Bridge operation risk targeted prevention and control method based on driver intention recognition

Abstract

The present application belongs to the field of intelligent transportation system and bridge safety prevention and control technology, and relates to a bridge operation risk targeted prevention and control method based on driver intention recognition, a bridge operation risk targeted prevention and control system, comprising: a radar-vision integrated machine for acquiring vehicle visual image data and radar point cloud data, a variable information display board, a sound column for emitting directional sound waves, a wireless radio, a fog lamp, a weather station, and a control unit for carrying out "one-to-one" accurate early warning on "careless" high-risk vehicles by combining the accurate identification and positioning of the radar-vision integrated machine, the weather conditions obtained by the weather station, the variable information display board, the sound column and the wireless radio. The present application can accurately identify driving intention by relying on the LSTM model, can actively predict risks and target alarm careless high-risk vehicles by fusing bridge multi-dimensional risks, has the advantages of identification intelligence and strong scene adaptability, and builds a bridge active safety prevention and control system integrating people-vehicle-road-environment.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation systems and bridge safety control technology, and relates to a targeted control method for bridge operation risks based on driver intent recognition. Specifically, it relates to an active control method that uses a deep learning model to identify driver intent and provides targeted warnings to drivers who are unaware of specific bridge risks. Background Technology

[0002] Bridges, as key nodes in road traffic, directly affect the smooth flow and safety of routes. Traditional bridge safety control methods mainly rely on fixed-point monitoring and variable message signs for macro-level information dissemination, but these methods have the following significant drawbacks: 1) Passive response: The system usually only issues warnings after events such as accidents or congestion occur, lacking the ability to proactively defend against risks; 2) Ignoring driver status: Existing systems rarely consider the driver's real-time intentions and behaviors; 3) Lack of targeted prevention and control: Traditional VMS broadcasts to all vehicles and cannot provide accurate warnings for specific high-risk vehicles.

[0003] If a secondary targeted warning can be issued to the highest-risk drivers based on the initial traffic flow warning, the effectiveness of the warning system can be significantly improved. This secondary targeted warning should target drivers who have not noticed the risk, i.e., those who have not shown any intention to slow down or change lanes. Although Long Short-Term Memory (LSTM) networks and Transformers have demonstrated powerful capabilities in time-series data processing and have been used to identify driving intentions, such as lane changes and overtaking, current technologies are primarily focused on Advanced Driver Assist Systems (ADAS) or behavioral analysis. Deep integration with active safety control systems for road infrastructure is still lacking to achieve targeted intervention for vehicles that fail to notice risks in specific risk scenarios.

[0004] Chinese patent document (application number 202510776648.6, publication date 2025.07.08) discloses a lane-change trajectory prediction model, which analyzes the current trajectory and historical trajectory of a vehicle and feeds back the prediction results to surrounding vehicles for safety warnings. This prediction model can inform surrounding vehicles of the lane-change intention of the vehicle, thereby achieving safe driving warnings. However, to achieve targeted prevention and control of bridge operation risks, this comparative document still has the following shortcomings: 1) This warning scheme only focuses on the interactive risks between vehicles and does not consider the inherent risks of the bridge (such as insufficient visibility, road surface water / snow / ice, crosswind interference, etc.) and superimposed risks (such as accidents blocking lanes, construction, debris, etc.), and cannot achieve "human-vehicle-road-environment" risk identification, resulting in the failure of the bridge risk warning strategy; 2) This scheme does not have targeted warning capabilities. Targeted bridge operation risk prevention and control not only needs to obtain the driver's lane-change intention, but also needs to obtain the deceleration intention, where the lane-change intention is used to assess the driver's attitude towards the road ahead. The response capability to lane-level risks such as lane closures, and the deceleration intention are used to assess the driver's response capability to road segment-level risks such as insufficient visibility on the bridge, sudden changes in strong crosswinds, reduced road surface adhesion due to bridge icing / slippery surfaces, traffic accidents or vehicle malfunctions on the bridge, bridge construction, lane closures caused by debris on the bridge, and slow-moving or congested queues of multiple vehicles ahead. Relying solely on lane-changing intentions cannot fully assess the driver's perception of risks; 3) Considering that existing warning methods target drivers, it is necessary to eliminate autonomous vehicles through trajectory analysis to reduce the number of targeted warning targets, but this patent does not disclose information in this regard.

[0005] Therefore, there is an urgent need for an active bridge control method that can accurately identify driver intentions, dynamically assess bridge operational risks, and provide targeted early warnings for specific vehicles, in order to solve the problems existing in current technologies. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a bridge operation risk targeted prevention and control method based on driver intention recognition.

[0007] To achieve the above objectives, the present invention provides the following technical solution: On one hand, this invention provides a bridge operation risk targeted prevention and control system based on driver intention recognition, including: a radar-visual integrated machine for acquiring vehicle visual image data and radar point cloud data, a variable information information board, a speaker emitting directional sound waves, a wireless broadcast, fog lights, and a weather station; and a control unit that uses the precise identification and positioning of the radar-visual integrated machine and the weather conditions acquired by the weather station, combined with the variable information information board, speaker, and wireless broadcast, to provide "one-to-one" precise early warning for high-risk vehicles that are "missed". The radar-visual integrated machine, variable information information board, and speaker are arranged every 150-200m in front of and in the middle of the bridge to ensure 100% system coverage; fog lights are arranged every 5-10m in the middle of the bridge; and the wireless broadcast and weather station are arranged in front of or in the middle of the bridge depending on the actual conditions of the bridge. The control unit is electrically connected to the radar-visual integrated machine, variable information information board, speaker, wireless broadcast, fog lights, and weather station respectively.

[0008] On the other hand, the present invention provides a prevention and control method based on the bridge operation risk targeted prevention and control system described above, comprising the following steps: Step 1: Use the integrated radar vision system to collect visual image data and radar point cloud data of vehicles on the target road section in real time. The control unit will extract the dynamic data sequence of the vehicle, align the visual image and radar point cloud in time and space, and fuse the visual image and radar point cloud at the target level in real time. Step 2: The control unit analyzes the vehicle dynamic data obtained in Step 1 to distinguish whether the vehicle is an autonomous vehicle or a manually driven vehicle by analyzing behavioral pattern characteristics, and performs differential filtering on the dynamic data to retain the dynamic data of manually driven vehicles. Step 3: Input the dynamic data sequence after vehicle type processing into the pre-trained LSTM neural network model to identify the driver's real-time intentions; Step 4: The control unit acquires the bridge's operational risk information in real time, matches the identified driver's real-time intentions with the operational risk information faced by the current vehicle in its lane, and determines whether the vehicle is in a safe or risky state; for vehicles in a risky state, a risk quantification assessment is performed to obtain different risk levels for the vehicles. Step 5: The control unit follows the progressive control and early warning principle of "first global notification, then targeted intervention". It first provides a global basic risk notification to vehicles on the bridge. If the control unit detects that the vehicle risk has not been eliminated, it then provides different levels of targeted early warning to the risky vehicles based on the vehicle risk level obtained in Step 4.

[0009] Specifically, step 1 is as follows: Step 1.1: The control unit continuously extracts the dynamic data sequence of the vehicle based on visual image data and radar point cloud data; the dynamic data sequence integrates visual features and radar features, including but not limited to: longitudinal displacement, relative distance to lane lines, and turn signal status extracted from visual features, and vehicle speed and longitudinal distance extracted from radar features. Step 1.2: Perform spatiotemporal alignment between the visual image and the radar point cloud: In the time dimension, data time synchronization is achieved through timestamp interpolation; on the basis of time alignment, spatial alignment is performed: using the pre-calibrated camera intrinsic and extrinsic parameter matrix and the projection matrix of the radar to the world coordinate system, the radar target is projected into the visual pixel coordinate system, and the two-dimensional bounding box of the vehicle recognized by the visual model is matched with the bounding box of the three-dimensional moving target extracted by the radar after projection by intersection-union ratio. Step 1.3: Target-level fusion of the matched visual image and radar point cloud: A fusion strategy is adopted to integrate the advantageous data of the two types of sensors. For target position information, the high precision of radar 3D positioning and the pixel-level details of the visual image are combined, and the position coordinates are fused and estimated using the Kalman filter algorithm to reduce positioning errors. For target motion state (velocity, direction), radar point cloud data is used as the main basis, combined with continuous frame visual images for correction to improve the smoothness and accuracy of motion state estimation. For target attributes (category, size), the visual image recognition results are used as the core, and radar point cloud data is used to supplement the target's 3D size information to improve the target description.

[0010] Specifically, step 2 also includes: real-time monitoring of all vehicle trajectories. If a vehicle exhibits strong autonomous driving characteristics (such as extremely high lane centering stability) in the previous time period, but suddenly experiences violent fluctuations in lateral displacement or a sudden drop in acceleration in a short period of time, the control unit determines that the vehicle has exited the autonomous driving system and is in the transition period of manual emergency takeover. For such high-risk vehicles with takeover failure, the control unit issues a warning mark.

[0011] Specifically, in step 2, the control unit distinguishes between autonomous vehicles and manually driven vehicles by analyzing the following behavioral pattern characteristics: Step 2.1: Calculate the stability of vehicle speed and lateral position during cruise: Using vehicle dynamic data, select a stable cruise trajectory segment where the vehicle does not change lanes, and calculate the change in lateral distance between the vehicle center and the lane centerline, as well as the change in speed. Autonomous driving systems, with their precise control capabilities, can maintain extremely low and stable centering errors and speed fluctuations, while manual driving exhibits continuous minor corrections and higher fluctuations. The specific calculation formula is as follows: (1) (2) in For the first i The lateral distance between the vehicle center and the lane centerline at the sampling point. v i For the first i Vehicle speed at sampling points, where N is the total number of samples and , The standard deviation of the lateral distance. The standard deviation of the velocity; or It is believed to be in autopilot mode at that time. or It was considered to be in manual driving mode at that time; Step 2.2: Assess the matching degree between turn signal use and steering behavior: This is determined by statistically analyzing the proportion of vehicles with a clear intention to move laterally but without activated turn signals, and the stability of the time difference between turn signal activation and the start of steering. Autonomous driving behavior typically strictly follows the rule of "prompt before action" with a fixed time difference, while manual driving behavior shows a higher proportion of forgetting or lag. The specific calculation formula is as follows: (3) Where R represents the probability of a turning action occurring without using the turn signal; Indicates the status of the turn signal. A value of 1 indicates that the turn signal is on. A value of 0 indicates that the turn signal was not activated; Indicates the state of turning behavior. =1 indicates that the vehicle has begun to deviate from its lane. =0 indicates that there is no deviation from the lane; At that time, it was assumed that the vehicle was in autonomous driving mode. At that time, it was assumed that the vehicle was in a manual driving state; Step 2.3: Analyze the consistency of the following process: Extract the longitudinal control features of the vehicle in the following scenario, and calculate the coefficient of variation of the headway and the root mean square of the rate of change of acceleration. The ACC system can stably maintain the preset headway through closed-loop control, and its headway coefficient of variation is usually less than 0.05, with a low root mean square value of acceleration and very few high-frequency components (the degree of change in the rate of change of acceleration). In contrast, due to perception delay and operational roughness, the headway coefficient of variation of manual driving is often greater than 0.15, and the error correction has overshoot and low-frequency oscillations, with a significantly larger root mean square value of acceleration. The control unit constructs a classification model based on the above multi-dimensional features, and classifies vehicles that exhibit typical stable ACC following characteristics as autonomous driving (including ACC) vehicles, and filters them from the subsequent warning process. The specific calculation formula is as follows: (4) (5) (6) in Indicates the first i The headway of the sampling point , For the first i The distance between the front of the vehicle at each sampling point For the first i Instantaneous velocity at each sampling point The coefficient of variation is 1. The standard deviation of the distance between the front ends of the vehicle. The root mean square of the rate of change of acceleration. Indicates the first i The acceleration of each sampling point; The vehicle was considered to be in autonomous driving mode at the time. It was considered to be in manual driving mode at that time.

[0012] Furthermore, when providing global basic risk notification, only for secondary warnings, the control unit is set to two modes: (1) if a warning is issued for a confirmed manually driven vehicle, then only the vehicles that meet the manual driving criteria are marked with a warning; (2) if a warning is not issued for a confirmed autonomous vehicle, then the vehicles that meet the autonomous driving criteria are removed from the detection vehicle set, and the remaining vehicles are marked with a warning.

[0013] Specifically, in step 3, the driving intention includes lane-changing intention and deceleration intention; the LSTM neural network model is trained with a large amount of historical driving data and can accurately predict the driving intention in the next few seconds based on the vehicle's behavior sequence in a short period of time. LSTM Input Feature Sequence Construction: For a target vehicle, the control unit extracts multi-dimensional features within a pre-defined time window of length T from the vehicle dynamic data obtained in step 1, forming an input sequence containing multiple time steps N. Each time step includes lateral motion features, longitudinal motion features, and state features. Lateral motion features include the vehicle's lateral velocity, lateral acceleration, and the distance between the vehicle's center and the lane centerline. Longitudinal motion features include the vehicle's longitudinal velocity and longitudinal acceleration. State features include the vehicle's turn signal status, headway to the vehicle ahead, and collision time. The LSTM network structure is as follows: Input layer: Receives a feature sequence of size (T, NL), where T is the sequence length and NL is the feature dimension at each time step.

[0014] The first LSTM layer has 128 hidden units and returns a complete time step sequence. This layer is mainly responsible for learning the underlying dynamic patterns of driving behavior, such as the temporal combination of short-term actions like acceleration, braking, and steering.

[0015] The second LSTM layer has 64 hidden units and returns the output of the last time step. This layer learns the temporal logic of higher-order intentions based on the first layer, such as recognizing whether a combination of actions is ultimately for lane changing or braking.

[0016] Dropout layer: A Dropout layer is set after each LSTM layer. This value needs to be adjusted after multiple training sessions. Its function is to randomly deactivate some neurons during training, which prevents the model from overfitting and enhances the robustness of the model.

[0017] Fully connected layer: Maps the final hidden state of the second LSTM layer to a dual-task classifier. Task 1 outputs the lateral intent probabilities: left and right lane change and lane keeping; Task 2 outputs the longitudinal intent probabilities: acceleration, deceleration and constant speed.

[0018] Loss function: The model uses a dual-task weighted loss function for optimization. The total loss is the sum of the horizontal intent loss and the vertical intent loss. Both tasks use the cross-entropy loss function, as shown in the following formula: (7) in For the calculated loss value, The total number of categories and , Encode the real label vector. Predict probabilities for the model.

[0019] Specifically, in step 4, the operational risk information includes: insufficient visibility on the bridge surface, sudden changes in strong crosswinds, icing / slippery bridge surface leading to a decrease in road surface adhesion coefficient, traffic accidents on the bridge surface, vehicle malfunctions, bridge surface construction, lane closures caused by debris on the bridge surface, and multiple vehicles slowing down or queuing ahead.

[0020] Specifically, in step 4, the identified driver's intent is matched with the risks faced by the lane where the vehicle is currently located, as follows: Step 4.1: Initial screening of risk intent and qualitative assessment of status Safety Status: The vehicle's driving intentions align with objective safety avoidance logic. In this state, both subjective and objective risks are in a safe equilibrium. The control unit only maintains background monitoring and does not trigger warnings. Safety status matching scenarios include: Longitudinal hazard avoidance matching: When there is a static lane occupant (accident, malfunction, construction, spilled material) or a low-speed moving bottleneck (multiple vehicles moving slowly, congestion and queuing) in front of this lane, a clear lane-changing intention or timely deceleration and braking intention is detected.

[0021] Lateral hazard avoidance matching: When faced with insufficient visibility on the bridge surface, crosswinds, icing / slippery conditions, etc., which result in insufficient driver visibility and reduced lateral stability of the vehicle, the system detects the intention to actively decelerate and maintain stable driving in the lane (without any intention to change lanes sharply or turn suddenly).

[0022] Risk Status: The vehicle's driving intention contradicts objective safety logic. For such mismatched vehicles, the control unit locks them as high-risk target vehicles and immediately triggers the subsequent quantitative assessment mechanism. High-risk status matching scenarios include: Ignoring longitudinal risks: When there is a static blockage (accident, breakdown, construction, spilled material) or a slow-moving bottleneck (multiple vehicles moving slowly, congestion and queuing) ahead of this lane, but the driver does not show any intention to slow down or change lanes, and continues to drive at the original speed and in the original lane, or even shows any intention to accelerate and follow.

[0023] Lateral loss of control: When visibility is poor on the bridge, there is crosswind, or the road is icy / waterlogged, drivers may still show signs of rapid acceleration, rapid deceleration, or violent and continuous lane changes, which can easily induce the vehicle to skid or overturn.

[0024] Incorrect evasive action: When one lane is blocked ahead, the driver attempts to change lanes towards the blocked lane.

[0025] Step 4.2, Real-time Risk Quantitative Assessment For high-risk vehicles, the control unit immediately initiates a real-time high-frequency recursive calculation mechanism based on a kinematic model. During the quantitative assessment process, the control unit reads the target vehicle's own speed and acceleration, relative speed and acceleration with other targets frame by frame at each time step based on the underlying dynamic data sequence, and combines it with its relative distance to the physical hazard source in front to calculate and update the time to collision (TTC) and safety margin (SM) in real time. TTC focuses on representing the urgency of collisions in the time dimension, and is calculated as follows: (8) in, This refers to the relative distance between the vehicle and the target ahead. This is the difference between the vehicle's speed and the speed of the target vehicle ahead.

[0026] SM, as a supplementary indicator to the spatial dimension, is used to characterize the physical distance difference between the vehicle's current position and its theoretically limit safe stopping position. This compensates for the near-infinite calculation problem of TTC in low-speed or near-close speed scenarios, and intuitively reflects whether there is still space margin for emergency avoidance steering or braking. The calculation method is as follows: (9) in, The relative distance between the vehicle and the target ahead (such as an accident vehicle, a stationary obstacle, or a slow-moving vehicle): The vehicle's current speed; The speed of the vehicle in front; Braking system coordination delay time typically includes the driver's reaction time and the pressure build-up response time of the mechanical braking system; This represents the maximum deceleration that the vehicle can achieve under the current road surface adhesion coefficient.

[0027] Furthermore, in step 4, the risk level is classified according to the possible response methods for each risk state (as shown in Table 1): if it includes deceleration and lane changing, indicating an intention to change lanes, it is judged as a level zero risk; otherwise, it is classified according to the case that only includes deceleration. Based on the dynamic attenuation degree of TTC and SM, the control unit continuously quantifies the risk level of the mismatched vehicle into three levels, so that the risk prediction of different levels has high real-time performance and distinguishability. Level 0 risk: TTC is greater than or equal to the preset Level 1 safety threshold. Level 1 safety threshold Defined as the time required for comfortable braking, its calculation formula is as follows: (10) in, The value is approximately 2.5 m / s 2 ; Potential risk (Level 1 risk): TTC is less than the preset Level 1 safety threshold. ; High risk (Level 2 risk): TTC is less than the Level 2 safety threshold. Furthermore, if the driver's intent remains unchanged, the probability of a collision increases significantly; Level 2 safety threshold. Defined as the time required for the vehicle's FCM braking, its calculation formula is as follows: (11) in, The value is approximately 5 m / s 2 ; Critical collision risk (Level 3 risk): TTC is less than the Level 3 safety threshold. and safety margin At this point, the target distance is less than the limit of physical braking distance, and the vehicle is in an extremely dangerous imminent collision. Conventional braking is no longer sufficient to avoid an accident, and the control unit determines it to be at the highest danger level; Level 3 safety threshold. Defined as the time required for the vehicle's AEB braking, its calculation formula is as follows: (12) in, The value is approximately 10 m / s 2 ; Note: If the vehicle is a heavy-duty truck, the first-level risk will be automatically upgraded to the second-level risk, and the second-level risk will be automatically upgraded to the third-level risk according to formula (10-12).

[0028] Specifically, in step 5, the early warning triggering is strictly divided into two stages: basic risk broadcasting and high-risk targeted intervention. Step 5.1, Mapping Early Warning Strategies and Broadcast Content: The control unit first extracts the corresponding standard response methods based on the specific risk factors of the current bridge; then, the control unit divides the early warning strategy into two progressive stages and maps different broadcast content to them: A single early warning for global risk notification: For all vehicles entering the upstream impact zone of the hazard source, a basic broadcast content template is extracted and widely disseminated to eliminate information asymmetry and guide drivers to proactively adjust their driving behavior; Secondary warning-specific instruction generation: For vehicles that fail to adjust their driving behavior according to the first warning reminder, the control unit extracts the characteristics of the target vehicle and dynamically splices them with a high-risk template to generate a set of exclusive warning instructions for the high-risk vehicle; The specific mapping relationship between typical bridge risk scenarios and the two-step early warning broadcast content is shown in Table 1. The first-step early warning adopts a combination of visual and auditory warnings. The visual warning is broadcast on VMS according to the first-step early warning content in Table 1; the auditory warning is based on directional sound wave arrays and wireless broadcasting to accurately send voice alarms to the specific area, and the wireless broadcasting improves the penetration of the warning information.

[0029] Step 5.2, Targeted Warning: The control unit continuously tracks the vehicle trajectory and movement status after the global risk notification in Step 5.1; when it is determined that a vehicle does not show a reasonable intention to avoid danger (i.e., is in a risky state), and is still locked as a high-risk or critical collision risk target (e.g., TTC is less than the level 2 safety threshold) after real-time risk quantification assessment, the vehicle is included in the targeted warning vehicle set; when the previous targeted warning vehicle passes the warning device, the control unit issues a targeted warning to the vehicle with the highest risk in the targeted warning vehicle set; the warning execution methods include: Visual warning: Force control the nearest variable message sign (VMS) in front of the target vehicle and highlight the exclusive warning information for that vehicle according to the secondary warning content in Table 1; Auditory warning: Activate the directional sound wave array deployed on the side of the bridge to issue a high-decibel directional sound wave warning towards the lane where the target vehicle is located.

[0030] It should be noted that the fog lights include yellow lights and red trail lights, as shown in Table 2. Based on the visibility identified by the weather station, visibility is categorized into four risk levels. When visibility is 100-200m, the control unit determines it to be a level four risk, and the yellow light remains on. When visibility is 50-100m, the control unit determines it to be a level three risk, and the yellow light flashes. When visibility is 30-50m, the control unit determines it to be a level two risk, and the yellow light flashes along with a red trail. When visibility is <30m, the control unit determines it to be a level one risk, and the yellow light flashes along with a red trail, reminding drivers to be aware of the foggy weather.

[0031] Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects: 1. Proactive Prevention and Control: This invention relies on the LSTM temporal deep learning model to accurately predict two key driving intentions of drivers: deceleration and lane changing. It combines multi-dimensional environmental risks such as bridge surface, weather, and road conditions to predict vehicle risk status in advance. It abandons the traditional delayed mode of passive alarm after a bridge accident occurs and intervenes before the driver notices the bridge surface hazard and is about to engage in dangerous driving behavior. It truly implements proactive defense against bridge traffic risks and shortens the risk response window period from the root.

[0032] 2. Targeted Warning: Relying on the integrated radar-visual system to achieve precise vehicle positioning on the road, and simultaneously linking multiple warning terminals such as VMS variable information signs and on-site sound columns; by identifying drivers' failure to slow down or change lanes, high-risk target vehicles are screened, while autonomous vehicles are eliminated to avoid invalid warnings. Unlike the traditional VMS's extensive mode of indiscriminate broadcasting across the entire road, warning information is only pushed to specific vehicles that have not responded to various risks on the bridge, greatly reducing the loss of invalid warnings and significantly improving the effectiveness of warning delivery.

[0033] 3. Intelligent Recognition: Employing an LSTM deep learning network with excellent time-series data processing capabilities to mine vehicle driving time-series features, it fully recognizes two types of driving behaviors: deceleration intention and lane-changing intention. Based on massive amounts of real bridge traffic data, it completes model iteration and optimization, resulting in higher accuracy in driver intention recognition. It is compatible with multiple complex risk scenarios such as sudden changes in crosswinds on the bridge surface, icy and slippery conditions, sudden drops in visibility, accidents involving lane occupancy, road construction, and road spills, adapting to the differentiated bridge environment and traffic flow changes, and ensuring stable and reliable assessment of driver risk perception status under multiple working conditions.

[0034] 4. System Integration: Breaking down the technical barriers between traditional vehicle-side ADAS driving analysis and road and bridge infrastructure early warning systems, it deeply integrates road surface radar perception and collection, bridge environmental risk monitoring, roadside facility communication transmission, and multi-channel early warning terminal execution full-link modules. It connects the data interaction of four types of risk elements: people (driver's intention), vehicles (vehicle operating status), roads (bridge road conditions), and environment (weather, emergencies), and constructs a closed-loop bridge active safety protection system of "perception-cognition-decision-execution", making up for the shortcomings of existing technologies that cannot coordinate multi-source risks. Attached Figure Description

[0035] The accompanying drawings are incorporated in and form part of this specification, and together with the description serve to explain the principles of the invention.

[0036] 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, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a diagram illustrating the overall system architecture and workflow of the present invention. Detailed Implementation

[0038] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples consistent with some aspects of the invention as detailed in the appended claims.

[0039] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0040] See Figure 1As shown, on one hand, the present invention provides a bridge operation risk targeted prevention and control system based on driver intention recognition, including: a radar-visual integrated machine for acquiring vehicle visual image data and radar point cloud data, a variable information information board, a speaker column emitting directional sound waves, a wireless broadcast, fog lights, a weather station, and a control unit that uses the precise identification and positioning of the radar-visual integrated machine and the weather conditions acquired by the weather station, combined with the variable information information board, speaker column, and wireless broadcast, to provide "one-to-one" precise early warning for high-risk vehicles that are "missed". The radar-visual integrated machine, variable information information board, and speaker column are arranged every 150-200m in front of and in the middle of the bridge to ensure 100% system coverage. Fog lights are arranged every 5-10m in the middle of the bridge. The wireless broadcast and weather station are arranged in front of or in the middle of the bridge depending on the actual situation of the bridge. The control unit is electrically connected to the radar-visual integrated machine, variable information information board, speaker column, wireless broadcast, fog lights, and weather station respectively.

[0041] On the other hand, the present invention provides a prevention and control method based on the bridge operation risk targeted prevention and control system described above, comprising the following steps: Step 1: Use the integrated radar vision system to collect visual image data and radar point cloud data of vehicles on the target road section in real time. The control unit will extract the dynamic data sequence of the vehicle, align the visual image and radar point cloud in time and space, and fuse the visual image and radar point cloud at the target level in real time. Step 2: The control unit analyzes the vehicle dynamic data obtained in Step 1 to distinguish whether the vehicle is an autonomous vehicle or a manually driven vehicle by analyzing behavioral pattern characteristics, and performs differential filtering on the dynamic data to retain the dynamic data of manually driven vehicles. Step 3: Input the dynamic data sequence after vehicle type processing into the pre-trained LSTM neural network model to identify the driver's real-time intentions; Step 4: The control unit acquires the bridge's operational risk information in real time, matches the identified driver's real-time intentions with the operational risk information faced by the current vehicle in its lane, and determines whether the vehicle is in a safe or risky state; for vehicles in a risky state, a risk quantification assessment is performed to obtain different risk levels for the vehicles. Step 5: The control unit follows the progressive control and early warning principle of "first global notification, then targeted intervention". It first provides a global basic risk notification to vehicles on the bridge. If the control unit detects that the vehicle risk has not been eliminated, it then provides different levels of targeted early warning to the risky vehicles based on the vehicle risk level obtained in Step 4.

[0042] Specifically, step 1 is as follows: Step 1.1: The control unit continuously extracts the dynamic data sequence of the vehicle based on visual image data and radar point cloud data; the dynamic data sequence integrates visual features and radar features, including but not limited to: longitudinal displacement, relative distance to lane lines, and turn signal status extracted from visual features, and vehicle speed and longitudinal distance extracted from radar features. Step 1.2: Perform spatiotemporal alignment between the visual image and the radar point cloud: In the time dimension, data time synchronization is achieved through timestamp interpolation; on the basis of time alignment, spatial alignment is performed: using the pre-calibrated camera intrinsic and extrinsic parameter matrix and the projection matrix of the radar to the world coordinate system, the radar target is projected into the visual pixel coordinate system, and the two-dimensional bounding box of the vehicle recognized by the visual model is matched with the bounding box of the three-dimensional moving target extracted by the radar after projection by intersection-union ratio. Step 1.3: Target-level fusion of the matched visual image and radar point cloud: A fusion strategy is adopted to integrate the advantageous data of the two types of sensors. For target position information, the high precision of radar 3D positioning and the pixel-level details of the visual image are combined, and the position coordinates are fused and estimated using the Kalman filter algorithm to reduce positioning errors. For target motion state (velocity, direction), radar point cloud data is used as the main basis, combined with continuous frame visual images for correction to improve the smoothness and accuracy of motion state estimation. For target attributes (category, size), the visual image recognition results are used as the core, and radar point cloud data is used to supplement the target's 3D size information to improve the target description.

[0043] Specifically, step 2 also includes: real-time monitoring of all vehicle trajectories. If a vehicle exhibits strong autonomous driving characteristics (such as extremely high lane centering stability) in the previous time period, but suddenly experiences violent fluctuations in lateral displacement or a sudden drop in acceleration in a short period of time, the control unit determines that the vehicle has exited the autonomous driving system and is in the transition period of manual emergency takeover. For such high-risk vehicles with takeover failure, the control unit issues a warning mark.

[0044] Specifically, in step 2, the control unit distinguishes between autonomous vehicles and manually driven vehicles by analyzing the following behavioral pattern characteristics: Step 2.1: Calculate the stability of vehicle speed and lateral position during cruise: Using vehicle dynamic data, select a stable cruise trajectory segment where the vehicle does not change lanes, and calculate the change in lateral distance between the vehicle center and the lane centerline, as well as the change in speed. Autonomous driving systems, with their precise control capabilities, can maintain extremely low and stable centering errors and speed fluctuations, while manual driving exhibits continuous minor corrections and higher fluctuations. The specific calculation formula is as follows: (1) (2) in For the first i The lateral distance between the vehicle center and the lane centerline at the sampling point. v i For the first i Vehicle speed at sampling points, where N is the total number of samples and , The standard deviation of the lateral distance. The standard deviation of the velocity; or It is believed to be in autopilot mode at that time. or It was considered to be in manual driving mode at that time; Step 2.2: Assess the matching degree between turn signal use and steering behavior: This is determined by statistically analyzing the proportion of vehicles with a clear intention to move laterally but without activated turn signals, and the stability of the time difference between turn signal activation and the start of steering. Autonomous driving behavior typically strictly follows the rule of "prompt before action" with a fixed time difference, while manual driving behavior shows a higher proportion of forgetting or lag. The specific calculation formula is as follows: (3) Where R represents the probability of a turning action occurring without using the turn signal; Indicates the status of the turn signal. A value of 1 indicates that the turn signal is on. A value of 0 indicates that the turn signal was not activated; Indicates the state of turning behavior. =1 indicates that the vehicle has begun to deviate from its lane. =0 indicates that there is no deviation from the lane; At that time, it was assumed that the vehicle was in autonomous driving mode. At that time, it was assumed that the vehicle was in a manual driving state; Step 2.3: Analyze the consistency of the following process: Extract the longitudinal control features of the vehicle in the following scenario, and calculate the coefficient of variation of the headway and the root mean square of the rate of change of acceleration. The ACC system can stably maintain the preset headway through closed-loop control, and its headway coefficient of variation is usually less than 0.05, with a low root mean square value of acceleration and very few high-frequency components (the degree of change in the rate of change of acceleration). In contrast, due to perception delay and operational roughness, the headway coefficient of variation of manual driving is often greater than 0.15, and the error correction has overshoot and low-frequency oscillations, with a significantly larger root mean square value of acceleration. The control unit constructs a classification model based on the above multi-dimensional features, and classifies vehicles that exhibit typical stable ACC following characteristics as autonomous driving (including ACC) vehicles, and filters them from the subsequent warning process. The specific calculation formula is as follows: (4) (5) (6) in Indicates the first i The headway of the sampling point , For the first i The distance between the front of the vehicle at each sampling point For the first i Instantaneous velocity at each sampling point The coefficient of variation is 1. The standard deviation of the distance between the front ends of the vehicle. The root mean square of the rate of change of acceleration. Indicates the first i The acceleration of each sampling point; The vehicle was considered to be in autonomous driving mode at the time. It was considered to be in manual driving mode at that time.

[0045] Furthermore, when providing global basic risk notification, only for secondary warnings, the control unit is set to two modes: (1) if a warning is issued for a confirmed manually driven vehicle, then only the vehicles that meet the manual driving criteria are marked with a warning; (2) if a warning is not issued for a confirmed autonomous vehicle, then the vehicles that meet the autonomous driving criteria are removed from the detection vehicle set, and the remaining vehicles are marked with a warning.

[0046] Specifically, in step 3, the driving intention includes lane-changing intention and deceleration intention; the LSTM neural network model is trained with a large amount of historical driving data and can accurately predict the driving intention in the next few seconds based on the vehicle's behavior sequence in a short period of time. LSTM Input Feature Sequence Construction: For a target vehicle, the control unit extracts multi-dimensional features within a pre-defined time window of length T from the vehicle dynamic data obtained in step 1, forming an input sequence containing multiple time steps N. Each time step includes lateral motion features, longitudinal motion features, and state features. Lateral motion features include the vehicle's lateral velocity, lateral acceleration, and the distance between the vehicle's center and the lane centerline. Longitudinal motion features include the vehicle's longitudinal velocity and longitudinal acceleration. State features include the vehicle's turn signal status, headway to the vehicle ahead, and collision time. The LSTM network structure is as follows: Input layer: Receives a feature sequence of size (T, NL), where T is the sequence length and NL is the feature dimension at each time step.

[0047] The first LSTM layer has 128 hidden units and returns a complete time step sequence. This layer is mainly responsible for learning the underlying dynamic patterns of driving behavior, such as the temporal combination of short-term actions like acceleration, braking, and steering.

[0048] The second LSTM layer has 64 hidden units and returns the output of the last time step. This layer learns the temporal logic of higher-order intentions based on the first layer, such as recognizing whether a combination of actions is ultimately for lane changing or braking.

[0049] Dropout layer: A Dropout layer is set after each LSTM layer. This value needs to be adjusted after multiple training sessions. Its function is to randomly deactivate some neurons during training, which prevents the model from overfitting and enhances the robustness of the model.

[0050] Fully connected layer: Maps the final hidden state of the second LSTM layer to a dual-task classifier. Task 1 outputs the lateral intent probabilities: left and right lane change and lane keeping; Task 2 outputs the longitudinal intent probabilities: acceleration, deceleration and constant speed.

[0051] Loss function: The model uses a dual-task weighted loss function for optimization. The total loss is the sum of the horizontal intent loss and the vertical intent loss. Both tasks use the cross-entropy loss function, as shown in the following formula: (7) in For the calculated loss value, The total number of categories and , Encode the real label vector. Predict probabilities for the model.

[0052] Specifically, in step 4, the operational risk information includes: insufficient visibility on the bridge surface, sudden changes in strong crosswinds, icing / slippery bridge surface leading to a decrease in road surface adhesion coefficient, traffic accidents on the bridge surface, vehicle malfunctions, bridge surface construction, lane closures caused by debris on the bridge surface, and multiple vehicles slowing down or queuing ahead.

[0053] Specifically, in step 4, the identified driver's intent is matched with the risks faced by the lane where the vehicle is currently located, as follows: Step 4.1: Initial screening of risk intent and qualitative assessment of status Safety Status: The vehicle's driving intentions align with objective safety avoidance logic. In this state, both subjective and objective risks are in a safe equilibrium. The control unit only maintains background monitoring and does not trigger warnings. Safety status matching scenarios include: Longitudinal hazard avoidance matching: When there is a static lane occupant (accident, malfunction, construction, spilled material) or a low-speed moving bottleneck (multiple vehicles moving slowly, congestion and queuing) in front of this lane, a clear lane-changing intention or timely deceleration and braking intention is detected.

[0054] Lateral hazard avoidance matching: When faced with insufficient visibility on the bridge surface, crosswinds, icing / slippery conditions, etc., which result in insufficient driver visibility and reduced lateral stability of the vehicle, the system detects the intention to actively decelerate and maintain stable driving in the lane (without any intention to change lanes sharply or turn suddenly).

[0055] Risk Status: The vehicle's driving intention contradicts objective safety logic. For such mismatched vehicles, the control unit locks them as high-risk target vehicles and immediately triggers the subsequent quantitative assessment mechanism. High-risk status matching scenarios include: Ignoring longitudinal risks: When there is a static blockage (accident, breakdown, construction, spilled material) or a slow-moving bottleneck (multiple vehicles moving slowly, congestion and queuing) ahead of this lane, but the driver does not show any intention to slow down or change lanes, and continues to drive at the original speed and in the original lane, or even shows any intention to accelerate and follow.

[0056] Lateral loss of control: When visibility is poor on the bridge, there is crosswind, or the road is icy / waterlogged, drivers may still show signs of rapid acceleration, rapid deceleration, or violent and continuous lane changes, which can easily induce the vehicle to skid or overturn.

[0057] Incorrect evasive action: When one lane is blocked ahead, the driver attempts to change lanes towards the blocked lane.

[0058] Step 4.2, Real-time Risk Quantitative Assessment For high-risk vehicles, the control unit immediately initiates a real-time high-frequency recursive calculation mechanism based on a kinematic model. During the quantitative assessment process, the control unit reads the target vehicle's own speed and acceleration, relative speed and acceleration with other targets frame by frame at each time step based on the underlying dynamic data sequence, and combines it with its relative distance to the physical hazard source in front to calculate and update the time to collision (TTC) and safety margin (SM) in real time. TTC focuses on representing the urgency of collisions in the time dimension, and is calculated as follows: (8) in, This refers to the relative distance between the vehicle and the target ahead. This is the difference between the vehicle's speed and the speed of the target vehicle ahead.

[0059] SM, as a supplementary indicator to the spatial dimension, is used to characterize the physical distance difference between the vehicle's current position and its theoretically limit safe stopping position. This compensates for the near-infinite calculation problem of TTC in low-speed or near-close speed scenarios, and intuitively reflects whether there is still space margin for emergency avoidance steering or braking. The calculation method is as follows: (9) in, The relative distance between the vehicle and the target ahead (such as an accident vehicle, a stationary obstacle, or a slow-moving vehicle): The vehicle's current speed; The speed of the vehicle in front; Braking system coordination delay time typically includes the driver's reaction time and the pressure build-up response time of the mechanical braking system; This represents the maximum deceleration that the vehicle can achieve under the current road surface adhesion coefficient.

[0060] Furthermore, in step 4, the risk level is classified according to the possible response methods for each risk state (as shown in Table 1): if it includes deceleration and lane changing, indicating an intention to change lanes, it is judged as a level zero risk; otherwise, it is classified according to the case that only includes deceleration. Based on the dynamic attenuation degree of TTC and SM, the control unit continuously quantifies the risk level of the mismatched vehicle into three levels, so that the risk prediction of different levels has high real-time performance and distinguishability. Level 0 risk: TTC is greater than or equal to the preset Level 1 safety threshold. Level 1 safety threshold Defined as the time required for comfortable braking, its calculation formula is as follows: (10) in, The value is approximately 2.5 m / s 2 ; Potential risk (Level 1 risk): TTC is less than the preset Level 1 safety threshold. ; High risk (Level 2 risk): TTC is less than the Level 2 safety threshold. Furthermore, if the driver's intent remains unchanged, the probability of a collision increases significantly; Level 2 safety threshold. Defined as the time required for the vehicle's FCM braking, its calculation formula is as follows: (11) in, The value is approximately 5 m / s 2 ; Critical collision risk (Level 3 risk): TTC is less than the Level 3 safety threshold. and safety margin At this point, the target distance is less than the limit of physical braking distance, and the vehicle is in an extremely dangerous imminent collision. Conventional braking is no longer sufficient to avoid an accident, and the control unit determines it to be at the highest danger level; Level 3 safety threshold. Defined as the time required for the vehicle's AEB braking, its calculation formula is as follows: (12) in, The value is approximately 10 m / s 2 ; Note: If the vehicle is a heavy-duty truck, the first-level risk will be automatically upgraded to the second-level risk, and the second-level risk will be automatically upgraded to the third-level risk according to formula (10-12).

[0061] Specifically, in step 5, the early warning triggering is strictly divided into two stages: basic risk broadcasting and high-risk targeted intervention. Step 5.1, Mapping Early Warning Strategies and Broadcast Content: The control unit first extracts the corresponding standard response methods based on the specific risk factors of the current bridge; then, the control unit divides the early warning strategy into two progressive stages and maps different broadcast content to them: A single early warning for global risk notification: For all vehicles entering the upstream impact zone of the hazard source, a basic broadcast content template is extracted and widely disseminated to eliminate information asymmetry and guide drivers to proactively adjust their driving behavior; Secondary warning-specific instruction generation: For vehicles that fail to adjust their driving behavior according to the first warning reminder, the control unit extracts the characteristics of the target vehicle and dynamically splices them with a high-risk template to generate a set of exclusive warning instructions for the high-risk vehicle; The specific mapping relationship between typical bridge risk scenarios and the two-step early warning broadcast content is shown in Table 1. The first-step early warning adopts a combination of visual and auditory warnings. The visual warning is broadcast on VMS according to the first-step early warning content in Table 1; the auditory warning is based on directional sound wave arrays and wireless broadcasting to accurately send voice alarms to the specific area, and the wireless broadcasting improves the penetration of the warning information.

[0062] Step 5.2, Targeted Warning: The control unit continuously tracks the vehicle trajectory and movement status after the global risk notification in Step 5.1; when it is determined that a vehicle does not show a reasonable intention to avoid danger (i.e., is in a risky state), and is still locked as a high-risk or critical collision risk target (e.g., TTC is less than the level 2 safety threshold) after real-time risk quantification assessment, the vehicle is included in the targeted warning vehicle set; when the previous targeted warning vehicle passes the warning device, the control unit issues a targeted warning to the vehicle with the highest risk in the targeted warning vehicle set; the warning execution methods include: Visual warning: Force control the nearest variable message sign (VMS) in front of the target vehicle and highlight the exclusive warning information for that vehicle according to the secondary warning content in Table 1; Auditory warning: Activate the directional sound wave array deployed on the side of the bridge to issue a high-decibel directional sound wave warning towards the lane where the target vehicle is located.

[0063] It should be noted that the fog lights include yellow lights and red trail lights, as shown in Table 2. Based on the visibility identified by the weather station, visibility is categorized into four risk levels. When visibility is 100-200m, the control unit determines it to be a level four risk, and the yellow light remains on. When visibility is 50-100m, the control unit determines it to be a level three risk, and the yellow light flashes. When visibility is 30-50m, the control unit determines it to be a level two risk, and the yellow light flashes along with a red trail. When visibility is <30m, the control unit determines it to be a level one risk, and the yellow light flashes along with a red trail, reminding drivers to be aware of the foggy weather.

[0064] To demonstrate the effectiveness of the present invention, the following embodiments are provided for verification. Example

[0065] To clearly illustrate the working principle of this invention, the following explanation will be based on a specific bridge risk scenario.

[0066] This embodiment uses a bridge as an example. This bridge has been equipped with the bridge operation risk targeted prevention and control system described in this invention, including a radar-visual integrated machine, variable message signs (VMS), directional acoustic arrays (sound columns), and equipment such as wireless broadcasting, a weather station, and fog lights. All equipment is interconnected with the bridge's local control unit. Each device in the bridge operation risk targeted prevention and control system is positioned appropriately according to the actual conditions of the bridge. Assume a two-vehicle rear-end collision occurs in the middle of the bridge, occupying the rightmost lane. For example... Figure 1 As shown, the bridge operation risk targeted prevention and control system begins with data collection by the radar-visual integrated device, and operates automatically throughout the entire process from risk detection to issuing early warning commands. The radar-visual integrated device captures the location and traffic status of vehicles involved in accidents in real time. The data is transmitted back to the control unit in a specific format. The LSTM model analyzes driving intentions, identifies potential risks of downstream vehicles being overlooked, and then triggers directional loudspeakers to broadcast warning information. Simultaneously, variable message signs dynamically display targeted warning prompts, achieving a rapid "perception-based warning" response. Specific aspects such as perception, monitoring, risk matching, and intention recognition are as follows: Vehicle target extraction and alignment fusion based on radar vision: A radar vision integrated machine deployed upstream of the accident site collects raw data in real time and extracts video features (such as relative distance to lane lines and turn signal status) and radar features (such as speed and acceleration) of the target vehicle. The control unit generates a structured dynamic data sequence containing information such as vehicle ID, position, speed, acceleration, and turn signal status through spatiotemporal alignment and target-level fusion.

[0067] Vehicle type screening: Based on the fused data, the control unit analyzes the behavioral characteristics of the target vehicle. It calculates the standard deviation of its lateral position during stable cruise. If this value exceeds a preset threshold, it exhibits typical human-driven fluctuation characteristics; simultaneously, no short-term changes from extreme stability to drastic fluctuations are detected, ruling out a takeover failure state. Based on this comprehensive assessment, the vehicle is determined to be a manually driven vehicle and is included in the subsequent warning process.

[0068] Driver Intent Recognition: The control unit inputs the dynamic data sequence of the target vehicle from the most recent few seconds into a pre-trained LSTM neural network model. The model outputs lateral intent (high probability of lane keeping, low probability of lane changing) and longitudinal intent (high probability of maintaining a constant speed). The recognition conclusion is that the vehicle's current primary intent is "to maintain a straight line and drive at a constant speed".

[0069] Bridge Risk Assessment: First, risk information is acquired. The bridge operation risk control system sends the risk information "the rightmost lane is closed due to an accident, requiring a lane change to the left" to the control unit. Next, an initial risk intent screening is performed. For vehicles traveling in lanes about to close, the correct avoidance logic should be "change lanes to the left" or "emergency deceleration," but this vehicle's intent is "to continue straight," a serious mismatch. The control unit determines it to be in a "risk state," indicating the driver is unaware of the impending risk. Finally, a real-time risk quantification assessment is conducted, updating the vehicle's Time to Collision (TTC) and Safety Margin (SM) in real time. Based on the risk type (accident blocking lanes) and the driver's lack of intention to change lanes or brake suddenly, the system classifies it as a potential risk and continuously tracks it.

[0070] Bridge risk warning triggering: Following the progressive principle of "global notification first, then targeted intervention," a global risk notification is first issued. The control unit uses VMS, wireless broadcasts, and loudspeakers upstream of the accident point to issue general risk warnings to all vehicles entering the affected area (such as "Accident ahead, please slow down and change lanes"). The system continuously tracks the vehicle and finds that it does not respond to the global broadcast and continues to drive at a constant speed in its original lane, with no change in intent. The control unit then generates a specific warning command, forcibly controlling the nearest VMS in front of the vehicle to highlight the warning information containing its license plate number, such as "***** (license plate number), accident ahead, please change lanes to the left," while simultaneously activating a directional sound wave array to issue a directional voice alarm. Finally, the radar-visual integrated machine detects that the vehicle activates its turn signal and begins to change lanes after receiving the warning, and the control unit removes it from the high-risk vehicle group.

[0071] It should also be noted that if the visibility is 50-100m as identified by the weather station, and the control unit determines that the visibility is at risk level three, then the control unit will control the fog lights to flash yellow to remind the vehicle to pay attention.

[0072] Table 1 Typical Risk Scenarios and Standard Response Strategies for Bridges Table 2. Fog Light Usage Risk Scenarios and Standard Response Strategies The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention.

[0073] It should be understood that the present invention is not limited to the content already described above, and various modifications and changes can be made without departing from its scope. The scope of the present invention is limited only by the appended claims.

Claims

1. A bridge operation risk targeted prevention and control system based on driver intent recognition, characterized in that, include: The system comprises a radar-visual integrated machine for acquiring vehicle visual image data and radar point cloud data, a variable message sign, a speaker emitting directional sound waves, a wireless broadcast, fog lights, and a weather station. It also includes a control unit that utilizes the radar-visual integrated machine's precise identification and positioning, along with weather conditions obtained from the weather station, combined with the variable message sign, speaker, and wireless broadcast to provide precise one-to-one warnings to high-risk vehicles that have been overlooked. The radar-visual integrated machine, variable message sign, and speaker are positioned every 150-200 meters in front of and in the middle of the bridge; fog lights are positioned every 5-10 meters in the middle of the bridge; and the wireless broadcast and weather station are positioned in front of or in the middle of the bridge. The control unit is electrically connected to the radar-visual integrated machine, variable message sign, speaker, wireless broadcast, fog lights, and weather station.

2. The prevention and control method based on the bridge operation risk targeted prevention and control system of claim 1, characterized in that, Includes the following steps: Step 1: Use the integrated radar vision system to collect visual image data and radar point cloud data of vehicles on the target road section in real time. The control unit will extract the dynamic data sequence of the vehicle, align the visual image and radar point cloud in time and space, and fuse the visual image and radar point cloud at the target level in real time. Step 2: The control unit analyzes the vehicle dynamic data obtained in Step 1 to distinguish whether the vehicle is an autonomous vehicle or a manually driven vehicle by analyzing behavioral pattern characteristics, and performs differential filtering on the dynamic data to retain the dynamic data of manually driven vehicles. Step 3: Input the dynamic data sequence after vehicle type processing into the pre-trained LSTM neural network model to identify the driver's real-time intentions; Step 4: The control unit acquires the bridge's operational risk information in real time, matches the identified driver's real-time intentions with the operational risk information faced by the current vehicle in its lane, and determines whether the vehicle is in a safe or risky state. For vehicles in a risky state, a risk quantification assessment is conducted to obtain different risk levels for the vehicles; Step 5: The control unit follows the progressive control and early warning principle of "first global notification, then targeted intervention". It first provides a global basic risk notification to vehicles on the bridge. If the control unit detects that the vehicle risk has not been eliminated, it then provides different levels of targeted early warning to the risky vehicles based on the vehicle risk level obtained in Step 4.

3. The bridge operation risk targeted prevention and control method according to claim 2, characterized in that, Step 1 is as follows: Step 1.1: The control unit continuously extracts dynamic data sequences of the vehicle based on visual image data and radar point cloud data; the dynamic data sequences integrate visual features and radar features. Including but not limited to: longitudinal displacement, relative distance to lane lines, and turn signal status extracted from visual features, as well as vehicle speed and longitudinal distance extracted from radar features; Step 1.2: Perform spatiotemporal alignment between the visual image and the radar point cloud: In the time dimension, data time synchronization is achieved through timestamp interpolation; on the basis of time alignment, spatial alignment is performed: using the pre-calibrated camera intrinsic and extrinsic parameter matrix and the projection matrix of the radar to the world coordinate system, the radar target is projected into the visual pixel coordinate system, and the two-dimensional bounding box of the vehicle recognized by the visual model is matched with the bounding box of the three-dimensional moving target extracted by the radar after projection by intersection-union ratio. Step 1.3: Perform target-level fusion of the matched visual image and radar point cloud: A fusion strategy is adopted to integrate the advantageous data of the two types of sensors. For target position information, the high precision of radar 3D positioning and the pixel-level details of visual images are combined, and the position coordinates are fused and estimated by Kalman filtering algorithm to reduce positioning error. For target motion state, radar point cloud data is used as the main basis, and continuous frame visual images are combined for correction to improve the smoothness and accuracy of motion state estimation. For target attributes, visual image recognition results are used as the core, and radar point cloud data is used to supplement the target's three-dimensional size information to improve the target description.

4. The bridge operation risk targeted prevention and control method according to claim 2, characterized in that, Step 2 also includes: real-time monitoring of all vehicle trajectories. If a vehicle exhibits strong autonomous driving characteristics in the previous time period, but suddenly experiences severe fluctuations in lateral displacement or a sudden drop in acceleration within a short period of time, the control unit determines that the vehicle has exited the autonomous driving system and is in the transition period of manual emergency takeover. For such high-risk vehicles with takeover failure, the control unit issues a warning mark.

5. The bridge operation risk targeted prevention and control method according to claim 2, characterized in that, In step 2, the control unit distinguishes between autonomous vehicles and manually driven vehicles by analyzing the following behavioral pattern characteristics: Step 2.1: Calculate the stability of the vehicle's speed and lateral position during cruising: Using vehicle dynamic data, select a stable cruising trajectory segment where the vehicle has not changed lanes, and calculate the changes in the lateral distance between the vehicle's center and the lane centerline, as well as the changes in speed; the specific calculation formula is as follows: （1） （2） in For the first i The lateral distance between the vehicle center and the lane centerline at the sampling point. v i For the first i Vehicle speed at sampling points, where N is the total number of samples and , The standard deviation of the lateral distance. The standard deviation of the velocity; or It is believed to be in autopilot mode at that time. or It was considered to be in manual driving mode at that time; Step 2.2: Assess the matching degree between turn signal use and steering behavior: This is determined by statistically analyzing the proportion of vehicles with a clear intention to move laterally but whose turn signals are not activated, and the stability of the time difference between the turn signal illuminating and the start of steering; the specific calculation formula is as follows: （3） Where R represents the probability of a turning action occurring without using the turn signal; Indicates the status of the turn signal. A value of 1 indicates that the turn signal is on. A value of 0 indicates that the turn signal was not activated; Indicates the state of turning behavior. =1 indicates that the vehicle has begun to deviate from its lane. =0 indicates that there is no deviation from the lane; At that time, it was assumed that the vehicle was in autonomous driving mode. At that time, it was assumed that the vehicle was in a manual driving state; Step 2.3: Analyze the consistency of the following process: Extract the longitudinal control features of the vehicle in the following scenario, and calculate the coefficient of variation of the headway and the root mean square of the rate of change of acceleration. The ACC system can stably maintain a preset headway through closed-loop control, with a headway variation coefficient of less than 0.05, and a low root mean square value for the rate of change of acceleration with very few high-frequency components. In contrast, due to perception delays and operational roughness, the headway variation coefficient of manual driving is often greater than 0.15, error correction suffers from overshoot and low-frequency oscillations, and the root mean square value of acceleration is significantly larger. Based on the above multi-dimensional characteristics, the control unit constructs a classification model to identify vehicles exhibiting typical stable ACC following characteristics as autonomous vehicles and filters them from the subsequent warning process. The specific calculation formula is as follows: （4） （5） （6） in Indicates the first i The headway of the sampling point , For the first i The distance between the front of the vehicle at each sampling point For the first i Instantaneous velocity at each sampling point The coefficient of variation is 1. The standard deviation of the distance between the front ends of the vehicle. The root mean square of the rate of change of acceleration. Indicates the first i The acceleration of each sampling point; The vehicle was considered to be in autonomous driving mode at the time. It was considered to be in manual driving mode at that time.

6. The bridge operation risk targeted prevention and control method according to claim 5, characterized in that, When providing global basic risk notification, the control unit is set to two modes: (1) If a warning is issued for a confirmed manually driven vehicle, only the vehicles that meet the requirements for manual driving will be marked with a warning. (2) If no warning is issued for a confirmed autonomous vehicle, the above-mentioned vehicles that meet the autonomous driving criteria will be removed from the detection vehicle set, and the remaining vehicles will be marked with a warning.

7. The bridge operation risk targeted prevention and control method according to claim 2, characterized in that, In step 3, the driving intention includes lane-changing intention and deceleration intention; the LSTM neural network model is trained with a large amount of historical driving data and can accurately predict the driving intention in the next few seconds based on the vehicle's behavior sequence in a short period of time.

8. The bridge operation risk targeted prevention and control method according to claim 2, characterized in that, In step 4, the identified driver's intent is matched with the risks faced by the current vehicle in its lane, specifically: Step 4.1: Initial screening of risk intent and qualitative assessment of status Safety status: The vehicle's driving intentions conform to objective safety avoidance logic. In this state, subjective and objective risks are in a safe equilibrium. The control unit only maintains background monitoring and does not trigger warnings. Risk status: The vehicle's driving intentions contradict objective safety logic. For such mismatched vehicles, the control unit locks them as high-risk target vehicles and immediately triggers the subsequent quantitative assessment mechanism. Step 4.2, Real-time Risk Quantitative Assessment For high-risk vehicles, the control unit immediately initiates a real-time high-frequency recursive calculation mechanism based on the kinematic model. During the quantitative assessment process, the control unit reads the target vehicle's own speed and acceleration and its relative speed and acceleration with other targets frame by frame at each time step based on the underlying dynamic data sequence. Combined with its relative distance to the physical hazard source in front, the control unit calculates and updates the collision time TTC and safety margin SM in real time. The formula for calculating TTC is: （8） in, This refers to the relative distance between the vehicle and the target ahead. This is the difference between the vehicle's speed and the speed of the target vehicle ahead; The formula for calculating SM is: （9） in, The relative distance between the vehicle and the target ahead; The vehicle's current speed; The speed of the vehicle in front; Braking system coordination delay time; This represents the maximum deceleration that the vehicle can achieve under the current road surface adhesion coefficient.

9. The bridge operation risk targeted prevention and control method according to claim 8, characterized in that, In step 4, the risk level is classified according to the possible response methods for each risk state: if it includes deceleration and lane changing, indicating an intention to change lanes, it is judged as a level zero risk; otherwise, it is classified according to the case that only includes deceleration. Based on the dynamic attenuation degree of TTC and SM, the control unit continuously quantifies the risk level of the mismatched vehicle into three levels, so that the risk prediction of different levels has high real-time performance and distinguishability. Level 0 risk: TTC is greater than or equal to the preset Level 1 safety threshold. Level 1 safety threshold Defined as the time required for comfortable braking, its calculation formula is as follows: （10） in, The value is approximately 2.5 m / s 2 ; Potential risk: TTC is less than the preset Level 1 safety threshold. ; High risk: TTC is less than the level 2 safety threshold Furthermore, if the driver's intent remains unchanged, the probability of a collision increases significantly; Level 2 safety threshold. Defined as the time required for the vehicle's FCM braking, its calculation formula is as follows: （11） in, The value is approximately 5 m / s 2 ; Critical collision risk: TTC less than Level 3 safety threshold and safety margin At this point, the target distance is less than the limit of physical braking distance, and the vehicle is in an extremely dangerous imminent collision. Conventional braking is no longer sufficient to avoid an accident, and the control unit determines it to be at the highest danger level; Level 3 safety threshold. Defined as the time required for the vehicle's AEB braking, its calculation formula is as follows: （12） in, The value is approximately 10 m / s 2 .

10. The bridge operation risk targeted prevention and control method according to claim 2, characterized in that, In step 5, the early warning triggering is divided into two stages: basic risk broadcasting and high-risk targeted intervention. Step 5.1: Basic Risk Notification and Mapping of Early Warning Strategies and Broadcast Content: The control unit first extracts the corresponding standard response methods based on the specific risk factors of the current bridge; then, the control unit divides the early warning strategy into two progressive stages and maps different broadcast content to them: One warning, global risk notification: For all vehicles entering the upstream impact zone of the hazard source, the basic broadcast content template is extracted and widely disseminated to eliminate information asymmetry and guide drivers to proactively adjust their driving behavior; Secondary warning and dedicated instruction generation: For vehicles that fail to adjust their driving behavior according to the first warning reminder, the control unit extracts the characteristics of the target vehicle and dynamically splices them with a high-risk template to generate a dedicated warning instruction set for the high-risk vehicle. Step 5.2, Targeted Warning: The control unit continuously tracks the vehicle trajectory and motion status after the global risk notification in Step 5.1; when it is determined that a vehicle does not show a reasonable intention to avoid danger, and is still locked as a high-risk or critical collision risk target after real-time risk quantification assessment, the vehicle is included in the targeted warning vehicle set. When the previous targeted warning vehicle passes the warning device, the control unit issues a targeted warning to the vehicle with the highest risk in the targeted warning vehicle set.

Images