Mountain area expressway secondary accident active prevention and control method and system

Abstract

The invention discloses a mountainous area expressway secondary accident active prevention and control method and system, and the method comprises the steps: carrying out the global monitoring of a mountainous area expressway continuous slope road section, extracting traffic scene features, and identifying an accident scene; in combination with the current traffic flow state, the road static characteristics, the environmental conditions and the vehicle behaviors, estimating the initial accident occurrence risk based on an accident risk estimation model of risk coupling; the method comprises the following steps: predicting whether a secondary accident of the continuous slope bending of the expressway occurs or not based on a Bayesian optimization random forest model, if so, predicting the occurrence time and place of the secondary accident, and then performing active prevention and control through a warning means and a forced control means. According to the invention, through air-ground integrated monitoring and intelligent control, rapid detection of a primary traffic accident and accurate assessment and active prevention control of a secondary accident risk are realized, so that the traffic accident frequency and damage degree of a high-risk road section are greatly reduced.

Description

Technical Field

[0001] This invention belongs to the field of traffic safety technology and relates to a method and system for proactive prevention and control of secondary accidents on mountain highways. Background Technology

[0002] Currently, mountain highways, due to their complex terrain, often face high-risk sections such as long downhill slopes and continuous curves. These sections are highly prone to traffic accidents during adverse weather or periods of increased traffic volume, especially serious accidents such as heavy trucks losing control and multi-vehicle pileups, resulting in significant casualties. In the Lan-Hai Expressway accident on November 3rd, a heavy truck lost control on a series of curves and slopes, crashing into 31 vehicles inside a toll station, resulting in 15 deaths and 44 injuries. Similar accidents include: a multi-vehicle pileup on the Fu-Yin Expressway during the National Day holiday caused by rainfall, resulting in 5 deaths and multiple injuries; a multi-vehicle pileup on the Lian-Huo Expressway in 2022 caused by heavy rain, resulting in 6 deaths and multiple injuries; a rear-end collision on the Hang-Rui Expressway in 2023 caused by heavy rain, resulting in 3 deaths and multiple injuries; and a multi-vehicle pileup involving over 60 vehicles on the Henan Yan-Luo Expressway in 2024 due to icy road conditions. Traditional traffic safety measures, such as adding warning signs and optimizing road markings, can remind drivers to pay attention to safety to some extent, but their effectiveness in responding to sudden accidents, especially in preventing secondary accidents after the initial one, is limited.

[0003] Specifically, existing technologies lack mechanisms for rapid response, effective traffic diversion, and guidance of upstream and downstream vehicles to avoid accidents, allowing unsuspecting vehicles to continue entering the accident area, exacerbating the chain of accidents and increasing the severity of casualties. Furthermore, traditional methods struggle to achieve real-time, comprehensive monitoring of road conditions, failing to provide sufficient information support for rapid decision-making. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a proactive method for preventing secondary accidents on mountainous highways. Through integrated air-ground monitoring and intelligent control, it enables rapid detection of initial traffic accidents, accurate assessment of secondary accident risks, and proactive prevention and control, thereby significantly reducing the frequency and severity of traffic accidents on high-risk road sections.

[0005] Another objective of this invention is to provide an active prevention and control system for secondary accidents on mountain highways.

[0006] The technical solution adopted in this invention is a method for proactive prevention and control of secondary accidents on mountain highways, comprising the following steps:

[0007] Step S1: The entire area of ​​continuous curves and slopes on mountain highways is monitored by an integrated air-ground accident scene perception device. Traffic scene features are extracted by image recognition using convolutional neural networks and recurrent neural networks, and accident scenes are identified based on a generative pre-trained model.

[0008] Step S2: Combining the current traffic flow status, road static characteristics, environmental conditions, and vehicle behavior, the risk of the first accident is predicted based on a risk-coupled accident risk prediction model; using the real-time characteristics of the first accident, environmental conditions, traffic flow status, the risk of the first accident, road static characteristics, and historical data as input, a Bayesian optimized random forest model is used to predict whether a secondary accident will occur on the continuous slope curve of the highway. If it does occur, the time and location of the secondary accident are predicted.

[0009] Step S3: Based on the predicted time and location of secondary accidents, proactive prevention and control measures are implemented through warning and mandatory control measures.

[0010] Furthermore, step S1 includes the following steps:

[0011] S11, the air-ground integrated accident scene perception equipment includes drones and roadside equipment. The drones acquire aerial visible light image data, infrared shortwave thermal imaging data, terrain and environmental information, meteorological data, and drone status data; the roadside equipment acquires millimeter radar point cloud data, roadside video streams, vehicle micro-state data, and road infrastructure status; the drones and roadside equipment work together to achieve full-area monitoring.

[0012] S12 extracts spatial feature sequences from multi-view images using a convolutional neural network to generate 3D vehicle detection boxes, accurately representing the vehicle's position, size, and attitude in three-dimensional space. A recurrent neural network processes the spatial feature sequences extracted by the convolutional neural network, outputting a temporal dynamic feature vector to capture the dynamic evolution of vehicle trajectory, speed changes, and behavior patterns. SLAM extracts ORB feature points from UAV aerial images and stable reflection points from millimeter-wave radar, aligning the spatiotemporal references of UAV images and radar point clouds in real time. A Transformer cross-modal attention mechanism fuses heterogeneous multi-source data.

[0013] S13, based on historical accident data, trains a generative pre-trained model. The multi-source fusion dataset is input into the trained generative pre-trained model to generate accident probability distribution and risk hotspot areas, thereby identifying typical accident patterns, including heavy truck loss of control and multi-vehicle rear-end collisions.

[0014] Furthermore, in S1, accident risk features are introduced into the decision layer of the Transformer, and scene semantic understanding is enhanced and feature representation is optimized through generative pre-trained models; Bayesian inference and Kalman filtering are embedded in the encoding layer of the Transformer to suppress radar point cloud noise and video blur; through a cloud-edge collaborative architecture, high-frequency, low-latency data is preprocessed at edge nodes and the results are uploaded to the cloud, where the cloud is responsible for batch processing of complex tasks and long-term prediction, improving the real-time performance and scalability of data processing.

[0015] Furthermore, in S2, the initial accident risk assessment includes the following steps:

[0016] S21 breaks down traffic collision risks into passive risks and active risks;

[0017] S22 calculates the total passive risk value by weighting the road boundary constraints, lane line constraints on vehicles, the impact of road surface slope and friction coefficient on vehicles, and the lateral or longitudinal risks posed by surrounding vehicles; when the real-time vehicle speed v > v max At that time, it was determined to be high risk, v max This indicates the threshold for maintaining a stable vehicle speed.

[0018] S23, the active risk is the active collision risk caused by abnormal driving behavior. The driving behavior time series data is modeled by a long short-term memory network to identify abnormal driving behavior and output the risk value brought about by the active behavior of the vehicle; the collision time is less than a preset threshold and is judged as high risk.

[0019] S24, Construct a dynamic risk field, the dynamic risk value is calculated using the following formula:

[0020]

[0021] Where R(x,t) represents the dynamic risk value at position x at time t; m i The risk quality of the i-th vehicle is represented by its speed v. i and load factor η i Decide, ;d i This represents the Euclidean distance between the target vehicle at position x at time t and the i-th vehicle; k represents the environmental correction factor, with a value ranging from 0.5 to 1.5.

[0022] S25 weights and integrates the total passive risk value, active risk value, and dynamic risk value into a comprehensive risk index to classify risk levels.

[0023] Furthermore, the degree of danger of the aforementioned abnormal driving behavior is quantified by the following formula:

[0024]

[0025] Among them, S abnormal α represents the quantification of the degree of danger of abnormal driving behavior. t For acceleration, θ t For steering wheel angle, LDW t denoted as the number of lane departures, α, β, and γ are weighting coefficients, t represents the time step, and T is the total number of time steps in the entire analysis period.

[0026] Furthermore, S25 also includes: inputting the comprehensive risk index into the reverse cloud generator to obtain cloud digital features Ex, En, and He; calculating the membership cloud value α of the current value of the comprehensive risk index to each risk level; constructing an evaluation matrix to obtain the dynamic risk index m; and then... Standardization to range Risk levels are divided according to intervals.

[0027] Furthermore, S3 includes the following steps:

[0028] With the goals of eliminating lane-level accident precursor features and improving the overall operational efficiency of highways, we construct basic algorithms for mainline speed control, ramp flow control, and lane management in basic scenarios.

[0029] In the case of secondary accidents, with the aim of eliminating secondary accidents, an emergency algorithm for secondary accidents is constructed, which includes lane interception, active lane change prompts, and emergency lane management.

[0030] Based on the location and scale of the secondary accident and the current traffic flow data, a solution from the control strategy library is selected, and the optimal parameter combination is calculated through a simulation platform.

[0031] Furthermore, the mandatory control measures include electronic gates, dynamic lane-level speed limits and closures, mandatory intervention by vehicle terminals, and emergency lane occupancy control.

[0032] Furthermore, the electronic barrier gate includes a flexible composite material crossbar and a collision response mechanism. When the barrier gate is hit by a vehicle, it bounces open laterally to reduce impact damage. At the same time, the electronic barrier gate is connected to the platform center to achieve lane-level precise traffic control based on the secondary accident prediction results.

[0033] A proactive prevention and control system for secondary accidents on mountainous highways, employing the aforementioned proactive prevention and control method for secondary accidents on mountainous highways, includes:

[0034] The device access and management module is used for device registration and authentication, device data uploading and control command issuance, monitoring device status, and remote fault diagnosis and configuration management.

[0035] The scene identification module is used to identify whether an accident scene exists based on the data returned by the device, and to optimize the control algorithm with the optimal parameters selected automatically or manually.

[0036] The big data warehouse and computing module is used to save, extract, analyze and process multi-source heterogeneous data according to a homogeneous data structure through preset storage modes.

[0037] The beneficial effects of this invention are:

[0038] 1. Improve accident response speed: Through integrated air-ground monitoring, the initial accident alarm time is significantly shortened, gaining valuable time for the prevention and control of secondary accidents.

[0039] 2. Accurate Risk Assessment: By using a large AI model coupled with a multi-factor risk assessment model, we can accurately predict accident risks and provide a scientific basis for decision-making.

[0040] 3. Proactive prevention and control of secondary accidents: By integrating digital and intelligent equipment and dynamically adjusting control strategies, we can effectively eliminate early warning signs of accidents and reduce the incidence of secondary accidents.

[0041] 4. Improve management efficiency: The remote prevention and control smart cloud platform enables information visualization and adjustable control parameters, thereby improving the efficiency and decision-making level of traffic safety management.

[0042] 5. This invention integrates technologies such as air-ground integrated accident scene perception, multi-source data fusion, traffic AI large model accident scene identification, accident risk assessment, digital intelligent equipment for active prevention and control of secondary accidents, and remote prevention and control smart cloud platform, aiming to improve traffic safety on mountain highways and reduce the occurrence of secondary accidents. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart illustrating an embodiment of the present invention.

[0045] Figure 2 This is the GPT model for traffic accident scene identification in this embodiment of the invention.

[0046] Figure 3 This is a flowchart illustrating the research path of an embodiment of the present invention.

[0047] Figure 4 This describes the secondary accident occurrence process in an embodiment of the present invention.

[0048] Figure 5 This is the secondary accident prediction process in the embodiments of the present invention.

[0049] Figure 6 This is a schematic diagram of the random forest algorithm for classification and prediction in an embodiment of the present invention.

[0050] Figure 7 This is a block diagram of the main architecture of an embodiment of the present invention. Detailed Implementation

[0051] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0052] Basic concept of the embodiments of the present invention:

[0053] This invention addresses the need for initial accident perception on continuous curved and sloping sections of mountain roads. It completes the selection, adaptation, and integration of drones and roadside radar-visual integrated equipment. By combining the image recognition effects of convolutional neural networks and recurrent neural networks, it enhances the ability to extract traffic scene features and, based on the scene recognition ability of a generative pre-trained model (GPT), achieves accident scene identification. Secondly, by combining current traffic flow status, driving scenarios, and climate characteristics, it uses a risk-coupled accident risk assessment model to predict the risk of initial accidents. Then, it integrates warning methods (such as variable speed limits, variable message signs, and strobe warning lights) with strong control measures (such as electronic gates) to study proactive traffic flow control technology. Finally, it integrates scene perception, data communication, and remote control methods into a remote control cloud platform.

[0054] Example 1,

[0055] A proactive method for preventing secondary accidents on mountain highways, such as Figure 1 As shown, it includes the following steps:

[0056] Step S1: Utilize integrated air-ground accident scene perception equipment and scene identification to achieve full-area monitoring of key accident sections on mountainous highways; extract traffic scene features through convolutional neural networks and recurrent neural network image recognition, and identify accident scenes based on generative pre-trained models.

[0057] In step S1, for the perception of accidents and abnormal events on mountain highways, multi-source traffic environment information on mountain highways is acquired collaboratively using monitoring equipment such as drones, millimeter-wave radar, and infrared / shortwave radar. The drones integrate multiple sensors, enabling real-time acquisition of terrain, obstacle, and meteorological data, and collaborate with roadside equipment to achieve data fusion and comprehensive monitoring. They are wind-resistant, rain- and snow-resistant, and adaptable to complex weather and terrain in mountainous areas. This ensures long-term monitoring on continuous curved and sloping road sections, avoiding monitoring interruptions due to insufficient power or signal loss.

[0058] An autonomous path planning algorithm for unmanned aerial vehicles (UAVs) is designed to adjust the flight area based on real-time monitoring requirements and eliminate blind spots caused by terrain occlusion. A 3D raster map is constructed based on mountain terrain elevation data (DEM) to mark potential occlusion areas. An improved path planning algorithm is adopted, incorporating risk weights (such as wind speed and obstacle distance).

[0059]

[0060] in, This indicates a weighted average of actual cumulative costs, heuristic costs, and overall risks.

[0061] From the starting point to the current node The cumulative cost, calculated based on distance, energy consumption, and terrain complexity, is known in this field. To start from the current node The estimated minimum cost to the target node, used to guide the search direction, is known in this field.

[0062] For nodes The overall risk value reflects dynamic environmental threats (such as wind speed, obstacle distance, weather, etc.).

[0063]

[0064] Indicates wind speed risk value, such as The risk value increases from 0. Indicates the obstacle risk value. , Distance to the obstacle, For a safe distance, is the sensitivity coefficient, and e represents a constant.

[0065] Indicates the terrain risk value, based on Data calculation of slope ;

[0066]

[0067] Weight Calibrated using historical accident data or expert experience.

[0068] By detecting sudden obstacles using millimeter-wave radar, the path is dynamically adjusted using the RRT* (Rapid Expanding Random Tree) algorithm to achieve real-time obstacle avoidance. Based on a multi-layered fusion design framework, data fusion is divided into a sensor layer, a feature layer, and a decision layer. The sensor layer performs preliminary data preprocessing and calibration, the feature layer extracts and fuses features from multiple data sources, and the decision layer combines the results of multi-modal data fusion for intelligent analysis and decision-making. To address the issues of timeliness and spatial consistency of multi-source data, an adaptive fusion analysis strategy is adopted, adaptively adjusting the weights of different data sources according to environmental changes to improve the stability and accuracy of the fusion results.

[0069] In step S1, heterogeneous data source synchronization and calibration employ a time synchronization algorithm to ensure time consistency. It utilizes global clock synchronization technology (such as GPS timestamps) and a local clock drift correction mechanism to ensure that data from different sensors on special-purpose drones and ground equipment are aligned at the same time. SLAM technology is used to update sensor spatial relationships in real time, handling viewpoint shifts caused by equipment movement (such as drones) or vibration.

[0070] Based on visual SLAM and radar point cloud matching, multimodal sensor fusion is employed: ORB feature points are extracted from UAV aerial images, stable reflection points are extracted from millimeter-wave radar, and an objective function is constructed to minimize the reprojection error and point cloud matching error.

[0071]

[0072] in, The transformation matrix from radar to camera, For projection function, , These are 3D points in the radar and camera coordinate systems, respectively; specifically, It represents the three-dimensional spatial point cloud of the target object detected based on the millimeter-wave radar coordinate system. Its physical information comes from the electromagnetic wave reflection signal and directly represents the absolute position and outline of the object in the radar coordinate system. It represents the three-dimensional spatial coordinates of the target object obtained based on the camera coordinate system. Its physical information comes from the back projection calculation of visual image features, and it represents the relative position and geometric structure of the object from the camera's perspective.

[0073] The total optimization error of the E characterization system is physically the weighted sum of visual reprojection error and radar point cloud matching error, serving as a comprehensive indicator for evaluating the spatial consistency of multi-source sensor data. () is the camera projection function, whose physical function is to map three-dimensional spatial points from the camera coordinate system to the two-dimensional image plane, and establish the conversion relationship between three-dimensional geometric information and two-dimensional pixel observation; The two-dimensional pixel coordinates of feature points observed in the camera image represent the physical meaning of the coordinates as the reference true value in the calculation of reprojection error, providing visual observation constraints for multimodal data fusion.

[0074] Dynamic Environment Adaptability: To address sensor viewpoint shifts caused by UAV movement and vibration, a real-time spatial relationship update mechanism is introduced. SLAM technology is used to dynamically calibrate sensor pose, ensuring spatiotemporal alignment of heterogeneous data sources. Specifically, based on visual inertial odometry or lidar odometry, feature points or point cloud features are extracted from sensor sequence data in real time. Graph optimization or filter methods are used to estimate sensor pose in real time. Simultaneously, a local environment map is constructed and maintained, associating current observation features with map landmarks. Nonlinear optimization techniques are used to minimize reprojection errors or point cloud matching errors, thereby achieving continuous correction and dynamic calibration of sensor pose, effectively compensating for positioning drift caused by equipment movement or vibration.

[0075] Enhanced anti-interference capability: Sensor noise is processed through Bayesian inference and Kalman filtering, combined with adaptive fusion strategies (such as adjusting the weights of different data sources), improving robustness in low visibility (fog, rain, snow) and high occlusion scenarios.

[0076] Multimodal sensor fusion provides a redundant and complementary environmental perception data foundation for the system through cross-modal data alignment and feature unification; dynamic environmental adaptability is based on SLAM technology to calibrate sensor pose in real time, ensuring the spatiotemporal consistency of the above multi-source data in dynamic scenes; enhanced anti-interference capability is achieved by suppressing noise and optimizing weights of fused data through Bayesian inference and adaptive filtering, thereby maintaining the robustness of the perception system under low visibility or high occlusion conditions.

[0077] On mountainous highways with continuous curves and slopes, drones and roadside sensors may experience sensor perspective shifts due to terrain obstruction, drone vibration, or airflow disturbances. SLAM technology ensures precise temporal and spatial alignment of multi-source data, including drone aerial images and radar point clouds, by updating sensor pose and spatial relationships in real time, eliminating data misalignment caused by equipment movement or dynamic environmental changes. Mountainous highways with curves and slopes present blind spots, dynamic obstructions, and adverse weather conditions. Through joint optimization of visual SLAM (ORB feature points) and radar point cloud matching, SLAM technology significantly reduces reprojection errors and enhances the detection accuracy of obstacles and accident vehicles. Addressing the differences in sensor noise and heterogeneous data sources, SLAM technology combines Bayesian inference and Kalman filtering to dynamically correct sensor errors and ensures reliable perception in low-visibility environments through adaptive weight adjustment strategies. SLAM also enables the real-time construction of 3D grid maps (based on DEM data), marking potential risk areas (such as slopes >20° and areas with dense obstacles), providing a dynamic environmental model for secondary accident risk assessment, and assisting in accurate decision-making by control equipment such as electronic barriers and variable message signs.

[0078] Traditional SLAM may perform well on flat roads, but it faces challenges such as high terrain complexity, frequent dynamic occlusion, and strong sensor heterogeneity in mountainous and winding sections. For example, drones are susceptible to crosswind interference during flight, leading to spatiotemporal inconsistencies between aerial images and ground radar data. This invention addresses positioning drift and data fusion deviations in such scenarios through a SLAM algorithm enhanced with modal sensor fusion, dynamic environmental adaptability, and anti-interference capabilities. Single sensors (such as cameras alone) are prone to failure under adverse weather or occlusion conditions. SLAM technology, through cross-modal data fusion (vision + radar + infrared) combined with the cross-modal attention mechanism of the Transformer architecture, achieves high-precision perception in all weather and multi-scene environments, overcoming the limitations of single sensors. The prevention of secondary accidents relies on a rapid response to the initial accident. SLAM technology, through real-time calibration of sensor spatial relationships, ensures rapid identification of accident scenarios and prediction of risk propagation, providing a data foundation for the timely execution of control strategies. Accident risks in mountainous areas are dynamically coupled from multiple factors (slope, wind speed, and traffic density). The dynamic environment model and risk weight coefficients constructed by SLAM technology provide real-time spatial data support for risk assessment models, enabling accurate quantification of risk levels.

[0079] Furthermore, by employing a feature-matching-based spatial alignment and calibration algorithm, the perspective differences between different sensors are addressed, ensuring accurate fusion of location information from different data sources to guarantee spatial consistency in the dynamic traffic environment.

[0080] In step S1, a multimodal Transformer is used to perform multimodal heterogeneous data fusion on sensor data such as video and radar signals, achieving efficient feature alignment and information exchange. Through automatic feature extraction and multimodal feature fusion, problems such as complex weather and environmental occlusion are addressed, enabling all-weather high-precision perception. Furthermore, Bayesian inference and Kalman filtering methods are used to handle noise and uncertainty in the sensor data, ensuring robustness and high-reliability tracking capabilities under low visibility and high occlusion conditions.

[0081] Specifically, when fusing multimodal sensor data such as UAV aerial images, roadside videos, millimeter-wave radar data, and infrared / shortwave data, the Transformer module primarily achieves efficient fusion through a cross-modal attention mechanism and a unified feature representation. The role of infrared / shortwave data is to compensate for the perception limitations of visible light and radar in low-light, foggy, or rainy / snowy conditions by detecting the thermal radiation characteristics of target objects. This enables effective all-weather detection and identification of targets such as vehicles and obstacles, and provides feature inputs with temperature-related information for multimodal data fusion, enhancing the system's overall perception robustness and scene understanding completeness in complex environments.

[0082] The efficient fusion process mainly consists of four layers: an embedding layer, a cross-modal attention layer, an encoding layer, and a decision layer. The embedding layer maps features from different modalities to a unified vector space, as shown in the following formula:

[0083]

[0084] For the features of the i-th mode, and For learnable weights and biases, These are the features after embedding.

[0085] The cross-modal attention layer dynamically captures the correlations between different modalities through a cross-modal attention mechanism. The formula is as follows:

[0086]

[0087] Where Q, K, and V represent query, key, and value, respectively. This is the dimensionality scaling factor. Cross-modal attention mechanisms enable the model to simultaneously focus on key information from different modalities.

[0088] The encoding layer uses the Transformer's encoder structure to further encode the features output by the cross-modal attention layer, extracting higher-level feature representations. The formula is as follows:

[0089]

[0090] Where X represents the input feature, FFN represents the feedforward network, and Z represents the encoded feature.

[0091] Finally, the decision-making layer establishes a real-time monitoring and risk assessment model for accidents or anomalies on mountain highways based on the encoded feature results, enabling rapid identification and risk assessment of accident scenarios.

[0092] Convolutional Neural Networks (CNNs) are responsible for extracting spatial features from multi-view images and generating 3D vehicle detection boxes. The spatial relationships and motion state information contained in these detection boxes serve as inputs to risk coupling models and dynamic risk field models, used to assess passive and active risks, and to support Bayesian optimized random forest models in predicting whether, when, and where secondary accidents will occur.

[0093] Recurrent Neural Networks (RNNs) are used to model the temporal dynamic features of traffic scenarios, outputting a temporal feature vector that provides the dynamic evolution of vehicle behavior. This enhances the understanding of the evolution process of accident scenarios by Generative Pre-trained Models (GPTs) and provides key dynamic behavioral parameters for accident risk prediction models based on risk coupling.

[0094] The Transformer module serves as the core of multimodal heterogeneous data fusion, aligning and interacting the spatiotemporal features extracted by CNN and RNN with data from sensors such as radar and infrared across modalities. Finally, the unified feature representation after Transformer fusion is input into the GPT generative pre-trained model to achieve deep semantic understanding and risk identification of accident scenarios.

[0095] In step S1, for the cloud-edge collaborative computing architecture, a cloud platform combining edge computing and cloud computing is adopted. High-frequency, low-latency data is preprocessed at edge nodes, and the results are uploaded to the cloud. The cloud is responsible for batch processing of complex tasks and long-term predictive analysis. Through collaborative optimization between the edge and cloud, the real-time performance and scalability of data processing are improved. Data storage adopts a layered storage and optimization approach. A layered storage architecture based on memory and disk is designed, with hot data (real-time data) stored in high-performance memory and cold data (historical data) stored on disk and in the cloud, ensuring efficient data access under high concurrency. Furthermore, data redundancy recovery and fault tolerance mechanisms are introduced into the cloud computing platform, using RAID or distributed storage to prevent data loss, and ensuring data reliability through regular backup and recovery mechanisms.

[0096] In step S1, leveraging the superior scene understanding and analysis capabilities of generative pre-trained (GPT) large models, a GPT large model for traffic accident scene identification is developed. Key features are extracted from historical traffic accident data, and deep learning techniques are used to construct an intelligent identification algorithm for real-time accident scene monitoring. By identifying and analyzing common traffic accident patterns, traffic flow, weather conditions, and road features are comprehensively processed to accurately distinguish various accident scenarios. The model architecture is as follows: Figure 2 As shown.

[0097] Specifically, the traffic accident identification GPT model consists of two main components. The first is the scene perception module, which utilizes image input from multiple terminals to construct an image extraction and analysis network. This network seamlessly integrates and extracts panoramic multi-image features from multiple vehicles and road testing equipment. By constructing a robust deep learning framework, it effectively generates accurate 3D detection boxes for vehicles. The second module is the GPT inference module, which includes the standard GPT-4V module, a multi-level priority sampling prompt module based on corpus fusion, an active-passive task prompt module, and a multi-output module for accident-related tasks. For input data from images, each image is preprocessed based on the image backbone to extract different information and features. The generative pre-trained model GPT is trained using a multi-source fusion dataset, and the model parameters are repeatedly adjusted and optimized to improve the model's accuracy and real-time performance. Simultaneously, regular testing and validation are conducted to ensure the model achieves the expected results in practical applications.

[0098] Step S2: Initial accident risk assessment of continuous curves and slopes on mountain highways.

[0099] Combining current traffic flow status, road static characteristics, environmental conditions, and vehicle behavior, a risk-coupled accident risk prediction model is used to predict the risk of initial accidents. This model is designed for dynamic interactions involving multiple elements—people, vehicles, roads, and the environment—on mountainous highways with continuous curves and slopes. Supported by accident causation theory and safety resilience theory, it integrates the nonlinear coupling relationship between passive risk, active risk, and dynamic risk fields to construct a comprehensive risk quantification and prediction model. By analyzing the inherent coupling mechanism between static constraints (roads, lanes, road surfaces) and dynamic behavioral factors (driving operations, vehicle interactions), and combining extension cloud theory and fuzzy set methods to handle the uncertainty of risk factors, this model transforms scattered single risk indicators into a systematic and comprehensive risk assessment result. It achieves accurate quantification of the risk of initial accidents and provides core input for Bayesian optimized random forest models, supporting dynamic prediction of whether secondary accidents occur and their spatiotemporal location. This forms a closed-loop logic of "risk quantification - coupling analysis - level determination - prediction support," providing a scientific decision-making basis for the generation of lane-level active control strategies.

[0100] The inputs of the risk-coupled accident risk prediction model cover five categories: passive risk indicators, active risk indicators, dynamic risk field parameters, real-time scenario parameters, and historical prior data.

[0101] 1. Passive risk indicators: including passive risk value U and vehicle stability speed threshold. Deviation from real-time vehicle speed v .

[0102] 2. Active Risk Indicators: These include risk values ​​arising from the vehicle's active behaviors. Collision Time Tolerance (TTC)

[0103] 3. Dynamic risk field parameters: including dynamic risk values .

[0104] 4. Real-time scene parameters: covering curve radius of curvature R, road slope i, and road surface friction coefficient. (For dynamic correction) The data, including lane line type (solid / dashed), traffic flow density, weather conditions (rain, icing), and visibility, are collected in real time through integrated air-ground sensing equipment.

[0105] 5. Historical prior data: including accident statistics for this road section over the past 5 years (accident type, vehicle speed at the time of occurrence, gradient, coefficient of friction, etc.), and risk factor weighting results (passive risks, etc.). Proactive risk , , (etc.), cloud digital features (Ex, En, He) obtained from training the extension cloud model, and the pre-trained parameter set of the Bayesian optimized random forest model.

[0106] The model output focuses on supporting the risk assessment of initial accidents and the prediction of secondary accidents, providing three core results: quantitative risk indicators, qualitative risk levels, and prediction auxiliary information, as detailed below:

[0107] 1. Quantitative risk indicators: including comprehensive risk index Dynamic risk index m, probability of secondary accidents (Output from a Bayesian optimized random forest model, value range) ).

[0108] 2. Qualitative risk level: based on dynamic risk index The risk level is divided into four levels, and at the same time, it is combined with and Forming a dual-indicator verification, or It was immediately classified as high risk.

[0109] 3. Predictive auxiliary information: including dynamic risk field distribution matrix (grid-like risk intensity distribution in lane units) and key risk coupling triggering factors (such as speeding magnitude, TTC value, and abnormal driving behavior type), providing core input parameters for secondary accident spatiotemporal prediction of Bayesian optimized random forest model, while dynamically correcting risk field distribution to support the accurate implementation of subsequent control strategies such as lane closure in the near impact zone and warning in the far impact zone.

[0110] This invention decomposes vehicle collision risk into two dimensions: passive risk and active risk. Through in-depth analysis of driving constraints in static scenarios and the simulation of dynamic events, this embodiment of the invention meticulously constructs a vehicle-level micro-collision risk assessment model based on artificial potential field theory. Furthermore, considering the inherent coupling mechanism between multiple factors such as people, vehicles, and roads, and the potential for out-of-control vehicles to induce major traffic accidents, this embodiment of the invention, based on accident causation theory and safety resilience theory, integrates three progressively deeper levels: causal characteristics, coupling mechanisms, and evolutionary laws, systematically exploring the specific scope of risk factors and their disaster-causing processes in vehicle out-of-control scenarios. On this basis, it cleverly utilizes extension cloud theory and fuzzy set methods to construct a set of safety risk evaluation index systems and corresponding assessment models suitable for continuous curved and sloping road sections, thereby achieving accurate quantitative assessment of accident risks.

[0111] Passive risk:

[0112] The factors that affect a vehicle during its journey, such as roads, lanes, and other traffic participants, are categorized as passive risks to the vehicle.

[0113]

[0114] U represents the passive risk value. Road Representing road boundary constraints, these constraints ensure that vehicles can travel within the road. They have the highest weight among all factors, and their magnitude decreases from the road boundary to the center. They are quantified by combining road boundary distance functions with digital elevation model (DEM) data.

[0115] U Lane The lane line constraint on the vehicle includes the constraint that keeps the vehicle moving along the center of the lane, and also takes into account the difference in the strength of the constraint between solid lane lines and dashed lane lines. In order to ensure that the vehicle can change lanes or drive over the lines in emergency situations, the lane line constraint should take into account the relative strength. The calculation is based on the lane line type (solid line / dashed line) and the lateral offset distance of the vehicle using a piecewise function.

[0116] U SurfaceRepresenting the impact of road surface slope and friction coefficient on vehicles, it is mainly composed of road section slope and road surface friction coefficient, and is calibrated based on real-time collected road surface slope and friction coefficient through physical relationship formula.

[0117] U Car Lateral or longitudinal risks posed by surrounding vehicles are key aspects of passive risk. To differentiate between lateral and longitudinal distances using different measurement standards, a virtual distance between vehicles was used as a variable in the quantification process. Euclidean distances were used for the forward and lateral directions of vehicles, while a triangular risk zone was added behind each vehicle. The Euclidean distances and relative speeds between vehicles were dynamically calculated using millimeter-wave radar and visual sensors.

[0118] The weighting coefficients represent the road boundary constraints. This represents the weighting coefficient of the lane lines on the vehicle constraint. This represents the weighting coefficients of the impact of road surface slope and friction coefficient on vehicles. The weighting coefficients represent the lateral or longitudinal risks posed by surrounding vehicles, with the sum of the four weighting coefficients being 1; each weighting coefficient is determined through optimization using historical accident data.

[0119] Based on the balance between centrifugal force and friction, the vehicle speed threshold v for maintaining stability is calculated using the following formula. max :

[0120]

[0121] μ represents the road surface friction coefficient, g represents the gravitational acceleration, R represents the radius of curvature of the curve, and i represents the road slope.

[0122] When the real-time vehicle speed v > v max When a high-risk warning is triggered, the warning signal is used as a key input parameter to be input into the accident risk prediction model based on risk coupling in real time. This dynamically corrects the distribution of the dynamic risk field and further triggers the Bayesian optimized random forest model to predict secondary accidents, thereby providing real-time decision-making basis for the generation and implementation of lane-level control strategies in the subsequent active prevention and control stage.

[0123] Proactive risk:

[0124] U BehaviorThis represents the risk value resulting from the vehicle's proactive behavior. In quantifying proactive risk, the vehicle's behavior at future moments is crucial. Based on an intent prediction model, actions such as lane changes, acceleration, and deceleration are predicted. Due to the time-varying nature of driving intent, the risk area extends towards the current lane or adjacent lanes (determined by whether or not to change lanes). Therefore, a triangle with variable vertices is used as the basic model for the risk area. The position of the vertex is related to the driving intent and the vehicle's driving state. Specifically, the lateral coordinate of the vertex is related to the vehicle's lateral intent (lane change), and the longitudinal coordinate is related to the vehicle's longitudinal intent (acceleration, deceleration, or following the vehicle). Behavior The driving operation time series data is used to perform intent recognition and risk area modeling through a Long Short-Term Memory (LSTM) network, and the weights of each parameter are calibrated through an optimization algorithm driven by historical accident data.

[0125] Driving behavior quantification model: This model uses an LSTM (Long Short-Term Memory) network to model time-series driving behavior data and identify abnormal behaviors (such as frequent sudden braking and weaving). Abnormal behavior scoring formula:

[0126]

[0127] Among them, S abnormal α represents the quantification of the degree of danger of abnormal driving behavior. t For acceleration, θ t For steering wheel angle, LDW t The number of lane departures is represented by α, β, and γ, which are weighting coefficients (determined through training with historical accident data); t represents the time step, and T is the total number of time steps in the entire analysis period; by weighting and summing the three key driving behavior indicators within the monitoring period (t from 1 to T), the overall risk level of abnormal driving behavior is quantified.

[0128] Calculate the collision time TTC using the collision time model:

[0129] d represents the distance to the vehicle in front, and Δv represents the relative speed, both calculated based on millimeter-wave radar data. When TTC < 2.5s, it is considered high risk. This judgment result serves as a key real-time dynamic risk indicator and is immediately input into the risk-coupled accident prediction model.

[0130] In step S2, a macro-level accident risk assessment is performed for out-of-control events.

[0131] Specifically, this invention analyzes the inducing factors of accidents and the fundamental laws and inherent patterns of their interactions, coupling effects, and dynamic evolution. Adopting this core concept from the field of safety systems engineering, and combining it with the risk evolution path between various accident-inducing factors and vehicle loss of control phenomena on continuous curves and slopes in mountainous highways, this invention constructs a safety risk assessment model for heavy-duty truck loss of control scenarios. This model will enable scientific analysis and accurate assessment of the potential severity of accidents.

[0132] In constructing this model, embodiments of the present invention consider a series of key factors, including the real-time status of the out-of-control heavy-duty truck, the potential collision risk between vehicles, the complexity of the road alignment and the completeness of its protective facilities, and the dynamic state of traffic flow, etc. Figure 3 As shown, through this systematic and refined research approach, this invention aims to provide scientific basis and technical support for the safety management of continuous curves and slopes on mountain highways.

[0133] Highway safety accidents are multi-dimensional and complex systems engineering projects, and the analysis covers three levels: causal characteristics, coupling mechanisms, and evolution patterns. In terms of causes, out-of-control heavy-duty trucks are the primary cause, with secondary factors including six aspects: people, vehicles, roads, environment, management, and goods. Regarding coupling mechanisms, accidents are caused by the dynamic interaction of multiple interconnected factors, representing risk coupling. A high degree of coupling leads to qualitative changes in the system, triggering accidents; therefore, a safety constraint model is needed to constrain risk factors. In terms of evolution patterns, resilience theory has advantages in describing system evolution, and exploring the resilience of highway traffic safety systems can help investigate the patterns of system change. Therefore, a hierarchical network model of highway traffic safety resilience is constructed using people, vehicles, goods, roads, environment, and management as elements.

[0134] The dynamic risk field model is a spatial risk assessment method based on real-time traffic flow data and road environmental parameters. It constructs a dynamic risk distribution map by quantifying the impact of vehicle interactions and environmental factors on accident risk. Its core formula is:

[0135]

[0136] R(x,t) represents the dynamic risk value at location x at time t; m i The risk quality of the i-th vehicle is represented by its speed v. i and load factor η i Decide, ;d i This represents the Euclidean distance between the target vehicle and the i-th vehicle. k represents the environmental correction factor, which takes into account the effects of visibility, road surface friction coefficient, and weather conditions (such as rain and snow), and its value ranges from 0.5 to 1.5.

[0137] Overall Risk Index Calculation: The dynamic risk value R(x,t) is integrated with the passive risk value and the active risk value.

[0138]

[0139] Among them, weight =0.5, =0.3, =0.2, optimized using historical accident data.

[0140] Overall Risk Index It quantifies and integrates the initial accident risk of the "people-vehicle-road-environment" coupling multiple factors in the continuous curves and slopes of mountain highways, outputs an intuitive single risk value, and provides core quantitative basis for subsequent risk level determination, secondary accident prediction and control strategy selection.

[0141] Real-time calculation of risk quality m for each vehicle i According to the Euclidean distance d i Based on the environmental coefficient k, calculate the risk contribution value of each vehicle to the target location, sum up the risk contribution values ​​of all vehicles, generate a dynamic risk field distribution map, and combine passive risk values ​​and active risk values ​​to output a risk heat map of the entire road segment, identifying high-risk areas.

[0142] This study combines extension cloud theory with set pair analysis to assess the accident risk level of continuous long downhill sections of mountain highways. To facilitate the conversion between qualitatively described risk factors and quantitatively described risk levels, a forward cloud generator (CG) and a reverse cloud generator (CG-1) are constructed. CG generates N cloud point maps of a normal cloud distribution based on cloud digital characteristics (Ex, En, He), with the resultant variable `drop(x)`. i ,μ(x i ), while CG-1 has enabled a certain amount of data samples to be transformed into cloud digital features (Ex, En, He).

[0143] Overall Risk Index It is the core input of extension cloud theory and set pair analysis, first through By integrating dynamic risk values, passive risk values, and active risk values, and then inputting them into a reverse cloud generator to obtain cloud digital features (Ex, En, He), the membership cloud value is calculated, an evaluation matrix is ​​constructed, and finally a dynamic risk index m is obtained. After standardization, it is divided into four levels of risk: high, relatively high, medium, and low. That is, the comprehensive risk index is the preliminary quantitative basis for accurate risk level determination.

[0144] The risk index score is limited to [0, 100], therefore the calculation process of CG-1 can be expressed as:

[0145] (2)

[0146] Where Ex, En, and He represent the sample expectation, entropy, and hyperentropy of the cloud model, respectively; n is the number of samples, and S... 2 Let Variance be the variance.

[0147] Specifically, the comprehensive risk cloud of the evaluation unit is obtained by transforming the weight matrix and the cloud digital feature matrix. The correspondence between different indicators and risk levels can be expressed as follows:

[0148] (3)

[0149] Where α represents the cloud to which index i belongs at the corresponding level j, and is a certain assessment index value (such as the probability distribution of slope belonging to high risk, the probability distribution of wind speed belonging to medium risk, etc.). Based on this principle, a membership evaluation matrix Q and a fuzzy set cloud connectivity evaluation matrix can be constructed, and a dynamic risk index m can be obtained. In addition, referring to the risk level classification method, the interval is divided equally, which can realize the assessment of the accident risk level in the scene, as shown in Table 1.

[0150] Table 1 Accident Risk Level Classification

[0151]

[0152] Calculation of the grade range in Table 1:

[0153] 1. Calculation of Dynamic Risk Index:

[0154] A dynamic risk index is calculated using an extension cloud model and set pair analysis, combined with a weight matrix and a membership evaluation matrix. This index quantifies the coupling effect of risk factors through fuzzy set-based cloud connectivity evaluation, as shown in the following formula:

[0155]

[0156] in, The weights of each risk indicator, This refers to the cloud value belonging to the corresponding indicator.

[0157] 2. Standardization process:

[0158] Will Standardize to interval Standardize the units. The standardized formula is:

[0159]

[0160] in, and These are the minimum and maximum values ​​of the dynamic risk index in historical data.

[0161] 3. Interval division: Divide the standardized intervals equally. Each area is divided into four zones, corresponding to four risk levels:

[0162] High risk (IV): [-1, -0.5)

[0163] Higher risk (III): [ -0.5, 0)

[0164] Medium risk (II): [0, 0.5)

[0165] Low risk (I): [0.5,1).

[0166] In the risk analysis process of step S2, after the initial accident on the continuous curved slope section, the traffic flow at the accident site changes and propagates upstream, affecting the normal driving of upstream vehicles. As the accident scene is cleared and dealt with, the traffic flow continues to change. Upstream vehicles, affected by single or multiple factors such as weather conditions and road conditions, are unable to maintain safe driving, ultimately leading to a secondary accident. The process of the secondary accident is as follows: Figure 4 As shown.

[0167] Specifically, the causes of secondary accidents on highways can be categorized into two types based on whether they are avoidable: controllable factors and uncontrollable factors. Controllable factors primarily refer to those that can be changed according to the driver's subjective awareness, such as vehicle speed, following distance, and the driver's emotions; these are the direct causes of secondary accidents on highways. Uncontrollable factors mainly refer to weather conditions, road conditions, and the driving status of surrounding vehicles. These factors and their effects cannot be changed according to the driver's subjective awareness or behavior, and they are a significant cause of secondary accidents on highways.

[0168] The prediction of secondary accidents on highways is divided into two stages: first, predicting whether a secondary accident will occur; and second, predicting the time and location of the secondary accident. A schematic diagram of the secondary accident prediction process is shown below. Figure 5 As shown.

[0169] A Bayesian optimized random forest model is used to predict secondary accidents on highways with continuous slopes and curves. Multiple decision trees are constructed for classification or regression, and the final classification or regression result is determined by voting or averaging. During the training process, each decision tree is constructed from randomly sampled training samples and randomly selected features, ensuring that each decision tree has a certain degree of difference and increasing the generalization ability of the random forest. Random forest classification prediction demonstration. Figure 6 .

[0170] Specifically, Bayesian optimization is combined with random forest algorithms. Bayesian optimization is used to optimize the hyperparameters of the random forest algorithm. The optimization process is as follows: ① In each iteration, the Bayesian Optimized Random Forest (BORF) model selects a new set of hyperparameters using Bayesian optimization and trains a random forest model based on these hyperparameters. ② BORF predicts the objective function output based on the random forest model, i.e., the cross-validation score on the training set or the prediction accuracy on the test set. ③ BORF updates the surrogate model and selects new hyperparameters in the search space for the next round of optimization.

[0171] Compared to the traditional Random Forest algorithm, the BORF algorithm can find the optimal hyperparameter combination faster and reduces the risk of overfitting to some extent. The prediction results are of a binary classification type; therefore, the BORF model is used to predict whether a secondary accident will occur after the initial accident, represented by 1 and 0 respectively: a model output of 1 indicates a secondary accident has occurred, while 0 indicates no secondary accident will occur. Traffic management departments can use the prediction results to guide drivers in developing reasonable driving plans to minimize the occurrence of secondary accidents.

[0172] Real-time characteristic data of the initial accident includes: accident type (rear-end collision, loss of control, rollover, etc.); vehicle information (vehicle type (e.g., heavy truck), load status, braking performance); spatiotemporal information (accident time, location (GPS coordinates)); road alignment (slope, curve radius); environmental conditions (weather (rainfall, icing), road surface condition (friction coefficient, water accumulation), visibility).

[0173] Traffic flow dynamics data includes: flow and density (traffic flow upstream and downstream of the accident section, lane occupancy, vehicle density); speed distribution (average vehicle speed, speed standard deviation, abnormal acceleration / deceleration behavior); vehicle behavior (following distance, lane changing frequency, emergency braking events).

[0174] Vehicle and driver status data include: out-of-control vehicle status (trajectory deviation, brake failure signal, abnormal steering); surrounding vehicle status (speed, acceleration, relative distance to the accident vehicle); and driver sentiment (emotional score (positive / negative / neutral) analyzed through social media texts (such as Weibo and Douyin).

[0175] Static road characteristics data: road alignment (slope, curve curvature, sight distance limitation); safety facilities (guardrail integrity, speed bump installation, emergency lane availability).

[0176] Multi-source sensor data: UAV monitoring (aerial images, video streams, infrared thermal imaging data); roadside equipment (millimeter-wave radar point clouds, traffic scene videos from integrated radar-visual systems); vehicle terminals (vehicle operating parameters fed back by OBD (On-Board Diagnostics) systems).

[0177] Historical data: historical accident records (accident frequency, type, and casualties on the same road segment); traffic management data (variable message board information, electronic gate status, and speed limit adjustment records).

[0178] The driver sentiment analysis in the predictive model is primarily achieved through social media sentiment analysis technology. First, data collection and cleaning: accident-related text data is collected from social media platforms (such as Weibo and Douyin text API data interfaces) and then cleaned and preprocessed. Second, sentiment classification and quantification: the SnowNLP sentiment analysis algorithm is used to classify the text into sentiment categories (positive, negative, neutral) and calculate sentiment scores. Finally, sentiment trend prediction: through time series analysis or deep learning models, future trends in sentiment are predicted to support risk warnings.

[0179] The vehicle sensor data includes steering wheel angle, braking frequency, acceleration, and the number of times lane departure warning (LDW) is triggered, which are collected in real time through the OBD interface.

[0180] Vehicle-to-everything (V2V) data: real-time speed, location, and emergency braking signals from vehicle-to-vehicle (V2V) communication.

[0181] Step S3: Implement the secondary accident proactive control system based on road risk level.

[0182] Based on the terrain characteristics and traffic flow patterns of mountainous highways, a functional requirements analysis of the electronic barrier gate system was conducted to determine its control logic, hardware configuration, and communication protocol. Prototype testing was performed in a laboratory environment to ensure its reliability and response speed under complex road conditions. Based on the lane-level traffic flow characteristics of highways, the installation locations, spacing, and information content of the gantry and associated variable message signs were determined. The optimal timing and accuracy of information dissemination under different weather conditions and traffic events were considered to ensure the effectiveness and timeliness of information transmission. Furthermore, to avoid unnecessary damage caused by collisions between the electronic barrier gate and vehicles during traffic control, this section focuses on the design of a highly buffered, flexible barrier gate. A high-polymer elastic material combined with a brightly colored film is used as the crossbar 2. A horizontal displacement device for the crossbar 2 is designed to ensure that it can spring open to one side in the event of a high-speed collision with a vehicle.

[0183] According to the established deployment plan, electronic barriers and variable message signs were installed on the test section. Through on-site debugging, the working parameters and linkage mechanism of the equipment were optimized to ensure the coordinated capabilities of each device in accident early warning and emergency response.

[0184] Specifically, this study first investigates the basic algorithms for mainline speed control, ramp flow control, and lane management under fundamental scenarios, using a multi-objective function that integrates lane-level accident precursor feature elimination and overall highway operational efficiency. By establishing a multi-objective optimization function that merges dynamic risk field output and real-time traffic flow status, the traditional single efficiency objective is extended into a dual-objective framework that coordinates accident risk elimination and traffic efficiency optimization. Based on the Bayesian optimization algorithm, control parameters such as mainline speed limit, ramp adjustment rate, and lane activation scheme are solved collaboratively, thereby achieving the quantitative generation and dynamic optimization of management strategies under fundamental scenarios.

[0185] Secondly, in secondary accident scenarios, with the aim of eliminating secondary accidents, an emergency algorithm for secondary accidents is constructed, comprising lane interception, proactive lane-changing prompts, and emergency lane management. Based on the secondary accident prediction results of dynamic risk field distribution and Bayesian optimized random forest, an adaptive lane interception triggering mechanism linked to the spatiotemporal characteristics of the accident is established. Vehicle-level precise lane-changing guidance signals are generated and disseminated through multimodal perception data fusion. Furthermore, dynamic opening and closing rules for emergency lanes are constructed with the secondary accident risk level as the decision variable, thus forming a proactive prevention and control algorithm system that evolves in tandem with the accident risk situation. Based on this, a control scheme system for different scenarios is constructed for automatic scheme allocation.

[0186] Because the precursory characteristics of accidents vary significantly under different traffic flow conditions, the spatiotemporal impact of secondary accidents is affected by the location, scale, and current traffic flow status of the accident. The system automatically selects solutions from the control strategy library (specifically including lane interception schemes based on dynamic risk field distribution, active lane-changing guidance schemes integrating real-time vehicle trajectories, dynamic opening and closing management schemes for emergency lanes, variable speed limit control schemes for mainlines, graded flow regulation schemes for ramps, and collaborative warning schemes combining variable message signs and flashing warning lights; these schemes are adaptively selected and their parameters optimized based on the predicted location, scale, and real-time traffic flow status of the secondary accident). Combined with a simulation platform, the system calculates the optimal solutions for key parameters of single or combined schemes, while simultaneously providing feedback on the expected effects of the control scheme implementation, thereby achieving the goal of reducing accident occurrences and mitigating their impact.

[0187] The electronic barrier gate includes a horizontal bar 2, with columns 1 at both ends. A flashing light 3 and a height restriction sign 4 are mounted on the horizontal bar 2. A crossbeam 5 is located above the horizontal bar 2, with both ends fixedly connected to the columns 1. A display screen 6 is mounted on the crossbeam 5. The horizontal bar 2, made of a highly buffered flexible material and a thin-film composite structure, combined with a horizontally displaceable collision response mechanism, allows the barrier gate to spring open laterally upon vehicle impact, reducing impact damage. Simultaneously, the electronic barrier gate is communicatively connected to a platform center, enabling lane-level precise traffic control based on secondary accident prediction results. This constitutes a new type of active road control device integrating mechanical buffering, dynamic response, and intelligent decision-making.

[0188] Example 2,

[0189] The remote intelligent cloud platform for proactive prevention and control of secondary accidents is based on cloud computing principles and employs technologies such as virtualization, distributed storage, and distributed computing. The platform's basic resources, including computing, storage, and network, can be scaled up and adjusted according to the scale of the application demonstration and specific workload. The main architecture includes device access and management, a scene identification module, and a big data warehouse and computing capabilities, such as... Figure 7 As shown. The device access and management module mainly interfaces with sensing and control equipment, including low-altitude systems and roadside systems. Its main responsibilities include: ① Device registration and authentication: creating products and registering devices through the cloud platform's console or API to obtain identity authentication information; ② Device data upload and control command issuance: after establishing a connection between the device and the cloud platform, data can be uploaded in real time, and control commands from the cloud can be received; ③ The cloud platform can monitor the device's status, including online status and fault information, and can perform remote fault diagnosis and configuration management. The scene identification module integrates a large-scale traffic AI model to identify the existence of accident scenarios based on data transmitted from the devices, and optimizes the control algorithm with optimal parameters selected automatically or manually, providing decision-makers with detailed implementation procedures for proactive and emergency control solutions. For data storage, data warehouse technology is used. Through preset storage modes, multi-source heterogeneous data is saved, extracted, analyzed, and processed according to a homogeneous data structure.

[0190] Application demonstration of secondary accident prevention and control system for continuous curved and sloping road sections:

[0191] To meet the needs of monitoring road traffic conditions, detecting initial accidents, and proactively preventing secondary accidents, the factors to be considered during the deployment of the proactive secondary accident prevention system based on the demonstration road section include at least: setting up special drone patrol flights, deploying dedicated drone nests, deploying integrated radar and vision systems, and deploying electronic barriers and other warning equipment.

[0192] Specifically, taking the multi-rotor unmanned aerial vehicle (UAV) system for emergency command and control as an example, the equipment is planned to be deployed in the vicinity of the service area, as shown in the figure. It can guarantee a cruising radius of 10km and a loiter time of 30 minutes. At the same time, the UAV can conduct fixed-point monitoring at any point (the farthest end of the operating radius) during flight, with a hovering time of no less than 5 minutes. In addition, the equipment can be equipped with a visible light pod and a remote 4G loudspeaker device to remotely command and warn the site in real time based on the data transmitted back by the UAV, ensuring application safety.

[0193] To achieve proactive prevention and tiered early warning for secondary accidents, an integrated system of early warning and enforcement equipment, including electronic barriers, variable speed limits, variable message signs, and strobe warning lights, is implemented. A tiered early warning approach is used, analyzing factors such as terrain, slope, curves, visibility limitations, and accident-prone areas along the demonstration road section. Traffic flow changes at different times are examined to determine peak and off-peak hours, as well as vehicle type distribution, providing a basis for the equipment's operating hours and modes. Based on the performance parameters of the electronic barriers (raising and lowering speed), variable speed limit sign display range, variable message sign clarity, and strobe warning light brightness and coverage, the optimal placement of the equipment is determined.

[0194] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for proactive prevention and control of secondary accidents on mountain highways, characterized in that, Includes the following steps: Step S1: The entire area of ​​continuous curves and slopes on mountain highways is monitored by an integrated air-ground accident scene perception device. Traffic scene features are extracted by image recognition using convolutional neural networks and recurrent neural networks, and accident scenes are identified based on a generative pre-trained model. Step S2: Combining the current traffic flow status, road static characteristics, environmental conditions, and vehicle behavior, the risk of the first accident is predicted based on the risk-coupled accident risk prediction model. Using the real-time characteristics of the initial accident, environmental conditions, traffic flow status, initial accident risk, road static characteristics, and historical data as inputs, a Bayesian optimized random forest model is used to predict whether a secondary accident will occur on a highway with continuous slopes and curves. If it does occur, the time and location of the secondary accident are predicted. Step S3: Based on the predicted time and location of secondary accidents, proactive prevention and control measures are implemented through warning and mandatory control measures.

2. The method for proactive prevention and control of secondary accidents on mountainous highways according to claim 1, characterized in that, Step S1 includes the following steps: S11, the air-ground integrated accident scene perception equipment includes drones and roadside equipment. The drones acquire aerial visible light image data, infrared shortwave thermal imaging data, terrain and environmental information, meteorological data, and drone status data; the roadside equipment acquires millimeter radar point cloud data, roadside video streams, vehicle micro-state data, and road infrastructure status; the drones and roadside equipment work together to achieve full-area monitoring. S12 extracts spatial feature sequences from multi-view images using a convolutional neural network to generate 3D vehicle detection boxes, accurately representing the vehicle's position, size, and attitude in three-dimensional space. A recurrent neural network processes the spatial feature sequences extracted by the convolutional neural network, outputting a temporal dynamic feature vector to capture the dynamic evolution of vehicle trajectory, speed changes, and behavior patterns. SLAM extracts ORB feature points from UAV aerial images and stable reflection points from millimeter-wave radar, aligning the spatiotemporal references of UAV images and radar point clouds in real time. A Transformer cross-modal attention mechanism fuses heterogeneous multi-source data. S13, based on historical accident data, trains a generative pre-trained model. The multi-source fusion dataset is input into the trained generative pre-trained model to generate accident probability distribution and risk hotspot areas, thereby identifying typical accident patterns, including heavy truck loss of control and multi-vehicle rear-end collisions.

3. The method for proactive prevention and control of secondary accidents on mountain highways according to claim 2, characterized in that, In S1, accident risk features are introduced into the decision layer of the Transformer, and scene semantic understanding is enhanced and feature representation is optimized through generative pre-trained models. Bayesian inference and Kalman filtering are embedded in the encoding layer of the Transformer to suppress radar point cloud noise and video blur. Through a cloud-edge collaborative architecture, high-frequency, low-latency data is preprocessed at the edge nodes and the results are uploaded to the cloud. The cloud is responsible for batch processing of complex tasks and long-term prediction, improving the real-time performance and scalability of data processing.

4. The method for proactive prevention and control of secondary accidents on mountain highways according to claim 1, characterized in that, In step S2, the initial accident risk assessment includes the following steps: S21 breaks down traffic collision risks into passive risks and active risks; S22 calculates the total passive risk value by weighting the road boundary constraints, lane line constraints on vehicles, the impact of road surface slope and friction coefficient on vehicles, and the lateral or longitudinal risks posed by surrounding vehicles; when the real-time vehicle speed v > v max At that time, it was determined to be high risk, v max This indicates the threshold for maintaining a stable vehicle speed. S23, the active risk is the active collision risk caused by abnormal driving behavior. The driving behavior time series data is modeled by a long short-term memory network to identify abnormal driving behavior and output the risk value brought about by the active behavior of the vehicle; the collision time is less than a preset threshold and is judged as high risk. S24, Construct a dynamic risk field, the dynamic risk value is calculated using the following formula: ； Where R(x,t) represents the dynamic risk value at position x at time t; m i The risk quality of the i-th vehicle is represented by its speed v. i and load factor η i Decide, ;d i This represents the Euclidean distance between the target vehicle at position x at time t and the i-th vehicle; k represents the environmental correction factor, with a value ranging from 0.5 to 1.

5. S25 weights and integrates the total passive risk value, active risk value, and dynamic risk value into a comprehensive risk index to classify risk levels.

5. The method for proactive prevention and control of secondary accidents on mountain highways according to claim 4, characterized in that, The degree of danger of the abnormal driving behavior is quantified by the following formula: ； Among them, S abnormal α represents the quantification of the degree of danger of abnormal driving behavior. t For acceleration, θ t For steering wheel angle, LDW t denoted as the number of lane departures, α, β, and γ are weighting coefficients, t represents the time step, and T is the total number of time steps in the entire analysis period.

6. The method for proactive prevention and control of secondary accidents on mountain highways according to claim 4, characterized in that, S25 further includes: inputting the comprehensive risk index into the reverse cloud generator to obtain cloud digital features Ex, En, and He; calculating the membership cloud value α of the current value of the comprehensive risk index to each risk level; constructing an evaluation matrix to obtain the dynamic risk index m; and then... Standardization to range Risk levels are divided according to intervals.

7. The method for proactive prevention and control of secondary accidents on mountain highways according to claim 1, characterized in that, S3 includes the following steps: With the goals of eliminating lane-level accident precursor features and improving the overall operational efficiency of highways, we construct basic algorithms for mainline speed control, ramp flow control, and lane management in basic scenarios. In the case of secondary accidents, with the aim of eliminating secondary accidents, an emergency algorithm for secondary accidents is constructed, which includes lane interception, active lane change prompts, and emergency lane management. Based on the location and scale of the secondary accident and the current traffic flow data, a solution from the control strategy library is selected, and the optimal parameter combination is calculated through a simulation platform.

8. The method for proactive prevention and control of secondary accidents on mountain highways according to claim 1, characterized in that, The mandatory control measures include electronic gates, dynamic lane-level speed limits and closures, mandatory intervention by vehicle terminals, and emergency lane occupation control.

9. The method for proactive prevention and control of secondary accidents on mountain highways according to claim 1, characterized in that, The electronic barrier gate includes a flexible composite material crossbar (2) and a collision response mechanism. When the barrier gate is hit by a vehicle, it will open laterally to reduce impact damage. At the same time, the electronic barrier gate is connected to the platform center to achieve lane-level precise traffic control based on the secondary accident prediction results.

10. A proactive prevention and control system for secondary accidents on mountain highways, characterized in that, The method for proactive prevention of secondary accidents on mountain highways as described in claim 1 includes: The device access and management module is used for device registration and authentication, device data uploading and control command issuance, monitoring device status, and remote fault diagnosis and configuration management. The scene identification module is used to identify whether an accident scene exists based on the data returned by the device, and to optimize the control algorithm with the optimal parameters selected automatically or manually. The big data warehouse and computing module is used to save, extract, analyze and process multi-source heterogeneous data according to a homogeneous data structure through preset storage modes.

Images