Road vehicle traffic safety accident risk identification and early warning method

Abstract

The invention relates to the technical field of vehicles, and provides a road vehicle traffic safety accident risk identification and early warning method, which comprises the steps of collecting, sorting and processing historical data, real-time data and external data to obtain to-be-analyzed data; extracting time features, spatial features and behavior features from the to-be-analyzed data; the time features, the spatial features and the behavior features are spliced into a three-dimensional fusion tensor, and the three-dimensional fusion tensor is decoupled into a spatial-temporal feature matrix and dynamic sequence data after being subjected to three-dimensional convolution and spatial attention modulation; inputting the spatial-temporal feature matrix and the dynamic sequence data into a hybrid model of a convolutional neural network and a recurrent neural network for training to obtain a trained risk prediction model; based on the risk prediction model, analyzing the safety risk when the road vehicle travels, and adjusting the pre-tightening force of the safety belt and the unfolding parameters of the safety air bag according to the risk level obtained through analysis. Based on the scheme, the safety of the driver can be improved.

Description

Technical Field

[0001] The present invention relates to the field of vehicle technology, and in particular to a method for identifying and warning of road vehicle traffic safety accident risks. Background Art

[0002] With the continuous increase in road traffic volume, the incidence of traffic accidents has increased year by year, causing huge economic losses and casualties to society. Traditional traffic safety management methods mainly rely on statistical analysis of historical accident data and lack the ability to dynamically monitor and warn of real-time traffic conditions. Summary of the Invention

[0003] In order to solve the above technical problems, the present invention provides a method for identifying and warning of road vehicle traffic safety accident risks.

[0004] The present invention provides a method for identifying and warning of road vehicle traffic safety accident risks, comprising: S1. Collect, organize and process historical data, real-time data and external data to obtain data to be analyzed; S2. Extracting temporal features, spatial features, and behavioral features from the data to be analyzed; S3, splicing the temporal features, the spatial features, and the behavioral features into a three-dimensional fusion tensor, and decoupling them into a spatiotemporal feature matrix and dynamic sequence data after three-dimensional convolution and spatial attention modulation; S4. Inputting the spatiotemporal feature matrix and the dynamic sequence data into a hybrid model of a convolutional neural network and a recurrent neural network for training to obtain a trained risk prediction model; wherein the output of the risk prediction model is a three-dimensional probability vector + a risk level label, wherein the three-dimensional probability vector corresponds to the predicted probabilities of the three risk levels of low / medium / high, and the sum of the probabilities is 1; S5. Analyze the safety risk of the road vehicle while traveling based on the risk prediction model, and adjust the seat belt preload and airbag deployment parameters according to the risk level obtained from the analysis.

[0005] Optionally, the S3 specifically includes: The temporal features, the spatial features, and the behavioral features are spliced along the channel dimension to construct the three-dimensional fusion tensor, which is then decoupled into the spatiotemporal feature matrix and the dynamic sequence data after being modulated by the three-dimensional convolution kernel spatiotemporal anomaly detection and the behavior-driven spatial attention features.

[0006] Optionally, the process of collecting historical data in S1 specifically includes: Collect traffic accident data for in-depth investigations of traffic accidents, including the time and location of the accident, weather conditions, road conditions, types of vehicles involved, and driving behavior information; The process of collecting real-time data in S1 specifically includes: Real-time collection of traffic flow, vehicle speed, and traffic signal status through on-board sensors, cameras, and roadside units; The external data in S1 specifically includes meteorological data and holiday information.

[0007] Optionally, the seat belt preload and airbag deployment parameters are adjusted according to the risk level obtained through analysis, specifically including: When the risk level is low, maintaining the preload force of the seat belt at the first preload force and providing a vibration reminder; When the risk level is medium, increasing the preload force of the seat belt to a second preload force, and advancing the deployment time of the airbag from the first time point to a second time point; When the risk level is high, increasing the preload force of the seat belt to a third preload force, deploying the airbag, and performing a braking operation; The third pre-tightening force is greater than the second pre-tightening force, and the second pre-tightening force is greater than the first pre-tightening force.

[0008] Optionally, The scenarios corresponding to the high risk level include training samples of actual collisions, or scenarios in the real-time data where the collision time is less than 1 second, or the collision probability is greater than 90%; The scenarios corresponding to the medium risk level include the frequency of sudden braking being greater than or equal to 3 times per minute, or the steering wheel angle change being greater than 5° and lasting for more than 5 seconds; The scenario corresponding to the low risk level includes the vehicle speed exceeding the speed limit by 10% and lasting for less than 3 seconds.

[0009] Optionally, after S5, the step further includes: The risk prediction model is optimized according to updated traffic accident data.

[0010] Optionally, the process of processing the historical data, the real-time data, and the external data in S1 specifically includes: Based on time series association and density clustering algorithms, the historical data, the real-time data and the external data are cleaned, duplicate data, invalid data and noise data are removed, missing values are processed, the historical data, the real-time data and the external data are standardized, and the historical data, the real-time data and the external data are normalized so that the historical data, the real-time data and the external data are in the same dimension. The historical data, the real-time data and the external data are formatted and converted into the input format corresponding to the hybrid model.

[0011] Optionally, the S2 specifically includes: Extracting the time features from the timestamp of the data to be analyzed, including hour, day of the week, and season; Extracting the spatial features, including road topology, slope, and curvature features, using a geographic information system; The behavioral features are extracted by analyzing the driver's driving behavior, including the vehicle speed change rate, the direction angle change, sudden braking, and frequent lane changes.

[0012] Compared with the prior art, the present invention has the following beneficial effects: The present invention constructs a traffic accident warning system by fusing historical data, real-time data and external data from multiple sources. First, the time-synchronized historical accident data, real-time vehicle data and external environment data are aligned and standardized in time and space, and spliced into a three-dimensional fusion tensor (time step × spatial grid × feature channel). The layered three-dimensional convolution kernel is used to synchronously extract local spatiotemporal correlation features (such as the acceleration mutation and angular offset coupling features of sudden braking). The fused tensor is then decoupled into a spatiotemporal feature matrix and dynamic sequence data, which are respectively input into the convolutional neural network and recurrent neural network branches for spatial topology modeling and behavioral temporal dependency analysis. The collaborative reasoning of risk factors is achieved through cross-dimensional feature cross-layers, and the combined pattern of abnormal driving behavior and environmental risks is dynamically identified. This drives the precise response of graded warnings and seat belt pretensioning adjustment, thereby improving driver safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0014] Figure 1 A schematic flow chart of a method for identifying and warning of road vehicle traffic safety accident risks provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0015] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention are described clearly and completely below. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are also within the scope of protection of the present invention.

[0016] Figure 1A schematic flow chart of a method for identifying and warning of road vehicle traffic safety accident risks provided by an embodiment of the present invention includes: S1. Collect, organize and process historical data, real-time data and external data to obtain data to be analyzed; In some implementations, the process of collecting historical data in S1 specifically includes: Collect traffic accident data for in-depth investigations of traffic accidents, including the time and location of the accident, weather conditions, road conditions, types of vehicles involved, and driving behavior information; The process of collecting real-time data in S1 specifically includes: Real-time collection of traffic flow, vehicle speed, and traffic signal status through on-board sensors, cameras, and roadside units; The external data in S1 specifically includes weather data and holiday information.

[0017] In some embodiments, the process of processing historical data, real-time data, and external data in S1 specifically includes: Based on time series association and density clustering algorithms, historical data, real-time data and external data are cleaned, duplicate data, invalid data and noise data are removed, missing values are processed, historical data, real-time data and external data are standardized, and historical data, real-time data and external data are normalized so that they are in the same dimension. The historical data, real-time data and external data are formatted and converted into the input format corresponding to the hybrid model.

[0018] This paper focuses on the DBSCAN density clustering algorithm. The key lies in the setting of the parameters eps (Epsilon, indicating the maximum distance between two samples considered neighbors) and min_samples (indicating the minimum number of samples in a cluster) to effectively identify and remove noisy data. On this basis, duplicate data in the data is identified and eliminated by comparing the unique identifiers of the records or using a deduplication algorithm, while invalid data in the data is identified and eliminated through logical checks, data validation rules, or evaluating missing key information. The data is then standardized, and data from different sources is normalized to ensure that the data is in the same dimension. The data format is also converted to convert the original data into a format suitable for hybrid model input. The specific cleaning method is as follows: Consider changes in key factors such as vehicle speed and distance in traffic flow, which will directly affect the distribution characteristics and density of the data. According to the needs of the early warning method, set an EPS range that can capture potential traffic accident risks. If the early warning method needs to identify close-range vehicle conflicts (micro risks), the EPS range can be set to 0.2-0.5 after standardization (corresponding to actual vehicle distances of approximately 10-30 meters on urban roads and approximately 30-50 meters on highways); if the early warning method needs to identify broader traffic situation changes (macro risks), the EPS range should be set to 0.8-1.5 after standardization (corresponding to regional vehicle distance fluctuations of 50-100 meters, such as congestion spread and convoy formation).

[0019] Taking into account multiple factors such as the number of vehicles involved in a traffic incident and the duration of the incident, these factors will have a direct impact on the number of data points and the size of the clusters formed. According to the needs of the early warning method, set a min_samples value that can capture a sufficient number of data points to form a valid cluster. If the early warning method needs to identify large-scale traffic congestion incidents (high-speed chain collisions, etc.), the min_samples value should be set to between 5-10 to ensure that the noise can be filtered out and only significant accident clusters are retained. If the early warning method needs to identify small-scale vehicle conflict incidents (such as bicycle scratches at intersections), the min_samples value should be set to between 2-3 to avoid missing these small-scale accident clusters.

[0020] Clustering experiments were conducted by setting different eps values and min_samples values. According to the stability and rationality of the clustering results, the eps value and min_samples value were gradually adjusted until a satisfactory clustering effect was obtained to ensure that the clustering results met the requirements of the early warning method.

[0021] After determining the key parameters, the collected multi-source heterogeneous data is aggregated into a dataset, and the data space is established and arranged. Noise data is usually located in sparse areas of the dataset, and the number of data points in its neighborhood is less than the threshold specified by the algorithm parameter min_samples. The DBSCAN algorithm can automatically identify and process duplicate data and invalid noise data in the dataset. The density of spatial data feature points is calculated according to formula (1).

[0022] (1) Where: θ is the density of data feature points, i represents data, k represents the spatial data dimension, a Represents the size of data distribution in space.

[0023] Taking into account the influence of historical data, real-time data, and external data, as well as the interaction between these factors, a data denoising threshold is set. Using the threshold as the dividing line, the data is divided into two spatial data groups: a group with data above the threshold and a group with data below the threshold. The data division boundary is iterated and adjusted multiple times based on the two-dimensional graph of the data feature points until the data meets the noise requirements in one space, that is, normal data points and noise points are accurately distinguished. The denoised data is normalized, as shown in Formula (2).

[0024] (2) Where: j represents the denoised data, z j Indicates the normalized standard format data, n j represents the feature quantity of the data after denoising, represents the mean of the data feature quantity, s represents the standard deviation of the data feature points, η represents the density coefficient, Indicates the density of feature points of the data after denoising.

[0025] In the normalization process of formula (2), it is necessary to differentiate the characteristics of multi-source data: for categorical data such as weather and vehicle type, first convert them into numerical feature quantities n through One-Hot encoding or semantic embedding (such as heavy rain → [0.1, 0.7, 0.2], truck → 0.3) j , and then substitute into the formula to calculate z j Time series data such as traffic signal status requires dynamic calculation of local means within a sliding window and standard deviation s (for example, take the vehicle speed data of the 5-second window before and after), and adjust the density coefficient η by the proportion of state duration (the proportion of red light duration increases the weight); for high-dimensional spatial data such as lidar point clouds, the three-dimensional coordinate components are first normalized independently Generate n j (where x represents the original coordinate component value, μ x is the mean of all data points in this dimension, σ x is the standard deviation of all data points in that dimension (such as the x dimension), and x' is the standardized component value), and then the global voxel grid density (number of points per unit volume) is calculated. and s, and finally the multimodal feature quantities are dimensionally fused in a unified space through formula (2), while retaining the duplicate identifier verification and logic anomaly elimination rules in the original denoising process to ensure that the normalized z j Compatible with spatiotemporal distribution characteristics and physical constraints.

[0026] In the calculation of formula (2), the value of the density coefficient is determined by the denoising threshold. The denoised data is mapped in complex space and the data features are extracted. A network module for data feature extraction is constructed and the data in the space is matched within this module. The data features are arranged in the form of scattered points to establish a scattered feature matrix. The features of multi-source heterogeneous data are classified and extracted by combining the scatter matrix design principle. The extracted data features are analyzed and the temporal correlation between the data features is calculated. A ternary temporal module is established to perform global temporal correlation on the original data features. Data features are coupled within this module and the ordered power function is calculated, as shown in formula (3).

[0027] (3) Where: f represents the ordered power function, G j Indicates the order of data j, G max and G min They represent the maximum and minimum order in the data space, respectively. This expresses the temporal association of multi-source heterogeneous data, as shown in formula (4).

[0028] (4) Where: F represents the time series of multi-source heterogeneous data, f1…f j For different f. Based on the time series of the data, the temporal correlation of the data features is calculated, as shown in formula (5), completing the calculation of the temporal correlation of multi-source heterogeneous data features.

[0029] (5) Where: δ represents the temporal correlation of data features, represents the mean of the time series, and L represents the length of the data feature.

[0030] The multi-source heterogeneous data is fed back in the original data space according to the calculated feature time series correlation. The cluster points of the data and its adjacent data are set to determine the cluster affiliation of the data. The Euclidean distance between the data and its cluster points is calculated, as shown in Formula (6).

[0031] (6) Where: d represents the Euclidean distance between data and cluster points in space, o represents the cluster points of data clustering, l j Indicates the location of the data in space, l o Represents the location of cluster points in space.

[0032] After clustering, the density of the data is calculated as shown in formula (7).

[0033] (7) Where: θ' represents the spatial density of the data after clustering, m represents the data sequence, and r represents the cluster connection point.

[0034] Based on the final data density calculation results, abnormal data is identified. The boundary value for abnormal identification is set as shown in formula (8). Data whose calculated data density values do not fall within the above range are screened one by one to identify abnormal data.

[0035] (8) Where: μ represents the density mean, σ represents the density standard deviation.

[0036] The identified multi-source heterogeneous abnormal data are cleaned, and the missing data points of the abnormal data are analyzed. According to the different temporal correlations of the abnormal data, the abnormal data are divided into two types of abnormalities: local missing and long-term missing. The matrix is used to solve them, as shown in Formula (9).

[0037] (9) Where h represents the length of the time series with missing data, and c represents the current dimension of the data. A missing function is established to fill the missing parts of the abnormal data, as shown in formula (10).

[0038] (10) Where: J represents the missing content filled in for the abnormal data, and Y represents the time series length of the normal data with the same temporal correlation as the abnormal data.

[0039] After data preprocessing is complete (followed by subsequent steps), the collected data is divided into two parts: a spatiotemporal feature matrix and dynamic sequence data. The spatiotemporal feature matrix primarily includes information such as the accident time, location, weather conditions, road conditions, vehicle type, driving behavior, and external data. This can be converted into a CSV file format and subsequently imported into the CNN model for training. The dynamic sequence data primarily includes real-time data such as traffic flow, vehicle speed, traffic signal status, and seatbelt pretension signals. This can be converted into a Pandas Series (one-dimensional array) format and subsequently imported into the RNN model for training.

[0040] S2, extracting temporal features, spatial features, and behavioral features from the data to be analyzed; In some embodiments, S2 specifically includes: Extract time features from the timestamps of the data to be analyzed, including hours, days of the week, and seasons; Use geographic information systems to extract spatial features, including road topology, slope, and curvature characteristics; Behavioral features are extracted by analyzing the driver's driving behavior, including vehicle speed change rate, direction angle change, sudden braking, and frequent lane changes.

[0041] The specific extraction method is as follows: 1) Basic time unit extraction: This function breaks down timestamps into independent features (year, month, day, hour, minute, etc.), which is suitable for analyzing periodic or seasonal patterns. It also converts the time format to Unix timestamps or standardized formats (such as ISO 8601) to ensure consistency across data sources.

[0042] 2) Road spatial topology extraction, with its core process of "data cleaning - skeleton extraction - node generation - rule verification," achieves efficient and accurate modeling through the integration of multi-task deep learning and geographic information technology. First, the RGBM-LinKnet model is used to simultaneously optimize the extraction of road boundaries and centerlines. Dynamic threshold denoising and directional morphological filtering are combined to eliminate breaks and redundant branches caused by shadow occlusion, generating a highly complete binary road skeleton. Second, an improved parallel-serial hybrid thinning algorithm is used to compress the road surface layer by layer into single-pixel centerlines. Line width parameters are dynamically adjusted using the ArcGIS ArcScan tool to achieve continuous raster-to-vector conversion, avoiding the breakage caused by sudden line width changes in traditional methods. Furthermore, GIS topology editing tools are used to automatically break line segments at road intersections, constructing a graph structure with nodes as vertices and road segments as edges. Adjacency matrix sparse storage and a depth-first search algorithm are combined to rapidly verify road network connectivity. A rule library, such as directional consistency and loop closure, is integrated to enable automated error correction. Furthermore, leveraging digital elevation model (DEM) data, coupled slope and curvature analysis automatically extracts road longitudinal slopes and turning radii, providing quantitative features for road network safety assessments. This technology enhances initial data quality through deep learning and constructs a structured model using graph theory algorithms and GIS rule verification. This creates a closed-loop process from data preprocessing to topology verification, significantly reducing the need for manual intervention while balancing robustness in complex scenarios with efficient processing of large-scale road networks.

[0043] 3) Dangerous driving behavior feature extraction: Dangerous driving behavior features are extracted using vehicle trajectory data and vehicle dynamics data collected through PC-Crash simulation. The vehicle speed change rate is used to identify unstable driving behaviors. The calculation method for the average vehicle speed is shown in formula (11), and the calculation method for the standard deviation of vehicle speed is shown in formula (12).

[0044] (11) (12) Where N represents the number of trajectory points in a continuous trajectory data, v i represents the velocity of the trajectory point at time i, Indicates average speed.

[0045] According to the actual driving state of the vehicle, a threshold S1 is set. If S is less than S1, it is determined that the vehicle is in an unstable driving state.

[0046] Frequent and sudden lane changes of vehicles are identified using the following features: ① Divide the data set into multiple non-zero speed segments according to speed; ② Retrieve trajectory data and calculate the angular velocity at each moment according to formula (13) ; ③Record angular velocity exceeds the predefined threshold The direction angle α1; At the specified time threshold T LC The difference △α between the direction angles α2 and α1 at the subsequent moment is calculated according to formula (14). If △α is lower than the preset angle threshold α min , it indicates that the vehicle has made an abrupt lane change.

[0047] (13) (14) Where: α is the vehicle head direction angle, Predefine thresholds for angular velocity.

[0048] The time and the corresponding direction angle at this time are recorded to indicate the start of the lane change operation. An evaluation is performed here to determine the threshold value T at the next time. LC The vehicle's heading angle is closely aligned with the initial direction, so the deviation △α must be below a predetermined threshold. The predetermined threshold can be 3-4° in conventional highway lane change operations; on complex roads with low friction coefficients, α can be min Set to 1.5-2.5°; in emergency obstacle avoidance scenarios, the predetermined threshold can be relatively loose and set to 5°-10°, but it needs to be combined with other stability control systems to ensure safety.

[0049] S3: splice temporal features, spatial features, and behavioral features into a three-dimensional fusion tensor, and decouple it into a spatiotemporal feature matrix and dynamic sequence data after three-dimensional convolution and spatial attention modulation; In some embodiments, S3 specifically includes: The temporal features, spatial features, and behavioral features are spliced along the channel dimension to construct a three-dimensional fusion tensor. After three-dimensional convolution kernel spatiotemporal anomaly detection and behavior-driven spatial attention feature modulation, it is decoupled into spatiotemporal feature matrix and dynamic sequence data.

[0050] Specifically, the three types of features, temporal, spatial, and behavioral, can be concatenated along the feature channel dimension to form a fused tensor containing complete information about spatiotemporal behavior. This fused tensor undergoes a three-dimensional convolution operation, with the convolution kernel sliding across both the spatial and temporal dimensions. This captures abnormal behavior patterns in local areas within a specific time period. The fused features are dynamically modulated using spatial attention weights generated by the behavioral features, enhancing the characteristic response of high-risk areas.

[0051] The fused features are decoupled into a spatiotemporal feature matrix and dynamic sequence data. The spatiotemporal feature matrix retains its spatial grid structure and serves as the input to a convolutional neural network (CNN) to identify the spatial distribution patterns of high-risk areas. The dynamic sequence data, organized by time step, serves as the input to a recurrent neural network (RNN) to model the temporal patterns of risk evolution.

[0052] S4. Input the spatiotemporal feature matrix and dynamic sequence data into a hybrid model of a convolutional neural network and a recurrent neural network for training to obtain a trained risk prediction model; wherein the output of the risk prediction model is a three-dimensional probability vector + a risk level label, where the three-dimensional probability vector corresponds to the predicted probabilities of the three risk levels of low / medium / high, and the sum of the probabilities is 1; In some embodiments, Scenarios with a high risk level include actual collision training samples, or scenarios with a collision time of less than 1 second in real-time data, or a collision probability greater than 90%. The medium risk level corresponds to scenarios where the frequency of sudden braking is greater than or equal to 3 times per minute, or the steering wheel angle changes by more than 5 degrees and lasts for more than 5 seconds; The corresponding scenarios for low risk level include the vehicle speed exceeding the speed limit by 10% and lasting less than 3 seconds.

[0053] This paper uses a hybrid enhanced convolutional neural network + recurrent neural network (CNN+RNN) deep learning model. The model training adopts a supervised learning method, and the target value is a three-level risk label (low / medium / high) based on historical accident data. The label generation rules are as follows: ① High-level risk labels: These correspond to training samples of actual collisions, or scenarios where a time to collision (TTC) of less than 1 second and a collision probability of more than 90% are detected in real-time data; ② Medium-level risk label: dangerous driving behavior is detected (such as sudden braking frequency ≥ 3 times / minute, or the direction angle change △α in formula (14) exceeds 5°) and lasts for more than 5 seconds; ③ Low-level risk label: There is a slight abnormality (such as the vehicle speed exceeds the speed limit by 10% for less than 3 seconds), but no accident is triggered.

[0054] First, build the CNN and RNN models and optimize the model parameters through cross-validation. In the training set, select 80% of the imported data as training objects, and the remaining 20% as testing objects. Next, import the preprocessed CSV file and PandasSeries (one-dimensional array) file into the CNN model and RNN model respectively for model training. The output is a combination of a three-dimensional probability vector and a risk label. The three-dimensional probability vector corresponds to the predicted probability of low / medium / high risk levels (the sum is 1). Finally, use a loss function to represent the degree of difference between the model's predicted value and the true value. Save the network weights with the best accuracy after each iteration in the convolutional neural network shared file weights_mymodel.mdl. Gradually optimize the model parameters (such as the convolution kernel size, number of RNN layers, etc.) to improve model accuracy. Then, perform model evaluation and assess the model performance using the accuracy, recall, and F1 score indicators.

[0055] During the training process, CNN provides RNN with contextual information in the spatial dimension by capturing local patterns and global structures in spatial features; RNN provides CNN with dynamic change information in the temporal dimension by modeling long-term dependencies in time series.

[0056] The model evaluation involves the precision rate P, recall rate R, and F1 value formulas (15), (16), and (17), respectively: (15) (16) (17) In the formula, TP is the number of positive classes predicted as positive classes (True Positive), FN is the number of positive classes predicted as negative classes (False Negative), and FP is the number of negative classes predicted as positive classes.

[0057] The trained model is the risk prediction model.

[0058] S5. Analyze the safety risks of road vehicles based on the risk prediction model, and adjust the seat belt preload and airbag deployment parameters according to the risk level obtained from the analysis.

[0059] In some embodiments, adjusting the seatbelt preload and airbag deployment parameters based on the risk level obtained through analysis may include: When the risk level is low, the seat belt preload is maintained at the first preload and a vibration reminder is provided; When the risk level is medium, the seat belt preload is increased to a second preload, and the airbag deployment time is advanced from the first time point to the second time point; When the risk level is high, the seat belt preload is increased to the third preload, the airbag is deployed, and the braking operation is performed; The third pre-tightening force is greater than the second pre-tightening force, and the second pre-tightening force is greater than the first pre-tightening force.

[0060] Specifically, different warning levels can be set according to the risk level of low, medium, and high, and corresponding countermeasures can be taken. The seat belt preload can be adjusted. According to the real-time monitoring of the collision risk, the seat belt preload and airbag deployment parameters can be dynamically adjusted to improve the safety protection of the occupants. In addition, through driver feedback and intervention, the system can provide real-time risk warnings and safety suggestions to the driver through the on-board display and voice prompts to help him correct bad driving behavior. The system builds a full-scene safety protection system through the coordinated control of risk level response and dynamic parameters: in low-level risk scenarios, the seat belt maintains the basic locking force F α (Using the occupant's body parameters and the seat pressure sensor to calibrate), the vibration module of the seat / steering wheel is triggered synchronously with an output intensity of A ꞷ The pulse signal (frequency and amplitude are dynamically matched by the ambient noise level and the driver's alertness) is generated. At this time, the main and passenger airbags, side air curtains and seat side airbags are all in low-power standby mode (only sensor network monitoring is maintained); when the medium-level risk scenario is activated, the seat belt preload is automatically increased to F β (The adjustment intensity is positively correlated with the occupant's inertial forward momentum, and is calculated by the ECU based on vehicle speed, collision vector angle, and road adhesion coefficient), and the airbag triggering sequence is advanced. (The advance amount is output by the vehicle speed-collision angle coupling model) and the inflation rate Deploy to a partially inflated state (using a graded inflation strategy to mitigate the risk of whiplash injury to the occupant's neck); if the risk level jumps to a high level, the seatbelt preload instantly reaches its peak (Triggered by pyrotechnic pretensioner in response to high deceleration gradient), the driver and passenger airbags are inflated at a rate curve Fully deployed (the inflation time is controlled by the coordinated response of the multi-stage gas generator), the side air curtain covers the area from the A-pillar to the C-pillar and activates the pressure holding delay module for rollover conditions (pressure holding time Dynamically extended with the vehicle body roll rate until the roll angular velocity is lower than the safety threshold), the seat side airbags are synchronously activated to form a three-dimensional restraint network (the triggering threshold is dynamically determined by the occupant torso-seat angle, and intervention occurs when the angle difference exceeds the safety margin); if the warning is triggered If no effective braking is detected in the rear (verified by the brake pedal displacement sensor and wheel speed difference), the system will activate the emergency brake assist function and optimize the airbag deployment strategy (such as triggering the side airbags in advance to compensate for braking delay, or adjusting the main airbag inflation rate to adapt to sudden deceleration). All protection parameters (including ) are iteratively optimized through the vehicle model database. The database integrates occupant body size distribution, collision form library, road environment characteristics and driver behavior characteristics, and uses reinforcement learning algorithm to continuously update the parameter mapping table to ensure a balance between constraint response speed and secondary damage suppression effectiveness under different working conditions.

[0061] In some embodiments, further comprising: Optimize risk prediction models based on updated traffic accident data.

[0062] Specifically, the present invention can also conduct online learning and optimization, continuously optimize the risk prediction model based on new traffic accident data and accident cases, improve the accuracy of early warning, and have a feedback closed-loop mechanism when in use. By collecting feedback information from drivers and vehicles, a closed-loop system is formed to continuously improve the early warning algorithm and safety strategy.

[0063] In addition, the present invention can also integrate the vehicle-road cooperative system, integrate the early warning system with the intelligent transportation system, realize vehicle-road collaboration, improve overall traffic safety, and carry out big data platform and visualization, build a traffic big data platform, integrate multi-source data, provide data storage, processing and analysis capabilities, and display traffic trends and early warning information through visualization tools. The present invention has the ability of continuous optimization, and through the two-way interaction of real-time data collection on the vehicle side and cloud model updates, the early warning algorithm can adapt to changes in traffic characteristics in different regions. The collaborative network formed after the vehicles are connected to the network can share road risk information and optimize the risk avoidance strategy of group vehicles. Through the data exchange between roadside equipment and on-board systems, potential collision risks in complex scenarios such as intersections and bad weather can be predicted in advance, realizing a safety upgrade from single-vehicle protection to group collaboration.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying and warning of road vehicle traffic safety accident risks, characterized in that: include: S1. Collect, organize and process historical data, real-time data and external data to obtain data to be analyzed; S2. Extracting temporal features, spatial features, and behavioral features from the data to be analyzed; S3, splicing the temporal features, the spatial features, and the behavioral features into a three-dimensional fusion tensor, and decoupling them into a spatiotemporal feature matrix and dynamic sequence data after three-dimensional convolution and spatial attention modulation; S4. Inputting the spatiotemporal feature matrix and the dynamic sequence data into a hybrid model of a convolutional neural network and a recurrent neural network for training to obtain a trained risk prediction model; wherein the output of the risk prediction model is a three-dimensional probability vector + a risk level label, wherein the three-dimensional probability vector corresponds to the predicted probabilities of the three risk levels of low / medium / high, and the sum of the probabilities is 1; S5. Analyze the safety risk of the road vehicle while traveling based on the risk prediction model, and adjust the seat belt preload and airbag deployment parameters according to the risk level obtained from the analysis.

2. A method for identifying and warning of road vehicle traffic safety accident risks according to claim 1, characterized in that: The S3 specifically includes: The temporal features, the spatial features, and the behavioral features are spliced along the channel dimension to construct the three-dimensional fusion tensor, which is then decoupled into the spatiotemporal feature matrix and the dynamic sequence data after being modulated by the three-dimensional convolution kernel spatiotemporal anomaly detection and the behavior-driven spatial attention features.

3. A method for identifying and warning of road vehicle traffic safety accident risks according to claim 1, characterized in that: The process of collecting historical data in S1 specifically includes: Collect traffic accident data for in-depth investigations of traffic accidents, including the time and location of the accident, weather conditions, road conditions, types of vehicles involved, and driving behavior information; The process of collecting real-time data in S1 specifically includes: Real-time collection of traffic flow, vehicle speed, and traffic signal status through on-board sensors, cameras, and roadside units; The external data in S1 specifically includes meteorological data and holiday information.

4. A method for identifying and warning of road vehicle traffic safety accident risks according to claim 1, characterized in that: Adjust seatbelt preload and airbag deployment parameters based on the risk level analyzed, including: When the risk level is low, maintaining the preload force of the seat belt at the first preload force and providing a vibration reminder; When the risk level is medium, increasing the preload force of the seat belt to a second preload force, and advancing the deployment time of the airbag from the first time point to a second time point; When the risk level is high, increasing the preload force of the seat belt to a third preload force, deploying the airbag, and performing a braking operation; The third pre-tightening force is greater than the second pre-tightening force, and the second pre-tightening force is greater than the first pre-tightening force.

5. A method for identifying and warning of road vehicle traffic safety accident risks according to claim 1, characterized in that: The scenarios corresponding to the high risk level include training samples of actual collisions, or scenarios in the real-time data where the collision time is less than 1 second, or the collision probability is greater than 90%; The scenarios corresponding to the medium risk level include the frequency of sudden braking being greater than or equal to 3 times per minute, or the steering wheel angle change being greater than 5° and lasting for more than 5 seconds; The scenario corresponding to the low risk level includes the vehicle speed exceeding the speed limit by 10% and lasting for less than 3 seconds.

6. A method for identifying and warning of road vehicle traffic safety accident risks according to claim 1, characterized in that: After S5, the following is also included: The risk prediction model is optimized according to updated traffic accident data.

7. A method for identifying and warning of road vehicle traffic safety accident risks according to claim 1, characterized in that: The process of processing the historical data, the real-time data and the external data in S1 specifically includes: Based on time series association and density clustering algorithms, the historical data, the real-time data and the external data are cleaned, duplicate data, invalid data and noise data are removed, missing values are processed, the historical data, the real-time data and the external data are standardized, and the historical data, the real-time data and the external data are normalized so that the historical data, the real-time data and the external data are in the same dimension. The historical data, the real-time data and the external data are formatted and converted into the input format corresponding to the hybrid model.

8. The method for identifying and warning of road vehicle traffic safety accident risks according to claim 1, characterized in that: The S2 specifically includes: Extracting the time features from the timestamp of the data to be analyzed, including hour, day of the week, and season; Extracting the spatial features, including road topology, slope, and curvature features, using a geographic information system; The behavioral features are extracted by analyzing the driver's driving behavior, including the vehicle speed change rate, the direction angle change, sudden braking, and frequent lane changes.

Images