A dynamic variable speed limit control system and method based on multi-modal perception and deep reinforcement learning

Abstract

This invention provides a dynamic variable speed limit control system and method based on multimodal perception and deep reinforcement learning. It includes a data acquisition module for acquiring location and multi-source heterogeneous traffic data; a preprocessing and feature fusion module that uses a time synchronization mechanism of DTW and adaptive Kalman filtering to output a fused feature map; a state vector construction module that calculates traffic parameters weighted by vehicle size and multi-dimensional comprehensive risk indicators based on the fused feature map to construct a state space vector; a decision-making module that uses an improved TD3 algorithm with SHAP-VIF dual-criteria pruning; a command issuance module that issues commands via variable speed limit signs and V2X communication; and a closed-loop optimization module that dynamically optimizes model parameters based on a multi-objective priority sampling and adaptive triggering update mechanism. This variable speed limit control system, which integrates multi-source heterogeneous perception data, possesses deep dynamic decision-making capabilities, and achieves self-closed-loop optimization through real-time feedback, effectively addresses the challenges of traffic efficiency and driving safety on bottleneck road sections.

Description

Technical Field

[0001] This invention relates to the field of intelligent traffic control technology, specifically to a dynamic variable speed limit control system and method based on multimodal perception and deep reinforcement learning. Background Technology

[0002] With the continued explosive growth in the number of motor vehicles, traffic congestion on highways is becoming increasingly severe, leading to frequent safety accidents. Especially at bottleneck sections such as merging areas and tunnel entrances / exits, the intense weaving of traffic flow easily causes turbulence, significantly reducing traffic efficiency and increasing the risk of collisions. However, current mainstream speed limit control methods exhibit clear limitations in addressing these complex scenarios. First, traditional fixed speed limits cannot respond to real-time changes in traffic conditions. Highways typically use static speed limits such as 120 km / h. This "one-size-fits-all" management approach becomes inflexible during traffic surges or severe weather such as heavy rain or snow, easily leading to serious rear-end collisions or vehicle conflicts in merging areas due to improper speed control.

[0003] Besides the rigidity of speed limit strategies themselves, existing perception systems and decision-making algorithms also have significant technical shortcomings. At the perception level, most existing variable speed limit systems rely on single sensors such as cameras or radar. This single-modal perception approach has inherent blind spots and environmental adaptability deficiencies. For example, the recognition accuracy of visual sensors drops drastically in heavy rain, fog, or low-light conditions, while traditional millimeter-wave radar, although possessing strong all-weather capabilities, struggles to accurately distinguish vehicle types and identify static obstacles, leading to distorted assessments of traffic conditions. At the decision-making level, traditional control algorithms such as PID control or fuzzy control primarily rely on preset fixed rules, lacking adaptability to complex nonlinear traffic flows. Especially in highway merging zones, traffic fluctuations caused by vehicle merging behavior are highly random and dynamic, making it difficult for fixed rules to accurately regulate them. Furthermore, most existing systems have not established a closed-loop optimization mechanism of "decision-execution-feedback." The actual traffic improvement effect after the speed limit strategy is implemented cannot be fed back to the decision model in real time, causing the system's regulatory effect to gradually diminish over long-term operation, failing to maintain long-term optimization capabilities.

[0004] The patent with publication number "CN120708423A" discloses a road segment risk warning method based on radar-visual fusion perception. It mainly relies on anti-collision time (TTC) as the core indicator and combines lane and relative speed to construct a longitudinal risk model. It has a relatively crude solution to the problem of assessing the occupation of road resources by vehicles. Moreover, the control logic that usually treats vehicles as point masses or uniform size for positioning and mapping tends to be unidirectional in warning and release, that is, triggering speed limit release after detecting risk, lacking feedback and optimization of the strategy effect.

[0005] The patent with publication number "CN106157650A" discloses a method for improving the traffic efficiency of expressways based on reinforcement learning-based variable speed limit control. It mainly focuses on the convergence of traffic efficiency and reward value. However, it has the problem of easily leading to aggressive control strategies. Furthermore, it does not mention optimization for sparse samples in model update sampling, which may result in insufficient learning for sudden events. At the same time, it adopts traditional reinforcement learning methods, which need to traverse all "state-action" combinations until the reward value converges. It also mainly relies on offline learning, which has the problem of a single decision objective.

[0006] The patent with publication number "CN117475634A" discloses a variable speed limit and dynamic lane management method, device, and computer equipment. It is mainly used to obtain macro data such as traffic flow to judge congestion. The fusion processing of multi-source data is relatively general and is only described as "fusion of multi-source sensing data to obtain traffic flow data", without addressing the time synchronization problem of heterogeneous data.

[0007] Therefore, in the face of complex and ever-changing traffic environments, there is an urgent need to develop a variable speed limit control system that can integrate multi-source heterogeneous sensing data, possess deep dynamic decision-making capabilities, and achieve self-closed-loop optimization through real-time feedback, so as to effectively solve the problems of traffic efficiency and driving safety on bottleneck road sections. Summary of the Invention

[0008] The purpose of this invention is to propose a dynamic variable speed limit control system and method based on multimodal perception and deep reinforcement learning. This system is capable of integrating multi-source heterogeneous perception data, possessing deep dynamic decision-making capabilities, and achieving self-closed-loop optimization through real-time feedback. This effectively solves the problems of traffic efficiency and driving safety on bottleneck road sections.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] This invention provides a dynamic variable speed limit control system based on multimodal perception and deep reinforcement learning, including: a data acquisition module for acquiring multi-source heterogeneous traffic data from Beidou high-precision positioning, millimeter-wave radar, LiDAR+Camera sensors, and smart cameras;

[0011] The preprocessing and feature fusion module, connected to the data acquisition module, is used for spatiotemporal alignment and feature-level fusion of multi-source heterogeneous data. This module is equipped with a time synchronization mechanism based on Dynamic Time Warping (DTW) and adaptive Kalman filtering. It eliminates sampling time differences between sensors by calculating nonlinear time offsets and using the Kalman state update equation for smooth correction. The module also features a spatiotemporal cross-attention mechanism, introducing a time decay coefficient when generating the fused feature map. Historical features are downweighted, and a feature freshness index is introduced. The channel weights are dynamically weighted to output a fused feature map that includes spatiotemporal correlation characteristics;

[0012] The state vector construction module constructs a state space vector for reinforcement learning based on the fused feature map; this module introduces a vehicle size coefficient when calculating lane occupancy. The physical lengths of different vehicle models are weighted and adjusted; simultaneously, a multi-dimensional comprehensive risk index is constructed, including vehicle model risk coefficient, environmental risk coefficient, and lane-inherent risk coefficient. This is used as part of the state vector to quantify the micro and macro risks of traffic flow;

[0013] The decision-making module employs an improved Deep Deterministic Policy Gradient (TD3) algorithm to generate lane-level variable speed limit instructions. At the algorithm's input, a SHAP-VIF dual-criteria pruning mechanism is implemented, using the variance inflation factor (VIF) to filter redundant features and the SHAP value to filter low-contribution features. At the Actor network's output, a lane coordination constraint term is introduced. By calculating the weighted difference between the average speed of the current lane and adjacent lanes, the original network output is smoothly corrected to generate the final speed limit instruction. ;

[0014] The instruction issuing module issues instructions via variable speed limit signs and V2X communication.

[0015] The closed-loop optimization module stores "state-action-reward-new state" samples and updates model parameters. This module employs a multi-objective priority sampling mechanism to calculate sample priority. At the same time, the absolute value of the time difference error is fused. With sample diversity indicators This is to increase the sampling probability of sparse samples and achieve real-time closed-loop optimization of the model.

[0016] Dynamic variable speed limit control systems based on multimodal perception and deep reinforcement learning are used to improve traffic efficiency and driving safety in complex traffic scenarios, and are especially suitable for bottleneck sections such as highway merging zones and tunnels.

[0017] As a further improvement, the main contents of the data acquisition module of the system of the present invention are as follows:

[0018] The data acquisition module is the sensing front end of this system and the starting point of the logical closed-loop control. As the sensing front end and physical interface of this invention, it is configured to capture multi-source heterogeneous traffic data of the target road segment in real time, around the clock. This module adopts a distributed sensor network architecture, establishing a digital mapping of the traffic environment through the collaborative work of multi-dimensional sensors. Specifically, the data acquisition module includes the following four core acquisition units:

[0019] BeiDou high-precision positioning unit:

[0020] This unit is configured to receive signals from the BeiDou Navigation Satellite System, providing a unified global nanosecond-level timestamp and centimeter-level spatial coordinate reference for all sensors (radar, camera, etc.) within the system. Through a unified clock source, it solves the clock drift problem between different sensors, provides an absolute time reference for the time synchronization mechanism in the subsequent data preprocessing module, and ensures strict alignment of multi-source data in the time domain.

[0021] Millimeter-wave radar acquisition unit:

[0022] This unit is primarily responsible for providing stable vehicle kinematics data under low visibility or adverse weather conditions. It employs frequency-modulated continuous wave (FMCH) technology to transmit and receive millimeter-wave signals, and calculates the target vehicle's motion state vector in real time, specifically including the target vehicle's longitudinal velocity (…). ), lateral velocity ( ), relative distance ( ), azimuth ( ) and acceleration ( This unit has extremely strong penetrability, ensuring that the system can accurately capture vehicle dynamics even in environments such as rain, fog, and dust.

[0023] Rayvision integrated sensor unit:

[0024] This unit is crucial for acquiring detailed environmental features. It integrates LiDAR and a high-resolution camera, employing a hardware-level synchronous triggering mechanism to simultaneously acquire 3D spatial and 2D visual information. The acquired data includes 3D contours, specifically high-density 3D point cloud data of the target vehicle and its surrounding environment obtained through LiDAR, accurately depicting the vehicle's length, width, height, and spatial volume; it also includes fine textures, specifically RGB texture information corresponding to the point cloud, acquired in conjunction with optical imaging, to compensate for the radar's limitations in recognizing object edges and semantic information.

[0025] Intelligent surveillance camera unit:

[0026] This unit is used for semantic understanding and visual verification of macro traffic scenes. It is configured at the high point of the road test gantry or pole to collect video streams with a wide field of view. It is responsible for capturing lane-level background information, including but not limited to lane markings, traffic sign status, and macro traffic flow density. At the same time, this unit is also used to identify non-kinematic features of vehicles (such as vehicle type classification and license plate information).

[0027] As a further improvement, the preprocessing and feature fusion modules of the system of the present invention mainly include the following:

[0028] This module is the center of system data processing and enhancement. It communicates with the data acquisition module and receives raw multi-source heterogeneous traffic data. It aims to solve the technical problems of inconsistent sampling frequencies of multiple sensors, nonlinear clock drift, and feature space mismatch. Finally, it outputs a spatiotemporally aligned fused feature map with a high signal-to-noise ratio.

[0029] Time synchronization mechanism based on DTW and adaptive Kalman filtering:

[0030] To address the nonlinear time shift caused by differences in sensor sampling frequencies and the time axis misalignment problem in high-frequency dynamic scenarios, this module abandons the traditional linear interpolation method and adopts a processing mechanism combining Dynamic Time Warping (DTW) and adaptive Kalman filtering. The specific time synchronization calculation formula is as follows:

[0031]

[0032] in:

[0033] - The revised version of the first... Time stamp;

[0034] - Sensor Original Number Time stamp;

[0035] - : The first calculated by DTW Time offset at any moment;

[0036] - The previous moment ( Kalman estimates of time offsets;

[0037] - Kalman filter function, used to smooth the offset;

[0038] Existing time synchronization technologies mostly use linear interpolation, which cannot handle nonlinear time offsets between sensors, such as time axis misalignment caused by differences in sampling frequencies between radar and camera sensors, especially in high-frequency dynamic scenarios. This invention uses dynamic time warping (DTW) calculation and Kalman filtering for smoothing, breaking through the limitations of traditional linear interpolation and significantly improving the time synchronization accuracy of heterogeneous sensors in high-frequency dynamic traffic scenarios.

[0039] Spatial coordinate transformation optimized based on the Expectation-Maximization (EM) algorithm:

[0040] To accurately map the latitude and longitude data provided by BeiDou to the local road coordinate system, this module performs spatial coordinate transformation. Addressing the systematic errors in traditional Gauss-Kruger projection caused by terrain undulations, curves, and other factors, this invention introduces a dynamic calibration vector optimized using the Expectation-Maximization (EM) algorithm. The formula for spatial coordinate transformation is:

[0041]

[0042] in:

[0043] - : Coordinates transformed to the local coordinate system of the road;

[0044] - Rotation matrix (obtained by calibrating the sensor mounting angle);

[0045] - UTM coordinates of road segment reference points (such as the zero point of the merging zone) are used to eliminate the accuracy loss caused by large coordinate values.

[0046] - : UTM plane coordinates after Gauss-Kruger projection transformation ( Longitude (latitude)

[0047] - Dynamic calibration vector (optimized using the EM algorithm based on historical trajectory errors);

[0048] - Translation vector (obtained by calibration of the sensor mounting position).

[0049] When converting BeiDou latitude and longitude coordinates to a local road coordinate system, traditional fixed parameters (such as Gauss-Krüger projection) may introduce systematic errors due to factors such as terrain undulations and road curves. By iteratively optimizing historical trajectory errors (based on expectation maximization, EM algorithm), the coordinate deviation after projection is dynamically corrected, thereby improving positioning accuracy.

[0050] Adaptive noise filtering for millimeter-wave radar data:

[0051] To address the speed measurement errors that millimeter-wave radar is prone to during rapid vehicle acceleration or lane changes, this module introduces a dynamically adjusted process noise variance matrix into the Kalman filtering stage. The formula for millimeter-wave radar noise filtering is as follows:

[0052] ;

[0053] in:

[0054] - : The process noise variance matrix at time k;

[0055] - Initial process noise variance matrix;

[0056] - Adjustment factor (value ranges from 0.1 to 0.3);

[0057] - : No. Time and the The absolute value of the velocity difference at time points, which is normalized before being used in the calculation.

[0058] Traditional fixed noise matrices are prone to filtering lag. This invention dynamically adjusts the noise matrix by monitoring changes in velocity. When the vehicle undergoes violent movement (large speed difference), the process noise variance is automatically increased, making the filter more dependent on the observation value, thereby solving the filtering lag problem during rapid acceleration / lane change and reducing speed measurement error.

[0059] Spatiotemporal and freshness-weighted deep feature fusion:

[0060] After completing spatiotemporal alignment and denoising, the module generates a fused feature map using an improved attention mechanism. This invention innovatively introduces two dynamic factors: temporal decay and feature freshness.

[0061] Spatial weight matrix (introducing spatiotemporal decay):

[0062] Traditional space-channel attention mechanisms only perform static weighting, which cannot adapt to the time-sensitivity of traffic flow. The formula for the spatial weight matrix (time decay) is:

[0063] ;

[0064] in:

[0065] - Spatial weight matrix;

[0066] - : Sigmoid activation function;

[0067] - : 7×7 convolution operation is used to extract spatial neighborhood association features;

[0068] - Average pooling along the channel dimension transforms the feature map from... Compress to Preserve spatial structure information;

[0069] - Multi-scale feature matrix;

[0070] - Time decay coefficient (value ranges from 0.05 to 0.1);

[0071] - : The difference between the feature generation time and the current time.

[0072] Traditional spatial-channel attention mechanisms only statically weight features based on their spatial region and channel importance, which cannot adapt to the temporal changes in features within traffic scenarios (such as vehicle dynamic trajectories and sudden road conditions). Therefore, a time decay term is introduced into the spatial weight matrix calculation. Historical features are assigned a weight that decreases over time, highlighting key real-time areas (such as current congestion points and merging zones).

[0073] Channel weight vector:

[0074] To address the issue of inconsistent update frequencies among different sensors, the channel weight vector (feature freshness) formula constructed in this invention is as follows:

[0075] ;

[0076] in:

[0077] - Channel weight vector;

[0078] - : Sigmoid activation function;

[0079] - Multilayer perceptron operation;

[0080] - Global max pooling operation;

[0081] - Multi-scale feature matrix;

[0082] - Weighting coefficient (value 1~2);

[0083] - Freshness of the c-th feature channel (0~1, updated in real time).

[0084] Feature freshness is introduced in the calculation of the channel matrix. By assigning higher weights to core feature channels with high time sensitivity, such as vehicle speed and distance, and suppressing interference from environmental noise (such as background trees and fixed buildings), the effectiveness of decision data is significantly improved.

[0085] As a further improvement, the main contents of the state vector construction module of the system of the present invention are as follows:

[0086] This module is primarily responsible for transforming the fused feature map output from the preprocessing and feature fusion modules into a structured state space vector that the decision layer can understand. Furthermore, through in-depth analysis of microscopic vehicle behavior and macroscopic traffic flow characteristics, it calculates high-precision lane load and multidimensional risk indicators, providing accurate input state vectors for subsequent reinforcement learning decisions. This module mainly includes the construction of indicators in the following two core dimensions:

[0087] Lane occupancy calculation based on vehicle size weighting:

[0088] To address the technical issue of traditional lane occupancy calculations neglecting vehicle model differences and leading to inaccurate lane load assessments, this module introduces a vehicle size coefficient. The system utilizes 3D vehicle contour information extracted from radar-visual fusion data to weight the spatiotemporal occupancy of different vehicle models. The specific weighted lane occupancy calculation formula is as follows:

[0089]

[0090] in:

[0091] - : No. Lane weighted time occupancy (in %);

[0092] - : No. The actual physical length of the vehicle (measured by Ravis sensors, unit: m);

[0093] - : Size coefficient of the k-th vehicle, which is set to a value for large vehicles (such as trucks and buses). The value for small cars is This is used to correct the actual impact of large vehicles on road resources;

[0094] - : No. The total number of vehicles in the lane;

[0095] - The preset safety clearance constant (e.g., 2m, optional) means that the vehicle occupies not only the body but also the necessary front and rear clearance.

[0096] - ;

[0097] Current occupancy calculations assign equal weight to different vehicle types and calculate occupancy solely based on vehicle presence time. This can easily underestimate traffic congestion when large vehicles are mixed in, leading to distorted lane load assessments. This invention introduces a vehicle size coefficient... The occupancy rate index has been corrected from the perspective of physical space occupation, so that the calculation results can more realistically reflect the actual load status of the lanes and avoid misjudgment of "false smooth flow" caused by the density of large vehicles.

[0098] Construction of multi-dimensional comprehensive risk indicators:

[0099] To overcome the limitations of traditional risk assessments that rely solely on a single indicator (such as TTC), this module constructs a multi-dimensional comprehensive risk index that couples kinematic characteristics, static environment, and dynamic interaction. To eliminate the influence of different physical dimensions (such as seconds, frequency, and dimensionless coefficients), the state vector construction module normalizes each sub-item before calculating the comprehensive index. The specific formula for the multi-dimensional comprehensive risk index is as follows:

[0100] ;

[0101] in:

[0102] - : Comprehensive traffic risk index for road segments at any time (dimensionless value, the larger the value, the higher the risk).

[0103] - The weighting coefficients of each risk component satisfy the following conditions: (Recommended value: );

[0104] - The minimum time to collision (Time To Collision) within the monitored road segment is calculated using the following formula: , The relative distance between the vehicle in front and the vehicle behind. The relative speed between the vehicle in front and the vehicle behind;

[0105] - Minimum constant (e.g.) This is used to prevent the denominator from being 0 when the relative speed is extremely high or the distance is extremely small;

[0106] - Vehicle type risk coefficient (for large vehicles) Small cars This is used to amplify the potential hazards of large vehicles under low TTC conditions;

[0107] - Environmental risk coefficient (normalized value, e.g., 1 / 3 for good visibility) Heavy rain / dense fog );

[0108] - Lane inherent risk coefficient (merging lane / emergency lane) Main driving lane );

[0109] - The frequency of vehicle merging and lane changing per unit time (unit: times / minute) is used to characterize the severity of traffic flow interaction conflicts.

[0110] Traditional risk assessments often focus on a single factor (such as using only TTC to assess collision risk), neglecting key influences such as vehicle model differences, lane characteristics, and traffic flow interaction. This indicator couples three parts: Part One ( ), combining vehicle model characteristics to assess the vehicle's own collision risk, and solving the problem of "one-size-fits-all" risk assessment for large vehicles and small vehicles; Part Two ( ), relating the environment and lane characteristics, highlighting the risks of dangerous lanes under adverse weather conditions (such as the combined risks of low visibility and dense traffic in tunnels); Part Three ( ): Quantify traffic flow interaction conflicts and improve risk sensitivity in frequently interacting scenarios such as merging zones.

[0111] Weighting coefficient It can be dynamically adjusted according to the scenario (e.g., increasing the merging area). To enhance the risk of merging, the tunnel scenario is enlarged. (To enhance environmental risk) and make risk assessment more relevant to the characteristics of the scenario.

[0112] As a core component of the state-space vector, this indicator provides accurate risk input for the subsequent decision-making module (improving the TD3 algorithm), enabling the generated speed limit instructions to specifically reduce high-risk factors (such as appropriately reducing speed limits in merging areas and reducing interaction conflicts), ultimately improving traffic safety and efficiency.

[0113] As a further improvement, the decision-making module of the system of the present invention mainly includes the following:

[0114] The decision-making module is the core control unit of the system of this invention. This module is configured to receive high-dimensional state space vectors from the state vector construction module, perform inference using the improved Deep Deterministic Policy Gradient (TD3) algorithm, and generate lane-level dynamic speed limit commands.

[0115] To address the shortcomings of traditional deep reinforcement learning algorithms in processing high-dimensional traffic data, such as large computational redundancy, lack of lane coordination in output instructions, and single optimization objective, this module adopts the following specific improvement implementation scheme:

[0116] Feature pruning mechanism based on SHAP-VIF dual criteria:

[0117] In the state input stage of the TD3 algorithm, to address the "curse of dimensionality" problem caused by multi-source heterogeneous data, this module pre-emptively applies a SHAP-VIF dual-criteria feature filter to the input layer. This filter aims to eliminate ineffective features with "high redundancy but low contribution" (such as environmental background noise), retaining only key states that have a substantial impact on decision-making. The SHAP-VIF dual-criteria pruning formula is as follows:

[0118] ;

[0119] in:

[0120] - : Binary mask vector (1 indicates that the first digit is retained) (1 feature, 0 indicates pruning).

[0121] - : No. Variance inflation factor of each feature (measures multicollinearity) to filter out highly redundant features;

[0122] - VIF threshold (usually set to 5);

[0123] - : No. The SHAP value of each feature, calculated based on game theory, is used to quantify the marginal contribution of the feature to the final decision of the model and filter out features that are meaningless to the decision.

[0124] - SHAP value threshold (value range: 0.01~0.05).

[0125] In the state feature vector input stage of TD3, traditional single-criterion pruning (such as using only VIF or only SHAP) may retain features with "high redundancy but low contribution" (such as certain environmental noise), resulting in high model computation and interference with decision accuracy. This invention effectively compresses the state space dimension through dual screening of "VIF redundancy removal" and "SHAP value preservation," reducing the model's computational burden while avoiding interference from invalid information in the decision network, thus improving inference speed and accuracy.

[0126] Actor network output with lane coordination constraints:

[0127] During the instruction generation phase, to address the problem that traditional algorithms only pursue local optima for a single lane (such as excessively increasing the speed of the main lane), leading to significant speed differences between adjacent lanes (such as the main lane and ramps), and consequently causing frequent lane changes and safety hazards, this module introduces a lane cooperation penalty term at the output layer of the Actor network. The specific Actor network output formula is as follows:

[0128] ;

[0129] in:

[0130] - ;

[0131] - Scaling and offset coefficients (map the output to a preset speed limit range, such as k_1=30, k_2=90 corresponding to 60~120km / h);

[0132] - The hyperbolic tangent activation function maps the unbounded raw output of the neural network to an open interval (-1, 1) to ensure the numerical stability of the control quantity.

[0133] - : Actor network output layer The original values ​​of each neuron, whose range is the real number field. ;

[0134] - Lane coordination coefficient (value 0.1~0.2), which determines the system's sensitivity to speed differences between lanes;

[0135] -

[0136]

[0137] ;

[0138] Traditional TD3 only optimizes single-lane efficiency (e.g., maximizing the speed of the lane itself), which may lead to excessively large speed limit differences between adjacent lanes (e.g., 120 km / h for the main lane and 60 km / h for the merging lane), causing frequent lane changes and overtaking. This new formula, through a negative feedback mechanism, forcibly constrains the speed gradient between adjacent lanes. When the model attempts to assign an extremely high speed to a lane while adjacent lanes have lower speeds, a co-operational term applies a penalty, automatically smoothing out the speed limit difference between lanes. This mechanism is particularly crucial in sensitive sections such as merging zones and tunnel entrances / exits, effectively suppressing dangerous lane-changing behavior caused by excessive speed limit differences and reducing the risk of lateral collisions.

[0139] Design of a reward function that balances multiple objectives:

[0140] To guide the reinforcement learning model towards "safety, efficiency, and stability," this module designs a multi-objective reward function that includes a penalty term for the vehicle speed standard deviation. The specific reward function formula is as follows:

[0141] ;

[0142] in:

[0143] - The reward value at time t;

[0144] - Weighting coefficients for each indicator;

[0145] - Average speed increase (the difference between the current vehicle speed and the historical average vehicle speed).

[0146] - Average flow increase (the difference between current flow and historical average flow).

[0147] - Changes in risk indicators (the difference between current risk and historical risk).

[0148] - Standard deviation of vehicle speed within a lane (measures speed smoothness).

[0149] Existing reward functions often focus on maximizing vehicle speed, which can easily lead to aggressive speed limit commands from the system, causing drastic fluctuations in traffic speed (i.e., stop-and-go oscillations). This invention innovatively incorporates a vehicle speed standard deviation penalty term into the reward function. This forces agents to choose speed limit schemes that allow traffic flow to maintain a constant and smooth speed when exploring strategies, thereby improving traffic efficiency while ensuring driving stability and comfort.

[0150] As a further improvement, the main contents of the instruction issuing module of the system of the present invention are as follows:

[0151] The instruction issuing module is the execution terminal of the system of this invention. It is configured to convert the digital speed limit instruction generated by the decision module into dual control signals of the physical world and the digital world, so as to ensure that vehicles of different intelligence levels can receive instructions in real time.

[0152] The specific implementation of this module includes two parallel distribution links:

[0153] Physical side visualization release:

[0154] The system communicates with roadside variable message signs (VMS) or actively illuminated speed limit signs via fiber optic networks, rounds down the received continuous speed limit values ​​(e.g., 83 km / h) to the standard speed limit level (e.g., 80 km / h), and displays them on LED screens, ensuring that drivers of traditional non-networked vehicles can visually obtain the best recommended speed for the current road segment.

[0155] Digital-side V2X broadcast:

[0156] The system utilizes roadside units (RSUs) deployed on the roadside to encode speed limit instructions into communication frames (such as SPAT or RSI message sets) in C-V2X or DSRC standard formats. These instructions are broadcast with low latency to the OBUs (On-Board Units) of connected and autonomous vehicles entering the road segment. Autonomous vehicles can directly use these digital instructions as constraints for the planning and control layer, achieving seamless inter-vehicle (I2V) cooperative control.

[0157] As a further improvement, the main contents of the closed-loop optimization module of the system of the present invention are as follows:

[0158] The closed-loop optimization module is crucial for maintaining the system's continuous evolution and adaptability to complex environments. This module periodically collects the "state-action-reward-new state" quadruple. The data is then stored in the experience replay pool. To address the poor model robustness caused by the extremely uneven distribution of traffic flow data, this module innovatively adopts a multi-objective priority sampling mechanism.

[0159] Traditional Priority Experience Replay (PER) sampling only considers temporal difference error (TD-error), prioritizing the learning of samples with "prediction errors." However, in traffic scenarios, over 90% of the samples are high-frequency "normal free-flow," while critical low-frequency samples such as "accident congestion" and "extreme weather," though potentially having small errors, are crucial for safety. Sampling only based on error leads to insufficient learning of low-frequency, long-tail scenarios. Therefore, this invention introduces a diversity index and constructs the following sampling formula.

[0160] ;

[0161] in:

[0162] - : No. Sampling priority of each sample;

[0163] - : No. The absolute value of the time difference error of each sample represents the learning value of the sample;

[0164] - The minimum constant (e.g., 1e-5) ensures that zero-error samples still have a non-zero probability of being sampled.

[0165] - Value Sensitivity Index (Value) ), adjusting the nonlinear effect of the error term on priority;

[0166] - Diversity weighting coefficient (value) This is used to increase the probability of sparse samples being selected;

[0167] - : Sample diversity index. Calculated by the first... The state vector of each sample is obtained by the feature space distance (such as Euclidean distance) between the state vector of each sample and the existing cluster centers in the experience pool. This is used to quantify the "uniqueness" of the sample and solve the problem of insufficient sampling in long-tail scenarios (such as accidents and extreme weather).

[0168] Traditional TD3 sampling is based only on (Prioritizing samples with large learning errors), but in traffic scenarios, "normal traffic flow" samples account for a high proportion (high-frequency scenarios), while "accidents and extreme weather" samples account for a low proportion (low-frequency but crucial scenarios). Sampling only based on value leads to insufficient learning of low-frequency scenarios, resulting in decision failure in emergencies. By adding... The parameters are optimized by considering both the learning value of the samples and the diversity of scenarios, avoiding overfitting the model to common scenarios, improving adaptability to unexpected situations, increasing the sampling probability of unique scenario samples, and ensuring that the model covers all scenarios. As the system runs longer, the more diverse the samples accumulated in the experience pool, the stronger the generalization ability of the decision model (TD3) becomes, ultimately forming a virtuous cycle of "improved perception accuracy -> safer decision-making -> higher value of feedback data -> more intelligent model".

[0169] The adaptive trigger update mechanism includes triggering the reinforcement learning model parameter update when the reward value is lower than a preset threshold for several consecutive decision cycles, or when the comprehensive risk index continues to rise, in order to avoid frequent ineffective training and reduce computational load.

[0170] Compared to the patent with publication number "CN120708423A", this invention constructs a multi-dimensional comprehensive risk index, taking a more comprehensive view. This invention not only includes TTC (Traffic Traction Control), but also introduces "vehicle type risk coefficient" (with differentiated weighting for large and small vehicles), "environmental risk coefficient" (such as visibility, rain, and fog), "lane inherent risk coefficient," and "merging / lane change frequency." By coupling microscopic kinematic characteristics with macroscopic environmental characteristics, this invention can more accurately quantify the comprehensive risk of complex scenarios (such as large vehicles in merging zones during heavy rain), avoiding the distortion of a single TTC index in low-speed congestion or lateral conflict scenarios. Furthermore, this invention proposes a lane occupancy calculation method based on vehicle size weighting, resulting in more refined modeling. This invention utilizes radar-based sensors to acquire... The invention utilizes the three-dimensional contours of vehicles and introduces a "size factor" to weight large vehicles (trucks, buses) and small vehicles, while also considering a constant safety clearance. This method corrects the road load assessment bias caused by ignoring vehicle size differences in traditional occupancy calculations, and can accurately reflect "false smooth traffic" or actual congestion. Furthermore, the invention establishes a real-time closed-loop control mechanism of "perception-decision-execution-optimization." It employs an improved TD3 deep reinforcement learning algorithm, using "multi-objective priority sampling" and "adaptive trigger update mechanism" to dynamically optimize model parameters based on traffic flow feedback (reward signal) after policy execution. This enables the system to have self-evolution capabilities and adapt to long-term changes in traffic flow distribution, rather than simply providing open-loop warnings based on fixed rules.

[0171] Compared to the patent with publication number "CN106157650A", this invention adopts a deep deterministic policy gradient (TD3) algorithm based on the Actor-Critic architecture, which is more suitable for high-dimensional continuous action spaces. Traditional methods are prone to the "curse of dimensionality" when dealing with high-dimensional state spaces involving multiple lanes and multiple sensor inputs, resulting in low computational efficiency. This invention not only employs deep reinforcement learning but also innovatively introduces the SHAP-VIF dual-criteria pruning mechanism to remove highly redundant and low-contribution features at the input, significantly improving the inference speed and training efficiency of the decision model. Furthermore, this invention introduces "lane coordination constraints" and a multi-objective smoothing mechanism at the decision output layer. It adds a negative feedback term for the speed of neighboring lanes to the Actor network output, forcibly smoothing the speed gradient of adjacent lanes and avoiding frequent lane-changing conflicts caused by local optima in a single lane. Simultaneously, it adds a "vehicle speed standard deviation penalty term" to the reward function, guiding the strategy towards a smoother and more comfortable evolution rather than simply pursuing speed maximization. Moreover, this invention proposes a multi-objective priority sampling mechanism based on "value and diversity." When sampling in the experience replay pool, this invention considers not only TD error (learning value) but also introduces a sample diversity index (feature space distance), ensuring that "long-tail" low-frequency scenarios such as accidents and extreme weather can be fully learned by the model, thereby significantly improving the system's robustness under extreme conditions.

[0172] Compared to the patent with publication number "CN117475634A", this invention addresses the problem of inconsistent sampling frequencies among multi-source heterogeneous sensors (radar, camera) by proposing a time synchronization mechanism based on a combination of DTW (Dynamic Time Warping) and adaptive Kalman filtering. This invention, through nonlinear warping and filtering smoothing, accurately solves the problem of time axis misalignment between sensors, ensuring precise alignment of high-speed moving targets across different sensor data, with accuracy far exceeding conventional linear interpolation or simple alignment methods. Furthermore, this invention introduces "time decay" and "feature freshness" during feature fusion. Two dynamic weighting factors enable the model to dynamically adjust its attention based on the timeliness of data generation and the importance of the channel. This mechanism is not simply a superposition of data, but rather an automatic extraction of the most valuable spatiotemporal features for decision-making through neural networks, significantly improving the signal-to-noise ratio of the fused feature map. Furthermore, this invention constructs a high-dimensional state space vector and utilizes a deep neural network for continuous control. This invention quantifies traffic states into a vector containing multi-dimensional information such as speed, risk, and occupancy, and uses a trained TD3 network to directly generate continuous speed limit commands, rather than simple threshold triggering. This allows the invention to cope with minor traffic flow fluctuations and achieve smoother, more precise dynamic control, rather than intervening only after congestion occurs.

[0173] Compared with existing technologies, the beneficial effects of this invention are: it has significant advantages in perception robustness, feature fusion accuracy, decision adaptability, and system long-term effectiveness. In multimodal perception, the system innovatively integrates multi-source data from BeiDou high-precision positioning, millimeter-wave radar, LiDAR (Light Detection and Ranging), and intelligent cameras, effectively compensating for the performance limitations of single sensors in specific environments. This multi-source fusion mechanism endows the system with extremely strong environmental adaptability. For example, in extreme scenarios such as heavy rain or dense fog, the high-precision point cloud of LiDAR and the penetrating characteristics of millimeter-wave radar can seamlessly replace the visual information of the camera, ensuring that the system's assessment of traffic conditions is always continuous and effective, thereby significantly improving the robustness of perception.

[0174] At the data processing and decision control level, this invention achieves efficient fusion of heterogeneous features through an improved spatial-channel cross-attention mechanism. This mechanism intelligently focuses on key spatial areas such as merging zones and core traffic features such as vehicle speed and spacing, resulting in a significantly lower information entropy of the fused feature map compared to traditional methods, providing high signal-to-noise ratio input for the decision model. Based on this, combined with the TD3 deep reinforcement learning algorithm with adaptive pruning, the system can respond to millisecond-level traffic state changes in real time. Especially in complex merging zone scenarios, the algorithm can generate fine-grained lane-level speed limit instructions, such as implementing differentiated speed limit strategies for main lanes and merging lanes, thereby significantly reducing the probability of collisions when vehicles merge and improving overall traffic efficiency. More importantly, this invention constructs a real-time closed-loop optimization mechanism of "decision-feedback-update," which can dynamically adjust model parameters based on actual traffic flow feedback after strategy execution. Actual tests show that this self-evolving system exhibits a much lower decay of its control effect during long-term continuous operation than traditional open-loop systems, demonstrating excellent long-term stability. Attached Figure Description

[0175] Figure 1 This is a diagram illustrating the architecture of a dynamic variable speed limit control system based on multimodal perception and deep reinforcement learning, as described in this invention.

[0176] Figure 2 This is a flowchart of the spatiotemporal cross-attention mechanism of the present invention;

[0177] Figure 3 This is a flowchart of the TD3 algorithm decision-making process of the present invention. Detailed Implementation

[0178] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0179] like Figure 1-3 As shown, a dynamic variable speed limiting control system based on multimodal perception and deep reinforcement learning includes:

[0180] The data acquisition module is used to acquire multi-source heterogeneous traffic data from BeiDou high-precision positioning, millimeter-wave radar, LiDAR+Camera sensors, and smart cameras.

[0181] The preprocessing and feature fusion module, connected to the data acquisition module, is used for spatiotemporal alignment and feature-level fusion of multi-source heterogeneous data. This module is equipped with a time synchronization mechanism based on Dynamic Time Warping (DTW) and adaptive Kalman filtering. It eliminates sampling time differences between sensors by calculating nonlinear time offsets and using the Kalman state update equation for smooth correction. The module also features a spatiotemporal cross-attention mechanism, introducing a time decay coefficient when generating the fused feature map. Historical features are downweighted, and a feature freshness index is introduced. The channel weights are dynamically weighted to output a fused feature map that includes spatiotemporal correlation characteristics;

[0182] The state vector construction module constructs a state space vector for reinforcement learning based on the fused feature map; this module introduces a vehicle size coefficient when calculating lane occupancy. The physical lengths of different vehicle models are weighted and adjusted; simultaneously, a multi-dimensional comprehensive risk index is constructed, including vehicle model risk coefficient, environmental risk coefficient, and lane-inherent risk coefficient. This is used as part of the state vector to quantify the micro and macro risks of traffic flow;

[0183] The decision-making module employs an improved Deep Deterministic Policy Gradient (TD3) algorithm to generate lane-level variable speed limit instructions. At the algorithm's input, a SHAP-VIF dual-criteria pruning mechanism is implemented, using the variance inflation factor (VIF) to filter redundant features and the SHAP value to filter low-contribution features. At the Actor network's output, a lane coordination constraint term is introduced. By calculating the weighted difference between the average speed of the current lane and adjacent lanes, the original network output is smoothly corrected to generate the final speed limit instruction. ;

[0184] The instruction issuing module issues instructions via variable speed limit signs and V2X communication.

[0185] The closed-loop optimization module stores "state-action-reward-new state" samples and updates model parameters. This module employs a multi-objective priority sampling mechanism to calculate sample priority. At the same time, the absolute value of the time difference error is fused. With sample diversity indicators This is to increase the sampling probability of sparse samples and achieve real-time closed-loop optimization of the model;

[0186] The system is designed to improve traffic efficiency and driving safety in complex traffic scenarios, and is particularly suitable for bottleneck sections such as highway merging zones and tunnels.

[0187] The time synchronization formula (DTW + Kalman filtering) for the preprocessing and feature fusion modules is as follows:

[0188] ;

[0189] in:

[0190] - The revised version of the first... Time stamp;

[0191] - Sensor Original Number Time stamp;

[0192] - : The first calculated by DTW Time offset at any moment;

[0193] - The previous moment ( Kalman estimates of time offsets;

[0194] - The estimated time offset, obtained from the Kalman filter state update equation, is used to smooth the DTW calculation results.

[0195] The spatial coordinate transformation formula for the preprocessing and feature fusion module is:

[0196]

[0197] in:

[0198] - : Coordinates transformed to the local coordinate system of the road;

[0199] - Rotation matrix (obtained by calibrating the sensor mounting angle);

[0200] - UTM coordinates of road segment reference points (such as the zero point of the merging zone) are used to eliminate the accuracy loss caused by large coordinate values.

[0201] - : UTM plane coordinates after Gauss-Kruger projection transformation ( Longitude (latitude)

[0202] - Dynamic calibration vector (optimized using the EM algorithm based on historical trajectory errors);

[0203] - Translation vector (obtained by calibration of the sensor mounting position);

[0204] The millimeter-wave radar noise filtering formula for the preprocessing and feature fusion module is as follows:

[0205] ;

[0206] in:

[0207] - : The process noise variance matrix at time k;

[0208] - Initial process noise variance matrix;

[0209] - Adjustment factor (value ranges from 0.1 to 0.3);

[0210] - : The absolute value of the velocity difference between time k and time k−1, wherein the velocity difference is normalized before being included in the calculation.

[0211] The formula for the spatial weight matrix (time decay) of the preprocessing and feature fusion modules is as follows:

[0212] ;

[0213] in:

[0214] - Spatial weight matrix;

[0215] - : Sigmoid activation function;

[0216] - : 7×7 convolution operation is used to extract spatial neighborhood association features;

[0217] - Average pooling along the channel dimension transforms the feature map from... Compress to Preserve spatial structure information;

[0218] - Multi-scale feature matrix;

[0219] - Time decay coefficient (value ranges from 0.05 to 0.1);

[0220] - The difference between the feature generation time and the current time;

[0221] The formula for the channel weight vector (feature freshness) of the preprocessing and feature fusion module is as follows:

[0222] ;

[0223] in:

[0224] - Channel weight vector;

[0225] - : Sigmoid activation function;

[0226] - Multilayer perceptron operation;

[0227] - Global max pooling operation;

[0228] - Multi-scale feature matrix;

[0229] - Weighting coefficient (value 1~2);

[0230] - Freshness of the c-th feature channel (0~1, updated in real time);

[0231] The lane occupancy formula of the state vector construction module is:

[0232] ;

[0233] in:

[0234] - : No. Lane weighted time occupancy (in %);

[0235] - : No. The actual physical length of the vehicle (measured by Ravis sensors, unit: m);

[0236] - : Size coefficient of the k-th vehicle, which is set to a value for large vehicles (such as trucks and buses). The value for small cars is This is used to correct the actual impact of large vehicles on road resources;

[0237] - : No. The total number of vehicles in the lane;

[0238] - The preset safety clearance constant (e.g., 2m, optional) means that the vehicle occupies not only the body but also the necessary front and rear clearance.

[0239] - It is used to normalize the space occupied by vehicles in the lane, and its value is determined according to the effective coverage of the sensor or the preset statistical interval.

[0240] The formula for the multi-dimensional comprehensive risk index of the state vector construction module is as follows:

[0241] ;

[0242] in:

[0243] - : Comprehensive traffic risk index for road segments at any time (dimensionless value, the larger the value, the higher the risk).

[0244] - The weighting coefficients of each risk component satisfy the following conditions: (Recommended value: );

[0245] - The minimum collision time within the monitored road segment is calculated using the following formula: , The relative distance between the vehicle in front and the vehicle behind. The relative speed between the vehicle in front and the vehicle behind;

[0246] - Minimum constant (e.g.) This is used to prevent the denominator from being 0 when the relative speed is extremely high or the distance is extremely small;

[0247] - Vehicle type risk coefficient (for large vehicles) Small cars This is used to amplify the potential hazards of large vehicles under low TTC conditions;

[0248] - Environmental risk coefficient (normalized value, e.g., 1 / 3 for good visibility) Heavy rain / dense fog );

[0249] - Lane inherent risk coefficient (merging lane / emergency lane) Main driving lane );

[0250] - : Frequency of vehicle merging and lane changing per unit time (unit: times / minute), used to characterize the severity of traffic flow interaction conflicts;

[0251] The SHAP-VIF dual-criteria pruning formula for the decision module is as follows:

[0252] ;

[0253] in:

[0254] - : Binary mask vector (1 indicates that the first digit is retained) (1 feature, 0 indicates pruning).

[0255] - : No. Variance inflation factor of each feature (measures multicollinearity) to filter out highly redundant features;

[0256] - VIF threshold (usually set to 5);

[0257] - : No. The SHAP value of each feature (which measures its contribution to the decision) is used to filter out features that are meaningless to the decision.

[0258] - SHAP value threshold (value range: 0.01~0.05).

[0259] The Actor network output formula of the decision module is:

[0260] ;

[0261] in:

[0262] - ;

[0263] - Scaling and offset coefficients (map the output to a preset speed limit range, such as k_1=30, k_2=90 corresponding to 60~120km / h);

[0264] - The hyperbolic tangent activation function maps the unbounded raw output of the neural network to an open interval (-1, 1) to ensure the numerical stability of the control quantity.

[0265] - : Actor network output layer The original values ​​of each neuron, whose range is the real number field. ;

[0266] - Lane coordination coefficient (value 0.1~0.2), which determines the system's sensitivity to speed differences between lanes;

[0267] - : No. All adjacent lanes of the lane (set) The current average detected vehicle speed. This feature introduces a negative feedback mechanism based on real-time traffic conditions to forcibly smooth the speed gradient between lanes, avoiding excessive differences between the speed limit of the current lane and the actual speed of vehicles in adjacent lanes;

[0268] The reward function formula for the decision-making module is:

[0269] ;

[0270] in:

[0271] - The reward value at time t;

[0272] - Weighting coefficients;

[0273] - Average speed increase (the difference between the current vehicle speed and the historical average vehicle speed).

[0274] - Average flow increase (the difference between current flow and historical average flow).

[0275] - Changes in risk indicators (the difference between current risk and historical risk).

[0276] - Standard deviation of vehicle speed within a lane (measures speed smoothness).

[0277] The historical average value is a sliding average or an exponentially weighted average based on a preset time window, used to smooth the impact of short-term traffic fluctuations on reward calculation.

[0278] The multi-objective priority sampling formula of the closed-loop optimization module is as follows:

[0279] ;

[0280] in:

[0281] - : No. Sampling priority of each sample;

[0282] - : No. The absolute value of the time difference error of each sample represents the learning value of the sample;

[0283] - The minimum constant (e.g., 1e-5) ensures that zero-error samples still have a non-zero probability of being sampled.

[0284] - Value Sensitivity Index (Value) ), adjusting the nonlinear effect of the error term on priority;

[0285] - Diversity weighting coefficient (value) This is used to increase the probability of sparse samples being selected;

[0286] - : Sample diversity index. Calculated by the first... The state vector of each sample is obtained by measuring the feature space distance (such as Euclidean distance) between the state vector and the existing cluster centers in the experience pool. This distance is used to quantify the "uniqueness" of the sample, addressing the problem of insufficient sampling in long-tail scenarios (such as accidents and extreme weather). Figure 1 Figure 2 shows the connection relationships of the data acquisition module, feature fusion module, state vector construction module, decision module, instruction issuance module, and closed-loop optimization module; Figure 3 shows the process of single-modal feature extraction, spatial / channel weight calculation, and cross-fusion; Figure 4 shows the interaction logic of the Actor network and the dual Critic network.

[0287] Specifically:

[0288] Step 1: System Deployment and Equipment Installation

[0289] (1) Site selection planning:

[0290] Based on the lane distribution of the merging zone of the expressway (2 lanes on the main road + 1 lane for merging), traffic flow direction and terrain characteristics, three monitoring nodes are set up at 500m, 300m upstream and 200m downstream of the merging zone to ensure coverage of the entire merging zone (500m section) and the upstream and downstream affected areas, and to avoid monitoring blind spots.

[0291] (2) Equipment installation:

[0292] Beidou high-precision positioning equipment: fixed on the top of a gantry at 500m, with the antenna facing due north, establishing a centimeter-level differential link with the base station, with a baseline error ≤5mm and a time synchronization accuracy ≤10ns.

[0293] Millimeter-wave radar (2 units): installed on both sides of a gantry at 300m, with a horizontal depression angle of 15° and an azimuth angle coverage of ±30°. After calibration with debugging software, the detection distance error is ≤0.5m and the azimuth angle error is ≤1°.

[0294] The LiDAR sensor (LiDAR+Camera) is installed in the middle of a gantry at a distance of 300m. The LiDAR (128 lines) has a scanning frequency of 10Hz and a field of view of 120°×25°. The Camera (2 megapixels) has a frame rate of 30Hz. The overlap area of ​​the field of view is ensured to be ≥95% through a time synchronizer (error ≤1ms).

[0295] Intelligent camera: installed on a gantry at 100m, focusing on lane lines and merging zone entrances, with lane line positioning error ≤5cm after lens distortion correction.

[0296] (3) Installation of supporting facilities:

[0297] - Variable speed limit signs: 3 lanes correspond to 3 displays (1920×1080 resolution), installed at a height of 2.5m, ensuring visibility from 200m away.

[0298] - V2X communication equipment: Deployed on a gantry at 500m, with a PC5 interface communication distance of ≥800m, a transmission rate of ≥256kbps, and a latency of ≤100ms.

[0299] Step 2: Data Acquisition and Transmission

[0300] (1) Sensor startup and parameter configuration:

[0301] Beidou equipment: update frequency 10Hz, output latitude and longitude ( ),altitude ,speed , heading angle θ and PTP timestamp.

[0302] Millimeter-wave radar: Detection mode set to "high-density traffic flow", output distance d and radial velocity. Azimuth (10Hz).

[0303] Rayvision sensor: LiDAR outputs point cloud (x, y, z, intensity), Camera outputs RGB image (1920×1080), and the two are synchronized by hardware trigger.

[0304] Time synchronization: Based on the BeiDou PTP timestamp, the clocks of each sensor are calibrated by a time synchronizer to ensure that the timestamp error is ≤5ms.

[0305] (2) Data collection:

[0306] BeiDou data: Records vehicle positioning trajectory in real time, outputting it every 100ms.

[0307] Radar data: Outputs a target list every 100ms (maximum 50 targets / frame).

[0308] Leishi Data: LiDAR outputs 20,000 points every 100ms, and Camera outputs one frame of image every 33ms.

[0309] (3) Data transmission:

[0310] Fiber optic Ethernet transmission (1Gbps bandwidth) is used, and the data encapsulation format is JSON, including device ID, timestamp and checksum (CRC32). The data processing center has a cache capacity of ≥10GB to ensure that the original data is not lost.

[0311] Step 3: Data Preprocessing

[0312] (1) Time synchronization (DTW + Kalman filtering), corresponding to the time synchronization (DTW + Kalman filtering) formula of the preprocessing and feature fusion modules

[0313] Time offset of radar and camera data Perform nonlinear correction:

[0314] 1. Using the Dynamic Time Warping (DTW) algorithm, the time series of the same vehicle trajectory identified by radar and camera are compared to calculate the nonlinear offset. (e.g., radar hysteresis 45ms ± 10ms);

[0315] 2. Kalman filter smoothing offset:

[0316] ;

[0317] Among them, the state equation ( (Gaussian white noise), observation equation (v_k represents observation noise). After synchronization, the time deviation of multi-source data is ≤3ms.

[0318] (2) Spatial coordinate transformation, including dynamic calibration, corresponding to the spatial coordinate transformation formula of the preprocessing and feature fusion modules:

[0319] Transform to the local road coordinate system (origin at the starting point of the merging zone, x-axis for the direction of travel, y-axis for the lateral direction, z-axis for elevation):

[0320] ;

[0321] 1. UTM projection: Projecting WGS84 latitude and longitude ( Convert to UTM coordinates with parameters: major semi-axis a = 6378137m, first eccentricity e² = 0.00669438.

[0322] 2. Dynamic calibration vector Based on the historical trajectory error over the past 3 months (optimized iteratively using the EM algorithm), the value is... =[0.8m, -0.5m]( Direction correction);

[0323] 3. Rotation matrix R (sensor mounting angle 3°): Translation vector

[0324] After conversion, the positioning error is ≤0.8m.

[0325] (3) Noise filtering formula for millimeter-wave radar in noise filtering, preprocessing and feature fusion modules:

[0326] Millimeter-wave radar velocity filtering (adaptive) ):

[0327] ;

[0328] in , , During rapid acceleration ( =15km / h), Increased to 10.04 for faster tracking; at constant speed ( =2km / h), =0.746 to smooth noise, speed error ≤2km / h.

[0329] LiDAR point cloud filtering: statistical filtering (k=16 nearest neighbors, remove ±3σ outliers), retain vehicle point cloud (approximately 8000 points / frame), ground point removal rate ≥99%.

[0330] Step 4: Feature Fusion and State Construction

[0331] (1) Single-modal feature extraction

[0332] Camera images: ResNet-18 extracts features, output (256 lanes, 50×125 grid corresponding to a 500m road segment).

[0333] LiDAR point cloud: Feature extraction using PointNet++, output .

[0334] Radar data: 3-layer MLP converted into velocity heatmap, output .

[0335] Concatenate into a multi-scale feature matrix .

[0336] (2) Spatiotemporal cross-attention fusion (including time decay and freshness weighting), corresponding to the spatial weight matrix (time decay) formula of the preprocessing and feature fusion modules and the channel weight vector (feature freshness) formula of the preprocessing and feature fusion modules.

[0337] Spatial weight matrix (time decay):

[0338] ;

[0339] in , For the feature generation time difference (e.g., feature weight decays to 0.787 3 seconds ago), the merging zone mesh is calculated. .

[0340] Channel weight vector (feature freshness):

[0341] ;

[0342] in , (Speed ​​channel 0.9, noise channel 0.2), the weight of the speed channel is increased by 30%-50%.

[0343] Fusion Feature Map: After convolution adjustment The signal-to-noise ratio is improved by 40%.

[0344] (3) State vector construction:

[0345] Lane occupancy, vehicle size weighted, corresponding to the lane occupancy formula in the state vector construction module:

[0346] ;

[0347] Effective physical length of the monitored road segment (Corresponding to the effective coverage area of ​​a single-node radar sensor). At the current statistical moment, the [number]th [node / location]... A total of 1,000 lanes were identified Vehicles. Vehicle type distribution and parameters are as follows: Large vehicles (4 vehicles) physical length Set size weight coefficient (Simulating the braking distance requirements of large vehicles and the psychological pressure on following vehicles), calculate the weighted length of a single vehicle as follows: Physical length of small cars (8 vehicles) Set size weighting coefficients. Calculate the weighted length of a single vehicle as follows: Set the necessary safety clearance constant between the front and rear of the vehicle. The calculated space occupancy rate is as follows: This accurately reflects the actual spatial saturation of the road section after weighting by vehicle volume. Compared to the unweighted proportion of pure physical length (i.e., This indicator more accurately reflects congestion and avoids the risk of underestimating road load due to the presence of large vehicles.

[0348] The multi-dimensional comprehensive risk indicator, and the corresponding formula for the multi-dimensional comprehensive risk indicator in the state vector construction module:

[0349] ;

[0350] Parameters α=0.4, β=0.3, γ=0.3; =3.2s, =1.5 (large vehicle)

[0351] ; =0.6 (light rain) =1.2 (merging lanes); =8 times / minute, calculated after normalization. =0.64.

[0352] - State vector: (9 dimensions, including 3-lane speed, occupancy rate, risk, merging frequency, and...) ).

[0353] Step 5: Decision Module (Improved TD3, namely the SHAP-VIF dual-criteria pruning formula for the decision module, the Actor network output formula for the decision module, and the reward function formula for the decision module):

[0354] (1) SHAP-VIF dual-criteria pruning

[0355] Pruning formula:

[0356] right Nine-dimensional feature screening: excluding those with VIF=6.2. (SHAP=0.02), retaining 8-dimensional features .

[0357] (2) Actor network output (lane coordination constraint)

[0358] ;

[0359] Assume the suggested speed for this lane after the Actor network output is mapped. (Right now The actual average speed of adjacent lanes (such as merging ramps) Coefficient of coordination (This indicates sensitivity to velocity gradients). Calculation results:

[0360] ;

[0361] By introducing the speed of adjacent lanes The weighting, the final instruction from Downgraded to This smooths out speed differences between lanes and reduces the risk of side collisions.

[0362] (3) Critic network reward function (smoothness constraint)

[0363] ;

[0364] in vehicles / minute =-0.5, =8km / h, calculated as follows =2.55.

[0365] Step 6: Issuance and Execution of Instructions:

[0366] (1) Command format: speed limit value (number) + lane marking, V2X message conforms to ETSI TS 103 301 standard.

[0367] (2) Delivery delay: sign update ≤ 50ms, V2X broadcast coverage ≥ 800m, vehicle OBU reception success rate ≥ 98%.

[0368] Step 7: Closed-loop optimization (multi-objective priority sampling), the multi-objective priority sampling formula for the corresponding closed-loop optimization module is:

[0369] (1) Sampling formula: ,in The Euclidean distance in the feature space.

[0370] (2) The sample pool contains 100,000 data points, and the sampling probability of low-frequency scenarios (merging conflicts) is significantly increased. The model is updated every 5 minutes.

[0371] Compared with traditional methods, the efficiency of merging zone traffic is significantly improved, the accident rate is significantly reduced, and the decision delay meets the real-time requirements.

[0372] This implementation scheme strictly corresponds to the technical features of the claims, and achieves dynamic speed limit control through multimodal fusion and improved TD3, which is suitable for bottleneck sections of highways.

[0373] 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 modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A dynamic variable speed limiting control system based on multimodal perception and deep reinforcement learning, characterized in that, include: The data acquisition module is used to acquire multi-source heterogeneous traffic data from Beidou high-precision positioning, millimeter-wave radar, radar vision sensors, and smart cameras; The preprocessing and feature fusion module, connected to the data acquisition module, is used for spatiotemporal alignment and feature-level fusion of multi-source heterogeneous data. This module is equipped with a time synchronization mechanism based on dynamic time warping and adaptive Kalman filtering, which calculates nonlinear time offsets and uses the Kalman state update equation for smooth correction, eliminating sampling time differences between sensors. The module also features a spatiotemporal cross-attention mechanism, introducing a time decay coefficient when generating the fused feature map. Historical features are downweighted, and a feature freshness index is introduced. The channel weights are dynamically weighted to output a fused feature map that includes spatiotemporal correlation characteristics; The state vector construction module constructs a state space vector for reinforcement learning based on the fused feature map; this module introduces a vehicle size coefficient when calculating lane occupancy. The physical lengths of different vehicle models are weighted and adjusted; simultaneously, a multi-dimensional comprehensive risk index is constructed, including vehicle model risk coefficient, environmental risk coefficient, and lane-inherent risk coefficient. This is used as part of the state vector to quantify the micro and macro risks of traffic flow; The decision module uses an improved Deep Deterministic Policy Gradient (TD3) algorithm to generate lane-level variable speed limit instructions. At the input of this algorithm, a SHAP-VIF dual-criteria pruning mechanism is set to filter redundant features using the variance inflation factor (VIF) and low-contribution features using the SHAP value. At the output of the Actor network, a lane coordination constraint term is introduced. By calculating the weighted difference between the average speed of the current lane and the adjacent lanes, the original network output is smoothly corrected to generate the final speed limit instruction. ; The instruction issuing module issues instructions via variable speed limit signs and V2X communication. The closed-loop optimization module stores "state-action-reward-new state" samples and updates model parameters. This module employs a multi-objective priority sampling mechanism to calculate sample priority. At the same time, the absolute value of the time difference error is fused. With sample diversity indicators This is to increase the sampling probability of sparse samples and achieve real-time closed-loop optimization of the model.

2. The dynamic variable speed limiting control system based on multimodal perception and deep reinforcement learning according to claim 1, characterized in that, The time synchronization formula for the preprocessing and feature fusion modules is as follows: ； in: - The revised version of the first... Time stamp; - Sensor Original Number Time stamp; - : The first calculated by DTW Time offset at any moment; - The previous moment, i.e. Kalman estimates of time offsets; - The estimated time offset, obtained from the Kalman filter state update equation, is used to smooth the DTW calculation results. The spatial coordinate transformation formula for the preprocessing and feature fusion module is: in: - : Coordinates transformed to the local coordinate system of the road; - Rotation matrix; - : UTM coordinates of road segment reference points, used to eliminate the accuracy loss caused by large coordinate values; - : UTM plane coordinates after Gauss-Kruger projection transformation, where Longitude Latitude; - : Dynamic calibration vector; - Translation vector; The millimeter-wave radar noise filtering formula for the preprocessing and feature fusion module is as follows: ； in: - : The process noise variance matrix at time k; - Initial process noise variance matrix; - Adjustment factor, ranging from 0.1 to 0.3; - : The absolute value of the velocity difference between time k and time k−1, wherein the velocity difference is normalized before being included in the calculation.

3. A dynamic variable speed limiting control system based on multimodal perception and deep reinforcement learning according to claim 2, characterized in that, The formula for the spatial weight matrix of the preprocessing and feature fusion modules is: ； in: - Spatial weight matrix; - : Sigmoid activation function; - : 7×7 convolution operation is used to extract spatial neighborhood association features; - Average pooling along the channel dimension transforms the feature map from... Compress to Preserve spatial structure information; - Multi-scale feature matrix; - Time decay coefficient, ranging from 0.05 to 0.1; - The difference between the feature generation time and the current time; The channel weight vector formula for the preprocessing and feature fusion module is as follows: ； in: - Channel weight vector; - : Sigmoid activation function; - Multilayer perceptron operation; - Global max pooling operation; - Multi-scale feature matrix; - The weighting coefficient is 1~2; - : Freshness of the c-th feature channel, ranging from 0 to 1.

4. A dynamic variable speed limiting control system based on multimodal perception and deep reinforcement learning according to claim 3, characterized in that, The lane occupancy formula of the state vector construction module is: ； in: - : No. Lane weighted time occupancy rate, in % - : No. The actual physical length of the vehicle, in meters (m). - : Size coefficient of the k-th vehicle, with a value of [value] for large vehicles. The value for small cars is This is used to correct the actual impact of large vehicles on road resources; - : No. The total number of vehicles in the lane; - The preset safety clearance constant means that the vehicle occupies not only the body of the vehicle, but also the necessary front and rear clearance. - It is used to normalize the space occupied by vehicles in the lane, and its value is determined according to the effective coverage of the sensor or the preset statistical interval.

5. A dynamic variable speed limiting control system based on multimodal perception and deep reinforcement learning according to claim 4, characterized in that, The formula for the multi-dimensional comprehensive risk index of the state vector construction module is as follows: ； in: - : The comprehensive traffic risk index for each time segment is a dimensionless value; the larger the value, the higher the risk. - The weighting coefficients of each risk component satisfy the following conditions: ; - The minimum collision time within the monitored road segment is calculated using the following formula: , This represents the relative distance between the vehicle in front and the vehicle behind. The relative speed between the vehicle in front and the vehicle behind; - : Minimum constant, used to prevent the denominator from being 0 when the relative velocity is extremely large or the distance is extremely small; - Vehicle risk factor, used to amplify the potential dangers of large vehicles under low TTC conditions; - Environmental risk coefficient; - Lane inherent risk coefficient; - The frequency of vehicle merging and lane changing per unit time, measured in times per minute, is used to characterize the severity of traffic flow conflicts.

6. A dynamic variable speed limiting control system based on multimodal perception and deep reinforcement learning according to claim 4, characterized in that, The SHAP-VIF dual-criteria pruning formula for the decision module is as follows: ； in: - : Binary mask vector, 1 indicates that the first digit is retained One feature, where 0 indicates pruning; - : No. Variance inflation factor of each feature to filter out highly redundant features; - VIF threshold; - : No. The SHAP value of each feature is used to filter out features that are meaningless for decision-making. - SHAP value threshold, which is 0.01~0.

05.

7. A dynamic variable speed limiting control system based on multimodal perception and deep reinforcement learning according to claim 4, characterized in that, The Actor network output formula of the decision module is: ； in: - ； - Scaling and offset factors; - The hyperbolic tangent activation function maps the unbounded raw output of the neural network to an open interval (-1, 1) to ensure the numerical stability of the control quantity. - : Actor network output layer The original values ​​of each neuron, whose range is the real number field. ; - Lane coordination coefficient, ranging from 0.1 to 0.2, determines the system's sensitivity to speed differences between lanes; - : No. The current average detected vehicle speed in all adjacent lanes of the lane.

8. A dynamic variable speed limiting control system based on multimodal perception and deep reinforcement learning according to claim 4, characterized in that, The reward function formula for the decision-making module is: ； in: - The reward value at time t; - Weighting coefficients; - Average speed increase, which is the difference between the current vehicle speed and the historical average vehicle speed; - Average flow increase, which is the difference between the current flow and the historical average flow; - The change in risk indicators represents the difference between current risk and historical risk. - Standard deviation of vehicle speed within the lane; The historical average value is a sliding average or an exponentially weighted average value based on a preset time window, used to smooth the impact of short-term traffic fluctuations on reward calculation.

9. A dynamic variable speed limiting control system based on multimodal perception and deep reinforcement learning according to claim 4, characterized in that, The multi-objective priority sampling formula of the closed-loop optimization module is as follows: ； in: - : No. Sampling priority of each sample; - : No. The absolute value of the time difference error of each sample represents the learning value of the sample; - The minimum constant ensures that samples with zero error still have a non-zero probability of being sampled. - Value sensitivity index, which is Adjusting the nonlinear effect of the error term on priority; - : Diversity weighting coefficient, for This is used to increase the probability of sparse samples being selected; - : Sample diversity index, calculated by the first The distance between the state vector of a sample and the feature space of the existing cluster centers in the experience pool is used to quantify the "uniqueness" of the sample.

10. A dynamic variable speed limiting control method based on multimodal perception and deep reinforcement learning, characterized in that, The method is applied to the system according to any one of claims 1-9, and the method includes the following steps: S1. Collect multimodal traffic data; S2. Execution time synchronization, spatial coordinate transformation, and noise suppression; S3. Generate a fused feature map using spatiotemporal cross-attention; S4. Construct the state space vector; S5. The improved TD3 algorithm with SHAP-VIF pruning is used to generate lane-level speed limit instructions, which are then issued through the instruction distribution module. S6. Network parameters are updated based on multi-target priority sampling and adaptive triggering mechanism to achieve real-time closed-loop optimization.

Images