A tunnel vehicle dynamic tracking and linkage induction control method

Abstract

This application discloses a method for dynamic vehicle tracking and coordinated guidance control in tunnels. The method includes: S1: multi-source information holographic perception and fusion; S2: tunnel digital twin construction and real-time synchronization; S3: situational analysis and risk prediction based on the digital twin; S4: dynamic collaborative guidance strategy generation, with the optimization objective of minimizing the total virtual cost of the system. Based on the output of S3, personalized vehicle guidance strategies are generated for vehicles in the tunnel, and tunnel environment control strategies are generated in conjunction; S5: precise guidance command issuance and execution feedback. This application, by deploying a multi-source heterogeneous perception network including radar-visual integrated machines, environmental sensors, and V2X communication, and employing advanced spatiotemporal registration and data fusion algorithms, constructs a unified real-time data base covering individual vehicle trajectories, traffic flow status, tunnel environment, infrastructure status, and driver micro-states. This breaks through the limitations of traditional single perception methods and provides reliable data support for refined management.

Description

Technical Field

[0001] This application belongs to the field of intelligent transportation technology, specifically relating to a method for dynamic tracking and linkage guidance control of vehicles in tunnels. Background Technology

[0002] As a critical node and special enclosed section of the road traffic network, tunnels have always presented a key challenge in traffic engineering: ensuring traffic safety and operational efficiency. Traditional tunnel monitoring and management systems typically rely on isolated detection equipment (such as coils and video cameras) and independently operating subsystems (such as lighting, ventilation, and traffic signals), which have the following limitations:

[0003] First, at the perception level, traditional methods rely heavily on single or a few types of sensors, making it difficult to achieve continuous, high-precision, and holographic tracking at the vehicle level. For example, video surveillance is susceptible to environmental factors such as changes in lighting and dust within tunnels, resulting in insufficient stability; fixed-section detection cannot provide the vehicle's entire behavioral trajectory. For non-connected vehicles, there is a lack of effective channels for acquiring individual vehicle states; and for connected vehicles, the data they provide has not been deeply fused and verified with roadside perception.

[0004] Secondly, at the decision-making and control level, existing systems often exhibit a "siloed" structure, with each subsystem (such as traffic guidance, lighting control, and ventilation) operating independently according to its own logic, lacking coordination. Traffic guidance is mostly based on macro-level traffic flow or simple event triggers, issuing static and uniform instructions (such as speed limits across the entire route), unable to provide dynamic and personalized guidance for individual vehicles. Environmental control (such as lighting and ventilation) often adopts timed, fixed-level, or on / off control based on simple environmental thresholds (such as illuminance and CO concentration), resulting in delayed response and a disconnect from real-time traffic flow conditions and driver visual needs, making it difficult to achieve precise energy conservation while ensuring safety.

[0005] Furthermore, in terms of risk assessment and early warning, existing technologies mostly focus on post-event alarms or simple warnings based on fixed rules (such as speeding and congestion), lacking the ability to integrate, proactively predict, and quantitatively assess multimodal risks such as traffic conflicts, environmental risks, facility status, and driver status. In particular, the direct impact of dynamic changes in the tunnel's lighting environment on the driver's visual perception effectiveness is generally ignored in existing control logic.

[0006] Therefore, existing technical solutions suffer from problems such as fragmented perception, isolated decision-making, extensive control, and insufficient risk foresight when dealing with the complex dynamic traffic environment inside tunnels, making it difficult to achieve the goal of safe, efficient, and energy-saving refined collaborative operation. Summary of the Invention

[0007] This application provides a method for dynamic vehicle tracking and linkage guidance control in tunnels, aiming to solve the problems of existing technologies that make it difficult to achieve continuous, high-precision, and holographic tracking at the vehicle level, and the lack of coordination among subsystems.

[0008] A method for dynamic vehicle tracking and linkage guidance control in a tunnel, the method comprising:

[0009] S1: Multi-source information holographic perception and fusion, utilizing the Internet of Things sensing network deployed in the tunnel to collect real-time vehicle individual information, traffic flow status information, tunnel infrastructure status information and tunnel environment information;

[0010] S2: Construction and real-time synchronization of tunnel digital twin. Based on the tunnel geometry and facility static model and the real-time sensing data of S1, a high-precision tunnel digital twin model including a dynamic traffic body model and an environmental field model is constructed and dynamically updated.

[0011] S3: Situational analysis and risk prediction based on digital twins, including micro-driving behavior analysis, macro-traffic situation assessment, multi-modal risk fusion prediction, and driver perception effectiveness evaluation based on the digital twin model;

[0012] S4: Dynamic collaborative guidance strategy generation, with the optimization objective of minimizing the total virtual cost of the system, generates personalized vehicle guidance strategies for vehicles in the tunnel based on the output of S3, and generates tunnel environment control strategies in conjunction with it.

[0013] S5: Precise guidance command issuance and execution feedback, which sends the collaborative control command package to the target vehicle and tunnel environmental control facilities through multimodal channels, and tracks the actual effect of vehicle response and environmental adjustment to form a closed-loop control.

[0014] Optionally, in S1, the Internet of Things sensing network includes a fixed roadside sensing unit, a distributed environmental sensing unit, and a mobile vehicle terminal.

[0015] The fixed roadside sensing unit includes a high-precision radar-visual integrated machine and a roadside unit. The high-precision radar-visual integrated machine adopts hardware-level fusion of millimeter-wave radar and high-definition video camera to output lane-level vehicle trajectory data.

[0016] Optionally, in step S2, the generation and real-time synchronization of the dynamic traffic model includes:

[0017] When the perception system first stably tracks the physical vehicle, a digital image of the vehicle is instantiated in the twin and assigned a globally unique identifier;

[0018] The vehicle digital image receives state updates from the perception fusion layer at a frequency of no less than 10Hz, and maintains a time difference of less than 100 milliseconds with the physical entity through timestamp alignment and interpolation algorithms.

[0019] Optionally, in S2, the environmental field model includes a light environment field model and an airflow field and pollutant diffusion model;

[0020] The optical environment field model is constructed by combining real-time monitoring data-driven and physical mechanism model simulation, and uses sensor networks for data assimilation and dynamic correction.

[0021] Optionally, in step S3, the multimodal risk fusion prediction includes:

[0022] Traffic conflict risk calculation based on short-term predicted trajectories of vehicle digital mirrors;

[0023] The environmental risk index is calculated from the output of the integrated environmental field model.

[0024] Receive infrastructure status information to assess the impact of facility failures;

[0025] The above risk values ​​are standardized and weighted to generate a real-time risk heatmap.

[0026] Optionally, in S3, the driver perception performance evaluation includes:

[0027] Based on the background brightness distribution provided by the light environment field model, and combined with the characteristics of the target object and the driver's state data, the visibility level is calculated.

[0028] The driver's average reaction time is estimated based on visibility level, scene complexity, and driver status.

[0029] Optionally, in S4, the total virtual cost of the system is modeled as a weighted sum of safety cost, efficiency cost, energy consumption cost, and driver psychological compliance cost;

[0030] The driver's psychological compliance cost is dynamically calculated using a probability model based on the vehicle's historical driving behavior characteristics and real-time contextual characteristics.

[0031] Optionally, in step S4, the optimization decision variables include vehicle-oriented control variables and facility-oriented control variables;

[0032] The optimization process is executed on a rolling basis with a fixed period, combining the latest twin state and prediction information to solve the optimal control problem in the finite time domain.

[0033] Optionally, in step S5, the multimodal hierarchical instruction issuance mechanism includes:

[0034] Personalized guidance instructions at the vehicle level can be issued to connected vehicles via C-V2X or RSU;

[0035] Variable information signs and lane indicators are used to issue group-level guidance and mandatory control instructions to all vehicles;

[0036] Environmentally enhanced guidance instructions are issued through intelligent road luminous markings or side wall projection devices;

[0037] The industrial control network sends coordinated control commands to lighting and ventilation facilities.

[0038] Optionally, in S5, the closed-loop control includes: after the instruction is executed, receiving feedback on the actual execution status and comparing it with the expected instruction;

[0039] A multi-dimensional quantitative evaluation of the induction control effect based on digital twins;

[0040] The execution feedback data will be used for twin state correction and model iterative optimization.

[0041] Compared with the prior art, this application has at least the following beneficial effects:

[0042] This application utilizes a multi-source heterogeneous sensing network, including integrated radar-visual systems, environmental sensors, and V2X communication, and employs advanced spatiotemporal registration and data fusion algorithms to construct a unified real-time data base covering individual vehicle trajectories, traffic flow status, tunnel environment, infrastructure status, and driver micro-states. This breaks through the limitations of traditional single sensing methods and provides reliable data support for refined management.

[0043] This application creates a virtual image of a tunnel by integrating and synchronously updating a static BIM model, a real-time dynamic traffic body, and a dynamic environmental field (light and air). This twin not only achieves "what you see is what you get" visual monitoring, but also serves as a unified sandbox for simulation, prediction, and optimization, providing a precise virtual experimental environment for subsequent analysis and decision-making.

[0044] This application integrates multi-dimensional information such as traffic conflicts, environmental hazards (low visibility, pollutants), and facility failures, performs fusion calculations and dynamic risk classification, generates a visualized real-time risk heat map, and integrates a light environment field model and a driver perception efficiency assessment model. It can quantify the driver's "visibility level" and theoretical reaction time at a specific location, and elevate the impact of environmental risks on safety from qualitative judgment to quantitative analysis, greatly enhancing the foresight and scientific nature of risk warning. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating a method for dynamic vehicle tracking and linkage guidance control in a tunnel, as provided in this application. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments.

[0047] This application provides a method for dynamic vehicle tracking and linkage guidance control in tunnels, comprising the following steps:

[0048] S1. Multi-source information holographic perception and fusion: Utilizing the Internet of Things sensing network deployed in the tunnel, real-time collection of vehicle individual information, traffic flow status information, tunnel infrastructure status information, and tunnel environment information;

[0049] Specifically, the Internet of Things (IoT) sensing network is composed of fixed roadside sensing units, distributed environmental sensing units, and mobile vehicle terminals working together.

[0050] Fixed roadside sensing units are deployed at tunnel entrances, exits, internal functional sections, and at predetermined intervals (recommended 200-300 meters) on cross-sections. Integrated sensing stations are deployed, and the core equipment of each sensing station includes:

[0051] High-precision radar-visual integrated device: This device integrates millimeter-wave radar and high-definition video cameras at the hardware level. The millimeter-wave radar operates in the 77GHz band, achieving a positional accuracy of 0.1 meters and a velocity accuracy of 0.1 meters per second, accurately measuring the distance, radial velocity, and angle information of target vehicles. The high-definition camera has a resolution of at least 1920×1080 and a frame rate of at least 25fps, using deep learning algorithms (such as the YOLO series and DeepSORT) to achieve vehicle detection, license plate / vehicle type recognition, and visual tracking. The radar-visual data undergoes spatiotemporal registration and target-level fusion through an embedded fusion processor, outputting lane-level vehicle trajectories (distinguishing between main lanes, overtaking lanes, and emergency lanes). Data attributes include: unique tracking ID, 3D coordinates (x, y, z), velocity vector (vx, vy), acceleration, vehicle outline dimensions, license plate number (anonymized), and vehicle type. The data refresh rate is at least 10Hz.

[0052] The roadside unit supports C-V2X or DSRC communication protocols and periodically broadcasts information such as local tunnel maps, traffic events, and guidance instructions. At the same time, it receives vehicle status data (such as BSM messages) sent by the on-board unit that conforms to the standards of connected vehicles, as a supplement and verification of the roadside perception data.

[0053] The deployment of distributed environmental sensing units includes:

[0054] An ambient light sensing network is constructed along the tunnel's longitudinal direction, with illuminance sensors spaced at intervals not exceeding 50 meters above the lane centerline and at specific heights on the sidewalls, to measure the horizontal and vertical illuminance of the road surface. Simultaneously, luminance meters are deployed at representative cross-sections to measure road surface brightness and brightness uniformity. All optical sensors require regular calibration to ensure data accuracy.

[0055] An air quality and disaster sensing network is established, deploying multi-gas detectors at 100-150 meter intervals along the tunnel ceiling to monitor parameters such as carbon monoxide (CO), nitrogen oxides (NOx), and visibility (transmittance method) in real time. Heat and smoke detectors are installed in critical locations such as cable trays and equipment boxes. Wind speed and direction sensors are installed near ventilation fans to monitor ventilation efficiency.

[0056] Infrastructure status awareness involves equipping each lighting fixture, fan, lane indicator, and fire pump with a smart controller and current / voltage sensor to report its on / off status, operating mode, energy consumption, fault codes, and brightness / speed adjustments in real time.

[0057] Meanwhile, for connected vehicles equipped with a combined vehicle-mounted global satellite navigation system and inertial navigation unit positioning system, high-precision positioning data (after encryption and anonymization), heading angle, yaw rate, and other data can be uploaded to the roadside unit or central platform via V2X communication. This data can be used for redundancy verification with the roadside sensing trajectory, or to provide continuous tracking in roadside sensing blind spots (such as areas of temporary equipment failure).

[0058] High-definition camera arrays based on near-infrared technology, equipped with high-performance image processing units, are deployed at key sections such as tunnel entrances, exits, and long downhill slopes. By capturing driver facial images, a localized, lightweight deep learning model is run to analyze the following metrics in real time:

[0059] Attention concentration: Based on head posture estimation (yaw, pitch angle) and gaze direction tracking, determine whether the gaze is continuously deviating from the road ahead;

[0060] Visual load and fatigue: Eye movement analysis was used to calculate blink frequency, average closure time (PERCLOS), and pupil diameter change rate per unit time. Continuous abnormal dilation of pupil diameter can indirectly reflect increased visual accommodative load caused by low illumination.

[0061] Physiological stress level: Using the principle of photoplethysmography, pulse wave signals are extracted from facial videos to estimate heart rate and heart rate variability, serving as a rough indicator of stress response;

[0062] All sensing data must be accurately tagged with UTC timestamps and device spatial coordinate labels. The system maintains a unified spatiotemporal benchmark, ensuring time consistency across all devices through a clock synchronization protocol (such as PTP). At the data fusion layer, Kalman filtering or more advanced deep learning-based multi-target tracking algorithms are used to correlate and fuse the trajectories of targets from different radar-visual integrated devices and at different times across devices and cycles, forming a continuous, smooth, and unique set of all vehicle trajectories within the tunnel. Environmental parameters and infrastructure status data are then correlated and aligned with traffic flow data in the spatiotemporal dimension, together constituting the native data stream required for digital twin synchronization.

[0063] S2. Construction and real-time synchronization of tunnel digital twin: Based on the tunnel geometry and facility static model and the real-time sensing data of step S1, a high-precision tunnel digital twin model containing a dynamic traffic body model and an environmental field model is constructed and dynamically updated to achieve synchronous mapping between the physical tunnel and the virtual model.

[0064] Specifically, the construction and synchronization process of the tunnel digital twin includes:

[0065] S2.1: Based on the Building Information Model (BIM) from the tunnel engineering design phase, construct a lightweight, structured tunnel information model. This process includes:

[0066] Model Lightweighting and Conversion: The original design BIM model is simplified through meshing, instantiated, and converted to a lightweight 3D model suitable for real-time rendering and spatial calculation. The format can be glTF or a dedicated engineering format. The model must accurately include all structural features such as the tunnel lining inner contour, road surface elevation, lane geometry, maintenance walkways, and cross passages.

[0067] Facility Asset Digitization and Semantic Association: All electromechanical facilities within the tunnel were digitally modeled and registered. Each light fixture, each fan, each variable message sign, and each fire hydrant was created as an independent "asset object." Each asset object was associated with its unique asset code, physical installation location (3D coordinates), design parameters (such as light distribution curves for light fixtures and airflow-pressure curves for fans), the system it belongs to (lighting system, ventilation system, etc.), and its component ID in the BIM model. This process established a semantic link between the facility's "geometric entity," "functional attributes," and "real-time status," providing a foundation for subsequent status mapping and control command issuance.

[0068] S2.2: Generation and real-time synchronization mechanism of dynamic traffic body model. The dynamic traffic body is the core dynamic element in the twin that corresponds one-to-one with the physical vehicle and evolves in real time.

[0069] When the S1 fusion perception system first stably tracks a physical vehicle, it instantiates a digital image of the vehicle in the twin. This image is given a globally unique identifier, which is bound to the tracking ID provided by the perception system and weakly correlated with the intercepted vehicle feature information (such as license plate hash value) to support re-identification. The life cycle of the image begins when the vehicle enters the tunnel perception domain and ends when it leaves the tunnel perception domain and is confirmed after a delay.

[0070] The core state attributes of the vehicle's digital mirror include its 3D position, velocity vector, heading angle, lane location, and vehicle size type. These attributes are received from the perception fusion layer via a dedicated data bus at a frequency of at least 10Hz. The synchronization process employs timestamp alignment and interpolation algorithms to ensure smooth motion of the virtual mirror and a time difference of less than 100 milliseconds with the physical entity. Furthermore, any state predictions or virtual test results made by the twin on the vehicle's digital mirror can be fed back to the physical vehicle via the V2X link as augmented information (e.g., predictive collision warning).

[0071] Modeling of Behavior and Intention (Short-Term Prediction): The vehicle's digital mirror not only reflects its current state but also embeds a simplified behavioral model. Based on its historical trajectory sequence, Kalman filtering, a constant-speed turning model, or a lightweight recurrent neural network are used to make short-term predictions of its trajectory within the next 2-5 seconds. This predicted position and speed will be used for conflict risk assessment in S3.

[0072] S2.3: Construction and Dynamic Update of the Environmental Field Model. The environmental fields within the tunnel, such as light, wind, and smoke, are crucial to safety and energy conservation. The twin model is constructed through a combination of "real-time monitoring data-driven" and "physical mechanism model simulation," including a light environment field model and an airflow field and pollutant diffusion model. Among these:

[0073] The construction of the lighting environment field model is based on the precise geometric structure in the tunnel information model, the digital assets of all lighting facilities, and their physical and optical characteristics. First, detailed optical parameters for each luminaire are extracted from the facility's digital assets, including but not limited to its light distribution curve data (usually stored in IES or LDT standard format), rated luminous flux, color temperature, and installation location (three-dimensional coordinates) and orientation (pitch angle, yaw angle). These parameters are combined with the real-time reported on / off status and dimming level (0-100%) of the luminaire asset objects to form a complete description set of all effective light sources at the current moment.

[0074] Using the light source description set and the material reflection characteristics parameters of the tunnel surface (such as the reflectivity of the road surface and sidewalls), a simplified radiative transfer algorithm is employed for calculation. One specific implementation involves discretizing the tunnel's inner surfaces (road surface, sidewalls, and ceiling) into a large number of tiny surface elements. Based on the visibility factor, the direct illumination contribution of all light sources to each surface element and the indirect illumination contribution after multiple diffuse reflections are pre-calculated or calculated in real-time, thereby generating a baseline brightness and illuminance distribution map of each sampling point on the road surface and sidewalls throughout the entire tunnel. This calculation can be performed offline, with pre-calculation and result storage in a database for all possible luminaire combinations and dimming levels to support real-time rapid querying; alternatively, it can be performed online in real-time using high-performance computing units and algorithms such as simplified photon mapping.

[0075] To achieve consistency between the virtual model and the real environment, the aforementioned baseline distribution map needs to be assimilated with the data collected in real time by the distributed illuminance / brightness sensor network deployed in S1. The system establishes a mapping relationship between the actual measurement locations of the sensors and their corresponding predicted points in the twin model. A state estimation algorithm, such as Kalman filtering or its variants (e.g., unscented Kalman filtering), is used to continuously correct the light environment model. The state vector of this filter can include latent variables such as the light decay coefficient of the lamps and local variations in the reflectivity of the tunnel surface. The filter uses real-time sensor readings as observations and the calculated values ​​from the baseline model as predictions, and through iterative calculations, dynamically estimates the optimal light field distribution that best explains all current observation data. This generates a high-confidence, continuous light environment distribution field throughout the tunnel, with a spatial resolution higher than the deployment density of physical sensors.

[0076] This dynamic light environment field model, as a core service module, can output multiple quantitative evaluation indicators in real time for any specified location within the tunnel (especially the field of view area in front of the vehicle's digital mirror location). These indicators include, but are not limited to, road surface brightness, overall and longitudinal uniformity of road surface brightness, and glare evaluation indicators such as threshold increments calculated according to CIE standards. More importantly, this model provides direct input for evaluating the driver's visual perception effectiveness. Based on the current vehicle position, driver's eye level, line of sight direction, and background brightness distribution provided by the model, combined with a standard small target visibility model (such as based on contrast thresholds), the system can calculate the theoretical visibility distance or "visibility level" required for the driver to detect obstacles ahead. This calculation result will serve as one of the key decision-making bases for risk assessment in S3 and guidance strategies and linked lighting dimming control in S4, enabling on-demand lighting and energy-saving optimization while ensuring safe visibility distance.

[0077] Based on computational fluid dynamics (CFD) principles, a simplified CFD simulation model of a tunnel is established. Real-time traffic flow data (vehicle number, speed, and location as mobile heat sources and pollutant source intensity) and fan operating status (start / stop, speed, and wind direction) are used as boundary conditions and source terms to drive the simulation model for rapid solution (a reduced-order model or a pre-trained surrogate model can be used to improve speed). The simulation results are compared and corrected with measured data from deployed CO / visibility sensors, dynamically outputting three-dimensional distribution cloud maps of wind speed and direction, temperature stratification, and CO and smoke concentrations at different cross-sections within the tunnel. This model is used for smoke spread prediction and ventilation strategy optimization in fire scenarios.

[0078] S3. Based on the digital twin, conduct micro-driving behavior analysis, macro-traffic situation assessment, multi-modal risk fusion prediction, and driver perception effectiveness evaluation, and output real-time risk map, traffic condition prediction, and perception effectiveness indicators.

[0079] The specific analysis steps for S3 are as follows:

[0080] S3.1: Microscopic Driving Behavior Anomaly Detection. The analysis engine processes the motion trajectory sequence of each vehicle's digital image in real time. By setting a series of rule models based on physical thresholds and statistical laws, it automatically identifies potential abnormal driving behaviors, including:

[0081] Sudden deceleration / sudden braking behavior: When the longitudinal deceleration of the vehicle's digital image continuously exceeds a preset threshold (e.g., -3.5m / s²) for a certain period of time, and is not caused by an obstacle or congestion in front, it is marked as abnormal sudden braking.

[0082] Snake-like driving behavior: Calculate the frequency and amplitude of lateral acceleration changes in the vehicle's digital image. If, within a short time window, the lateral acceleration direction frequently alternates between positive and negative and the amplitude exceeds a threshold, while the heading angle continuously and rapidly oscillates with small amplitude, it is determined to be snake-like driving.

[0083] Abnormal parking behavior: When a vehicle's digital mirror image slows to zero in a non-permitted parking area (such as a non-emergency parking lane) for more than a set threshold (such as 60 seconds), it is marked as abnormal parking. The system will classify the risk based on its location (whether it is blocking the lane);

[0084] All identified abnormal behaviors are encapsulated into "event objects" with tags for time, location, vehicle ID, and behavior type, and pushed to the event bus for subsequent modules to subscribe to;

[0085] S3.2: Macro-level traffic situation analysis and short-term forecasting. Based on aggregated information from the digital mirroring of vehicles throughout the tunnel, macro-level traffic flow parameters are calculated, including but not limited to cross-sectional flow rate, average speed, lane density, and time occupancy. These real-time parameters drive an integrated traffic flow model (such as a cellular automata model or a simplified fluid dynamics model) to provide rolling forecasts of traffic conditions for the next 5-15 minutes. This model is capable of:

[0086] Congestion identification and bottleneck location: Identify tunnel sections where the current speed is consistently below a critical value (e.g., 50% of the design speed) and define them as congestion bottlenecks. Combine vehicle trajectory data to trace the starting point of the congestion;

[0087] Congestion Evolution Prediction: Based on current inflow, bottleneck capacity, and predicted outflow, the system simulates the evolution trends of congestion queue length, duration, and dissipation time. The prediction results are visualized in a twin as a "future traffic state heatmap."

[0088] S3.3: Multimodal risk fusion prediction, specifically generating a dynamic risk map through the following steps:

[0089] Traffic conflict risk calculation: Based on the short-term predicted trajectories of vehicle digital mirrors (from step S2), a time-collision algorithm is used to detect conflicts between pairs of vehicles. The estimated minimum distance and occurrence time for each pair of vehicles are calculated. If the minimum distance is less than a safety threshold and the occurrence time is within the next few seconds, a potential conflict is determined. The severity of all conflicts within a geographical unit (e.g., each unit is 20 meters) is cumulatively calculated (weighted by subsequent time) to obtain the real-time traffic conflict risk value for that unit.

[0090] Environmental risk index calculation: This is the output of an integrated environmental field model. Visibility and CO concentration values ​​are compared to safety standard thresholds, and the environmental risk index is calculated through normalization and weighting. For example, the environmental risk index is higher in areas with visibility below 200 meters or CO concentration exceeding 100 ppm.

[0091] Facility Failure Impact Assessment: Receives infrastructure status information. If a large-scale lighting failure in a certain area results in severely insufficient illumination, or if a critical fan failure affects ventilation, the facility risk level of that area will be directly increased.

[0092] Risk fusion and classification: After standardizing the traffic conflict risk value, environmental risk index, and facility risk level, a weighted fusion algorithm (weights can be calibrated based on historical accident data) is used to calculate the comprehensive risk value of each geographical unit. Based on preset thresholds, the risks are classified into "low, medium, high, and extremely high" levels, and are overlaid with different colors on the 3D tunnel model in the digital twin to form a real-time risk heat map;

[0093] S3.4: Driver perception effectiveness assessment, which quantifies the impact of the current lighting environment on the driver's ability to acquire information at a specific location. Its core is to run a "visual effectiveness calculation model", as follows;

[0094] Input parameters: The inputs to this model include: (a) the background brightness distribution along the visual path in front of the driver's eye point (calculated based on the vehicle's digital mirror position and a standard human body model) from the ambient light field model; (b) the characteristics of the target object, such as the size, reflectance, and color of the guidance sign, or the size and reflectance of the assumed obstacle; and (c) optionally, real-time pupil size data from the driver's state perception (if available).

[0095] Visibility level calculation: Based on relevant CIE standards, a small target visibility model is used. The brightness contrast between the target and the background is calculated, and the just-perceptible contrast threshold is obtained by looking up a table or calculating based on assumed parameters such as the current background brightness and the driver's age. The visibility level is defined as the ratio of the actual contrast to the threshold contrast. VL > 1 indicates that the target is theoretically visible;

[0096] Reaction time estimation: An empirical model is established to model the relationship between reaction time and visibility level, scene complexity (e.g., traffic flow density), and driver state (e.g., estimated attention level). The basic principle is that the lower the visibility level (VL) or the more distracted the driver's attention, the longer the estimated reaction time. The system estimates a "theoretical average reaction time under the current environment" for each primary vehicle digital mirror.

[0097] Output Application: The output visibility level and estimated reaction time are input as key context parameters to the guidance strategy generation module in step S4. For example, in low visibility road sections, the system will tend to issue earlier and more conservative guidance instructions; or when the assessment finds that the VL value of a guidance sign is below the safety threshold, it will automatically trigger an instruction to increase the lighting in that area;

[0098] S4. Dynamic collaborative guidance strategy generation, with the optimization objective of minimizing the total virtual cost of the system, generates personalized vehicle guidance strategies for vehicles in the tunnel based on the output of S3, and generates tunnel environment control strategies in conjunction with these strategies, forming a "vehicle-facility" collaborative control instruction package; the total virtual cost of the system shall at least comprehensively consider safety cost, efficiency cost and energy consumption cost.

[0099] System total virtual cost modeling: Total virtual cost J is the security cost ( ), efficiency cost ( Energy consumption cost ) and driver psychological compliance costs ( The weighted sum of ), i.e. Among them, the weighting coefficient ( , , , This can be dynamically adjusted based on operational strategies (such as safety-first, efficiency-first, or energy-saving-first modes). The specific definitions of each cost are as follows:

[0100] Security costs ( ): Calculate the total risk that all vehicles may face in the next decision cycle (e.g., the next 30 seconds). For vehicle i, its safety cost is calculated by integrating the comprehensive risk values ​​of each point on its planned path provided in step S3, with special weighting of the risk values ​​at points where its trajectory conflicts with the predicted trajectories of other vehicles. The probability of entering a "high-risk" or "extremely high-risk" area is directly converted into a higher cost term;

[0101] Efficiency Cost ( ): Calculates the sum of the deviations between the estimated travel time of all vehicles and the travel time under ideal smooth conditions. It also includes a penalty term for the overall traffic flow stability, such as penalizing excessive future speed variance in the traffic flow prediction model to suppress traffic oscillations.

[0102] Energy consumption cost ( ): Calculates the incremental energy consumption of the tunnel environmental control system in implementing the guidance strategy. This mainly includes: a) the additional power consumption of the lighting system due to dynamic dimming or enhanced lighting in specific areas relative to the baseline operating conditions; b) the additional power consumption of the ventilation system due to increased fan speed or activation of additional fans to cope with specific traffic events or pollutant accumulation. Energy consumption is estimated based on equipment power models and expected operating times.

[0103] Driver psychological compliance cost ( This cost is designed to improve the acceptability and actual compliance rate of guidance instructions. The system assigns a base "psychological resistance" value to different types of guidance instructions (such as "suggest adjusting speed to X km / h", "suggest changing lanes", "mandatory lane control"). This value is calibrated using historical instruction compliance data and represents the average compliance willingness of the driver group. For an individual vehicle, if its historical behavior shows low compliance with a certain type of instruction (such as changing lanes), or if its current driver state perception module assesses that its attention is distracted, the psychological cost of issuing that type of instruction will increase accordingly. Issuing instructions deemed "unacceptable" will cause this cost to increase sharply, thus prompting the optimization algorithm to find alternatives;

[0104] Optimize decision variables and constraints. Decision variables are divided into two categories: vehicle-oriented control variables and facility-oriented control variables.

[0105] Vehicle control variables ( For each induced vehicle (connected vehicle or vehicle that can be controlled by lane signals), a series of discrete speed setpoints and lane selection suggestions are recommended over several decision cycles in the future.

[0106] Facility control variables ( Set the dimming level for each lighting zone in the tunnel, set the speed level or start / stop status for each ventilation fan, and control the display content of variable message signs and lane indicators;

[0107] The optimization process is subject to strict physical and safety constraints, including: vehicle dynamics constraints (speed and acceleration limits), lane capacity constraints, inter-vehicle safety distance constraints, minimum illuminance / brightness constraints for lighting specifications, and upper limits for ventilation pollutant concentrations.

[0108] The collaborative strategy is generated and output on a rolling basis, with the decision engine running optimization calculations at fixed intervals (e.g., every 5 seconds). Each calculation uses the optimal solution from the previous cycle as the initial point, combining the latest twin state and prediction information to solve a finite-time optimal control problem. The solution algorithm can employ gradient-based methods, heuristic algorithms (such as genetic algorithms), or deep reinforcement learning agents trained offline.

[0109] The optimization result is output as a structured cooperative control instruction package, which specifies the execution time window, the target object, and the specific instructions. For example, a typical instruction package might include:

[0110] Instruction to vehicle A: "Time T to..." At any given time, between chainage K1+100 and K1+600, it is recommended to maintain a speed of 65-70 km / h, and to complete the lane change to the left lane before chainage K1+150. This instruction is issued to the onboard unit of vehicle A via V2X communication or from a specific RSU location.

[0111] The coordinated instruction for environmental facilities is as follows: "Starting from time T, increase the dimming level of the lighting above the left lane in the section from chainage K1+000 to K1+800 from 70% to 90% for 300 seconds; at the same time, start the fan numbered FAN-12 and run it at 80% of its rated speed for 600 seconds."

[0112] All instruction packages are virtually rehearsed and validated in the digital twin before being sent to the actual physical execution units. The system continuously monitors traffic flow and environmental feedback after instruction execution for iterative improvement of cost models and optimization algorithms.

[0113] Personalized dynamic modeling of driver psychological compliance costs is a key module for achieving accurate delivery of guidance instructions and improving the overall efficiency of the system. This module achieves fine-grained and dynamic quantification of psychological compliance costs by constructing and continuously updating the "driving behavior fingerprint" and the corresponding "instruction acceptance tendency model" for each identifiable vehicle.

[0114] Specifically, the system maintains a dynamic profile for each vehicle with a stable identity (such as an anonymized but re-identifiable digital token). The core of this profile is a driving behavior feature vector extracted from its long-term historical trajectory data. Feature extraction is performed offline in the cloud or edge computing units, using historical travel data of vehicles in tunnels and connected open road sections to calculate multiple statistical indicators, such as: the distribution of the ratio of average speed to the road segment speed limit, the frequency and intensity thresholds of acceleration and deceleration events, the aggressiveness of lane-changing behavior (such as the minimum time difference of lane-changing gaps), and the average response compliance rate to historical guidance instructions (such as speed limit reminders). After standardization, these indicators are used to classify vehicles into preset driving style prototypes (such as "conservative," "normal," and "aggressive") using unsupervised clustering algorithms (such as K-means).

[0115] Based on this, the system establishes a prior probability of "baseline acceptance rate" for different types of guidance commands for each type of driving style prototype. This prior probability is derived from statistical analysis of historical command execution feedback data from a large number of vehicles with similar driving styles. When the system needs to issue a new command to a specific vehicle, the personalized modeling process is initiated, as follows:

[0116] Real-time contextual feature fusion: The system captures the current real-time context, including the traffic density level of the area where the vehicle is located in the tunnel, weather conditions (such as whether it is raining, obtained from the weather station at the tunnel entrance), lighting conditions (day / night / dusk), and the real-time state estimate of the vehicle driver (such as attention level, if available). These contextual parameters are encoded into feature vectors.

[0117] The Bayesian posterior probability update uses the baseline acceptance rate of the vehicle's driving style as the prior probability and the current real-time context features as observational evidence. The system employs a pre-trained Naive Bayes classifier or a more complex probabilistic graphical model to calculate the posterior probability P(accept | driving style, context) of the driver accepting the upcoming specific instruction (e.g., "change lanes after X meters") in the current specific context. The model pre-training phase utilizes historical acceptance / rejection records covering different styles, contexts, and instruction types.

[0118] Psychological cost conversion, ultimately leading to psychological compliance cost. For vehicle i and instruction a, it is defined as a monotonically increasing function of the expected non-compliance probability. One direct mapping method is: ,in It is a cost scaling factor used to place this psychological cost on a comparable scale to physical costs such as safety and efficiency. Therefore, issuing a "forced lane change" instruction to a vehicle identified as "aggressive" in a congested rainy night will result in a very high calculated psychological cost. This may lead the optimization algorithm to choose to generate a milder "suggested speed adjustment" instruction for it, or to assist its decision-making by enhancing environmental guidance (such as brightening the target lane lighting), rather than directly issuing an instruction with high psychological resistance.

[0119] The model has online learning capabilities. Whenever an instruction is issued, the system automatically generates an acceptance / rejection label sample by comparing the actual behavior of the vehicle's digital mirror with the instruction content. These new samples are safely used to periodically and incrementally update the driving style classification model and the acceptance rate prediction model, thereby making the psychological cost prediction increasingly closer to the group and individual behavioral characteristics of real traffic participants over time.

[0120] S5. Precise guidance command issuance and execution feedback: The collaborative control command package is issued to the target vehicle and tunnel environmental control facilities through multimodal channels for execution, and the actual effect of vehicle response and environmental adjustment is continuously tracked. The execution feedback data is sent back to the digital twin to form a closed-loop control.

[0121] Specifically, S5 includes:

[0122] S5.1: Multimodal hierarchical instruction issuance mechanism. The system selects the optimal issuance channel based on the instruction's target audience, urgency, and execution device. Specifically, this includes:

[0123] Personalized guidance instructions at the vehicle level are issued for connected vehicles or vehicles equipped with in-vehicle intelligent terminals. These instructions (such as precise speed guidance and lane change suggestions) are sent directly to the target vehicle via the C-V2XPC5 interface or RSU forwarding in a low-latency, high-reliability message format (such as SPAT / MAP / RSM messages conforming to SAE J2735 or custom application layer messages). The message clearly specifies the effective start and end times of the instruction, the affected road segment, the suggested action, and the reason code (such as "accident ahead").

[0124] Group-level guidance and mandatory control commands are issued, targeting all vehicles or non-connected vehicles for group guidance (such as lane closures and speed limits), through variable message signs, lane indicators, and traffic lights deployed within the tunnel. For critical mandatory lane control (such as opening / closing lanes), the red / green arrow status of lane indicators is directly controlled via hardwired or secure wireless network.

[0125] Enhanced environment guidance release: To assist driving decisions or enhance guidance effects, the system can control intelligent road luminous markings (such as LED embedded road studs) or side wall projection devices to dynamically generate visual guidance patterns (such as lane change arrows, speed prompt numbers, and warning area highlight boxes) on specific road sections, forming an immersive guidance environment.

[0126] The facility linkage control command is issued, which adjusts environmental control systems such as lighting and ventilation. This command is sent via industrial control networks (e.g., based on the OPCUA protocol) to each area controller or individual smart lighting fixtures / fans. The command includes the target device ID, setpoints (such as dimming level and fan speed), and execution schedule.

[0127] S5.2: Command execution and status synchronization. After executing the command, all physical devices receiving the command (vehicles, signs, lights, etc.) report the actual execution status (such as actual vehicle speed, actual lane indicator display status, and actual light brightness) as "execution feedback" to the central platform through their own sensors or controllers. The platform compares the reported "actual execution status" with the "expected command" issued in step S4 in real time. If a significant deviation is found (such as the vehicle not changing lanes as suggested, or the light malfunction not reaching the dimming brightness), the system will trigger an alarm and may activate backup guidance strategies or equipment redundancy schemes.

[0128] S5.3: Multi-dimensional closed-loop effect evaluation. The system uses a digital twin as a benchmark to quantitatively evaluate the actual effect of the induction control, forming a closed loop.

[0129] Traffic behavior response assessment evaluates the effectiveness of traffic guidance by comparing the micro-behavioral changes in the target vehicles and surrounding traffic flow before and after the instruction is issued (such as improved speed consistency, reduced lane-changing conflicts, and accelerated queue dissipation). For example, it calculates the compliance rate of vehicles with speed recommendations, the actual execution rate of recommended lane-changing actions, and the smoothness of execution.

[0130] The environmental regulation effect verification involves comparing the measured data of environmental sensors (illuminance meter, visibility meter, CO sensor) after the facility linkage command is executed with the predicted values ​​of the twin environmental field model under linkage conditions to verify whether the environmental regulation has achieved the expected goals (such as whether the illuminance of the target area has increased to the safe threshold and whether the pollutant concentration has effectively decreased).

[0131] Post-evaluation of safety and efficiency indicators: After a induced event handling cycle ends, the system comprehensively evaluates key performance indicators such as the number of actual risk events, changes in traffic throughput, and average travel time during the cycle, and compares them with the results of twin simulations assuming a "no-induction" scenario to quantify the safety gains and efficiency improvements brought about by this collaborative induction.

[0132] Cost model calibration involves using actual energy consumption data and observed average driver compliance as feedback signals to input into the virtual cost model in step S4, which is used to dynamically calibrate cost weights or model parameters, so that future optimization decisions are more in line with the actual system response.

[0133] S5.4: Feedback-driven twin and model iteration. All execution feedback and effect evaluation data are synchronously fed back into the digital twin. On the one hand, this is used to correct the state of the twin in real time, making it more consistent with the physical world; on the other hand, this labeled real data (instruction-response-result) is stored in a historical case library for offline training and optimization of the risk prediction model in S3, the traffic flow prediction model, and the optimization decision model in S4, thereby achieving a continuous and gradual improvement in the overall intelligence level of the system.

[0134] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for dynamic vehicle tracking and linkage guidance control in a tunnel, characterized in that, The method includes: S1: Multi-source information holographic perception and fusion, utilizing the Internet of Things sensing network deployed in the tunnel to collect real-time vehicle individual information, traffic flow status information, tunnel infrastructure status information and tunnel environment information; S2: Construction and real-time synchronization of tunnel digital twin. Based on the tunnel geometry and facility static model and the real-time sensing data of S1, a high-precision tunnel digital twin model including a dynamic traffic body model and an environmental field model is constructed and dynamically updated. S3: Situational analysis and risk prediction based on digital twins, including micro-driving behavior analysis, macro-traffic situation assessment, multi-modal risk fusion prediction, and driver perception effectiveness evaluation based on the digital twin model; S4: Dynamic collaborative guidance strategy generation, with the optimization objective of minimizing the total virtual cost of the system, generates personalized vehicle guidance strategies for vehicles in the tunnel based on the output of S3, and generates tunnel environment control strategies in conjunction with it. S5: Precise guidance command issuance and execution feedback, which sends the collaborative control command package to the target vehicle and tunnel environmental control facilities through multimodal channels, and tracks the actual effect of vehicle response and environmental adjustment to form a closed-loop control.

2. The method for dynamic vehicle tracking and linkage guidance control in tunnels according to claim 1, characterized in that, In S1, the Internet of Things sensing network includes a fixed roadside sensing unit, a distributed environmental sensing unit, and a mobile vehicle terminal. The fixed roadside sensing unit includes a high-precision radar-visual integrated machine and a roadside unit. The high-precision radar-visual integrated machine adopts hardware-level fusion of millimeter-wave radar and high-definition video camera to output lane-level vehicle trajectory data.

3. The method for dynamic vehicle tracking and linkage guidance control in tunnels according to claim 1, characterized in that, In step S2, the generation and real-time synchronization of the dynamic traffic body model includes: When the perception system first stably tracks the physical vehicle, a digital image of the vehicle is instantiated in the twin and assigned a globally unique identifier; The vehicle digital image receives state updates from the perception fusion layer at a frequency of no less than 10Hz, and maintains a time difference of less than 100 milliseconds with the physical entity through timestamp alignment and interpolation algorithms.

4. The method for dynamic vehicle tracking and linkage guidance control in tunnels according to claim 1, characterized in that, In S2, the environmental field model includes a light environment field model and an airflow field and pollutant diffusion model; The optical environment field model is constructed by combining real-time monitoring data-driven and physical mechanism model simulation, and uses sensor networks for data assimilation and dynamic correction.

5. The method for dynamic vehicle tracking and linkage guidance control in tunnels according to claim 1, characterized in that, In S3, the multimodal risk fusion prediction includes: Traffic conflict risk calculation based on short-term predicted trajectories of vehicle digital mirrors; The environmental risk index is calculated from the output of the integrated environmental field model. Receive infrastructure status information to assess the impact of facility failures; The above risk values ​​are standardized and weighted to generate a real-time risk heatmap.

6. The method for dynamic vehicle tracking and linkage guidance control in tunnels according to claim 1, characterized in that, In S3, the driver perception performance assessment includes: Based on the background brightness distribution provided by the light environment field model, and combined with the characteristics of the target object and the driver's state data, the visibility level is calculated. The driver's average reaction time is estimated based on visibility level, scene complexity, and driver status.

7. The method for dynamic vehicle tracking and linkage guidance control in tunnels according to claim 1, characterized in that, In S4, the total virtual cost of the system is modeled as a weighted sum of safety cost, efficiency cost, energy consumption cost, and driver psychological compliance cost; The driver's psychological compliance cost is dynamically calculated using a probability model based on the vehicle's historical driving behavior characteristics and real-time contextual characteristics.

8. The method for dynamic vehicle tracking and linkage guidance control in tunnels according to claim 1, characterized in that, In S4, the optimization decision variables include vehicle-oriented control variables and facility-oriented control variables; The optimization process is executed on a rolling basis with a fixed period, combining the latest twin state and prediction information to solve the optimal control problem in the finite time domain.

9. The method for dynamic vehicle tracking and linkage guidance control in tunnels according to claim 1, characterized in that, In S5, the multimodal hierarchical instruction issuance mechanism includes: Personalized guidance instructions at the vehicle level can be issued to connected vehicles via C-V2X or RSU; Variable information signs and lane indicators are used to issue group-level guidance and mandatory control instructions to all vehicles; Environmentally enhanced guidance instructions are issued through intelligent road luminous markings or side wall projection devices; The industrial control network sends coordinated control commands to lighting and ventilation facilities.

10. The method for dynamic vehicle tracking and linkage guidance control in tunnels according to claim 1, characterized in that, In S5, the closed-loop control includes: after the instruction is executed, receiving feedback on the actual execution status and comparing it with the expected instruction; A multi-dimensional quantitative evaluation of the induction control effect based on digital twins; The execution feedback data will be used for twin state correction and model iterative optimization.

Images