Tunnel environment-oriented mixed traffic flow vehicle following control method

Abstract

The invention provides a mixed traffic flow vehicle following control method for a tunnel environment. The method comprises the following steps: acquiring environmental parameters in a tunnel, traffic flow state data and vehicle type information; generating environmental factors based on the environmental parameters; a three-layer coupling car-following model is constructed and comprises a basic physical layer, an intelligent driver model is adopted as a dynamic core, and car-following key parameters of the basic physical layer are dynamically adjusted through an environment adaptation layer; the environment adaptation layer is used for dynamically adjusting car-following key parameters according to the environment factors, and the car-following key parameters comprise the response time and comfortable deceleration of a human driving vehicle and the minimum safety distance and maximum acceleration of an automatic driving vehicle; the behavior decision-making layer is used for executing a differential interaction strategy according to the vehicle type of the preceding vehicle; based on the three-layer coupling car-following model, a reinforcement learning algorithm is adopted to generate a longitudinal acceleration control instruction of the automatic driving vehicle, and mixed traffic flow collaborative car-following is achieved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent transportation systems and intelligent networked vehicle control, and particularly relates to a mixed traffic flow vehicle car-following control method for a tunnel environment. BACKGROUND

[0002] Among numerous road scenes, highway tunnels become the concentrated outbreak point of mixed traffic flow safety and efficiency problems due to their special physical structure. Tunnel environments have typical characteristics of being closed, dynamic, and complex: the light intensity inside the tunnel experiences a sharp decay within a few seconds at the entrance section, and the uniformity is poor inside the tunnel which relies on artificial lighting; limited air circulation can lead to the accumulation of exhaust smoke, causing random fluctuations in visibility; to adapt to the terrain, the tunnel often contains a combination of slope and curvature changes, forming a complex spatial alignment. These environmental parameters are not static backgrounds, but dynamic disturbance sources that change over time, traffic volume, and external weather. Research shows that the above factors can significantly affect the physiological and psychological state of drivers, leading to problems such as visual adaptation delay, speed perception distortion, and following distance misjudgment, which are important reasons for the higher accident rate in tunnels than in open road sections.

[0003] However, the current mainstream vehicle car-following models, such as the intelligent driver model and its many improved variants, are designed primarily for homogeneous and ideal traffic flow. When applied to mixed traffic flow, their limitations become apparent: First, the models usually assume that all vehicles have uniform dynamic characteristics and decision-making logic, which cannot capture the randomness and hysteresis of human-driven vehicles (HDVs) in response to environmental and psychological influences, as well as the fundamental differences between autonomous vehicles (AVs) based on sensors and algorithms and the deterministic and forward-looking control of AVs. This difference makes it difficult for the model to accurately reproduce key interaction phenomena such as the amplification of AV disturbances on HDVs or the inadaptation of HDV conservative strategies to AVs when simulating or controlling mixed vehicle platoons.

[0004] Second, existing models generally lack the ability to structure and quantify dynamic and adverse environmental factors such as tunnels. Most studies treat the environment as a fixed or ideal condition, or introduce individual factors (such as slope) in the form of simple parameter modifications. However, tunnel safety risks are the result of the nonlinear superposition and joint action of multiple factors such as light, visibility, and alignment. Existing technologies lack a mechanism that can perceive and fuse multi-source heterogeneous environmental data in real time, and convert it into a quantitative index that can uniformly represent the overall environmental risk level and be directly understood and responded to by the car-following model. This makes vehicle control strategies either insensitive to environmental changes, using fixed parameters that increase risk rapidly in adverse conditions, or relying on coarse-grained and segmented rule switching, with poor response and poor smoothness.

[0005] Furthermore, at the control optimization level, research on car-following strategies for AV (Action Vehicle) often focuses on improving a single objective or using multi-objective optimization with fixed weights. In the rapidly changing environment of tunnels, the priority between safety and efficiency is not static. For example, in the clear, well-lit sections of the tunnel, the focus should be on improving traffic efficiency; while in the "black hole effect" area at the entrance or in sections with curves and slopes, safety must be given absolute priority. Existing methods struggle to achieve this smooth, adaptive optimization objective trade-off based on real-time environmental risk perception, resulting in control strategies that are either generally conservative with low throughput or unable to promptly strengthen safety guarantees when risks escalate.

[0006] In summary, existing technologies have shortcomings when facing complex scenarios of mixed traffic flow in tunnel environments, such as accurate modeling of heterogeneous vehicle behavior, quantitative perception and fusion of dynamic environmental risks, and adaptive adjustment of control parameters and optimization objectives based on real-time risks. Summary of the Invention

[0007] To address the shortcomings and deficiencies of existing technologies, such as poor environmental adaptability, insufficient modeling of heterogeneous vehicle interactions, and rigid control strategies, this invention provides a vehicle car-following control method and system for mixed traffic flow in tunnel environments. First, an integrated vehicle-road-cloud perception system is used to acquire environmental parameters such as light intensity, visibility, road slope, and curvature within the tunnel, as well as traffic flow status data and information distinguishing between human-driven and autonomous vehicles. Preprocessing ensures data usability. Then, the aforementioned environmental parameters are transformed into normalized influence indices, which are weighted and fused using nonlinear mapping to generate environmental factors that quantify the safety level of the tunnel environment. These factors range from (0,1) and are positively correlated with the environmental safety level. A dynamic update mechanism avoids instantaneous noise interference. Next, a three-layer coupled car-following model is constructed. The basic physical layer uses an intelligent driver model as its dynamic core, and its key parameters are dynamically adjusted by the environmental adaptation layer. The environmental adaptation layer adapts the reaction time of human-driven vehicles based on environmental factors. The system determines the minimum safe distance and maximum acceleration for autonomous vehicles, and the behavioral decision layer executes differentiated interaction strategies based on the type of vehicle ahead. A deep deterministic policy gradient algorithm is then used to train the autonomous vehicle controller, integrating environmental factors and traffic flow characteristics into the state space. Through a multi-objective reward function and dynamic weighting mechanism, it achieves intelligent switching between safety priority and efficient operation. Finally, the model and control strategy are deployed on an onboard or cloud-based management platform to complete longitudinal vehicle control, supporting multi-vehicle collaborative scheduling and model iterative optimization. This invention, through environmental adaptive parameter adjustment, differentiated interaction among heterogeneous vehicles, and intelligent algorithm optimization, synergistically improves the safety, stability, traffic efficiency, and comfort of mixed traffic flow in tunnels. Furthermore, its modular design facilitates engineering implementation and widespread adoption.

[0008] The specific technical solution adopted by this invention to solve its technical problem is as follows:

[0009] A vehicle following control method for mixed traffic flow in tunnel environments includes:

[0010] The system acquires environmental parameters, traffic flow data, and vehicle type information within the tunnel. The environmental parameters include at least light intensity, visibility, road slope, and curvature. The vehicle type information is used to distinguish between human-driven vehicles and autonomous vehicles.

[0011] Environmental factors are generated based on the environmental parameters. The environmental factors are obtained by normalizing and weighting each environmental parameter to obtain a comprehensive influence degree, and then generated by nonlinear mapping. The values ​​are in (0,1] and are positively correlated with the safety level of the tunnel environment.

[0012] Construct a three-layer coupled car-following model, including:

[0013] The basic physical layer uses an intelligent driver model as the core of dynamics, and its key car-following parameters are dynamically adjusted through the environmental adaptation layer.

[0014] The environmental adaptation layer dynamically adjusts key parameters of the following vehicle based on the environmental factors, including the reaction time and comfort deceleration of human-driven vehicles and the minimum safe distance and maximum acceleration of autonomous vehicles.

[0015] The behavioral decision-making layer executes differentiated interaction strategies based on the type of vehicle preceding it;

[0016] Based on the aforementioned three-layer coupled car-following model, a reinforcement learning algorithm is used to generate longitudinal acceleration control commands for autonomous vehicles, thereby achieving coordinated car-following of mixed traffic flows.

[0017] Furthermore, the methods for acquiring data include:

[0018] By using sensing devices, vehicle-to-everything (V2X) communication, and geographic information data deployed along the tunnel route, environmental parameters, vehicle type information, traffic flow status data, and infrastructure parameters can be obtained.

[0019] The raw data is preprocessed by filtering and denoising, coordinate unification, and time synchronization to ensure that the data is usable under the same spatiotemporal reference.

[0020] Furthermore, the generation process of the environmental factors includes:

[0021] Each environmental parameter is converted into a normalized index characterizing its negative impact on driving safety, with a value range of [0,1].

[0022] The comprehensive influence degree is obtained by weighting and integrating each normalized indicator according to a preset weight.

[0023] The comprehensive impact is transformed into an environmental factor with a value in (0,1) by a nonlinear mapping function;

[0024] A sliding update mechanism with a time window is set up to avoid large fluctuations in the environmental factors caused by instantaneous noise.

[0025] Furthermore, the parameter adjustment method of the environmental adaptation layer is as follows: when the environmental risk increases, the reaction time of human-driven vehicles is increased and their comfort deceleration is reduced, while the minimum safe distance of autonomous vehicles is increased and their maximum acceleration is reduced; when the environmental risk decreases, the parameters are adjusted in the opposite way to improve traffic efficiency.

[0026] The differentiated interaction strategies of the behavioral decision layer include: when an autonomous vehicle follows a human-driven vehicle, a predictive buffer strategy is activated to actively and smoothly decelerate; a shorter headway is used for the following time between autonomous vehicles; and the effective reaction time is shortened when a human-driven vehicle follows an autonomous vehicle.

[0027] Furthermore, the key parameters for following the car are dynamically determined by the environmental adaptation layer based on environmental factors, including at least one of reaction time, comfort deceleration, minimum safe distance, and maximum acceleration.

[0028] Furthermore, the reinforcement learning algorithm is a deep deterministic policy gradient algorithm, specifically including:

[0029] The state space includes the vehicle's own motion state, environmental factors, the type of the vehicle in front, and local traffic flow characteristics. All state variables are normalized and then input into the neural network.

[0030] The reward function includes objectives related to safety, stability, efficiency, and comfort. The weight of each objective is adaptively adjusted according to environmental factors to achieve intelligent switching between safety priority and efficient operation.

[0031] The motion space consists of continuous acceleration commands used to control the longitudinal movement of autonomous vehicles.

[0032] Furthermore, the reward function is a multi-objective reward function, and its sub-items include:

[0033] Safety bonuses are calculated based on the difference between the actual net distance between the vehicle and the vehicle in front and the dynamic safe distance threshold.

[0034] Stability bonus is negatively correlated with the absolute value of vehicle acceleration;

[0035] The efficiency reward is negatively correlated with the square of the difference between the vehicle's current speed and the expected speed.

[0036] Comfort reward is negatively correlated with the absolute value of vehicle acceleration.

[0037] Furthermore, the implementation methods for executing the control instructions include:

[0038] The trained algorithm model parameters are solidified into an executable module and embedded into the onboard decision-making system or edge computing node of an autonomous vehicle.

[0039] When the vehicle is running, sensor data is collected in real time to construct the current state vector. After being input into the model, the optimal acceleration control command is generated and sent to the vehicle's power or braking system via the CAN bus to complete longitudinal control.

[0040] And, a hybrid traffic flow vehicle following control system for tunnel environments, comprising:

[0041] The sensing module is used to acquire environmental parameters, traffic flow status data and vehicle type information inside the tunnel. The environmental parameters include at least light intensity, visibility, road slope and curvature.

[0042] The environmental factor calculation module is used to generate environmental factors based on the environmental parameters through normalization, weighted fusion and nonlinear mapping. The environmental factors take values ​​in (0,1) and are positively correlated with the tunnel environmental safety level.

[0043] The model building module is used to build a three-layer coupled car-following model, which includes a basic physical layer, an environmental adaptation layer, and a behavioral decision layer, respectively realizing the description of the basic laws of car-following, dynamic adjustment of parameters, and execution of differentiated interaction strategies.

[0044] The control module is used to generate longitudinal acceleration control commands for autonomous vehicles based on the three-layer coupled car-following model using reinforcement learning algorithms.

[0045] The execution module is used to execute the acceleration control command to achieve coordinated following of mixed traffic flow.

[0046] And a computer device including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described above.

[0047] A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0048] Compared with the prior art, the present invention and its preferred embodiments have at least the following beneficial effects:

[0049] 1. Significantly improves tunnel driving safety

[0050] By dynamically quantifying the comprehensive risks of tunnels through real-time generated environmental factors, and adaptively adjusting the key parameters of the car-following model accordingly, vehicles can automatically adopt a more conservative car-following strategy in high-risk scenarios such as sudden drop in light, deterioration of visibility, or complex road conditions, effectively reducing the risk of rear-end collisions.

[0051] 2. Enhance the stability of mixed traffic flow

[0052] By designing differentiated strategies to address the interaction characteristics of heterogeneous vehicles, autonomous vehicles can predict and buffer the unstable behavior of human-driven vehicles. At the same time, their own smooth driving provides positive guidance to vehicles behind them, thereby suppressing the propagation of disturbances and reducing traffic flow oscillations and "phantom traffic jams".

[0053] 3. Optimize traffic efficiency

[0054] Under the premise of ensuring safety, the model can automatically optimize the following parameters and target weights when environmental conditions are good, allowing for shorter safe distances and higher operating speeds, thereby improving the overall traffic capacity of the tunnel and alleviating congestion bottlenecks.

[0055] 4. Improve driving and riding comfort

[0056] The reinforcement learning-based controller can smoothly generate acceleration commands, avoiding sudden acceleration and deceleration, providing a more comfortable driving experience for passengers. At the same time, its parameter adaptation mechanism also conforms to the psychological expectations of human drivers, reducing the tension of drivers behind.

[0057] 5. Enhance environmental adaptability

[0058] The innovative approach integrates multi-dimensional environmental parameters into a single environmental factor, enabling the control system to have a holistic perception and response capability to the dynamically changing environment of the tunnel, overcoming the shortcomings of existing models that are insensitive to environmental disturbances or have a delayed response.

[0059] 6. Improve system usability and deployability

[0060] The solution adopts a modular design, and the model can be deployed on the vehicle for real-time decision-making or integrated into the cloud for collaborative management and control. It supports online monitoring and feedback optimization, and has good prospects for engineering implementation and promotion value. Attached Figure Description

[0061] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0062] Figure 1 This is a schematic diagram illustrating the implementation process of the vehicle following model and optimization strategy for mixed traffic flow in tunnel environments according to an embodiment of the present invention.

[0063] Figure 2This is a schematic diagram illustrating the process of acquiring multi-dimensional information within a tunnel using multi-source sensors, as described in an embodiment of the present invention.

[0064] Figure 3 A flowchart illustrating the establishment of a three-layer coupling model for an embodiment of the present invention;

[0065] Figure 4 This is a schematic diagram of the process for calculating tunnel environmental factors according to an embodiment of the present invention;

[0066] Figure 5 A flowchart illustrating the design of an AV optimization strategy based on DDPG for an embodiment of the present invention;

[0067] Figure 6 This is a schematic diagram illustrating the process of model deployment and practical application in an embodiment of the present invention. Detailed Implementation

[0068] To make the features and advantages of the present invention more apparent and understandable, specific embodiments are described below in detail:

[0069] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0070] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0071] To overcome the problems of poor environmental adaptability, insufficient modeling of heterogeneous vehicle interactions, and rigid control strategies in existing technologies, and to achieve a synergistic improvement in safety, stability, and efficiency, this invention provides a well-structured, physically meaningful, and engineering-applicable vehicle car-following model and optimization strategy for hybrid traffic flow in tunnel environments. This includes five main stages: multi-source perception data acquisition, hierarchical car-following model construction, tunnel environment factor calculation, autonomous vehicle (AV) optimization strategy design based on Deep Deterministic Policy Gradient (DDPG), and system deployment and application. Among them, such as... Figure 1As shown, the stages are divided as follows: Stage S1 acquires vehicle status, tunnel characteristics, and traffic flow information through roadside and vehicle-mounted sensor networks; Stage S2 establishes a three-layer coupled model including the improved Intelligent Driver Model (IDM) base layer, environmental adaptation layer, and behavioral decision layer; Stage S3 generates nonlinear environmental factors based on the normalized fusion of parameters such as illumination, visibility, slope, and curvature. In stage S4, the DDPG algorithm is used to train an AV controller with environmental perception capabilities; in stage S5, the model is integrated and deployed on the vehicle-mounted or cloud-based management platform. This method significantly improves the safety, stability, and traffic efficiency of mixed traffic flow in tunnel environments, and has good practicality and promotional value.

[0072] like Figure 2 As shown, in a preferred embodiment, step S1 specifically includes the following steps:

[0073] Step S11: Deploy a multi-source sensor network in the tunnel. Install light sensors, visibility meters, high-definition cameras, millimeter-wave radar, lidar and inertial navigation aids along the tunnel and connect them to a high-precision map database to form a "vehicle-road-cloud integrated" perception system.

[0074] Step S12: Through communication and video tracking technology between the Onboard Unit (OBU) and Road Side Unit (RSU), the basic dynamic and physical attributes of each vehicle are acquired in real time, including driving speed, acceleration, driving direction angle, vehicle mass, center of gravity position, and vehicle type identification. The vehicle type identification distinguishes between Human-Driven Vehicles (HDVs) and AVs.

[0075] Step S13: Collect static and semi-static infrastructure parameters through regular inspection reports and road surface monitoring devices, including: tunnel alignment (slope G, curvature C), lane width, lateral clearance, marking wear, reflectivity, contrast, actual width, road surface water film thickness, and friction coefficient.

[0076] Step S14: Extract group behavior features within a certain range centered on the target vehicle through vehicle-to-everything (V2X) communication and roadside perception fusion systems, including: local traffic density. Average speed and the estimated acceleration of the vehicle in front Among them, the estimated acceleration of the preceding vehicle is used to predict the behavior of the following vehicle and judge the trend of relative speed change, providing support for the acceleration calculation of the car-following model and the state perception of reinforcement learning.

[0077] Step S15: Data preprocessing and spatiotemporal alignment. The raw data is filtered and denoised, coordinates are unified, and time is synchronized to ensure that data from different sources are available under the same spatiotemporal reference.

[0078] like Figure 3 As shown, in this preferred embodiment, step S2 is implemented as follows:

[0079] Step S21: Construct the basic physical layer—improved IDM model;

[0080] An improved intelligent driver model is adopted as the core framework of the dynamics, and its acceleration update formula is:

[0081]

[0082] In the formula, The acceleration of the vehicle at the current moment (unit: m / s²) 2 Positive values ​​indicate acceleration, and negative values ​​indicate deceleration. Maximum acceleration of the vehicle (unit: m / s²) 2 HDV values ​​are typically 1.0 to 1.5, and AV values ​​are typically 1.5 to 3.0, representing the upper limit of power performance. The current vehicle speed (unit: m / s) is measured. HDV values ​​are typically 0-30 (approximately 108 km / h), and AV values ​​are also typically 0-30 (approximately 108 km / h), obtained through real-time sensing. The HDV value is typically 25-30 (90-108 km / h), representing the driver's desired free-flow speed or speed limit (unit: m / s). The AV value can be dynamically adjusted to approximate the road design speed. This is the speed sensitivity index, typically set to 4. It controls the degree of nonlinearity in the speed term and affects low-speed response. The expected minimum safety distance (in meters) calculated for the model varies dynamically. This represents the current net distance to the vehicle in front (in meters), and its value is usually greater than 0. .

[0083] The desired spacing function is:

[0084]

[0085] In the formula, The desired safe distance (unit: m) changes dynamically and is a core parameter of IDM, determining the following distance. The minimum safe distance when stationary (unit: m), HDV values ​​are typically 2~4, and AV values ​​are typically 1.5~3.0. This represents the vehicle's current speed (in meters). HDV values ​​are typically 2-4, and AV values ​​are typically 1.5-3.0, determining the reaction distance. Driver reaction time (unit: seconds) is typically measured in HDV values, which are usually between 1.0 and 2.0, while AV values ​​are usually between 0.1 and 0.5. HDV is limited by human perception, while AV is extremely short. for (Unit: m / s), values ​​are typically positive for approaching and negative for moving away, representing the speed difference between the vehicles in front and behind. For comfortable deceleration (unit: m / s) 2 HDV values ​​are typically 1.5 to 2.0, and AV values ​​are typically 2.0 to 3.0. Higher values ​​indicate more "aggressive" braking while still providing comfort.

[0086] Step S22: Introduce the environmental perception and adaptation layer, and define comprehensive environmental factors. This value is used to quantify the impact of the current environment on driving behavior. A value closer to 1 indicates a favorable environment, while a value closer to 0 indicates a higher risk.

[0087] Step S23: Normalize the single environmental factor and convert each original environmental variable into an influence index in the [0,1] interval;

[0088] Light Influence :

[0089]

[0090] In the formula, This represents the degree of negative impact of lighting on driving behavior, typically ranging from 0 to 1, with a higher value indicating poorer lighting. The measured light intensity (unit: lux) is obtained in real time from sensors inside the tunnel. The ideal light level for normal driving (unit: lux) typically ranges from 300 to 500, and is the standard for indoor lighting during the day. The minimum safe light intensity (unit: lux) is typically between 50 and 100; anything below this value is considered dangerous.

[0091] visibility Similar definitions:

[0092]

[0093] In the formula, The value represents the negative impact of visibility on driving; it is typically between 0 and 1, with a higher value indicating poorer visibility. The current visibility distance (unit: m) is measured in real time from a visibility meter or a visual algorithm. Ideal visibility (unit: meters), typically ranging from 100 to 200, represents the visible distance under normal weather conditions. The minimum safe visibility (unit: m) is typically 30 to 50. Visibility below this value is dangerous in dense fog / smoke.

[0094] Slope influence :when Linear growth over time;

[0095]

[0096] In the formula, The risk impact of slope is typically rated between 0 and 1, meaning that the risk increases on both uphill and downhill slopes. The longitudinal slope of the road (unit: %) is usually taken as follows. Positive sign indicates uphill, negative sign indicates downhill. The safe slope threshold (unit: %) is typically set at 3%, and is less than... It is considered to have no significant effect. The maximum design gradient for the tunnel (unit: %) is usually taken as 6%, and the denominator is used for normalization.

[0097] curvature influence After exceeding the safety threshold, the percentage increases proportionally based on the sharpest bend.

[0098]

[0099] In the formula, The risk level associated with a curve is typically measured in increments of 0 to 1, with sharper curves posing a higher risk. The road curvature (unit: 1 / m) is typically taken as... , As the radius increases, the greater the curvature, the sharper the turn. The safety curvature threshold (unit: 1 / m) is typically set to 0.005. The impact of gentle bends is negligible. The steepest curvature of the tunnel (unit: 1 / m), typically ranging from 0.01 to 0.02. ), used for normalization.

[0100] Step S24: Weighted fusion to generate comprehensive influence;

[0101]

[0102] In the formula, This is the weighted comprehensive impact of all environmental factors, typically ranging from 0 to 1, with values ​​closer to 1 indicating a more severe environmental impact. These are the weighting coefficients of each factor, summing to 1, and typically taking values ​​of 1. , , , The weight w can be set according to the tunnel type (e.g., mountain tunnels). High altitude, long tunnel (Higher)

[0103] Step S25, Nonlinear mapping output Using the Sigmoid function to Mapped to (0,1];

[0104]

[0105] In the formula, This is the tunnel environmental factor, typically ranging from (0,1], where closer to 1 represents good and closer to 0 represents poor. This is a shape parameter (steepness), typically ranging from 5 to 10. The larger it is, the more sensitive it is to change. This is the neutral threshold point, typically taking values ​​in the range [0,1]. hour, , Controlling sensitivity, Set a neutral level (0.5 is recommended).

[0106] Step S26, based on Adjust HDV model parameters, including:

[0107] Reaction time:

[0108]

[0109] In the formula, This is the adjusted reaction time (in seconds), typically 1.5, but can reach 2.5 in adverse conditions. The worse the environment, the slower the reaction. The baseline reaction time (in seconds) is typically 1.0 to 1.5 seconds. This is the adjustment coefficient, typically ranging from 0.5 to 1.0, which controls the adjustment range. The value is usually [0,1) to represent the degree of environmental degradation; the larger the difference, the stronger the regulation.

[0110] Comfort deceleration:

[0111]

[0112] In the formula, Adjusted comfort deceleration (unit: m / s²) 2 The value is typically 2.0, but can be reduced to 1.5 in severe conditions. A smaller value indicates a "gentler" braking effect. Reference comfort deceleration (unit: m / s) 2 The value is typically between 1.8 and 2.2. This is an adjustment coefficient, typically ranging from 0.2 to 0.5, to prevent... A negative value reflects the behavior of human drivers who react more slowly and brake more cautiously in harsh environments.

[0113] Step S27, based on Adjust AV security policy;

[0114] Safe distance:

[0115]

[0116] In the formula, This is the adjusted minimum safe distance (unit: m), typically set to 2.0, 3.0 in adverse conditions, and a greater distance should be maintained in poor environments. The baseline safety distance (unit: m) is typically 1.5~2.5, and AV can be set even smaller. This is an adjustment coefficient, typically ranging from 0.3 to 0.8, to control the level of conservatism.

[0117] Maximum acceleration:

[0118]

[0119] In the formula, The adjusted maximum acceleration (unit: m / s²) 2 The value is usually 3.0, and 2.4 in adverse conditions. Avoid sharp acceleration on complex road sections. Reference maximum acceleration (unit: m / s²) 2 The value is typically between 2.5 and 3.5. This is an adjustment factor, typically ranging from 0.1 to 0.3, which can be fine-tuned to prevent instability and ensure that AV adopts a conservative strategy in high-risk areas.

[0120] Step S28: Construct the behavioral decision layer: Differentiated interaction strategies, implementing different strategies based on the type of vehicle in front.

[0121] AV→HDV: Enable the "predictive buffering strategy" to identify potential unstable behaviors of HDV in advance (such as sudden deceleration), actively smooth deceleration, and suppress disturbance propagation;

[0122] AV→AV: Allows for shorter headway, improving formation efficiency under regulatory and comfort constraints;

[0123] HDV→AV: Considering the positive impact of AV driving stability on the psychology of following vehicles, the effective reaction time is appropriately shortened to enhance convoy stability.

[0124] The above-mentioned differentiated interaction strategy is achieved by adjusting the core parameters of the basic physical layer (improved IDM model): 1) When the AV follows the HDV, an additional buffer term based on the estimated acceleration of the preceding vehicle is introduced into the expected minimum safe distance to improve the conservatism of the strategy; 2) When the AVs follow each other, a smaller reference reaction time and a smaller reference minimum safe distance are used to shorten the headway; 3) When the HDV follows the AV, the reaction time is shortened by optimizing the perception threshold of the driver's effective reaction time.

[0125] like Figure 4 As shown, as a preferred embodiment, step S3 is implemented as follows:

[0126] Step S31: Collect raw environmental data in real time and obtain real-time values ​​of L, V, G, and C from the sensor network;

[0127] Step S32: Perform single-factor influence transformation, following the method in S23. calculate;

[0128] Step S33: Perform weighted fusion to obtain Call the preset weight Perform a weighted summation;

[0129] Step S34: Apply the Sigmoid function to generate Output the final environmental assessment value for subsequent model use;

[0130] Step S35: Caching and Update Mechanism. Set up a time window sliding update mechanism to avoid instantaneous noise interference. The fluctuations are too large.

[0131] like Figure 5 As shown, as a preferred embodiment, step S4 is implemented as follows:

[0132] Step S41: Construct the state space The state vector is defined to include the following dimensions;

[0133]

[0134] All state variables are normalized and then input into the neural network.

[0135] Step S42: Define the motion space The action is a continuous spatial acceleration command;

[0136]

[0137] Step S43: Design a multi-objective reward function;

[0138]

[0139] The sub-items are as follows:

[0140]

[0141] In the formula, For immediate safety rewards, the value is usually negative or 0; the smaller the value, the more dangerous the situation. The actual net distance (unit: m) is taken as a value. Actual measurement The dynamic safety distance threshold (unit: m) has a value of Related to the environment This is the penalty intensity coefficient, typically ranging from 10 to 50. The larger the distance, the heavier the penalty when the distance is insufficient. The desired travel speed (unit: m / s) is usually the speed limit or free-flow speed. This is the vehicle's current speed (unit: m / s), measured in real time. As an efficiency reward, values ​​close to 0 indicate high efficiency. Current acceleration (unit: m / s²) 2 The values ​​can be positive or negative. For comfort, the lower the acceleration, the more comfortable it is.

[0142] Step S44: Set up a dynamic weighting mechanism, based on... Adaptive weight adjustment;

[0143]

[0144]

[0145] Achieve intelligent switching from prioritizing safety to efficient operation.

[0146] Step S45: Build the DDPG network architecture:

[0147] Actor Network Input state, output action;

[0148] Critic Network Estimate the Q value;

[0149] Target Network Soft update tracking main network .

[0150] Step S46: During the offline training phase, construct a traffic flow simulation environment containing the above three-layer coupled model (used to simulate HDV and background traffic flow) for training the AV's DDPG controller. Initialize the training environment and parameters, initialize the Actor / Critic network parameters, initialize the experience replay buffer D, and set the discount factor. Learning rate .

[0151] Step S47: Interactive sampling and experience storage. The AV runs in the simulation environment and executes actions. (Ornstein-Uhlenbeck noise added for exploration), viewing bonus and new status and tuple Store in D.

[0152] Step S48: Randomly sample a small batch of data, randomly select N samples from D for training, and break the data correlation.

[0153] Step S49: Update the Critic network;

[0154] Calculate the target Q value:

[0155]

[0156] Minimize MSE loss:

[0157]

[0158] Step S410: Update the Actor network to maximize the Q value;

[0159]

[0160] The Actor parameters are updated using the policy gradient method.

[0161] Step S411: Soft update the target network;

[0162]

[0163] Step S412: Repeat the training until convergence;

[0164] Repeat steps S47-S412 until the cumulative reward is stable or the maximum number of training rounds is reached.

[0165] like Figure 6 As shown, as a preferred embodiment, step S5 is implemented as follows:

[0166] Step S51: Strategy solidification and encapsulation

[0167] The trained Actor network parameters are solidified into an executable module and embedded into an AV vehicle decision system or edge computing node;

[0168] Step S52: Real-time state awareness and input construction

[0169] During vehicle operation, sensor data is collected in real time to construct the current state vector. ;

[0170] Step S53: Infer and output the optimal acceleration.

[0171] Will Input the Actor network and output the optimal acceleration command. ;

[0172] Step S54: Perform vehicle control

[0173] Longitudinal control is achieved by sending commands to the power / braking system via the CAN bus;

[0174] Step S55: Support centralized collaborative management and control

[0175] When conditions permit, the strategy will be deployed in the tunnel intelligent traffic control center to achieve multi-vehicle collaborative scheduling and macro-flow control;

[0176] Step S56: Online performance monitoring and feedback optimization

[0177] Record actual running data for model iteration optimization and anomaly detection.

[0178] In summary, the solutions provided in the embodiments of this invention effectively prevent rear-end collisions through Etun-driven parameter adaptation and autonomous vehicle prediction buffering mechanisms; enhance stability by suppressing disturbance propagation and reducing "phantom traffic jams"; improve efficiency by optimizing vehicle speed and distance while ensuring safety, thereby increasing traffic capacity; improve comfort by enabling smooth driving of autonomous vehicles and reducing abrupt acceleration / deceleration; demonstrate strong environmental adaptability, capable of handling various complex conditions such as lighting, visibility, and road alignment; and are easy to implement in engineering, with modular design facilitating integration into existing autonomous driving systems or smart tunnel platforms. This invention provides a comprehensive evaluation perspective for vehicle following models and optimization strategies for mixed traffic flow in tunnel environments, overcoming the shortcomings of existing models that lack a quantitative characterization mechanism for dynamic environmental disturbances (such as sudden drops in entrance lighting, low visibility due to smoke, and sharp downhill curves), and are unable to achieve adaptive parameter adjustment.

[0179] Compared with the prior art, the beneficial effects of the present invention include:

[0180] (1) Enhanced security: through The adaptive parameters of the driving system and the predictive buffer mechanism of autonomous vehicles effectively prevent rear-end collisions.

[0181] (2) Enhanced stability: suppresses disturbance propagation and reduces the occurrence of "phantom traffic jams".

[0182] (3) Improved efficiency: Optimize vehicle speed and distance while ensuring safety to improve traffic capacity.

[0183] (4) Improved comfort: Smooth driving of autonomous vehicles reduces sudden acceleration / deceleration.

[0184] (5) Strong environmental adaptability: It can cope with various complex working conditions such as lighting, visibility, and line shape.

[0185] (6) Easy to implement in engineering: The modular design makes it easy to integrate into existing autonomous driving systems or smart tunnel platforms.

[0186] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0187] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0188] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0189] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

[0190] This invention is not limited to the preferred embodiment described above. Anyone inspired by this invention can derive other forms of vehicle following control methods for mixed traffic flow in tunnel environments. All equivalent variations and modifications made within the scope of the claims of this invention shall fall within the scope of this invention.

Claims

1. A vehicle following control method for mixed traffic flow in tunnel environments, characterized in that, include: The system acquires environmental parameters, traffic flow data, and vehicle type information within the tunnel. The environmental parameters include at least light intensity, visibility, road slope, and curvature. The vehicle type information is used to distinguish between human-driven vehicles and autonomous vehicles. Environmental factors are generated based on the environmental parameters. The environmental factors are obtained by normalizing and weighting each environmental parameter to obtain a comprehensive influence degree, and then generated by nonlinear mapping. The values ​​are in (0,1] and are positively correlated with the safety level of the tunnel environment. Construct a three-layer coupled car-following model, including: The basic physical layer uses an intelligent driver model as the core of dynamics, and its key car-following parameters are dynamically adjusted through the environmental adaptation layer. The environmental adaptation layer dynamically adjusts key parameters of the following vehicle based on the environmental factors, including the reaction time and comfort deceleration of human-driven vehicles and the minimum safe distance and maximum acceleration of autonomous vehicles. The behavioral decision-making layer executes differentiated interaction strategies based on the type of vehicle preceding it; Based on the aforementioned three-layer coupled car-following model, a reinforcement learning algorithm is used to generate longitudinal acceleration control commands for autonomous vehicles, thereby achieving coordinated car-following of mixed traffic flows.

2. The vehicle following control method for mixed traffic flow in tunnel environments according to claim 1, characterized in that: The methods for obtaining data include: By using sensing devices, vehicle-to-everything (V2X) communication, and geographic information data deployed along the tunnel route, environmental parameters, vehicle type information, traffic flow status data, and infrastructure parameters can be obtained. The raw data is preprocessed by filtering and denoising, coordinate unification, and time synchronization to ensure that the data is usable under the same spatiotemporal reference.

3. The vehicle following control method for mixed traffic flow in tunnel environments according to claim 1, characterized in that: The generation process of the environmental factors includes: Each environmental parameter is converted into a normalized index characterizing its negative impact on driving safety, with a value range of [0,1]. The comprehensive influence degree is obtained by weighting and integrating each normalized indicator according to a preset weight. The comprehensive impact is transformed into an environmental factor with a value in (0,1) by a nonlinear mapping function; A sliding update mechanism with a time window is set up to avoid large fluctuations in the environmental factors caused by instantaneous noise.

4. The vehicle following control method for mixed traffic flow in tunnel environments according to claim 1, characterized in that: The parameter adjustment method of the environmental adaptation layer is as follows: when the environmental risk increases, the reaction time of human-driven vehicles is increased and their comfort deceleration is reduced, while the minimum safe distance of autonomous vehicles is increased and their maximum acceleration is reduced; when the environmental risk decreases, the parameters are adjusted in the opposite way to improve traffic efficiency. The differentiated interaction strategies of the behavioral decision layer include: when an autonomous vehicle follows a human-driven vehicle, a predictive buffer strategy is activated to actively and smoothly decelerate; a shorter headway is used for the following time between autonomous vehicles; and the effective reaction time is shortened when a human-driven vehicle follows an autonomous vehicle.

5. The vehicle following control method for mixed traffic flow in tunnel environments according to claim 4, characterized in that: The key parameters for following the car are dynamically determined by the environmental adaptation layer based on environmental factors, including at least one of reaction time, comfort deceleration, minimum safe distance, and maximum acceleration.

6. The vehicle following control method for mixed traffic flow in tunnel environments according to claim 1, characterized in that: The reinforcement learning algorithm is a deep deterministic policy gradient algorithm, specifically including: The state space includes the vehicle's own motion state, environmental factors, the type of the vehicle in front, and local traffic flow characteristics. All state variables are normalized and then input into the neural network. The reward function includes objectives related to safety, stability, efficiency, and comfort. The weight of each objective is adaptively adjusted according to environmental factors to achieve intelligent switching between safety priority and efficient operation. The motion space consists of continuous acceleration commands used to control the longitudinal movement of autonomous vehicles.

7. The vehicle following control method for mixed traffic flow in tunnel environments according to claim 6, characterized in that: The reward function is a multi-objective reward function, and its sub-items include: Safety bonuses are calculated based on the difference between the actual net distance between the vehicle and the vehicle in front and the dynamic safe distance threshold. Stability bonus is negatively correlated with the absolute value of vehicle acceleration; The efficiency reward is negatively correlated with the square of the difference between the vehicle's current speed and the expected speed. Comfort reward is negatively correlated with the absolute value of vehicle acceleration.

8. The vehicle following control method for mixed traffic flow in tunnel environments according to claim 1, characterized in that: The implementation methods for executing the control commands include: The trained algorithm model parameters are solidified into an executable module and embedded into the onboard decision-making system or edge computing node of an autonomous vehicle. When the vehicle is running, sensor data is collected in real time to construct the current state vector. After being input into the model, the optimal acceleration control command is generated and sent to the vehicle's power or braking system via the CAN bus to complete longitudinal control.

9. A vehicle following control system for mixed traffic flow in tunnel environments, characterized in that, include: The sensing module is used to acquire environmental parameters, traffic flow status data and vehicle type information inside the tunnel. The environmental parameters include at least light intensity, visibility, road slope and curvature. The environmental factor calculation module is used to generate environmental factors based on the environmental parameters through normalization, weighted fusion and nonlinear mapping. The environmental factors take values ​​in (0,1) and are positively correlated with the tunnel environmental safety level. The model building module is used to build a three-layer coupled car-following model, which includes a basic physical layer, an environmental adaptation layer, and a behavioral decision layer, respectively realizing the description of the basic laws of car-following, dynamic adjustment of parameters, and execution of differentiated interaction strategies. The control module is used to generate longitudinal acceleration control commands for autonomous vehicles based on the three-layer coupled car-following model using reinforcement learning algorithms. The execution module is used to execute the acceleration control command to achieve coordinated following of mixed traffic flow.

10. A non-transitory computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the vehicle following control method for mixed traffic flow in a tunnel environment according to any one of claims 1-8.

Images