An accident warning analysis method and system based on driving data

Abstract

The present invention discloses an accident early warning analysis method and system based on driving data, which relates to the technical field of accident early warning analysis. By extracting the road condition complexity coefficient Cen, the line-of-sight chaos coefficient Cvs, and the power interference coefficient Cdn, the system of the present invention models the external environment not only considering the current state, but also analyzing its dynamic interference characteristics, so that the system still has high reliability in scenarios such as rain and fog, night, and congestion. The comprehensive score value Srisk is no longer a single binary determination, but combines information such as operation disturbance, attitude deviation, and fatigue index for multi-factor fusion interpretation, and outputs a response result WIN with hierarchical characteristics, which can realize multi-level early warning from voice reminder to active intervention. The dynamic feedback optimization module constructs a risk response parameter Ψstab according to the score fluctuation trend, and dynamically adjusts the multi-modal model weight coefficient ωj to generate a new coefficient nω, realizing self-parameter adjustment and adaptive reinforcement learning.

Description

Technical Field

[0001] The present invention relates to the technical field of accident warning analysis, and specifically provides an accident warning analysis method and system based on driving data. Background Art

[0002] In active safety control, modules such as vehicle collision warning, lane keeping assistance, and pedestrian recognition reminder are particularly critical subsystems in advanced driver assistance systems. The accident warning analysis system based on driving data is exactly one of the key core modules embedded in such ADAS architectures. It identifies potential risk factors through real-time analysis of multi-modal information such as vehicle dynamics, environmental perception, and driver behavior, and is the core bridge to promote the closed-loop of autonomous driving from perception to prediction and then to control.

[0003] Currently, most warning systems rely on single-scenario modeling or hard rule triggering mechanisms, and it is difficult to accurately model the coupling relationship between complex driving behaviors and dynamic environments. For example, in the FCW system, a fixed time threshold is generally used for judgment, but the driver's operation intention or environmental complexity at that time is not considered, which is prone to false alarms or missed alarms; the recognition accuracy of the PCW decreases when facing pedestrians in multiple postures, especially in night scenes, and the recognition accuracy of traditional models highly depends on lighting conditions; the HMW system also has the problem that blind spots cannot monitor distant targets. The phenomena of "modal fragmentation", "behavior isolation", and "prediction lag" existing in these existing systems make the triggering of warning signals lack flexibility and accuracy, and it is difficult to adapt to the increasingly complex and changeable road traffic environment.

[0004] The fundamental reasons behind these current situations are, on the one hand, the lack of a unified modeling mechanism for driving behaviors, environmental changes, and risk patterns in the warning system, resulting in a narrow understanding of the scenario by the system, especially the lack of the ability to predict the trend of behavior evolution; on the other hand, the utilization of driving data in traditional systems only stays at the surface threshold judgment, and intelligent analysis means such as deep feature extraction and time series prediction are not introduced, and it is impossible to dynamically adapt to the state changes of drivers and environmental interferences. These defects may lead to serious consequences in actual operation. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention provides an accident warning analysis method and system based on driving data, which solves the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention is realized through the following technical solutions: An accident warning analysis system based on driving data includes a data collection and preprocessing module, an abnormal behavior recognition module, an environmental complexity recognition module, a multi-mode risk fusion and assessment module, a comprehensive assessment and warning decision module, and a dynamic feedback and optimization module;

[0007] The data acquisition and preprocessing module collects multi-dimensional driving data of the vehicle through the CV2328 four-in-one driving recorder and sensors, fits them into the original dataset WQ, and performs preprocessing to obtain the driving dataset WS;

[0008] The abnormal behavior recognition module extracts abnormal behavior features from the obtained driving dataset WS, including the driving behavior fluctuation index Dar and the operation rhythm imbalance index Dos, and fits them into the abnormal behavior parameter YC;

[0009] The environmental complexity recognition module extracts road condition complexity features from the obtained driving dataset WS, including the road condition complexity coefficient Cen, the line-of-sight confusion coefficient Cvs, and the power interference coefficient Cdn, and fits them into the complexity parameter YZ;

[0010] The multi-mode risk fusion evaluation module fuses the abnormal behavior parameter YC and the complexity parameter YZ obtained to obtain the early warning risk set R, and calculates the early warning risk set R through the use of a multi-modal risk aggregation model to obtain the risk score Rag;

[0011] The comprehensive evaluation and early warning decision-making module comprehensively interprets the obtained risk score Rag to obtain the comprehensive score value Srisk, and issues an early warning and generates corresponding measures WIN;

[0012] The dynamic feedback optimization module interprets the change trend of the comprehensive score value Srisk within a fixed period to obtain the risk response parameter Ψstab, and adjusts the coefficient ωj of the multi-modal risk aggregation model to obtain the new coefficient nω.

[0013] Preferably, the data acquisition and preprocessing module includes a multi-source data acquisition unit and a standardized preprocessing unit;

[0014] The multi-source data acquisition unit collects multi-dimensional driving data of the vehicle, including the vehicle vertical vibration frequency Fv, the driver's steering action micro-response time Ft, the line-of-sight image blur Fb, the micro-speed change rate Fp of the front target, and the driver's blink rate Fe, and fits them to obtain the original dataset WQ;

[0015] Among them, the vehicle vertical vibration frequency Fv and the instantaneous value alt of the vehicle lateral acceleration are collected through the IMU sensor built in the driving recorder;

[0016] The driver's steering action micro-response time Ft is collected through a corner sensor;

[0017] The line-of-sight image blur Fb and the distance df to the vehicle ahead are collected through the high-definition camera in the driving recorder;

[0018] The micro-speed change rate Fp of the front target is collected through a millimeter-wave radar;

[0019] The driver's blinking rate Fe is collected by an in-vehicle camera;

[0020] The standardized preprocessing unit eliminates outliers and normalizes the original data set WQ to obtain the driving data set WS;

[0021] Outlier elimination includes processing the original data set WQ using the three-standard-deviation method to eliminate outliers;

[0022] Normalization processing uses the maximum-minimum normalization method to process the original data set WQ to obtain the driving data set WS;

[0023] The driving data set WS is obtained through the following formula:

[0024] ;

[0025] In the formula, WSb represents the b-th data in the driving data set WS, WQb represents the b-th data in the original data set WQ, minWQb represents the valley value of the b-th data in the original data set WQ, and maxWQb represents the peak value of the b-th data in the original data set WQ.

[0026] Preferably, the abnormal behavior recognition module includes an abnormal behavior feature extraction unit and an abnormal feature fusion unit;

[0027] The abnormal behavior feature extraction unit extracts abnormal behavior features from the driving data set WS, including the driving behavior fluctuation index Dar and the operation rhythm imbalance index Dos;

[0028] The driving behavior fluctuation index Dar is obtained through the following formula:

[0029] ;

[0030] In the formula, Tp represents the standardized reference value of the micro velocity change rate of the front target, and Te represents the standardized reference value of the driver's blinking rate

[0031] The operation rhythm imbalance index Dos is obtained through the following formula:

[0032] ;

[0033] In the formula, alt represents the instantaneous value of the vehicle's lateral acceleration, palt represents the average value of the lateral acceleration, and Fv represents the vehicle's vertical vibration frequency;

[0034] The abnormal feature fusion unit fuses the driving behavior fluctuation index Dar and the operation rhythm imbalance index Dos, and calculates to obtain the abnormal behavior parameter YC;

[0035] The abnormal behavior parameter YC is obtained through the following formula:

[0036] ;

[0037] Wherein, respectively represent the preset weight values of the driving behavior fluctuation index Dar and the operation rhythm imbalance index Dos.

[0038] Preferably, the environmental complexity recognition module includes a road condition complexity analysis unit and a complexity parameter fusion unit;

[0039] The road condition complexity analysis unit extracts road condition complexity features from the driving data set WS, including the road condition complexity coefficient Cen, the line-of-sight confusion coefficient Cvs, and the power interference coefficient Cdn;

[0040] The road condition complexity coefficient Cen is obtained through the following formula:

[0041] ;

[0042] Wherein, Fb represents the line-of-sight image blur, Fp represents the micro speed change rate of the front target, df represents the distance to the vehicle in front, obtained from the driving recorder, and alt represents the instantaneous value of the vehicle lateral acceleration;

[0043] The line-of-sight confusion coefficient Cvs is obtained through the following formula:

[0044] ;

[0045] Wherein, E represents a constant to prevent division by zero error, and Fb represents the line-of-sight image blur;

[0046] The power interference coefficient Cdn is obtained through the following formula:

[0047] ;

[0048] Wherein, Ft represents the micro response time of the driver's steering action, and Fe represents the driver's blinking rate;

[0049] The complexity parameter fusion unit fuses the obtained road condition complexity coefficient Cen, line-of-sight confusion coefficient Cvs, and power interference coefficient Cdn, and calculates to obtain the complexity parameter YZ;

[0050] The complexity parameter YZ is obtained through the following formula:

[0051] ;

[0052] Wherein, respectively represent the preset weight values of the road condition complexity coefficient Cen, the line-of-sight confusion coefficient Cvs, and the power interference coefficient Cdn.

[0053] Preferably, the multi-mode risk fusion evaluation module includes a risk score fusion unit and a multi-modal risk aggregation unit;

[0054] The risk score fusion unit fuses the obtained abnormal behavior parameter YC and complexity parameter YZ to obtain an early warning risk set R;

[0055] The early warning risk set R is obtained through the following formula:

[0056] ;

[0057] In the formula, R YC represents the score of the abnormal behavior parameter, R YZ represents the score of the complexity parameter, maxYC represents the maximum value of the abnormal behavior parameter, and maxYZ represents the maximum value of the complexity parameter.

[0058] Preferably, the multi-modal risk aggregation unit calculates the early warning risk set R according to the obtained early warning risk set R by using a multi-modal risk aggregation model to obtain a risk score Rag, and compares it with a preset risk threshold TRa to judge the risk status;

[0059] The risk score Rag is obtained through the following formula:

[0060] ;

[0061] In the formula, Z represents a normalization factor, N represents the number of modalities in the early warning risk set R, including the score R of the abnormal behavior parameter YC and the score R of the complexity parameter YZ , Rj represents the j-th modality in the early warning risk set R, ln represents the logarithmic function, represents the weight coefficient of the j-th modality, and Pj represents the occurrence probability of the j-th modality;

[0062] The risk status is obtained by matching in the following way:

[0063] When the risk score Rag ≤ the risk threshold TRa, it means there is no risk;

[0064] When the risk score Rag > the risk threshold TRa, it means there is a risk.

[0065] Preferably, the comprehensive evaluation and early warning decision-making module includes a risk score intelligent interpretation unit and a risk determination and response control unit;

[0066] The risk score intelligent interpretation unit comprehensively interprets according to the obtained risk score Rag, and combines the line-of-sight image blur Fb, the driver's steering action micro-response time Ft, and the driver's blink rate Fe to calculate and obtain a comprehensive score value Srisk;

[0067] The comprehensive score value Srisk is obtained through the following formula:

[0068] ;

[0069] In the formula, ln represents the logarithmic function, Fb represents the blurriness of the line-of-sight image, represents the normalized index of driver response delay, represents the blink rate deviation rate;

[0070] Normalized index of driver response delay is obtained through the following formula:

[0071] ;

[0072] In the formula, Ft represents the micro-response time of the driver's steering action, and Tt represents the reference steering response time

[0073] Blink rate deviation rate is obtained through the following formula:

[0074] ;

[0075] In the formula, pFe represents the normal reference rate of the driver's blink.

[0076] Preferably, the risk determination and response control unit compares the obtained comprehensive score value Srisk with a preset comprehensive threshold TSr, and triggers different levels of warning measures according to the judgment result;

[0077] The judgment result is obtained through the following matching method:

[0078] When the comprehensive score value Srisk ≤ the comprehensive threshold TSr, no alarm is triggered;

[0079] When the comprehensive score value Srisk > the comprehensive threshold TSr, an alarm is triggered, the alarm is divided into three levels, and corresponding measures WIN are generated;

[0080] The corresponding measures WIN are obtained through the following formula:

[0081] ;

[0082] In the formula, In1 and In2 represent the threshold adjustable increment intervals, and In2 is greater than In1.

[0083] Preferably, the dynamic feedback optimization module analyzes the fluctuation trend of the comprehensive score value Srisk within a fixed period, establishes a quantitative stability evaluation index: risk response parameter Ψstab, and a preset false alarm threshold Tst, and judges the false alarm state;

[0084] The risk response parameter Ψstab is obtained through the following formula:

[0085] ;

[0086] In the formula, m represents the number of scoring times within the observation period, Srisk,k represents the comprehensive scoring value of the k-th period, Srisk,k - 1 represents the comprehensive scoring value of the (k - 1)-th period, Δtk represents the time interval of the k-th scoring change, and ηk represents the k-th scoring error rate;

[0087] The false alarm state is obtained through the following matching method:

[0088] When the risk response parameter Ψstab ≤ the false alarm threshold Tst, it indicates that the model state is normal, the false alarm rate is normal, and the current configuration is maintained;

[0089] When the risk response parameter Ψstab > the false alarm threshold Tst, it indicates that the model state is abnormal, the scoring fluctuates, the false alarm rate is abnormal, and the coefficient ωj of the multi-modal risk aggregation model is readjusted to obtain the new coefficient nω;

[0090] The new coefficient nω is obtained through the following formula:

[0091] ;

[0092] In the formula, μ represents the adjustment coefficient and E represents a constant.

[0093] An accident warning analysis method based on driving data includes the following steps:

[0094] Step 1: The data acquisition and preprocessing module collects multi-dimensional driving data of the vehicle through the CV2328 four-in-one driving recorder and sensors, fits them into the original data set WQ, and performs preprocessing to obtain the driving data set WS;

[0095] Step 2: The abnormal behavior recognition module extracts abnormal behavior characteristics from the obtained driving data set WS, including the driving behavior fluctuation index Dar and the operation rhythm imbalance index Dos, and fits them into the abnormal behavior parameter YC;

[0096] Step 3: The environment complexity recognition module extracts road condition complexity characteristics from the obtained driving data set WS, including the road condition complexity coefficient Cen, the line-of-sight chaos coefficient Cvs, and the power interference coefficient Cdn, and fits them into the complexity parameter YZ;

[0097] Step 4: The multi-mode risk fusion evaluation module fuses the risk scores of the obtained abnormal behavior parameter YC and complexity parameter YZ to obtain the warning risk set R, and calculates the warning risk set R through the use of a multi-modal risk aggregation model to obtain the risk score Rag;

[0098] Step 5: The comprehensive evaluation and early warning decision-making module comprehensively interprets the obtained risk score Rag, obtains the comprehensive score value Srisk, and issues an early warning and generates corresponding measures WIN;

[0099] Step 6: The dynamic feedback optimization module interprets the change trend of the comprehensive score value Srisk within a fixed period, obtains the risk response parameter Ψstab, and adjusts the coefficient ωj of the multi-modal risk aggregation model to obtain the new coefficient nω.

[0100] The present invention provides an accident early warning analysis method and system based on driving data, having the following beneficial effects:

[0101] (1) During the operation of the system, by extracting the road condition complexity coefficient Cen, the line-of-sight chaos coefficient Cvs, and the power interference coefficient Cdn, the system of the present invention not only considers the current state in the modeling of the external environment, but also analyzes its dynamic interference characteristics, making the system still have high reliability in scenarios such as rain and fog, night, and congestion. The comprehensive score value Srisk is no longer a single binary determination, but combines information such as operation disturbance, attitude deviation, and fatigue index for multi-factor fusion interpretation, and outputs a response result WIN with hierarchical characteristics, which can achieve multi-level early warnings from voice reminders to active interventions.

[0102] The dynamic feedback optimization module constructs the risk response parameter Ψstab according to the score fluctuation trend, and dynamically adjusts the multi-modal model weight coefficient ωj to generate the new coefficient nω, realizing model self-tuning and adaptive reinforcement learning, and effectively solving the problem of "weak adaptability" of the system in different scenarios. Using the triple fusion strategy of driving behavior + environmental perception + aggregation algorithm, it realizes the recognition of a driving risk scenario closer to the real situation, improves the early warning accuracy and sensitivity of the system under complex conditions, and effectively reduces false alarms and missed alarms.

[0103] (2) Adopting an explicit linear fusion structure, reasonable weights are respectively assigned to each feature to realize the modeling of the differential contributions of different dimensional indicators. This method can better reflect the sensitivity and importance of each indicator in different situations compared with "average addition" or "hard threshold comparison", enhancing the decision-making flexibility and interpretability of the system.

[0104] Combining volatility analysis and rhythm imbalance analysis, the system can not only identify obvious abnormal operations such as sudden braking and yawing, but also identify hidden dangerous behaviors such as frequent blinking, response delay, and micro-amplitude correction, realizing earlier and more concealed risk prediction. By integrating multiple factors such as line-of-sight quality, target dynamics, acceleration perturbation, and driver state, the system can construct a comprehensive judgment model for complex environments, and still maintain stable complexity recognition performance especially in extreme driving situations such as rain and fog, high speed, and night.

[0105] (3) By setting up two - level functional modules of the risk score fusion unit and the multi - modal risk aggregation unit, a closed - loop process is realized from the scoring normalization fusion of the abnormal behavior parameter YC and the environmental complexity parameter YZ, to the weighted calculation aggregation, and then to the risk status output. This breaks through the problems of traditional systems relying only on a single threshold trigger and having a rough judgment mechanism, and constructs a fully structured risk fusion and response mechanism.

[0106] Through the standardization scoring and probability weighting mechanism, parameters with significantly different physical dimensions can be uniformly modeled, improving the system's understanding accuracy of multi - modal risk signals, and significantly reducing the fusion error and judgment deviation. By introducing the modal occurrence probability Pj, the system can adjust the influence of different modal risks in real - time according to the actual risk evolution situation, realize the dynamic capture of the "dominant risk factor" in complex scenarios, and enhance the agility of the overall risk perception.

[0107] (4) Through the risk score intelligent interpretation unit, on the basis of the original risk score, indicators such as the blurriness of the line - of - sight image, the driver's reaction delay, and the blink deviation are introduced to construct a comprehensive score value of multi - factor fusion. This structure enables the system to rise from a single - numerical judgment to a comprehensive interpretation of multi - modal behaviors and states, solving the problems of the traditional model's scoring results lacking semantic support and having weak interpretability. During the scoring calculation process, by combining the driver's micro - response time, physiological fatigue state, and the current visual environment quality, a construction method of collaborative scoring in the three domains of behavior - physiology - environment is realized, making the risk results more context - understandable and closer to the actual driving state. This effectively makes up for the modeling shortcoming of the existing warning system's lack of consideration of the driver's subjective factors. Description of the Drawings

[0108] Figure 1 It is a schematic diagram of the block - diagram process of an accident warning analysis system based on driving data according to the present invention;

[0109] Figure 2 It is a schematic diagram of the steps of an accident warning analysis method based on driving data according to the present invention;

[0110] Figure 3 It is a schematic diagram of the system process for judging whether the model of the present invention is normal;

[0111] Figure 4 It is a line graph of the abnormal behavior parameters of the present invention;

[0112] Figure 5 It is a schematic diagram of the structure of the driving recorder used in the present invention. Detailed Embodiments

[0113] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0114] Embodiment 1

[0115] The present invention provides an accident warning analysis system based on driving data. Please refer to Figures 1 to 5 , which includes a data collection and preprocessing module, an abnormal behavior recognition module, an environmental complexity recognition module, a multi-mode risk fusion evaluation module, a comprehensive evaluation and warning decision module, and a dynamic feedback optimization module;

[0116] The data collection and preprocessing module collects multi-dimensional driving data of the vehicle through a CV2328 four-in-one driving recorder and sensors, fits them into an original data set WQ, and performs preprocessing to obtain a driving data set WS;

[0117] The abnormal behavior recognition module extracts abnormal behavior features from the obtained driving data set WS, including a driving behavior fluctuation index Dar and an operation rhythm imbalance index Dos, and fits them into abnormal behavior parameters YC;

[0118] The environmental complexity recognition module extracts road condition complexity features from the obtained driving data set WS, including a road condition complexity coefficient Cen, a visual distance confusion coefficient Cvs, and a power interference coefficient Cdn, and fits them into complexity parameters YZ;

[0119] The multi-mode risk fusion evaluation module fuses the obtained abnormal behavior parameters YC and complexity parameters YZ to obtain a warning risk set R, and calculates the warning risk set R through the use of a multi-modal risk aggregation model to obtain a risk score Rag;

[0120] The comprehensive evaluation and warning decision module comprehensively interprets the obtained risk score Rag to obtain a comprehensive score value Srisk, and issues a warning and generates corresponding measures WIN;

[0121] The dynamic feedback optimization module interprets the change trend of the comprehensive score value Srisk within a fixed period to obtain a risk response parameter Ψstab, and adjusts the coefficient ωj of the multi-modal risk aggregation model to obtain a new coefficient nω.

[0122] In this embodiment, the two-dimensional fusion of the abnormal behavior parameter YC and the complexity parameter YZ is introduced to construct the early warning risk set RRR, and the multi-modal risk aggregation model is used for integrated scoring, effectively avoiding the problems of modal fragmentation and local judgment, and improving the comprehensive understanding ability of the system. In the abnormal behavior recognition module, by analyzing the driving behavior fluctuation index Dar and the operation rhythm imbalance index Dos, the driving behavior of the driver is analyzed for trend rather than instantaneous judgment, enabling the system to perform time-series prediction on potential risk behaviors and improving the defect of the lag in prediction of the traditional system.

[0123] By extracting the road condition complexity coefficient Cen, the line-of-sight chaos coefficient Cvs, and the power interference coefficient Cdn, the system of the present invention not only considers the current state in the modeling of the external environment, but also analyzes its dynamic interference characteristics, enabling the system to still have high reliability in scenarios such as rain, fog, night, and congestion. The comprehensive score value Srisk is no longer a single binary determination, but rather a multi-factor fusion interpretation by combining information such as operation perturbation, attitude deviation, and fatigue index, and a response result WIN with hierarchical characteristics is output, enabling multi-level early warning from voice reminder to active intervention.

[0124] The dynamic feedback optimization module constructs the risk response parameter Ψstab according to the scoring fluctuation trend and dynamically adjusts the weight coefficient ωj of the multi-modal model to generate a new coefficient nω, realizing self-parameter adjustment and adaptive reinforcement learning of the model, and effectively solving the problem of "weak adaptability" of the system in different scenarios. Using the triple fusion strategy of driving behavior + environmental perception + aggregation algorithm, the identification of a driving risk scenario closer to the real situation is realized, improving the early warning accuracy and sensitivity of the system under complex conditions and effectively reducing false alarms and missed alarms.

[0125] The analysis of the abnormal behavior modeling and the scoring change trend enables the system to identify the risk evolution trend in advance, and the early warning intervenes one step earlier, truly possessing the timeliness characteristics of "prediction + early warning". The environmental complexity identification module enables the system to have the ability to cope with complex, variable, and uncertain environments, no longer being limited to ideal or single scenarios. The risk scoring fluctuation analysis and the dynamic weight optimization mechanism can ensure that the system adjusts the decision-making logic in a timely manner when the scoring fluctuates abnormally, improving the robustness of the overall risk perception and the stability of the system operation. Through the dynamic feedback mechanism, the system has the ability to automatically optimize the parameter structure according to historical operation data, can perform adaptive adjustment according to different driver habits and road types, and realizes customized and personalized safety assistance. From data collection, behavior recognition, environmental modeling, to risk assessment, early warning response, and then to model optimization feedback, an intelligent early warning analysis system with high coupling, high closed-loop, and high controllability is constructed, promoting the intelligent assisted driving system to move towards high-order automatic decision-making.

[0126] Embodiment 2

[0127] This embodiment is an explanatory description carried out in Embodiment 1. Please refer to Figure 1 and Figure 5 , specifically: The data acquisition and preprocessing module includes a multi-source data acquisition unit and a standardization preprocessing unit;

[0128] The multi-source data acquisition unit collects multi-dimensional driving data of the vehicle, including the vehicle vertical vibration frequency Fv, the driver's steering action micro-response time Ft, the line-of-sight image blur degree Fb, the micro-speed change rate Fp of the front target, and the driver's blink rate Fe, and performs fitting to obtain the original data set WQ;

[0129] Among them, the vehicle vertical vibration frequency Fv and the instantaneous value alt of the vehicle lateral acceleration are collected through the IMU sensor built in the tachograph;

[0130] The driver's steering action micro-response time Ft is collected through a steering angle sensor;

[0131] The line-of-sight image blur degree Fb and the distance df to the vehicle ahead are collected through the high-definition camera in the tachograph;

[0132] The micro-speed change rate Fp of the front target is collected through a millimeter-wave radar;

[0133] The driver's blink rate Fe is collected through an in-vehicle camera;

[0134] The standardization preprocessing unit performs outlier removal and normalization processing on the original data set WQ to obtain the driving data set WS;

[0135] Outlier removal includes using the three-standard-deviation method to process the original data set WQ and removing outliers;

[0136] The normalization processing uses the maximum-minimum normalization method to process the original data set WQ to obtain the driving data set WS;

[0137] The driving data set WS is obtained through the following formula:

[0138] ;

[0139] In the formula, WSb represents the b-th data in the driving data set WS, WQb represents the b-th data in the original data set WQ, minWQb represents the valley value of the b-th data in the original data set WQ, and maxWQb represents the peak value of the b-th data in the original data set WQ.

[0140] In this embodiment, by setting up a multi-source data acquisition unit, five major types of parameters, namely the vehicle vertical vibration frequency Fv, the driver's steering action micro-response time Ft, the line-of-sight image blur degree Fb, the micro-speed change rate Fp of the front target, and the driver's blink rate Fe, are introduced simultaneously at the data level for the first time. Compared with the traditional single structure that only focuses on vehicle speed or acceleration, the richness of the original data and the risk description ability are greatly improved, realizing the three-domain co-acquisition of "behavior - physiology - environment", and providing more expressive data support for subsequent risk modeling. The data acquisition device clearly corresponds to highly adaptable physical sensors, realizing the structural closure and clear source of the acquisition link, avoiding data deviation, improving the stability and controllability of the original data, and providing sensor-level support for building a highly reliable model in a real scenario.

[0141] In the standardized preprocessing unit of this embodiment, the three-sigma method is used to remove outliers from the original data set WQWQWQ. This statistical noise filtering mechanism can effectively shield interference factors such as equipment false alarms and sampling jumps, and is more robust than the traditional "taking the mean + direct modeling" method, laying a foundation for high-precision modeling. The maximum-minimum normalization method is used to standardize various types of driving data, enabling parameters with different physical units to be compared and fused on a unified scale, improving the model training efficiency and the interpretability of the risk analysis results, and enhancing the cross-scenario adaptation ability of the system.

[0142] By introducing a multi-source and multi-modal acquisition mechanism, a cross-level and multi-dimensional panoramic driving state data is constructed, providing high-density and high-dimensional original information support for abnormal behavior analysis and environmental complexity modeling. The outlier removal and normalization mechanisms are introduced at the acquisition stage, effectively excluding extreme errors, environmental noise, and hardware anomalies, significantly improving the stability, effectiveness, and fitting accuracy of the modeling data. The normalized data set WS enables the model to perform unified risk score calculations based on different types and dimensions of data, avoiding risk judgment errors caused by inconsistent units or numerical scale deviations.

[0143] Embodiment 3

[0144] This embodiment is an explanatory description based on Embodiment 2. Please refer to Figure 1 and Figure 4 , specifically: The abnormal behavior recognition module includes an abnormal behavior feature extraction unit and an abnormal feature fusion unit;

[0145] The abnormal behavior feature extraction unit extracts abnormal behavior features from the driving data set WS, including the driving behavior fluctuation index Dar and the operation rhythm imbalance index Dos;

[0146] The driving behavior fluctuation index Dar is obtained through the following formula:

[0147] ;

[0148] Wherein, Tp represents the standardized reference value of the micro velocity change rate of the front target, and Te represents the standardized reference value of the driver's blinking rate.

[0149] The operation rhythm imbalance index Dos is obtained through the following formula:

[0150] ;

[0151] Wherein, alt represents the instantaneous value of the vehicle's lateral acceleration, palt represents the average value of the lateral acceleration, and Fv represents the vehicle's vertical vibration frequency;

[0152] The abnormal feature fusion unit fuses the driving behavior fluctuation index Dar and the operation rhythm imbalance index Dos, and calculates and obtains the abnormal behavior parameter YC, as shown in Table 1 specifically;

[0153] The abnormal behavior parameter YC is obtained through the following formula:

[0154] ;

[0155] Wherein, respectively represent the preset weight values of the driving behavior fluctuation index Dar and the operation rhythm imbalance index Dos.

[0156] Specific example:

[0157] Set to be 0.6 and 0.4 respectively;

[0158] Obtain the driving behavior fluctuation index Dar = 0.50;

[0159] The operation rhythm imbalance index Dos = 0.20;

[0160] Calculate and obtain the abnormal behavior parameter YC:

[0161] ;

[0162] Table 1: Calculation sample table of abnormal behavior parameters;

[0163]

[0164] The environmental complexity recognition module includes a road condition complexity analysis unit and a complexity parameter fusion unit;

[0165] The road condition complexity analysis unit extracts road condition complexity features from the driving data set WS, including the road condition complexity coefficient Cen, the sight distance confusion coefficient Cvs, and the power interference coefficient Cdn;

[0166] The road condition complexity coefficient Cen is obtained through the following formula:

[0167] ;

[0168] In the formula, Fb represents the line-of-sight image blur, Fp represents the micro velocity change rate of the front target, df represents the distance to the vehicle ahead, and alt represents the instantaneous value of the vehicle's lateral acceleration;

[0169] The line-of-sight confusion coefficient Cvs is obtained through the following formula:

[0170] ;

[0171] In the formula, E represents a constant, and Fb represents the line-of-sight image blur;

[0172] The dynamic interference coefficient Cdn is obtained through the following formula:

[0173] ;

[0174] In the formula, Ft represents the micro response time of the driver's steering action, and Fe represents the driver's blink rate;

[0175] The complexity parameter fusion unit fuses the obtained road condition complexity coefficient Cen, line-of-sight confusion coefficient Cvs, and dynamic interference coefficient Cdn, and calculates and obtains the complexity parameter YZ;

[0176] The complexity parameter YZ is obtained through the following formula:

[0177] ;

[0178] In the formula, respectively represent the preset weight values of the road condition complexity coefficient Cen, line-of-sight confusion coefficient Cvs, and dynamic interference coefficient Cdn.

[0179] In this embodiment, the abnormal behavior recognition module and the environmental complexity recognition module are further structured, and two-level logics of a feature extraction unit + a fusion unit are respectively set, making the extraction and fusion processes of each type of information more systematic and accurate, effectively avoiding problems such as information dispersion, structural duplication, and isolated judgment, and strengthening the module cooperation efficiency.

[0180] The extraction of abnormal behavior features is no longer limited to driving operation data, and the micro velocity change rate Fp of the front target and the driver's blink rate Fe are also introduced. Starting from two perspectives of the dynamic response of the driving environment and the physiological state of the driver respectively, the driving behavior fluctuation index Dar and the operation rhythm imbalance index Dos are constructed. Compared with the traditional model that only focuses on operation behaviors such as steering and acceleration, this design is more three-dimensional and physiologically relevant, enabling the system to identify potential abnormal states such as fatigue driving, slow operation, and vision judgment deviation in advance.

[0181] The environmental complexity parameter no longer only considers traditional distance and traffic flow density, but also proposes: the road condition complexity coefficient Cen: considering the image blur Fb, the micro velocity change rate Fp of the front target, the lateral acceleration alt, and the distance df to the vehicle ahead, forming a three - element fusion modeling of driving vision - dynamic behavior - space - time distance; the line - of - sight chaos coefficient Cvs: aiming at scenarios such as image defocus, night, rain and fog weather, strengthening the dynamic interference assessment of visual blur;

[0182] The power interference coefficient Cdn: the micro response time Ft of the driver's steering action and the driver's blink rate Fe, introducing a non - linear modeling perspective of the interaction between the driver's state and vehicle dynamics. These coefficients are fused through different dimensions to establish a complexity recognition system that better conforms to the fluctuations of the actual driving scenario. In the calculation of the abnormal behavior parameter YC and the environmental complexity parameter YZ, this embodiment adopts an explicit linear fusion structure, respectively assigning reasonable weights to each feature to realize the differential contribution modeling of different - dimension indicators. This method can better reflect the sensitivity and importance of each indicator in different situations compared with "average addition" or "hard - threshold comparison", enhancing the decision - making flexibility and interpretability of the system.

[0183] Combining volatility analysis and rhythm imbalance analysis, the system can not only identify obvious abnormal operations such as sudden braking and yawing, but also identify hidden dangerous behaviors such as frequent blinking, response delay, and micro - amplitude correction, realizing earlier and more concealed risk prediction. By fusing multiple factors such as line - of - sight quality, target dynamics, acceleration perturbation, and driver state, the system can construct a comprehensive judgment model for complex environments, and still maintain stable complexity recognition performance especially in extreme driving scenarios such as rain and fog, high - speed, and night.

[0184] Embodiment 4

[0185] This embodiment is an explanatory note based on Embodiment 3. Please refer to Figure 1 and Figure 3 , specifically: the multi - mode risk fusion assessment module includes a risk score fusion unit and a multi - modal risk aggregation unit;

[0186] The risk score fusion unit fuses the obtained abnormal behavior parameter YC and complexity parameter YZ to obtain an early - warning risk set R;

[0187] The early - warning risk set R is obtained through the following formula:

[0188] ;

[0189] In the formula, R YC represents the score of the abnormal behavior parameter, R YZ represents the score of the complexity parameter, maxYC represents the maximum value of the abnormal behavior parameter, and maxYZ represents the maximum value of the complexity parameter.

[0190] Based on the obtained early warning risk set R, the multimodal risk aggregation unit calculates the early warning risk set R by using the multimodal risk aggregation model to obtain the risk score Rag, and compares it with the preset risk threshold TRa to judge the risk status;

[0191] The risk score Rag is obtained through the following formula:

[0192] ;

[0193] In the formula, Z represents the normalization factor, N represents the number of modalities in the early warning risk set R, including the score R of the abnormal behavior parameter YC and the score R of the complexity parameter YZ , Rj represents the j-th modality in the early warning risk set R, ln represents the logarithmic function, represents the weight coefficient of the j-th modality, and Pj represents the occurrence probability of the j-th modality;

[0194] The risk status is obtained through the following matching method:

[0195] When the risk score Rag ≤ the risk threshold TRa, it means there is no risk;

[0196] When the risk score Rag > the risk threshold TRa, it means there is a risk.

[0197] In this embodiment, by setting two-level functional modules of the risk score fusion unit and the multimodal risk aggregation unit, a closed-loop process from score normalization fusion, weighted calculation aggregation to risk status output of the abnormal behavior parameter YC and the environmental complexity parameter YZ is realized, breaking through the problems of traditional systems relying only on a single threshold trigger and having a rough judgment mechanism, and building a fully structured risk fusion and response mechanism.

[0198] Through the standardized scoring and probability weighting mechanism, parameters with significantly different physical dimensions can be uniformly modeled, improving the system's understanding accuracy of multimodal risk signals, and significantly reducing the fusion error and judgment deviation. By introducing the modality occurrence probability Pj, the system can dynamically adjust the influence of different modality risks according to the actual risk evolution situation, realize the dynamic capture of the "dominant risk factor" in complex scenarios, and improve the agility of overall risk perception.

[0199] The risk scoring mechanism has changed from a "single indicator" to a "weighted fusion + threshold matching" mode, enabling the system to have a more delicate risk identification logic, support richer response strategies, and be further integrated with risk level models (such as mild / moderate / high risk). The structure of the multi-modal risk aggregation formula is open, and the parameters are extensible (such as introducing more risk modalities), supporting the access of new modal variables such as fatigue driving score and attention status score, making the system have good iterative upgrade and cross-scenario application capabilities. The risk score Rag, as the final fusion judgment result of the system, can be compared with the actual triggering event in the dynamic feedback optimization module, serving as the basic reference index for weight correction and threshold dynamic adjustment, and helping to build a closed-loop self-learning early warning system.

[0200] Embodiment 5

[0201] This embodiment is an explanatory description carried out in Embodiment 4. Please refer to Figure 1 and Figure 3 , specifically: The comprehensive evaluation and early warning decision-making module includes a risk score intelligent interpretation unit and a risk determination and response control unit;

[0202] The risk score intelligent interpretation unit comprehensively interprets according to the obtained risk score Rag, and combines the line-of-sight image blurriness Fb, the driver's micro-response time Ft of the steering action, and the driver's blinking rate Fe to calculate and obtain the comprehensive score value Srisk.

[0203] The comprehensive score value Srisk is obtained through the following formula:

[0204] ;

[0205] In the formula, ln represents the logarithmic function, Fb represents the line-of-sight image blurriness, represents the driver's reaction delay normalization index, represents the blinking rate deviation rate;

[0206] Driver's reaction delay normalization index is obtained through the following formula:

[0207] ;

[0208] In the formula, Ft represents the driver's micro-response time of the steering action, and Tt represents the reference steering response time

[0209] Blinking rate deviation rate is obtained through the following formula:

[0210] ;

[0211] In the formula, pFe represents the normal reference blinking rate of the driver.

[0212] The risk determination and response control unit compares the obtained comprehensive score value Srisk with the preset comprehensive threshold TSr, and triggers warning measures of different levels according to the judgment result;

[0213] The judgment result is obtained by matching in the following way:

[0214] When the comprehensive score value Srisk ≤ the comprehensive threshold TSr, no alarm is triggered;

[0215] When the comprehensive score value Srisk > the comprehensive threshold TSr, an alarm is triggered. The alarm is divided into three levels, and corresponding measures WIN are generated;

[0216] The corresponding measures WIN are obtained by the following formula:

[0217] ;

[0218] In the formula, In1 and In2 represent the threshold adjustable increment interval, and In2 is greater than In1.

[0219] The dynamic feedback optimization module analyzes the fluctuation trend of the comprehensive score value Srisk within a fixed period, establishes a quantitative stability evaluation index: the risk response parameter Ψstab, and a preset false alarm threshold Tst to judge the false alarm state;

[0220] The risk response parameter Ψstab is obtained by the following formula:

[0221] ;

[0222] In the formula, m represents the number of scoring times within the observation period, Srisk,k represents the comprehensive score value in the k-th period, Srisk,k - 1 represents the comprehensive score value in the (k - 1)-th period, Δtk represents the time interval of the k-th score change, and ηk represents the k-th score error rate;

[0223] The false alarm state is obtained by matching in the following way:

[0224] When the risk response parameter Ψstab ≤ the false alarm threshold Tst, it means that the model state is normal, the false alarm rate is normal, and the current configuration is maintained;

[0225] When the risk response parameter Ψstab > the false alarm threshold Tst, it means that the model state is abnormal, the score fluctuates, the false alarm rate is abnormal, and the coefficient ωj of the multi-modal risk aggregation model is readjusted to obtain the new coefficient nω;

[0226] The new coefficient nω is obtained by the following formula:

[0227] ;

[0228] In the formula, μ represents the adjustment coefficient, and E represents a constant.

[0229] In this embodiment, through the risk score intelligent interpretation unit, on the basis of the original risk score, the line-of-sight image blur degree, driver response delay index, and blink offset index are introduced to construct a comprehensive score value with multi-factor fusion. This structure enables the system to rise from single-value judgment to the comprehensive interpretation of multi-modal behaviors and states, solving the problems of the lack of semantic support and weak interpretability of the scoring results of traditional models. During the scoring calculation process, the driver's micro-response time, physiological fatigue state, and the current visual environment quality are combined to realize a construction method for the collaborative scoring of the three domains of behavior-physiology-environment, making the risk results more contextually understandable and closer to the actual driving state. This approach effectively makes up for the modeling shortcoming of the lack of consideration of the driver's subjective factors in existing warning systems.

[0230] The risk determination and response control unit of this embodiment sets a three-level warning trigger mechanism, which can dynamically adjust the response intensity according to the amplitude of the comprehensive score value to achieve a hierarchical control method from "voice prompt" to "visual reminder" and then to "automatic control". This progressive response mechanism is more user-friendly and scene-adaptive in human-computer interaction than the traditional binary trigger model, reducing driving interference and improving the response matching degree. The system sets up a risk response parameter Ψstab to measure the change trend and false alarm state of the score value within a fixed period, and constructs a stability index system for the scoring curve. This mechanism realizes the evaluation and intervention of the performance of the risk scoring system for the first time, provides the system with the ability of "self-perception", and breaks through the technical limitation that traditional models cannot judge the source of scoring errors.

[0231] The comprehensive score value is calculated jointly by multiple key parameters, combined with the driver's state and environmental perception, making the risk score no longer an isolated value, but having context information and behavioral causal relationships, improving the system transparency and driver trust. Through the hierarchical risk determination mechanism, the system can adapt to the scene requirements under different risk intensities, effectively avoid false alarms and over-warnings, and improve the system usage experience and the rationality of response strategies. Through the evaluation of the scoring fluctuation trend and the false alarm rate judgment, the system can timely identify the abnormal performance of the scoring model, thereby avoiding misjudgment results under abnormal scoring conditions, and significantly enhancing the system robustness and reliability. When there is an error offset in the scoring system, the risk aggregation model can be automatically corrected through the dynamic feedback mechanism to ensure that the system continuously maintains high performance operation under different driving habits and environmental conditions, and has strong generalization ability and intelligent evolution ability.

[0232] Embodiment 6

[0233] An accident warning analysis method based on driving data, please refer to Figure 2 , specifically: including the following steps:

[0234] Step 1: The data acquisition and preprocessing module collects multi-dimensional driving data of the vehicle through the CV2328 four-in-one driving recorder and sensors, fits them into the original dataset WQ, and performs preprocessing to obtain the driving dataset WS;

[0235] Step 2: The abnormal behavior recognition module extracts abnormal behavior features from the obtained driving dataset WS, including the driving behavior fluctuation index Dar and the operation rhythm imbalance index Dos, and fits them into the abnormal behavior parameter YC;

[0236] Step 3: The environment complexity recognition module extracts road condition complexity features from the obtained driving dataset WS, including the road condition complexity coefficient Cen, the visual distance confusion coefficient Cvs, and the power interference coefficient Cdn, and fits them into the complexity parameter YZ;

[0237] Step 4: The multi-mode risk fusion evaluation module fuses the abnormal behavior parameter YC and the complexity parameter YZ obtained to obtain the early warning risk set R, and calculates the early warning risk set R through the use of a multi-modal risk aggregation model to obtain the risk score Rag;

[0238] Step 5: The comprehensive evaluation and early warning decision-making module comprehensively interprets the obtained risk score Rag to obtain the comprehensive score value Srisk, and issues an early warning and generates corresponding measures WIN;

[0239] Step 6: The dynamic feedback optimization module interprets the change trend of the comprehensive score value Srisk within a fixed period to obtain the risk response parameter Ψstab, and adjusts the coefficient ωj of the multi-modal risk aggregation model to obtain the new coefficient nω.

[0240] In this embodiment, the method systematically divides the driving accident early warning process into six major steps: data acquisition → abnormal recognition → environmental analysis → risk fusion → early warning decision-making → dynamic optimization, forming a full-process closed-loop design from perception, evaluation to response and feedback. It solves the problems of process fragmentation, single evaluation dimension, and missing feedback link in traditional methods.

[0241] In the data acquisition stage, this method is based on the CV2328 four-in-one driving recorder, integrates multi-class data such as vehicle dynamics, driver status and external environment, and constructs a multi-dimensional original dataset. This multi-source information integrated acquisition mode significantly improves the breadth and fine-grainedness of the original data, providing a high-dimensional and highly expressive input basis for the subsequent model. By extracting the driving behavior fluctuation index and the operation rhythm imbalance index, in-depth modeling of the driver's control stability and physiological rhythm changes is realized, which can not only identify obvious abnormalities, but also identify hidden behavior abnormalities, significantly improving the system's early prediction ability for potential risks.

[0242] In the risk assessment stage, a two-layer structure of first normalizing and fusing and then aggregating and modeling is adopted. The risk score is calculated by combining the occurrence probability and weighted contribution of risk modes, avoiding the problems of "opaque scoring source and unbalanced modal influence" in traditional methods, and improving the interpretability and logical reliability of the score.

[0243] In the early warning decision-making stage, a hierarchical output mechanism is introduced. Different levels of response measures are intelligently matched according to the comprehensive score value, realizing a progressive intervention from mild reminder to system control, greatly improving the human-computer interaction experience and driver acceptance, and at the same time avoiding driving dependence or boredom caused by over-intervention. It realizes the fusion processing and deep evaluation of multi-dimensional risk information, improves the overall risk identification ability and response accuracy of the system; constructs a complete closed-loop intelligent early warning method architecture, which is highly efficient and collaborative in the whole link from bottom-layer data collection to upper-layer intelligent control, and has strong engineering feasibility and commercial implementation value.

[0244] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An accident warning analysis system based on driving data, characterized in that: It includes data collection and preprocessing module, abnormal behavior recognition module, environmental complexity recognition module, multi-mode risk fusion assessment module, comprehensive assessment and early warning decision module and dynamic feedback optimization module; The data acquisition and preprocessing module collects the multi-dimensional driving data of the car through the CV2328 four-in-one driving recorder and sensors, fits it into the original data set WQ, and performs preprocessing to obtain the driving data set WS; The abnormal behavior recognition module extracts abnormal behavior features from the acquired driving data set WS, including the driving behavior fluctuation index Dar and the operating rhythm imbalance index Dos, and fits them into abnormal behavior parameters YC; The environment complexity recognition module extracts the road condition complexity features from the acquired driving data set WS, including the road condition complexity coefficient Cen, the sight range confusion coefficient Cvs and the dynamic interference coefficient Cdn, and fits them into the complexity parameter YZ; The road condition complexity analysis unit extracts road condition complexity features from the driving data set WS, including the road condition complexity coefficient Cen, the sight range confusion coefficient Cvs and the dynamic interference coefficient Cdn; The road complexity coefficient Cen is obtained by the following formula: ； In the formula, Fb represents the blur of the visual range image, Fp represents the rate of change of the front target micro-speed, df represents the distance to the front vehicle, and alt represents the instantaneous value of the vehicle's lateral acceleration; The line-of-sight confusion coefficient Cvs is obtained by the following formula: ； In the formula, E represents a constant, and Fb represents the blur of the viewing image; The dynamic interference coefficient Cdn is obtained by the following formula: ； Where, Ft represents the micro-response time of the driver's steering action, and Fe represents the driver's blinking rate; The complexity parameter fusion unit fuses the acquired road condition complexity coefficient Cen, sight range confusion coefficient Cvs and power interference coefficient Cdn to calculate and obtain the complexity parameter YZ; The complexity parameter YZ is obtained by the following formula: ； In the formula, Respectively represent the preset weight values ​​of the road complexity coefficient Cen, the sight range confusion coefficient Cvs and the dynamic interference coefficient Cdn; The multi-modal risk fusion assessment module fuses the acquired abnormal behavior parameter YC and complexity parameter YZ with risk scores to obtain the warning risk set R, and calculates the warning risk set R using the multi-modal risk aggregation model to obtain the risk score Rag; The comprehensive assessment and early warning decision module comprehensively interprets the acquired risk score Rag, obtains the comprehensive score value Srisk, issues an early warning and generates corresponding measures WIN; The dynamic feedback optimization module interprets the changing trend of the comprehensive score value Srisk within a fixed period, obtains the risk response parameter Ψstab, and adjusts the coefficient ωj of the multimodal risk aggregation model to obtain a new coefficient nω.

2. The accident warning analysis system based on driving data according to claim 1, characterized in that: The data acquisition and preprocessing module includes a multi-source data acquisition unit and a standardized preprocessing unit; The multi-source data acquisition unit collects multi-dimensional driving data of the car, including the vehicle vertical vibration frequency Fv, the driver's steering action micro-response time Ft, the visual range image ambiguity Fb, the front target micro-speed change rate Fp, the driver's blink rate Fe, the front vehicle distance df and the vehicle's lateral acceleration instantaneous value alt, and performs fitting to obtain the original data set WQ; Among them, the vehicle vertical vibration frequency Fv and the vehicle lateral acceleration instantaneous value alt are collected by the IMU sensor built into the driving recorder; The driver's steering action micro-response time Ft is collected through the steering angle sensor; The blur of the sight image Fb and the distance to the preceding vehicle df are collected by a high-definition camera in the driving recorder; The micro-speed change rate Fp of the target ahead is collected by millimeter-wave radar; The driver’s blink rate Fe is collected through the in-car camera; The standardization preprocessing unit removes outliers and normalizes the original data set WQ to obtain the driving data set WS; Outlier removal includes processing the original data set WQ using the triple standard deviation method to remove outliers; Normalization processing uses the maximum and minimum normalization method to process the original data set WQ to obtain the driving data set WS; The driving dataset WS is obtained by the following formula: ； Wherein, WSb represents the b-th data item in the driving dataset WS, WQb represents the b-th data item in the original dataset WQ, minWQb represents the valley value of the b-th data item in the original dataset WQ, and maxWQb represents the peak value of the b-th data item in the original dataset WQ.

3. The accident warning analysis system based on driving data according to claim 2, characterized in that: The abnormal behavior recognition module includes an abnormal behavior feature extraction unit and an abnormal feature fusion unit; The abnormal behavior feature extraction unit extracts abnormal behavior features from the driving data set WS, including the driving behavior fluctuation index Dar and the operating rhythm imbalance index Dos; The driving behavior fluctuation index Dar is obtained by the following formula: ； Where Tp represents the standardized reference value of the speed change rate of the target ahead, and Te represents the standardized reference value of the driver's blink rate; The operating rhythm imbalance index Dos is obtained by the following formula: ； In the formula, alt represents the instantaneous value of the vehicle's lateral acceleration, palt represents the average value of the lateral acceleration, and Fv represents the vertical vibration frequency of the vehicle; The abnormal feature fusion unit fuses the driving behavior fluctuation index Dar and the operation rhythm imbalance index Dos to calculate and obtain the abnormal behavior parameter YC; The abnormal behavior parameter YC is obtained by the following formula: ； In the formula, They respectively represent the preset weight values ​​of the driving behavior fluctuation index Dar and the operating rhythm imbalance index Dos.

4. The accident warning analysis system based on driving data according to claim 1, characterized in that: The multimodal risk fusion assessment module includes a risk score fusion unit and a multimodal risk aggregation unit; The risk score fusion unit fuses the acquired abnormal behavior parameter YC and complexity parameter YZ to obtain the warning risk set R; The early warning risk set R is obtained by the following formula: ； In the formula, R YC represents the score of abnormal behavior parameters, R YZ represents the score of the complexity parameter, maxYC represents the maximum value of the abnormal behavior parameter, and maxYZ represents the maximum value of the complexity parameter.

5. The accident warning analysis system based on driving data according to claim 4, characterized in that: The multimodal risk aggregation unit calculates the warning risk set R based on the obtained warning risk set R by using the multimodal risk aggregation model to obtain the risk score Rag, and compares it with the preset risk threshold TRa to determine the risk status; The risk score Rag is obtained by the following formula: ； Where Z represents the normalization factor, N represents the number of modes in the early warning risk set R, including the score R of the abnormal behavior parameter YC and the score R of the complexity parameter YZ , Rj represents the jth mode in the early warning risk set R, ln represents the logarithmic function, represents the weight coefficient of the jth mode, and Pj represents the occurrence probability of the jth mode; The risk status is obtained by matching: When the risk score Rag ≤ the risk threshold TRa, it means there is no risk; When the risk score Rag>risk threshold TRa, it indicates that there is a risk.

6. The accident warning analysis system based on driving data according to claim 5, characterized in that: The comprehensive assessment and early warning decision module includes a risk scoring intelligent interpretation unit and a risk determination and response control unit; The risk score intelligent interpretation unit performs a comprehensive interpretation based on the obtained risk score Rag, combined with the blur of the sight image Fb, the driver's steering action micro-response time Ft and the driver's blink rate Fe, and calculates the comprehensive score value Srisk; The comprehensive score Srisk is obtained by the following formula: ； In the formula, ln represents the logarithmic function, Fb represents the blur of the viewing distance image, represents the normalized index of driver reaction delay, represents the blink rate deviation rate; Normalized driver reaction delay index Obtained by the following formula: ； Where Ft represents the micro-response time of the driver's steering action, and Tt represents the reference steering response time. Blink rate offset Obtained by the following formula: ； Where pFe is the normal reference rate of the driver's blinking.

7. The accident warning analysis system based on driving data according to claim 6, characterized in that: The risk determination and response control unit compares the obtained comprehensive score value Srisk with the preset comprehensive threshold TSr, and triggers different levels of early warning measures according to the judgment results; The judgment result is obtained by matching in the following ways: When the comprehensive score value Srisk ≤ the comprehensive threshold TSr, no alarm is triggered; When the comprehensive score value Srisk>comprehensive threshold TSr, an alarm is triggered, the alarm is divided into three levels, and the corresponding measures WIN are generated; The corresponding measure WIN is obtained by the following formula: ； Wherein, In1 and In2 represent the adjustable increment range of the threshold, and In2 is greater than In1.

8. The accident warning analysis system based on driving data according to claim 7, characterized in that: The dynamic feedback optimization module analyzes the fluctuation trend of the comprehensive score value Srisk within a fixed period, establishes a quantitative stability evaluation indicator: the risk response parameter Ψstab, and the preset false alarm threshold Tst to judge the false alarm state; The risk response parameter Ψstab is obtained by the following formula: ； Where m represents the number of ratings in the observation period, Srisk,k represents the comprehensive rating value of the kth period, Srisk,k-1 represents the comprehensive rating value of the k-1th period, Δtk represents the time interval of the kth rating change, and ηk represents the kth rating error rate; The false alarm status is obtained by matching the following methods: When the risk response parameter Ψstab ≤ the false alarm threshold Tst, it means that the model status is normal, the false alarm rate is normal, and the current configuration is maintained; When the risk response parameter Ψstab>false alarm threshold Tst, it means that the model state is abnormal, the score fluctuates, the false alarm rate is abnormal, and the coefficient ωj of the multimodal risk aggregation model is readjusted to obtain a new coefficient nω; The new coefficient nω is obtained by the following formula: ； Wherein, μ represents the adjustment coefficient and E represents the constant.

9. An accident warning analysis method based on driving data, applied to an accident warning analysis system based on driving data as claimed in any one of claims 1 to 8, characterized in that: The following steps are involved: Step 1: The data acquisition and preprocessing module collects the multi-dimensional driving data of the car through the CV2328 four-in-one driving recorder and sensors, fits it into the original data set WQ, and performs preprocessing to obtain the driving data set WS; Step 2: The abnormal behavior recognition module extracts abnormal behavior features from the acquired driving data set WS, including the driving behavior fluctuation index Dar and the operating rhythm imbalance index Dos, and fits them into abnormal behavior parameters YC; Step 3: The environment complexity recognition module extracts the road condition complexity features from the acquired driving data set WS, including the road condition complexity coefficient Cen, the sight range confusion coefficient Cvs and the dynamic interference coefficient Cdn, and fits them into the complexity parameter YZ; The road condition complexity analysis unit extracts road condition complexity features from the driving data set WS, including the road condition complexity coefficient Cen, the sight range confusion coefficient Cvs and the dynamic interference coefficient Cdn; The road complexity coefficient Cen is obtained by the following formula: ； In the formula, Fb represents the blur of the visual range image, Fp represents the rate of change of the front target micro-speed, df represents the distance to the front vehicle, and alt represents the instantaneous value of the vehicle's lateral acceleration; The line-of-sight confusion coefficient Cvs is obtained by the following formula: ； In the formula, E represents a constant, and Fb represents the blur of the viewing image; The dynamic interference coefficient Cdn is obtained by the following formula: ； Where, Ft represents the micro-response time of the driver's steering action, and Fe represents the driver's blinking rate; The complexity parameter fusion unit fuses the acquired road condition complexity coefficient Cen, sight range confusion coefficient Cvs and power interference coefficient Cdn to calculate and obtain the complexity parameter YZ; The complexity parameter YZ is obtained by the following formula: ； In the formula, Respectively represent the preset weight values ​​of the road complexity coefficient Cen, the sight range confusion coefficient Cvs and the dynamic interference coefficient Cdn; Step 4: The multi-modal risk fusion assessment module fuses the acquired abnormal behavior parameter YC and complexity parameter YZ with risk scores to obtain the warning risk set R, and calculates the warning risk set R using the multi-modal risk aggregation model to obtain the risk score Rag; Step 5: The comprehensive assessment and early warning decision module comprehensively interprets the obtained risk score Rag, obtains the comprehensive score value Srisk, issues an early warning and generates corresponding measures WIN; Step 6: The dynamic feedback optimization module interprets the changing trend of the comprehensive score value Srisk within a fixed period, obtains the risk response parameter Ψstab, and adjusts the coefficient ωj of the multimodal risk aggregation model to obtain a new coefficient nω.

Images