior data, use a sequence analysis model based on a long short-term memory network to predict the driver's distracted driving behavior, and calculate the distracted driving score DSS:

[0027]

[0028] Among them, x k is the distracted behavior feature at the k-th moment, w k is the weight coefficient, y m is the reaction time of the driver at the m-th moment, α m is the weight coefficient of the reaction time, z n is the driver's physiological signal at the n-th moment, β n is the weight coefficient of the physiological signal, σ is the activation function, K is the number of distracted behavior features, L is the number of reaction time features, and P is the number of physiological signal features. When the distracted driving score is greater than the set threshold, it is determined as a distracted driving behavior;

[0029] S25. Based on the analysis results, judge the abnormality degree of the driving behavior through the set threshold, and respectively determine it as severe fatigue driving, emergency braking or distracted driving, and then perform corresponding feedback.

[0030] Optionally, the S3 specifically includes:

[0031] S31. According to the identified abnormal behavior type, based on the fatigue index FI, the emergency braking probability EBS and the distracted driving score DSS, select a suitable network architecture for further analysis;

[0032] S32. When the fatigue index exceeds the preset threshold, select an adaptive visual recursive network architecture, and by combining a vision transformer and a recursive neural network, deeply analyze the temporal changes of the fatigue state and facial features, and adjust the influence degree of the facial features in real time through a weighted impulse function:

[0033]

[0034] where w i is the facial feature weighting coefficient at the i-th moment, and FI i is the fatigue index, Δt i is the time interval, α1 is the time decay coefficient, t is the current moment, and T min is the fatigue start time, N is the number of samples, and FDS t is the fatigue state score, and exp() is the exponential function;

[0035] S33. When the hard braking probability exceeds the set threshold, select the graph convolutional temporal network architecture. By introducing an event-driven graph convolutional model, analyze the spatio-temporal dynamic characteristics during hard braking, construct an event-driven graph through vehicle speed, acceleration, and spatio-temporal position information, and dynamically capture the hard braking risk index:

[0036]

[0037] where a j is the acceleration at the j-th moment, Δv j is the vehicle speed change, Δt j is the time interval, γ and β are adjustment parameters, d j is the vehicle spatial displacement, M is the number of sample data points, and EBS t is the hard braking risk index;

[0038] S34. When the distracted driving score exceeds the set threshold, select the spatio-temporal attention-autoregressive network architecture. By combining the spatio-temporal attention mechanism with the autoregressive model, comprehensively analyze the temporal changes in the driver's attention state and reaction time, introduce the spatio-temporal attention fusion mechanism, and combine physiological signals and behavioral data to calculate the attention state score ASR t :

[0039]

[0040] where x k is the distracted behavior feature, w k is the weight coefficient, y m is the reaction time, α m is the weighting coefficient, z p is the physiological signal feature, φ p is the weight coefficient of the physiological signal, γ k is the weighted adjustment coefficient, and K, L, and P are the numbers of behavioral features, reaction time features, and physiological signal features;

[0041] S35. According to the calculated fatigue state score, hard braking risk index, and attention state score, adjust the feedback strategy in real time through the adaptive feedback optimization mechanism.

[0042] Optionally, the S4 specifically includes:

[0043] S41: Optimize the adaptive visual recursive network architecture through hybrid neural architecture search and make adaptive adjustments based on the characteristics and behavioral data of fatigue driving;

[0044] S42. Optimize the graph convolutional temporal network architecture with meta-learning training, use the sudden braking risk score to optimize sudden braking detection, and evaluate changes in vehicle speed, acceleration, and spatial position in real time to determine the risk of sudden braking behavior:

[0045]

[0046] Among them, a j is the acceleration at the jth moment, v j is the vehicle speed change, Δt j is the time interval, λ q and ω are adjustment coefficients, dist j is the spatial distance change of the vehicle, ν is the weighted index, ABRI t represents the sudden braking risk score, N is the number of samples;

[0047] S43. Optimize the calculation method of distracted driving risk factors for the spatiotemporal attention-autoregressive network architecture, and adjust the network's response strength in real time through multimodal signal fusion:

[0048]

[0049] Among them, x k For behavioral characteristics, is the prediction feature, δ k is the weighting coefficient, p m is the physiological signal feature, τ m is a constant, φ m is the weighting coefficient, s is the feature difference weight, K and P are the number of behavioral features and physiological signal features respectively, DRF t represents the distracted driving risk factor, log2 is the logarithmic function, and abs is the absolute value operation;

[0050] S44: Through the optimized adaptive visual recurrent network architecture, graph convolutional temporal network architecture, and spatiotemporal attention-autoregressive network architecture, the system adjusts the feedback strategy in real time:

[0051]

[0052] Among them, w i is the weight coefficient, FI i is the fatigue index, θ jis the weight coefficient, ABRI j is the hard braking risk score, β k is the weight coefficient, DRF k is the distracted driving risk factor, α1, α2, and α3 are the feedback strength coefficients for dynamic adjustment, Feedback t represents the feedback type and strength dynamically adjusted according to the scores of three driving abnormal behaviors, sigmoid() is the activation function,

[0053] S45. Based on the feedback results in each monitoring process, the system makes real-time adjustments to the optimized network architecture.

[0054] Optionally, the S41 specifically includes:

[0055] S411. Use the hybrid neural architecture search algorithm to optimize the adaptive vision recurrent network architecture. First, the system generates multiple candidate network architectures through neural architecture search. These architectures combine the vision transformer and the recurrent neural network in their design. The hybrid neural architecture search selects the best architecture for training and deployment by evaluating the performance of each candidate architecture:

[0056]

[0057] where FI i is the fatigue index at the i-th moment, is the fatigue index predicted by the model, w i is the weighting coefficient, Δt i is the time variation, p and q are the adjustment exponents, τ is a constant, N is the number of samples, TRS t represents the fatigue risk status score;

[0058] S412. Through the optimized adaptive vision recurrent network architecture, the system makes adaptive adjustments based on the driver's fatigue characteristic data. The system combines the fatigue characteristics extracted by the vision transformer with the vehicle dynamic data, and gradually improves the detection accuracy of fatigue driving through the network adaptive adjustment mechanism:

[0059]

[0060] where FI i is the fatigue index, w i is the weight coefficient, Δt i is the time interval, U o is the time decay coefficient, t is the current moment, T min is the fatigue start time, N is the number of samples, Adaptive FDS t represents the adaptive fatigue status score, exp() is the exponential function;

[0061] S413. By analyzing the driver's dynamic behavior, the system adjusts the parameters of the adaptive visual recursive network architecture in real time:

[0062]

[0063] where Δv j is the vehicle speed change at the j-th moment, a j is the acceleration, δ2 and γ2 are weighting coefficients, d k is the change in the vehicle's spatial position, θ2 is the spatial weight coefficient, N is the number of samples, FDS Adjustment t represents the fatigue state score adjusted according to the driver's dynamic behavior data, δ j is the weighting coefficient for adjusting the fatigue state score by vehicle speed, and α2 is the spatial acceleration influence coefficient;

[0064] S414. Based on the optimized network architecture and adaptive adjustment, the system dynamically adjusts the feedback strategy according to the driver's fatigue state, behavior data, and real-time monitoring results.

[0065] Optionally, the S5 specifically includes:

[0066] S51. According to the calculated fatigue risk state score, hard braking risk score, and distracted driving risk factor, set the severity level of each driving behavior, and set the feedback level in layers according to the risk score:

[0067] If the fatigue risk state score TRS t is higher than the preset threshold, it is determined as severe fatigue driving;

[0068] If the hard braking risk score ABRI t exceeds the preset threshold, it is determined as a high hard braking risk;

[0069] If the distracted driving risk factor DRF t exceeds the set threshold, it is determined as a severe distracted driving behavior;

[0070] S52. According to the severity level of the driver's behavior, the system selects an appropriate feedback level:

[0071] First-layer feedback: The fatigue risk state score is greater than the preset threshold or the hard braking risk score is higher than the threshold ω1, and the system provides a voice prompt;

[0072] Second-layer feedback: The fatigue risk state score is greater than the preset threshold or the distracted driving risk factor is higher than the threshold ε1, and the system provides a vibration prompt;

[0073] Third-layer feedback: The fatigue risk state score is greater than the preset threshold If the hard braking risk score is higher than the threshold ω3, or the distracted driving risk factor is higher than the threshold ε2, the system provides a strong audible warning and a visual warning;

[0074] S53. If the driver fails to respond to the first - layer or second - layer feedback in a timely manner, the system will dynamically increase the feedback intensity:

[0075] Feedback t+1 = Feedback t ·(1 + κ1·|ΔTRS t | p + κ2·|ΔABRI t | q );

[0076] where ΔTRS t and ΔABRI t are the changes in the fatigue risk status score and the hard braking risk score at the current moment respectively, κ1 and κ2 are adjustment coefficients, p and q are weighting exponents, Feedback t is the current feedback intensity, and Feedback t+1 is the feedback intensity at the next moment;

[0077] S54. If the driver adjusts the driving behavior according to the feedback, resulting in a decrease in the score, the system will gradually weaken the feedback intensity:

[0078] Feed t = Feedback t ·(1 - λ t ·(|TRS t -TRS min |+|ABRI t -ABRI min |+

[0079] |DRF t -DRF min |));

[0080] where TRS min 、ABRI min and DRF min are the minimum safety thresholds for fatigue risk, hard braking, and distracted driving respectively, λ t is the weakening coefficient, and Feed t is the weakened feedback intensity;

[0081] S55. After each monitoring session, the system learns based on the driver's response and optimizes the feedback strategy.

[0082] The beneficial effects of the present invention are:

[0083] By introducing a multi-level feedback mechanism and an adaptive adjustment strategy based on deep learning, the present invention greatly improves the accuracy and response ability of the driving abnormal behavior monitoring system. Compared with the prior art, the intelligent feedback mechanism of the present invention can dynamically adjust the feedback intensity and method according to the personalized behavior pattern and real-time feedback response of the driver. This feature enables the system to provide more accurate and personalized warnings when facing the behavior responses of different drivers, thus improving driving safety.

[0084] A remarkable advantage of the present invention is its multi-level feedback mechanism. By monitoring the driver's behavior in real time, the system can layer-select different intensities and types of feedback according to different types of abnormal behaviors such as fatigue driving, sudden braking, and distracted driving. In this way, the driver can receive timely and accurate feedback in case of different degrees of behavior abnormalities, avoiding the ineffective situation that may occur in the simple and single feedback method of the prior art. When the driver fails to respond to the mild feedback in time, the system can quickly increase the feedback intensity to ensure that the driver can receive sufficient warnings, thus effectively avoiding traffic accidents.

[0085] In addition, the adaptive adjustment mechanism of the present invention enables the system to dynamically adjust the feedback strategy according to the driver's historical behavior data and real-time reaction. This dynamic adaptive adjustment greatly improves the adaptability and real-time performance of the system, ensuring the efficient operation of the system in complex driving environments. Especially in various driver behaviors and driving environments, the system can automatically optimize the feedback strategy according to the scoring results and feedback responses, avoiding the limitations of the traditional fixed feedback method.

[0086] By introducing deep learning and meta-learning technologies, the present invention further improves the intelligent level of the system in processing complex driving data. The system can not only evaluate the driver's behavior in real time but also continuously learn and improve according to the feedback effect, thus enhancing the accuracy and adaptability of the feedback mechanism. Compared with the prior art, the present invention can better meet the diverse needs of drivers, provide safety reminders while avoiding excessive interference with drivers, and thus optimize the driving experience.

[0087] Generally speaking, through intelligent monitoring of driving abnormal behaviors and personalized feedback mechanisms, the present invention effectively improves the accuracy and real-time response ability of driver monitoring, has higher accuracy, adaptability, and safety than the prior art, can significantly reduce the risk of traffic accidents, and provides important technical support for the development of intelligent transportation systems and autonomous driving assistance systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0089] Figure 1 This is a flowchart of the driving abnormal behavior monitoring method for a driving recorder based on a hierarchical dynamic feedback mechanism proposed by the present invention;

[0090] Figure 2 This is a schematic diagram of the dynamic adjustment of the adaptive feedback strategy of the driving abnormal behavior monitoring method for a driving recorder based on a hierarchical dynamic feedback mechanism proposed by the present invention. Detailed implementation manners

[0091] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0092] Referring to Figure 1 and Figure 2 , the driving abnormal behavior monitoring method for a driving recorder based on a hierarchical dynamic feedback mechanism includes the following steps:

[0093] S1. Collect the facial images and vehicle dynamic data of the driver through the driving recorder;

[0094] S2. Identify the types of driving abnormalities, analyze the facial images and vehicle dynamic data of the driver, and predict the distracted driving behavior of the driver;

[0095] S3. Based on the identified types of driving abnormalities, select a task-specific network architecture;

[0096] S4. According to the current task, optimize the network architecture through hybrid neural architecture search and perform meta-learning training;

[0097] S5. Perform hierarchical dynamic feedback according to the severity of the driver's behavior. The first layer of feedback is voice prompt, the second layer of feedback is vibration prompt, and the third layer of feedback is strong sound warning and visual warning;

[0098] S6. Dynamically adjust the feedback intensity and method according to the driver's reaction to the feedback, monitor the driver's behavior in real time, and automatically switch tasks and adjust the feedback strategy when an abnormality occurs;

[0099] S7. Through continuous optimization and real-time adjustment, the system provides personalized driving behavior monitoring according to different drivers and driving environments.

[0100] In this embodiment, the specific content of S2 includes:

[0101] S21. Use the facial images and vehicle dynamic data collected by the driving recorder to perform preliminary feature extraction;

[0102] S22. Analyze the driver's facial expression and eye movement state using a vision transformer based on the features of the facial image, and calculate the fatigue index FI:

[0103]

[0104] Among them, E i is the eye closure degree of the i-th frame image, E mean is the average value of the closure degree, E std is the standard deviation of the closure degree, A i is the facial movement amplitude of the i-th frame, A mean is the average value of the movement amplitude, γ1 is the non-linear weighting coefficient of the movement amplitude, α is the exponential weight of the movement amplitude, δ i is the interaction influence coefficient of the closure degree and the movement amplitude, X is the number of frames of the collected image, exp() is the exponential function. When the fatigue index FI exceeds the preset threshold, it is determined as fatigue driving;

[0105] S23. According to the dynamic data of the vehicle, including vehicle speed change and acceleration change, use a temporal convolutional neural network to process the temporal data of acceleration and vehicle speed, and calculate the emergency braking probability EBS:

[0106]

[0107] Among them, Δv j represents the vehicle speed change at the j-th moment, Δt j is the time interval of the vehicle speed change, γ2 is the weighting exponent of the vehicle speed change, β2 is the weighting exponent of the time interval, a k is the acceleration at the k-th moment, α2 is the weighting exponent of the acceleration, γ3 is the non-linear activation function of the acceleration, ω1 is the interaction coefficient between the acceleration and the vehicle speed change, N is the number of samples, Δv l represents the vehicle speed change at the l-th moment, a l is the acceleration at the l-th moment. When the emergency braking probability EBS exceeds the set threshold, it is determined as an emergency braking behavior;

[0108] S24.Combine the driver's historical behavior data, and use a sequence analysis model based on a long short-term memory network to predict the driver's distracted driving behavior, and calculate the distracted driving score DSS:

[0109]

[0110] Among them, x k is the distracted behavior feature at the k-th moment, w k is the weight coefficient, y m is the reaction time of the driver at the m-th moment, α m is the weight coefficient of the reaction time, z nis the driver's physiological signal at the nth moment, and β n is the weight coefficient of the physiological signal, σ is the activation function, K is the number of distracted behavior features, L is the number of reaction time features, and P is the number of physiological signal features. When the distracted driving score is greater than the set threshold, it is determined as distracted driving behavior;

[0111] S25. Based on the analysis results, judge the abnormality degree of the driving behavior through the set threshold, and determine it as severe fatigue driving, hard braking or distracted driving respectively, and then give corresponding feedback.

[0112] In this embodiment, S3 specifically includes:

[0113] S31. According to the identified abnormal behavior type, based on the fatigue index FI, hard braking probability EBS and distracted driving score DSS, select a suitable network architecture for further analysis;

[0114] S32. When the fatigue index exceeds the preset threshold, select the adaptive vision recurrent network architecture. By combining the vision transformer and the recurrent neural network, deeply analyze the temporal changes of the fatigue state and facial features, and adjust the influence degree of facial features in real time through the weighted impulse function:

[0115]

[0116] where w i is the facial feature weighting coefficient at the ith moment, FI i is the fatigue index, Δt i is the time interval, α1 is the time decay coefficient, t is the current moment, T min is the fatigue start time, N is the number of samples, FDS t is the fatigue state score, and exp() is the exponential function;

[0117] S33. When the hard braking probability exceeds the set threshold, select the graph convolutional temporal network architecture. By introducing the event-driven graph convolutional model, analyze the spatio-temporal dynamic features during hard braking, and construct an event-driven graph through vehicle speed, acceleration and spatio-temporal position information to dynamically capture the hard braking risk index:

[0118]

[0119] where a j is the acceleration at the jth moment, Δv j is the vehicle speed change, Δt j is the time interval, γ and β are adjustment parameters, d j is the vehicle spatial displacement, N is the number of samples, and EBS t is the hard braking risk index;

[0120] S34. When the distracted driving score exceeds the set threshold, select the spatio-temporal attention-autoregressive network architecture. By combining the spatio-temporal attention mechanism and the autoregressive model, comprehensively analyze the temporal changes in the driver's attention state and reaction time, introduce the spatio-temporal attention fusion mechanism, and combine physiological signals and behavioral data to calculate the attention state score ASR. t :

[0121]

[0122] Among them, x k is the distracted behavior feature, w k is the weight coefficient, y m is the reaction time, α m is the weighting coefficient, z p is the physiological signal feature, φ p is the weight coefficient of the physiological signal, γ k is the weighted adjustment coefficient, and K, L, and P are the numbers of behavioral features, reaction time features, and physiological signal features;

[0123] S35. According to the calculated fatigue state score, hard brake risk index, and attention state score, adjust the feedback strategy in real time through the adaptive feedback optimization mechanism.

[0124] In this embodiment, the specific content of S4 includes:

[0125] S41. For the adaptive visual recurrent network architecture, optimize it through hybrid neural architecture search and make adaptive adjustments according to the characteristics of fatigue driving and behavioral data;

[0126] S42. For the graph convolutional temporal network architecture, optimize it by combining meta-learning training, use the hard brake risk score to optimize hard brake detection, and real-time evaluate the changes in vehicle speed, acceleration, and spatial position to judge the risk of hard brake behavior:

[0127]

[0128] Among them, a j is the acceleration at the j-th moment, v j is the change in vehicle speed, Δt j is the time interval, λ q and ω are adjustment coefficients, dist j is the change in the spatial distance of the vehicle, ν is the weighting exponent, and ABRI t represents the hard brake risk score, and N is the number of samples;

[0129] S43. For the spatio-temporal attention-autoregressive network architecture, optimize the calculation method of the distracted driving risk factor, and adjust the response intensity of the network in real time through multi-modal signal fusion:

[0130]

[0131] Among them, x k is the behavior feature, is the prediction feature, δ k is the weighting coefficient, p m is the physiological signal feature, τ m is a constant, φ m is the weighting coefficient, s is the feature difference weighting, K and P are the numbers of behavior features and physiological signal features respectively, DRF t represents the distracted driving risk factor, log2 is the logarithmic function, and abs is the absolute value operation;

[0132] S44. Through the optimized adaptive visual recursive network architecture, graph convolutional temporal network architecture, and spatio-temporal attention-autoregressive network architecture, the system adjusts the feedback strategy in real time:

[0133]

[0134] Among them, w i is the weight coefficient, FI i is the fatigue index, θ j is the weight coefficient, ABRI j is the hard braking risk score, β k is the weight coefficient, DRF k is the distracted driving risk factor, α1, α2, and α3 are the feedback intensity coefficients adjusted dynamically, Feedback t represents the feedback type and intensity adjusted dynamically according to the scores of three driving abnormal behaviors, and sigmoid() is the activation function,

[0135] S45. Based on the feedback results in each monitoring process, the system adjusts the optimized network architecture in real time.

[0136] In this embodiment, the S41 specifically includes:

[0137] S411. Optimize the adaptive visual recursive network architecture using the hybrid neural architecture search algorithm. First, the system generates multiple candidate network architectures through neural architecture search. These architectures combine the vision transformer and the recurrent neural network in design. The hybrid neural architecture search selects the best architecture for training and deployment by evaluating the performance of each candidate architecture:

[0138]

[0139] Among them, FI i is the fatigue index at the i-th moment, is the fatigue index predicted by the model, w iis the weighting coefficient, Δt i is the time variation, p and q are adjustment exponents, τ is a constant, N is the number of samples, TRS t represents the fatigue risk status score;

[0140] S412. Through the optimized adaptive visual recursive network architecture, the system makes adaptive adjustments based on the driver's fatigue feature data. The system combines the fatigue features extracted by the vision transformer and the vehicle dynamic data, and through the network adaptive adjustment mechanism, gradually improves the detection accuracy of fatigue driving:

[0141]

[0142] Among them, FI i is the fatigue index, w i is the weight coefficient, Δt i is the time interval, U o is the time decay coefficient, t is the current moment, T min is the fatigue start time, N is the number of samples, Adaptive FDS t represents the adaptive fatigue status score, exp() is the exponential function;

[0143] S413. By analyzing the driver's dynamic behavior, the system adjusts the parameters of the adaptive visual recursive network architecture in real time:

[0144]

[0145] Among them, Δv j is the vehicle speed change at the j-th moment, a j is the acceleration, δ2 and γ2 are the weighting coefficients, d k is the change in the vehicle's spatial position, θ2 is the spatial weight coefficient, N is the number of samples, FDS Adjustment t represents the fatigue status score adjusted according to the driver's dynamic behavior data, δ j is the weighting coefficient for adjusting the fatigue status score by the vehicle speed, α2 is the spatial acceleration influence coefficient;

[0146] S414. Based on the optimized network architecture and adaptive adjustment, the system dynamically adjusts the feedback strategy according to the driver's fatigue status, behavior data and real-time monitoring results.

[0147] In this embodiment, the S5 specifically includes:

[0148] S51. According to the calculated fatigue risk status score, hard braking risk score and distracted driving risk factor, set the severity level of each driving behavior, and set the feedback level in layers according to the risk score:

[0149] If the fatigue risk status score TRS t is higher than the preset threshold, it is determined as severe fatigue driving;

[0150] If the emergency braking risk score ABRI t exceeds the preset threshold, it is determined as a high risk of emergency braking;

[0151] If the distracted driving risk factor DRF t exceeds the set threshold, it is determined as a severe distracted driving behavior;

[0152] S52. According to the severity level of the driver's behavior, the system selects an appropriate feedback level:

[0153] The first-level feedback: The fatigue risk status score is greater than the preset threshold or the emergency braking risk score is higher than the threshold ω1, and the system provides a voice prompt;

[0154] The second-level feedback: The fatigue risk status score is greater than the preset threshold or the distracted driving risk factor is higher than the threshold ε1, and the system provides a vibration prompt;

[0155] The third-level feedback: The fatigue risk status score is greater than the preset threshold or the emergency braking risk score is higher than the threshold ω3, or the distracted driving risk factor is higher than the threshold ε2, and the system provides a strong sound warning and a visual warning;

[0156] S53. If the driver fails to respond to the first-level or second-level feedback in a timely manner, the system will dynamically increase the feedback intensity:

[0157] Feedback t+1 = Feedback t ·(1 + κ1·|ΔTRS t | p + κ2·|ΔABRI t | q );

[0158] where, ΔTRS t and ΔABRI t are respectively the changes in the fatigue risk status score and the emergency braking risk score at the current moment, κ1 and κ2 are adjustment coefficients, p and q are weighting exponents, Feedback t is the current feedback intensity, and Feedback t+1 is the feedback intensity at the next moment;

[0159] S54. If the driver adjusts the driving behavior according to the feedback and the score decreases, the system will gradually weaken the feedback intensity:

[0160] Feed t= Feedback t ·(1 - λ t ·(|TRS t - TRS min | + |ABRI t - ABRI min | +

[0161] |DRF t - DRF min |));

[0162] Wherein, TRS min , ABRI min and DRF min are the minimum safety thresholds for fatigue risk, hard braking, and distracted driving respectively, λ t is the attenuation coefficient, and Feed t is the attenuated feedback strength;

[0163] S55. After each monitoring ends, the system learns according to the driver's response and optimizes the feedback strategy.

[0164] Example 1:

[0165] To verify the feasibility of the present invention in implementation, the present invention is applied to a long - distance driver. The long - distance driver drives an ordinary small car on the highway during the morning rush hour. Since the continuous driving exceeds 4 hours, the driver's energy begins to decline and signs of fatigue driving appear. During this process, the system can monitor the driver's behavior in real time and give feedback reminders in a timely manner.

[0166] The vehicle is equipped with an intelligent driving recorder, which is equipped with a high - precision camera, an accelerometer, a GPS sensor, and a biosignal sensor, and can collect data such as the driver's facial expression, eye movement state, vehicle speed, and acceleration in real time. Based on these data, the system monitors and intervenes in abnormal driving behaviors through the multi - level feedback mechanism proposed by the present invention. Specifically, after the system detects that the driver has slight signs of fatigue, it guides the driver to take appropriate rest through low - intensity feedback (such as voice reminder). When the signs of fatigue worsen or a hard braking behavior occurs, it reminds the driver to take immediate action through strong visual warnings and sound alarms.

[0167] In this specific application, the intelligent monitoring system first determines whether the driver is fatigued by real-time collecting the driver's facial expressions and eye states. In the collected facial images of the driver, the system uses a vision transformer model to analyze features such as the degree of eye closure and subtle changes in facial muscles, so as to calculate the fatigue risk status score in real time. If the fatigue risk score exceeds the set threshold, the system will issue the first layer of feedback - a voice prompt "Driver, you seem a bit fatigued. Please take a proper rest." If the driver fails to respond in time, the system will judge whether other serious abnormal behaviors occur based on the emergency braking risk score calculated by the system and the distracted driving risk factor.

[0168] The system will also monitor the driver's acceleration and vehicle speed changes in real time, especially the occurrence of emergency braking behaviors. The system processes the vehicle speed and acceleration data through a temporal convolutional neural network to calculate the emergency braking risk score. When the driver's emergency braking behavior exceeds the set threshold, the system will immediately issue the second layer of feedback - a strong vibration warning, and give an emergency reminder through the vehicle display screen: "Emergency braking, decelerate and maintain a safe distance!" At this time, the system will also check the driver's distracted state. If the distracted driving risk factor exceeds the set threshold, the system will switch to the third layer of feedback, issue a strong sound warning, and simultaneously flash visual warning lights to ensure that the driver can be aware of and correct the driving behavior in time.

[0169] During the feedback process, the system will also dynamically adjust the feedback strategy according to the driver's reaction to the feedback. If the driver does not make any response after receiving the first layer of feedback, the system will automatically increase the feedback intensity and improve the warning effect through the second layer of feedback or the third layer of feedback. According to the actual driver response situation, the intensity of the feedback can be gradually weakened to ensure that the driver's driving experience is not overly disturbed.

[0170] To verify the effectiveness of the present invention in actual driving scenarios, we conducted a 10-hour field test on the highway and collected the driver's behavior data, feedback response data, and relevant safety data. During the test, we used facial expression analysis based on vision transformers, acceleration data, and vehicle speed data to monitor the driver's fatigue level, emergency braking behavior, and distracted driving situation in real time. The test data included the scores and feedback response situations of the driver in different behavior states, as well as the intervention effects of different levels of feedback.

[0171] Table 1 Comparison of Driver Behavior Responses and Safety Improvement Effects

[0172]

[0173]

[0174] Based on the data in the above table, the following analysis can be carried out:

[0175] The table shows the driver's behavioral responses and safety improvement effects at different feedback levels. First, in the absence of feedback, the incidence of abnormal driver behavior is 10%, and after feedback, the incidence of abnormal behavior increases to 12%. This indicates that in the absence of feedback, the driver's safety condition is poor and abnormal behavior cannot be corrected in a timely manner. The driver does not receive effective intervention in this state, so the driver's abnormal behavior is not effectively improved.

[0176] When the first layer of feedback (voice prompt) is adopted, the driver's feedback reaction time is 1.5 seconds. The driver can respond in a timely manner, and the incidence of abnormal behavior decreases from 12% to 8%. This shows that voice feedback can help the driver identify and correct minor abnormal behaviors to a certain extent. At the same time, the driver's behavior improvement ratio is 20%, indicating that voice feedback can promote the improvement of the driver's behavior and reduce their abnormal driving behavior to a certain extent. However, there are still certain safety hazards because the accident incidence rate fails to drop to a low level.

[0177] When the feedback intensity is further increased to the second layer of feedback (vibration warning), the driver's reaction delay increases to 2.3 seconds. This indicates that the feedback of vibration warning may be rather abrupt, but it can significantly improve the driver's alertness to abnormal behavior. Under this feedback, the incidence of the driver's abnormal behavior drops significantly, from 20% to 5%. The abnormal behavior correction time increases to 30 seconds, indicating that vibration warning can effectively make the driver respond more quickly and reduce the occurrence of high-risk driving behaviors. In addition, the driver's behavior improvement ratio is as high as 50%, significantly improving safety.

[0178] Finally, when the third layer of feedback (visual warning and strong sound warning) is adopted, the driver's reaction delay is shortened to 0.8 seconds, indicating that strong visual and sound warnings can quickly attract the driver's attention and force them to take corrective measures. Under this feedback, the incidence of abnormal behavior drops significantly from 30% to 1%, showing extremely high effectiveness and almost completely eliminating the driver's abnormal behavior. The driver's behavior improvement ratio also increases to 70%, indicating that such feedback can not only correct the driver's abnormal behavior in a timely manner but also greatly improve their driving safety. At the same time, the driver's overall sense of safety is improved by 60%, indicating that the effect of strong feedback is very significant in enhancing the driver's sense of security.

[0179] From the above analysis, it can be seen that as the feedback level increases, the effects of driver behavior improvement and safety enhancement are gradually strengthened. The greater the feedback intensity, the shorter the driver's reaction time and the faster the correction of abnormal behaviors. Especially in the case of the third-level feedback, the system can greatly reduce the accident rate and increase the proportion of driver behavior improvement and safety. The data in this table clearly demonstrates the effectiveness and superiority of the multi-level dynamic feedback mechanism of the present invention in improving driving safety and enhancing driver behavior.

[0180] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes should be covered within the protection scope of the present invention.

Claims

1. A method for monitoring abnormal driving behaviors of a driving recorder based on a hierarchical dynamic feedback mechanism, characterized in that It includes the following steps: S1. Collect the driver's facial images and vehicle dynamic data through a dash cam; S2. Identify the types of abnormal driving, analyze the driver's facial images and vehicle dynamic data, and predict the driver's distracted driving behavior; S3. Based on the identified types of abnormal driving, select a task-specific network architecture; S4. According to the current task, optimize the network architecture through hybrid neural architecture search and conduct meta-learning training; S5. According to the severity of the driver's behavior, conduct hierarchical dynamic feedback. The first layer of feedback is voice prompt, the second layer of feedback is vibration prompt, and the third layer of feedback is strong sound warning and visual warning; S6. According to the driver's reaction to the feedback, dynamically adjust the feedback intensity and method, monitor the driver's behavior in real time, and automatically switch tasks and adjust the feedback strategy when an abnormality occurs; S7. Through continuous optimization and real-time adjustment, the system provides personalized driving behavior monitoring according to different drivers and driving environments.

2. The method for monitoring abnormal driving behaviors of a driving recorder based on a hierarchical dynamic feedback mechanism according to claim 1, characterized in that The specific content of S2 includes: S21. Use the facial images and vehicle dynamic data collected by the dash cam to conduct preliminary feature extraction; S22. Based on the features of the facial image, use a vision transformer to analyze the driver's facial expression and eye movement state, and calculate the fatigue index FI; Among them, E i is the eye closure degree of the i-th frame image, E mean is the average value of the closure degree, E std is the standard deviation of the closure degree, A i is the facial movement amplitude of the i-th frame, A mean is the average value of the movement amplitude, γ1 is the non-linear weighting coefficient of the movement amplitude, α is the exponential weight of the movement amplitude, δ i is the interaction influence coefficient of the closure degree and the movement amplitude, X is the number of frames of the collected images, exp() is the exponential function, and when the fatigue index FI exceeds the preset threshold, it is determined as fatigue driving; S23. According to the vehicle's dynamic data, including vehicle speed change and acceleration change, use a temporal convolutional neural network to process the temporal data of acceleration and vehicle speed, and calculate the emergency braking probability EBS; where, Δv j represents the vehicle speed change at the j-th moment, Δt j is the time interval of the vehicle speed change, γ2 is the weighted index of the vehicle speed change, β2 is the weighted index of the time interval, a k is the acceleration at the k-th moment, α2 is the weighted index of the acceleration, γ3 is the non-linear activation function of the acceleration, ω1 is the interaction coefficient between the acceleration and the vehicle speed change, N is the number of samples, Δv l represents the vehicle speed change at the l-th moment, a l is the acceleration at the l-th moment. When the emergency braking probability EBS exceeds the set threshold, it is determined as an emergency braking behavior; S24. Combine the driver's historical behavior data, use a sequence analysis model based on a long short-term memory network to predict the driver's distracted driving behavior, and calculate the distracted driving score DSS; Among them, x k is the distraction behavior feature at the k-th moment, w k is the weight coefficient, y m is the reaction time of the driver at the m-th moment, α m is the weight coefficient of the reaction time, z n is the physiological signal of the driver at the n-th moment, β n is the weight coefficient of the physiological signal, σ is the activation function, K is the number of distraction behavior features, L is the number of reaction time features, P is the number of physiological signal features. When the distracted driving score is greater than the set threshold, it is determined as a distracted driving behavior; S25. Based on the analysis results, judge the abnormal degree of the driving behavior through a set threshold, and respectively determine it as severe fatigue driving, emergency braking or distracted driving, and then conduct corresponding feedback.

3. The method for monitoring abnormal driving behaviors of a driving recorder based on a hierarchical dynamic feedback mechanism according to claim 1, wherein The specific content of S3 includes: S31. According to the identified types of abnormal behavior, based on the fatigue index FI, the emergency braking probability EBS, and the distracted driving score DSS, select a suitable network architecture for further analysis; S32. When the fatigue index exceeds the preset threshold, select an adaptive vision recurrent network architecture. By combining a vision transformer and a recurrent neural network, deeply analyze the temporal changes of the fatigue state and facial features, and adjust the influence degree of facial features in real time through a weighted impulse function; where, w i is the facial feature weighting coefficient at the i-th moment, FI i is the fatigue index, Δt i is the time interval, α1 is the time decay coefficient, t is the current moment, T min is the fatigue start time, N is the number of samples, FDS t is the fatigue state score, exp() is the exponential function; S33. When the emergency braking probability exceeds the set threshold, select a graph convolutional temporal network architecture. By introducing an event-driven graph convolutional model to analyze the spatio-temporal dynamic features during emergency braking, construct an event-driven graph through vehicle speed, acceleration, and spatio-temporal position information, and dynamically capture the risk index of emergency braking; where a j is the acceleration at the j-th moment, Δv j is the vehicle speed change, Δt j is the time interval, γ and β are adjustment parameters, d j is the vehicle spatial displacement, N is the number of samples, and EBS t is the hard braking risk index; S34. When the distracted driving score exceeds the set threshold, select the spatio-temporal attention-autoregressive network architecture. By combining the spatio-temporal attention mechanism and the autoregressive model, comprehensively analyze the temporal changes in the driver's attention state and reaction time, introduce the spatio-temporal attention fusion mechanism, and combine physiological signals with behavioral data to calculate the attention state score ASR t : Among them, x k is the distraction behavior feature, w k is the weight coefficient, y m is the reaction time, α m is the weighting coefficient, z p is the physiological signal feature, φ p is the weight coefficient of the physiological signal, γ k is the weighting adjustment coefficient, and K, L, and P are the numbers of the behavior feature, reaction time feature, and physiological signal feature; S35. According to the calculated fatigue state score, emergency braking risk index, and attention state score, adjust the feedback strategy in real time through an adaptive feedback optimization mechanism.

4. The method for monitoring abnormal driving behaviors of a driving recorder based on a hierarchical dynamic feedback mechanism according to claim 1, wherein The specific content of S4 includes: S41. For the adaptive vision recurrent network architecture, conduct optimization through hybrid neural architecture search, and make adaptive adjustments according to the characteristics and behavior data of fatigue driving; S42. Optimize the graph convolutional temporal network architecture by combining meta-learning training, and use the emergency braking risk score to optimize emergency braking detection. Real-time evaluate the changes in vehicle speed, acceleration, and spatial position to judge the risk of emergency braking behavior: where a j is the acceleration at the j-th moment, v j is the vehicle speed change, Δt j is the time interval, λ q and ω are adjustment coefficients, dist j is the change in the spatial distance of the vehicle, ν is the weighting exponent, ABRI t represents the hard braking risk score, and N is the number of samples; S43. For the spatio-temporal attention-autoregressive network architecture, optimize the calculation method of the distracted driving risk factor, and adjust the response intensity of the network in real time through multi-modal signal fusion: Among them, x k is a behavioral feature, is a prediction feature, δ k is a weighting coefficient, p m is a physiological signal feature, τ m is a constant, φ m is a weighting coefficient, s is the weighted feature difference, K and P are the numbers of behavioral features and physiological signal features respectively, DRF t represents the distracted driving risk factor, log2 is the logarithmic function, and abs is the absolute value operation; S44. Through the optimized adaptive visual recurrent network architecture, graph convolutional temporal network architecture, and spatio-temporal attention-autoregressive network architecture, the system adjusts the feedback strategy in real time: Among them, w i is the weight coefficient, FI i is the fatigue index, θ j is the weight coefficient, ABRI j is the hard braking risk score, β k is the weight coefficient, DRF k is the distracted driving risk factor, α1, α2, and α3 are the feedback intensity coefficients for dynamic adjustment, Feedback t represents the feedback type and intensity dynamically adjusted according to the scores of three types of driving abnormal behaviors, sigmoid() is the activation function, S45. Based on the feedback results during each monitoring process, the system makes real-time adjustments to the optimized network architecture.

5. The health consultation method based on general artificial intelligence according to claim 4, wherein, The specific steps of S41 are as follows: S411. Use the hybrid neural architecture search algorithm to optimize the adaptive visual recurrent network architecture. First, the system generates multiple candidate network architectures through neural architecture search. These architectures combine vision transformers and recurrent neural networks in their design. The hybrid neural architecture search selects the best architecture for training and deployment by evaluating the performance of each candidate architecture: Among them, FI i is the fatigue index at the i-th moment, is the fatigue index predicted by the model, w i is the weighting coefficient, Δt i is the time change, p and q are adjustment indices, τ is a constant, N is the number of samples, and TRS t represents the fatigue risk status score; S412. Through the optimized adaptive visual recurrent network architecture, the system makes adaptive adjustments based on the driver's fatigue feature data. The system combines the fatigue features extracted by the vision transformer and the vehicle dynamic data, and gradually improves the detection accuracy of fatigue driving through the network adaptive adjustment mechanism: Among them, FI i is the fatigue index, w i is the weight coefficient, Δt i is the time interval, U o is the time decay coefficient, t is the current time, T min is the fatigue start time, N is the number of samples, Adaptive FDS t represents the adaptive fatigue state score, and exp() is the exponential function; S413. By analyzing the driver's dynamic behavior, the system adjusts the parameters of the adaptive visual recurrent network architecture in real time: where, Δv j is the vehicle speed change at the j-th moment, a j is the acceleration, δ2 and γ2 are weighting coefficients, d k is the change in the vehicle's spatial position, θ2 is the spatial weight coefficient, N is the number of samples, FDS Adjustment t represents the fatigue state score adjusted according to the driver's dynamic behavior data, δ j is the weighting coefficient for adjusting the fatigue state score by the vehicle speed, and α2 is the spatial acceleration influence coefficient; S414. Based on the optimized network architecture and adaptive adjustment, the system dynamically adjusts the feedback strategy according to the driver's fatigue state, behavior data, and real-time monitoring results.

6. The method for monitoring abnormal driving behaviors of a driving recorder based on a hierarchical dynamic feedback mechanism according to claim 1, characterized in that The specific steps of S5 are as follows: S51. According to the calculated fatigue risk status score, emergency braking risk score, and distracted driving risk factor, set the severity level of each driving behavior, and set the feedback level in layers according to the risk score: If the fatigue risk status score TRS t is higher than the preset threshold, it is determined as severe fatigue driving; If the emergency braking risk score ABRI t exceeds the preset threshold, it is determined to be a high risk of emergency braking; If the distracted driving risk factor DRF t exceeds the set threshold, it is determined to be a serious distracted driving behavior; S52. According to the severity level of the driver's behavior, the system selects the appropriate feedback level: First-layer feedback: The fatigue risk status score is greater than the preset threshold Or when the hard braking risk score is higher than the threshold ω1, the system provides a voice prompt; Second-layer feedback: The fatigue risk status score is greater than a preset threshold Or when the distracted driving risk factor is higher than the threshold ε1, the system provides a vibration prompt; Third-layer feedback: The fatigue risk status score is greater than the preset threshold Or the hard braking risk score is higher than the threshold ω3, or the distracted driving risk factor is higher than the threshold ε2, and the system provides a strong audible warning and a visual warning; S53. If the driver fails to respond to the first or second layer of feedback in a timely manner, the system will dynamically increase the feedback intensity: Feedback t+1 = Feedback t ·(1 + κ1·|ΔTRS t | p + κ2·|ΔABRI t | q ); where, ΔTRS t and ΔABRI t are the changes in the fatigue risk status score and the hard braking risk score at the current moment, κ1 and κ2 are adjustment coefficients, p and q are weighting exponents, Feedback t is the current feedback intensity, and Feedback t+1 is the feedback intensity at the next moment; S54. If the driver adjusts the driving behavior according to the feedback and the score decreases, the system will gradually weaken the feedback intensity: Feed t = Feedback t ·(1 - λ t ·(|TRS t - TRS min | + |ABRI t - ABRI min | + |DRF t - DRF min |)); Among them, TRS min , ABRI min and DRF min are the minimum safety thresholds for fatigue risk, hard braking, and distracted driving respectively, λ t is the attenuation coefficient, and Feed t is the attenuated feedback strength; S55. After each monitoring ends, the system learns according to the driver's response and optimizes the feedback strategy.

Images