A Traffic Signal Optimization Method Based on Model Predictive Control and Trust Region Bayesian Optimization

Abstract

This invention belongs to the field of traffic signal control and relates to a traffic signal optimization method based on model predictive control and trust-region Bayesian optimization. First, based on intersection video data, this invention uses target detection and tracking algorithms to automatically collect traffic flow data for different directions. Then, based on three types of heterogeneous data—radar data, video data, and map data—it designs methods for estimating road segment traffic flow and intersection turning ratios based on traffic flow theory. Finally, based on the traffic state reconstruction results, it establishes a microscopic traffic simulation model and designs a combined framework for traffic signal optimization based on centralized model predictive control and trust-region Bayesian optimization. According to the dynamic changes in traffic flow, it solves for the real-time optimal regional traffic signal timing scheme. This invention can effectively reduce average vehicle delay in the network and has high efficiency and scalability in solving large-scale signal optimization problems.

Description

Technical Field

[0001] This invention belongs to the field of traffic signal control and relates to traffic data acquisition based on computer vision, traffic state reconstruction based on multi-source heterogeneous data, and regional traffic signal optimization based on model predictive control. Specifically, it is a traffic signal optimization method based on model predictive control and trust region Bayesian optimization. Background Technology

[0002] Traffic congestion is becoming increasingly serious, significantly reducing the travel efficiency of urban residents. Traffic signal control, as an effective traffic management and control method, has become an important means to solve traffic congestion problems. Traditional traffic signal control methods can be divided into three categories: The first category is fixed-time control methods, which artificially set the duration of traffic signals by summarizing historical traffic operation patterns, such as the Webster method; the second category is inductive control methods, where traffic signals switch based on artificially set rules and specific conditions, such as the self-organizing traffic signal control method proposed by Cools et al. in "Self-organizing traffic lights: A realistic simulation," which switches the light color according to the number of arriving vehicles; the third category is adaptive control methods, among which the maximum pressure method is a representative method. This method selects the signal phase with the highest traffic pressure each time. For example, Varaiya designed a maximum pressure controller for arbitrary traffic networks in "The max-pressure controller for arbitrary networks of signalized intersections," which selects the green light duration for each phase based on the queuing at each intersection in the network. In recent years, the development of machine learning and intelligent transportation systems has provided new ideas for the research of traffic signal control optimization. These methods, by deploying sensors, cameras and wireless sensor networks, can acquire information such as traffic flow, vehicle speed and road conditions in real time, and adaptively optimize traffic signals to better cope with the ever-changing traffic system environment. For example, in "Research and Implementation of Multi-Intersection Traffic Signal Control Strategy Based on Deep Reinforcement Learning", Zhang Haopeng proposed a real-time traffic signal control method based on near-end policy optimization and multi-agent reinforcement learning.

[0003] The patent "A Method and System for Traffic Signal Control at Regional Boundaries Based on Model Predictive Control" (CN119723912A) proposes to partition the road network based on a macroscopic basic map and use model predictive control to optimize the signal timing at regional boundaries to regulate cross-regional traffic flow. However, this method relies on the assumption of uniform and homogeneous traffic flow and accurate MFD model fitting, making it insufficiently adaptable to situations where detectors are sparse and data is heterogeneous in actual road networks. Furthermore, its control granularity is relatively coarse, mainly focusing on traffic flow balance between regions, lacking fine-grained estimation and signal optimization of turning traffic at each intersection within the region, thus limiting its ability to alleviate local congestion. The patent "A Traffic Signal Model Predictive Control Method Based on Uncertain Parameter Prediction" (CN117334048A) proposes to use an improved Louvain algorithm for road network partitioning to reduce computational complexity and to construct an uncertain parameter prediction model that combines historical data and online prediction, aiming to improve the accuracy of traffic disturbance response. However, this method is highly dependent on the integrity and quality of historical data. Its complex time series model is difficult to initialize and update when the data is sparse or missing. Moreover, the optimization is mainly focused on signal timing, which does not fully solve the problem of incomplete perception caused by uneven detector distribution, nor does it form a closed loop with automated front-end data acquisition and multi-source data fusion. Therefore, its feasibility and data reliability in real sparse data environments face challenges.

[0004] In summary, existing traffic signal control optimization methods suffer from the following main problems: First, traditional model-driven signal control methods cannot adapt well to dynamically changing traffic environments, exhibiting insufficient model generalization ability and poor adaptability. Second, machine learning-based signal control methods rely on high-quality traffic data, failing to address the data sparsity caused by uneven detector distribution in real-world road networks. Furthermore, when solving large-scale traffic signal optimization problems, the solution efficiency gradually decreases as the road network size increases. Finally, most existing traffic signal optimization studies use simulated randomly generated traffic flow data to verify the effectiveness of signal optimization methods, neglecting practical issues such as the difficulty of regional traffic data collection and the sparse nature of the collected traffic data, thus overlooking the feasibility of the methods. Summary of the Invention

[0005] To address the problems of existing methods, this invention proposes a traffic signal optimization method based on model predictive control and trust-region Bayesian optimization. Based on intersection video data, target detection and tracking algorithms are used to automatically collect traffic flow data for different directions. Based on three types of heterogeneous data—radar data, video data, and Gaode Map data—methods for estimating road segment traffic flow and intersection turning ratios based on traffic flow theory are designed. A microscopic traffic simulation model is established based on traffic state reconstruction results, and a combined framework for traffic signal optimization based on centralized model predictive control and trust-region Bayesian optimization is designed. According to the dynamic changes in traffic flow, the real-time optimal regional traffic signal timing scheme is solved.

[0006] The technical solution of the present invention:

[0007] A traffic signal optimization method based on model predictive control and trust-region Bayesian optimization comprises three steps: traffic data acquisition, traffic state reconstruction, and traffic signal optimization. The traffic data acquisition section uses target detection algorithms (such as YOLO and SSD) for vehicle detection and target tracking algorithms (such as DeepSORT and FairMOT) for vehicle tracking, extracting traffic flow at different turns at intersections from video feeds. The traffic state reconstruction section uses partial radar data, video data, and map data of the study area to construct a macroscopic traffic flow model, estimating missing road segment traffic flow and intersection turning ratios to reconstruct the overall traffic flow operation state of the study area. The traffic signal optimization section uses microscopic traffic simulation software (such as VISSIM and SUMO) to establish a traffic simulation model of the study area based on the reconstructed traffic flow operation state, uses macroscopic traffic flow models (such as store-and-forward models and cellular automata models) to predict the traffic flow operation state, and combines centralized model predictive control to construct a signal optimization model. The model is then solved using a trust-region Bayesian optimization algorithm to obtain the optimal signal control scheme.

[0008] The specific steps are as follows:

[0009] Step 1: Traffic Data Collection

[0010] Traffic data collection mainly consists of three parts: data preparation, vehicle identification based on target tracking algorithms, and vehicle tracking based on target tracking algorithms.

[0011] Step 1.1, Data Preparation:

[0012] Acquire video data of monitored intersection entrances in the study area; perform data augmentation and normalization on the intersection images, etc.

[0013] Step 1.2, Vehicle detection based on object detection algorithm:

[0014] The target detection algorithm is executed to detect and label all vehicles in each frame of the image.

[0015] Step 1.3, Vehicle tracking based on target tracking algorithm:

[0016] First, based on the vehicle detection results in step 1.2, the target tracking algorithm is executed to extract and track motion features, output the tracked vehicle motion trajectory, and obtain the traffic flow at different directions at the intersection.

[0017] Step 2, Traffic Status Reconstruction:

[0018] Traffic condition reconstruction mainly consists of three parts: data preparation, traffic flow estimation, and intersection turning ratio estimation.

[0019] Step 2.1, Data Preparation:

[0020] First, based on step 1, traffic flow data for different turns at monitored intersections in the study area is obtained. Second, radar data for radar-equipped road sections in the study area is acquired and processed to obtain road section traffic flow data. Finally, a map is used to obtain built environment data (including the number of lanes and traffic hotspots), average speed data, and queue length data for all road sections in the study area.

[0021] Step 2.2, Traffic Flow Estimation for Road Sections:

[0022] Based on radar data of some road segments in the study area, average speed data of all road segments, queue length data of all intersections, built environment data, and intersection queue length data, combined with a basic traffic graph model, a road segment traffic estimation method is proposed to reconstruct the road segment traffic information of the entire network. The specific steps are as follows:

[0023] Step 2.2.1, Road Segment Classification. Based on the built environment data, clustering algorithms (such as K-means, SVM) are used to classify all road segments in the study area.

[0024] Step 2.2.2, Flow-velocity model construction. For each road segment category, flow-velocity models for Greenshields, Greenberg, and Underwood are established respectively. The specific calculation formulas are shown in (1)-(3):

[0025] (1)

[0026] (2)

[0027] (3)

[0028] in, Indicates road segment r Traffic; Indicates road segment r The blocking density; Indicates road segment r The average speed; Indicates road segment r Critical density; Indicates road segment r The free-flow velocity is determined based on the road grade.

[0029] Step 2.2.3: Dataset Partitioning. The average speed data and traffic flow data of the road segments are divided into two parts: one part is used as the training set for model parameter calibration; the other part is used as the validation set for model accuracy verification.

[0030] Step 2.2.4, Model Parameter Calibration. This involves calibrating the traffic data from the training set. and average speed Substitute the parameters into three flow-velocity models and then adjust the parameters in each model. , Perform calibration.

[0031] Step 2.2.5, Model Selection. For each road segment category, the average speed in the validation set will be selected. Substitute the calibrated flow-velocity model into the data to calculate the corresponding flow rate. Compare the calculated results with the actual flow rate in the validation set, calculate the error of each of the three models, and select the model with the smallest error for flow rate estimation of this road segment category.

[0032] Step 2.2.5: Estimate the flow rate of the remaining road segments in the study area. For the remaining road segments in the study area, select the model of the corresponding category based on the classification results, substitute its average speed data into the flow-speed model, and estimate the flow rate.

[0033] Step 2.3, Intersection turning ratio estimation:

[0034] Considering the various lane types at intersections, this paper proposes a complete method for estimating intersection turning ratios using video and radar data from some intersections, as well as signal timing and queue length data from all intersections. The specific steps are as follows:

[0035] Step 2.3.1: Determine the values ​​of the model parameters.

[0036] First, based on the traffic flow at different directions at the intersection obtained in step 1, determine the left-turn ratio of the shared straight and left-turn lanes. Right turn ratio of shared straight and right lanes The import channel to be estimated , The value can be selected from channels with similar channelization types, such as video inlet channels. , The values ​​are then determined through traffic surveys. The remaining parameters include green light loss time, headway for each lane, and phase green light duration.

[0037] Step 2.3.2: Estimate lane capacity. Based on different lane types, estimate lane capacity using the following formula:

[0038] Step 2.3.2.1, Formula for calculating vehicles passing through the straight lane during the green light period:

[0039] When setting up a dedicated signal for straight-through traffic:

[0040] (4)

[0041] in, This indicates vehicles passing through the straight-ahead lane during the green light period when a dedicated straight-ahead signal is in place. This indicates the duration of the green light for the straight-ahead phase of that approach lane; This indicates that the lost time during a green light is mainly due to the time lost due to vehicle starting. Indicates the headway of a straight-moving vehicle.

[0042] When no dedicated signal for straight-through traffic is set up:

[0043] (5)

[0044] in, This indicates vehicles passing through the straight-ahead lane during the green light period when no dedicated straight-ahead signal is provided. This represents the straight-ahead traffic flow reduction factor. Due to interference from left-turning traffic, the straight-ahead traffic flow will be less than the flow when a dedicated straight-ahead signal is set up.

[0045] Step 2.3.2.2, Formula for calculating vehicles passing through the left-turn lane during the green light period:

[0046] When setting up a dedicated left-turn signal:

[0047] (6)

[0048] in, This indicates the vehicles passing through the left-turn lane during the green light period when a dedicated left-turn signal is set up; This indicates the duration of the green light for the left-turn phase at that entrance lane; This indicates the distance between the front of the car when turning left.

[0049] When no dedicated left-turn signal is set up:

[0050] (7)

[0051] in, This indicates vehicles passing through the left-turn lane during the green light period when there is no dedicated left-turn signal; This represents the left-turn traffic flow reduction factor. Due to interference from through traffic, the left-turn traffic flow will be less than the flow flow when a dedicated left-turn signal is set up.

[0052] Step 2.3.2.3, Formula for calculating vehicles passing through the right-turn lane during the green light period:

[0053] (8)

[0054] in, This indicates vehicles passing through the right-turn lane during the green light period; This indicates the duration of the green light for the right turn phase at that entrance lane; This indicates the distance to the front of the car when turning right; other parameters are the same as before.

[0055] Step 2.3.2.4: Vehicles passing through the shared straight / left lane during the green light period. Calculation formula:

[0056] When the left turn ratio When the time is right, the calculation formula is:

[0057] (9)

[0058] When the left turn ratio When the time is right, the calculation formula is:

[0059] (10)

[0060] When the left turn ratio When the time is right, the calculation formula is:

[0061] (11)

[0062] Therefore, it can be concluded that left-turning vehicles passing through the combined straight and left-turn lanes... Calculation formula:

[0063] (12)

[0064] Formula for calculating the number of vehicles going straight through a shared straight-left lane:

[0065] (13)

[0066] in, This indicates vehicles proceeding straight through the shared straight / left-hand lane during the green light period; This indicates the duration of the green light for the straight-ahead and left-turn phases at this approach lane; Indicates the headway of vehicles in the combined straight and left lane; This indicates the proportion of left-turning traffic in a shared straight-left lane; This represents the reduction factor.

[0067] Step 2.3.2.5: Vehicles passing through the shared straight / right lane during the green light period. Calculation formula:

[0068] When the right turn ratio When the time is right, the calculation formula is:

[0069] (14)

[0070] When the right turn ratio When the time is right, the calculation formula is:

[0071] (15)

[0072] When the right turn ratio When the time is right, the calculation formula is:

[0073] (16)

[0074] Therefore, the formula for calculating the number of straight-going vehicles passing through the shared straight-and-right lane is as follows:

[0075] (17)

[0076] in, This indicates vehicles going straight through the shared straight / right lane during the green light period; This indicates the duration of the green light for the straight-ahead and right-turn phases at this approach lane; Indicates the headway of vehicles in the shared straight-and-right lane; This indicates the proportion of turning traffic in a shared straight-to-right lane; This represents the reduction factor.

[0077] Step 2.3.3: Estimate the number of vehicles in the queue. Based on the queue length data on the map, calculate the number of vehicles queuing at each turn of the entrance lane. The calculation formula is as follows:

[0078] (18)

[0079] in, This indicates the queue of vehicles turning at each direction on the entrance lane; Indicates the queue length; Indicates the distance between vehicles; This indicates the length of a standard car.

[0080] Step 2.3.4: Estimate the flow rate for a single signal cycle. Based on the capacity of each turn. Combined with the number of vehicles in the queue The flow rate for each turning point in a single cycle is calculated using the following formula:

[0081] when The flow rate calculation formula is as follows:

[0082] (19)

[0083] in, These represent left turn, straight ahead, and right turn, respectively. This represents the flow rate of a single signal cycle; This is the reduction factor.

[0084] when The flow rate calculation formula is as follows:

[0085] (20)

[0086] Step 2.3.5: Calculate the total flow rate during the study period. Based on the single-cycle flow rate, calculate the total flow rate at each turning point during the study period. The calculation formula is as follows:

[0087] (twenty one)

[0088] in, Indicates the total flow during the research period; Indicates the total time of the research period; This indicates the signal cycle duration at the intersection where the approach lane is located.

[0089] Step 2.3.6: Calculate the turning ratio of each approach lane. Calculate the turning ratio of each approach lane at the intersection, that is, the ratio of left-turn, straight-through, and right-turn traffic flows.

[0090] Step 3, Traffic Signal Optimization:

[0091] Based on the traffic state data obtained in step 2, a regional traffic signal optimization model is constructed using model predictive control methods, and the optimal signal control scheme is obtained by solving the model using the trust region Bayesian optimization algorithm. The specific steps are as follows:

[0092] Step 3.1: Construct a traffic simulation model:

[0093] Based on the signal timing scheme of the study area and the traffic state data obtained in step 2, a traffic simulation model of the study area is constructed using microscopic traffic simulation software to simulate the actual traffic flow process.

[0094] Step 3.2: Construct a traffic flow prediction model:

[0095] The regional traffic flow operation status is simulated based on the macro traffic flow model in order to predict the traffic flow status of road segments after the current signal timing scheme is implemented.

[0096] Step 3.3: Construct a regional traffic signal optimization model:

[0097] Assume the set of road segments is I The set of intersections is J intersection j The phase set is The cycle duration is The current control time step of the model predictive controller is k The time step is Control time domain is NAmong them, the control time step index time step index within the prediction time domain Belonging to different time scales: This represents the sequence points where the optimization decision is actually executed, while For use in each When constructing the optimization model, it refers to the predicted future step. Time period ( ). Road section i Controlling time steps k The traffic is intersection j phase p The green light time is The minimum green light time is The maximum green light time is The green light interval is The cycle duration is The control input (i.e., the decision variable) is the green light duration of all phases at each time step within the control time domain. In each control step, define and solve the following problem:

[0098] (twenty two)

[0099] (twenty three)

[0100] In formula (23) Let represent the objective function, which aims to minimize the sum of remaining traffic flow in the road segments at all time steps within the control time domain. Equation (23) represents the constraints. Given the known initial network state constraints, each control time step is obtained from the traffic simulation model constructed in step 3.1; intersection phase The green light time is The green light interval is ; , These are the upper and lower bound constraints for the decision variables, namely the minimum green light duration and the maximum green light duration; The cycle duration constraint means that the sum of the green light duration and the green light interval time for each phase at each intersection is equal to the cycle duration of that intersection.

[0101] Step 3.4: Solve the regional signal optimization model:

[0102] Trust-region Bayesian optimization is a highly efficient high-dimensional black-box optimization algorithm suitable for complex optimization problems, especially addressing the computational challenges faced in regional traffic signal optimization. This method utilizes historical data to construct a local surrogate model of the objective function, effectively reducing the dimensionality of decision variables without relying on gradient information, thereby improving optimization efficiency. The core of this algorithm includes the trust region, local Gaussian process regression, and Thompson sampling, which work together to determine the optimal trust region for each iteration. By dynamically adjusting the trust region, this method can intelligently select the sampling point with the greatest benefit, thus avoiding direct computation of the computationally complex objective function and significantly reducing computational costs. Therefore, this invention uses the trust-region Bayesian optimization algorithm to solve the regional signal optimization model constructed in step 3.3 to obtain the current control time step. k The optimal green light duration for each phase is determined by the following steps:

[0103] Step 3.4.1: Construct the prior dataset The Latin hypercube sampling method was used to randomly select... n Group signal timing scheme as initial sample Formula (22) is used to evaluate the initial sampling points and obtain the corresponding objective function values. and the initial prior dataset The initial training sample dataset is .

[0104] Step 3.4.2: Set algorithm parameters. The initial trust region's base radius length is... The maximum radius of the trust region is The minimum radius is Optimize the variable dimension to The batch sample size is The trust domain update success counter is The failure counter is The maximum number of iterations is .

[0105] Step 3.4.3: Initialize the trust domain space The initial trust region's base radius length is... The center point of the initial trust region is ,in Let the number of iterations be... .

[0106] Step 3.4.4: Fit the objective function using local GP. Based on the sample dataset. Local generalized physics (GP) is used to fit the nonlinear relationship between green light duration and the sum of traffic flow in the predicted time domain, thus obtaining the next batch of sampling points. The posterior probability distribution.

[0107] Steps 3.4.5: In the trust domain Thompson sampling is performed. First, in the trust region space of the current iteration... In, generate a size of random samples Then, through Repeat the screening to select the next batch q A number of candidate points (i.e., candidate schemes for predicting the green light duration at each time step in the time domain) are used to predict the green light duration at each time step in the time domain. ,in, , ; Indicates the first The objective function values ​​corresponding to the candidate green light durations follow a Gaussian process. ) distribution, in which For the current control time step The One green light duration candidate scheme It is local The fitted average predicted value, It is the uncertainty of prediction.

[0108] Step 3.4.6: Evaluate candidate points. Based on the traffic flow prediction model and objective function formula (22) in Step 3.2, calculate the corresponding objective function value to obtain the next batch of sampling dataset. .

[0109] Step 3.4.6: Determine if the iteration has terminated. If... If the condition is met, proceed to step 3.4.9; otherwise, continue to the next step.

[0110] Step 3.4.7, Trust Region Update. First, determine if a better solution can be found in the current iteration. If so, let... , ,in, Record the number of times a better solution is found consecutively. Record the number of consecutive times no better solution is found; otherwise, let , Then, determine whether the update condition for the trust region radius is met. Then let , ;if Then let , ,in, The threshold for the number of successful attempts. Define the search range based on the current base radius of the trust region. The maximum base radius of the trust region is used. Finally, the trust region is updated. That is, the range of candidate points for the optimal green light duration. It is the central location of the trust domain, in which, , For the green light duration scheme of the trust domain center point, This is the initial green light duration sample set. This serves as a candidate set for the next batch of green light durations.

[0111] Step 3.4.8: Update the dataset with the next sample set, let Return to step 3.4.4.

[0112] Step 3.4.9, obtain Optimal green light duration for each time step Choose the first k The result of the step is used as the optimal signal timing scheme for the current control time step.

[0113] Step 3.5: Determine whether to terminate the optimization. For the maximum control time step, if the current control time step If the signal optimization fails, terminate the signal optimization; otherwise, update the initial state of the road network and proceed to the next time step, letting... Return to step 3.1 and input the signal timing scheme obtained in step 3.4 into the traffic simulation model in step 3.1.

[0114] Compared with the prior art, the present invention has the following advantages:

[0115] (1) A full-process traffic signal optimization method is proposed. It considers three stages: traffic data collection, traffic state reconstruction, and regional traffic signal optimization. It can realize full-process signal optimization based on real sparse traffic data, ensuring that the proposed method has good feasibility.

[0116] (2) A traffic state reconstruction method based on multi-source heterogeneous data is proposed. This method integrates three types of heterogeneous data: video data, radar data, and data from the Gaode Maps platform. Then, based on a macroscopic traffic flow model, it estimates the segmental traffic flow and intersection turning ratios of the road network. This method effectively utilizes the spatiotemporal complementarity of multi-source heterogeneous data to accurately reconstruct regional traffic flow states, providing strong data support for subsequent traffic signal optimization.

[0117] (3) A regional traffic signal optimization method based on centralized model predictive control is proposed. A combined framework of model predictive control and trust-region Bayesian optimization is designed. A regional traffic signal optimization model is established through a centralized model predictive control strategy, and the optimal signal control strategy for the region is determined by solving the model using the trust-region Bayesian optimization algorithm. This method extends the traditional centralized model predictive control method to a higher dimension and has good adaptability to large-scale traffic signal optimization problems. Attached Figure Description

[0118] Figure 1 This is a flowchart of the traffic signal optimization method based on model predictive control and trust region Bayesian optimization of the present invention;

[0119] Figure 2 This is a logical schematic diagram of the method of the present invention;

[0120] Figure 3 This is a flowchart of the trust region Bayesian optimization algorithm in this invention;

[0121] Figure 4 This is a diagram illustrating the iterative process of the trust-region Bayesian optimization algorithm in the embodiment. Detailed Implementation

[0122] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings and technical solutions.

[0123] The basic process of the traffic signal optimization method based on model predictive control and trust region Bayesian optimization of this invention is as follows: Figure 1 and Figure 2 As shown in the figure. This embodiment selects a portion of a city as the study area to verify the effectiveness of the full-process regional signal optimization method proposed in this invention. The main roads in this area from north to south are Chongyang Road, Wenyang Road, and Xingyang Road, and the main roads from west to east are Xiucheng Road, Heilongjiang North Road, Zhongchuan Road, and Changcheng Road, with a total of 31 signal-controlled intersections and more than 200 road segments.

[0124] The specific implementation steps are as follows:

[0125] Step 1, Traffic Data Collection:

[0126] First, video data of the four approach lanes (east, south, west, and north) of four cross-shaped intersections were acquired from 7:00 to 8:00 on December 26, 2023. Second, the video data of the four intersections was processed using object detection and tracking algorithms (YOLOv8 and DeepSORT algorithms were used in this embodiment) to extract the left-turn and straight-ahead traffic flow of each approach lane. It should be noted that right-turn traffic flow is not counted in this patent because traffic signal control in my country typically does not consider right turns. Finally, manual counting was performed and compared with the algorithmic statistics to verify the accuracy of the algorithm's data collection. Table 1 shows the algorithmic and manual statistics results for the traffic flow of two intersections over 10 minutes. As can be seen from the table, for left-turn traffic flow, the algorithmic statistics showed mostly 0% error, with a maximum error within 15%; for straight-ahead traffic flow, the algorithmic statistics showed mostly 5% error, with a maximum error within 15%. Therefore, the data collection method proposed in this invention has high accuracy and can effectively complete traffic data collection.

[0127] Table 1 Accuracy Verification of Traffic Flow Data Acquisition Algorithm

[0128]

[0129] Step 2, Traffic Status Reconstruction:

[0130] Step 2.1, Data Preparation. First, radar data, intersection signal timing scheme data, and intersection channelization data for 36 road segments (i.e., the east, south, west, and north entrance road segments of 9 cross-shaped intersections) from 7:00 to 8:00 on April 1-30, 2023 and December 26, 2023 were obtained. Then, the average speed data of all road segments and the queue length data of all intersections from 7:00 to 8:00 on December 26, 2023 were calculated using Gaode Maps. The built environment data (including the number of lanes and the number of traffic hotspots) of all road segments in the study area were also calculated.

[0131] Step 2.2, Road Segment Traffic Estimation. First, a clustering algorithm (K-means algorithm in this embodiment) is used to classify all road segments, dividing the entire road network into three categories. Second, Greenshields, Greenberg, and Underwood traffic-speed models are constructed using average speed data and radar traffic data of representative road segments from April 1st to April 30th, 2023, and the model parameters are calibrated. Then, the average speed data from 7:00 to 8:00 on December 26th, 2023, is substituted into the calibrated models to estimate the traffic flow of the corresponding road segments. The estimated traffic flow is then compared with the actual radar traffic flow during that period to analyze the errors of the three models. The model with the lowest error is selected as the traffic-speed model for that type of road segment. Table 2 shows the traffic estimation errors for some road segments. Among all road segment errors, the maximum error percentage is 9.03%, and the minimum is 0.04%. The error percentage of the estimation results is less than 10%. Finally, the average speed of the remaining road sections in the study area between 7:00 and 8:00 on December 26, 2023, was substituted into the corresponding classification model to complete the traffic flow estimation for the entire area.

[0132] Table 2. Traffic flow estimation errors for some road sections

[0133]

[0134] Step 2.2, Intersection Turning Ratio Estimation. First, determine the model parameter values ​​based on the different turning traffic volumes of all approach road segments of the four intersections obtained in Step 1. , The remaining parameters were obtained through traffic surveys, and their specific values ​​are shown in Table 3. Then, based on the channelization and signal timing scheme of the intersection approach road sections, the steering ratio was estimated using formulas (4)-(21). Finally, the actual steering ratio was calculated based on the radar traffic data of 36 road sections, and compared with the estimated steering ratio to verify the accuracy of the method. Table 4 shows the steering ratio estimation errors of some intersections. As can be seen from the table, the estimation error of the steering ratio of the approach road of the Wenyang-Zhongchuan intersection is mostly within 10%, with a maximum error of 10.920%; the estimation error of the steering ratio of the approach road of the Wenyang-Ningcheng intersection is also mostly within 10%, with a maximum error of -15.819%; the estimation error of the steering ratio of the approach road of the Wenyang-Mincheng intersection is slightly larger than that of the other two intersections, with a maximum error of -18.564%, but the overall error is also within 15%. It can be seen that the overall estimation error of the intersection steering ratio estimation method proposed in this invention is relatively small.

[0135] Table 3. Main parameter values ​​for intersection turning ratio estimation method

[0136]

[0137] Table 4. Estimation error of turning ratio at some intersections

[0138]

[0139] Note: "-" indicates missing radar data.

[0140] Step 3, Traffic Signal Optimization:

[0141] Step 3.1: Construct a traffic simulation model. Based on the signal timing scheme of the study area and the traffic state data obtained in Step 2, a traffic simulation model of the study area is constructed using microscopic traffic simulation software (SUMO is used in this embodiment) to simulate the actual traffic flow process.

[0142] Step 3.2: Constructing the traffic flow prediction model. Based on the intersection signal timing scheme and road network topology, this embodiment uses a store-and-forward model as the traffic flow prediction model for the study area.

[0143] Step 3.3: Construct a regional traffic signal optimization model. Based on the traffic flow prediction model in Step 3.2, use formulas (22)-(23) to establish an optimization model for the green light duration of all phases in the study area. The parameter values ​​in the model are shown in Table 5.

[0144] Table 5. Parameter settings for the regional signal optimization model

[0145]

[0146] Step 3.4, Solve the signal optimization model. Use the trust-region Bayesian optimization algorithm to solve the signal optimization model constructed in step 3.3, as follows: Figure 3 As shown, the current control time step is obtained. k The optimal green light duration for each phase and the algorithm parameter values ​​are shown in Table 6. Figure 4 The figure illustrates the algorithm optimization iteration process for some control time steps, where the horizontal axis represents the number of iterations and the vertical axis represents the objective function value, i.e., the sum of the remaining traffic flow of road segments at each time step within the predicted time domain. Blue dots represent the objective function value corresponding to the sampling points in each evaluation, and the red curve represents the optimal solution for the current iteration. It is clearly observed from the figure that for each control time step, the objective function value decreases rapidly in the first 20 iterations or so, indicating that the algorithm can significantly improve the green light duration allocation of the network within a relatively small number of iterations. As the number of iterations further increases, the objective function value tends to stabilize, indicating that the algorithm has gradually converged and found a stable optimal solution. This demonstrates the efficiency of trust-region Bayesian optimization in solving large-scale signal optimization problems.

[0147] Table 6. Parameter settings for the trust region Bayesian optimization algorithm

[0148]

[0149] To further verify the effectiveness of the optimal solution obtained by the trust-region Bayesian optimization algorithm, this embodiment uses average vehicle delay as an evaluation index and compares the index values ​​before and after optimization. Specific results are shown in Table 7. The average vehicle delay of the original scheme is 96.84s, while the average vehicle delay of the method proposed in this invention is 83.46s. The optimization effect of the proposed method compared to the original scheme is 13.82%. It should be noted that the original scheme was obtained by engineers from Qingdao Hisense Network Technology Co., Ltd. using their signal timing method and combining it with practical experience. This scheme's phase structure better matches the traffic flow characteristics of each intersection and considers arterial coordination control, reasonably setting the phase difference according to road priority and channelization. However, the optimization effect of the full-process regional traffic signal optimization method proposed in this invention is superior to the original signal timing scheme, further reducing the average vehicle delay in the network, fully demonstrating that the optimal solution obtained by the proposed method is of high quality.

[0150] Table 7 Comparison of Optimization Results

[0151] .

Claims

1. A traffic signal optimization method based on model predictive control and trust region Bayesian optimization, characterized in that, The process comprises three steps: traffic data acquisition, traffic state reconstruction, and traffic signal optimization. The traffic data acquisition section uses object detection algorithms to detect vehicles and object tracking algorithms to track them, extracting traffic flow at different turns at intersections from the video. The traffic state reconstruction section uses partial radar data, video data, and map data of the study area to construct a macroscopic traffic flow model, estimating missing road segment traffic flow and intersection turning ratios to reconstruct the overall traffic flow operation status of the study area. The traffic signal optimization section uses microscopic traffic simulation software to build a traffic simulation model of the study area based on the reconstructed traffic flow operation status. It uses the macroscopic traffic flow model to predict the traffic flow operation status and combines it with centralized model predictive control to construct a signal optimization model. The model is then solved using a trust-region Bayesian optimization algorithm to obtain the optimal signal control scheme. The specific steps are as follows: Step 1: Traffic data collection, as detailed below: Step 1.1, Data Preparation: Acquire video data of monitored intersection entrances in the study area; perform data augmentation and normalization on the intersection images; Step 1.2, Vehicle detection based on object detection algorithm: Execute the object detection algorithm to detect and label all vehicles in each frame of the image; Step 1.3, Vehicle tracking based on target tracking algorithm: First, based on the vehicle detection results in step 1.2, the target tracking algorithm is executed to extract and track motion features, output the tracked vehicle motion trajectory, and obtain the traffic flow at different directions at the intersection. Step 2, Traffic Status Reconstruction, as detailed below: Step 2.1, Data Preparation: First, based on step 1, traffic flow data for different turns at monitored intersections in the study area is obtained; second, radar data for radar-equipped road sections in the study area is acquired and processed to obtain road section traffic flow data; finally, a map is used to obtain built environment data, average speed data, and queue length data for all road sections in the study area; the environmental data includes the number of lanes and the number of traffic hotspots. Step 2.2, Traffic Flow Estimation for Road Sections: Step 2.2.1, Road segment classification; Based on the built environment data, a clustering algorithm is used to classify all road segments in the study area; Step 2.2.2, Flow-speed model construction; For each road segment category, Greenshields, Greenberg, and Underwood flow-speed models are established respectively, and the specific calculation formulas are shown in (1)-(3): (1) (2) (3) in, Indicates road segment r Traffic; Indicates road segment r The blocking density; Indicates road segment r The average speed; Indicates road segment r Critical density; Indicates road segment r The free-flow velocity is determined based on the road grade. Step 2.2.3: Dataset partitioning; Divide the average speed data and traffic flow data of road segments into two parts. One part is used as the training set for model parameter calibration; the other part is used as the validation set for model accuracy verification. Step 2.2.4, Model parameter calibration; The traffic data in the training set... and average speed Substitute the parameters into three flow-velocity models and then adjust the parameters in each model. , Perform calibration; Step 2.2.5, Model Selection: For each road segment category, the average speed in the validation set will be selected. Substitute the calibrated flow-velocity model into the flow rate and calculate the corresponding flow rate. Compare the calculation results with the actual flow rate in the validation set, calculate the error of each of the three models, and select the model with the smallest error for flow rate estimation of this road segment category. Step 2.2.5: Estimate the flow rate of the remaining road segments in the study area; For the remaining road segments in the study area, select the model of the category according to the classification results, substitute its average speed data into the flow-speed model, and estimate the flow rate. Step 2.3, Intersection turning ratio estimation: Step 2.3.1: Determine the values ​​of the model parameters; First, based on the traffic flow at different directions at the intersection obtained in step 1, determine the left-turn ratio of the shared straight and left-turn lanes. Right turn ratio of shared straight and right lanes The import channel to be estimated , The value selection channel type is similar to that of the video import channel. , The values ​​are then determined through traffic surveys; the remaining parameters, including green light loss time, headway for each lane, and phase green light duration, are then determined. Step 2.3.2: Estimate lane capacity; Estimate lane capacity according to different lane types, using the following formula: Step 2.3.2.1, Formula for calculating vehicles passing through the straight lane during the green light period: When setting up a dedicated signal for straight traffic: (4) in, This indicates vehicles passing through the straight-ahead lane during the green light period when a dedicated straight-ahead signal is in place. This indicates the duration of the green light for the straight-ahead phase of that approach lane; This indicates that the lost time during a green light is mainly due to the time lost due to vehicle starting. Indicates the headway of a straight-going vehicle; When no dedicated signal for straight traffic is set up: (5) in, This indicates vehicles passing through the straight-ahead lane during the green light period when no dedicated straight-ahead signal is provided. This indicates the straight-through traffic flow reduction factor. Due to interference from left-turning traffic, the straight-through traffic flow will be less than the flow when a dedicated straight-through signal is set up. Step 2.3.2.2, Formula for calculating vehicles passing through the left-turn lane during the green light period: When setting up a dedicated left-turn signal: (6) in, This indicates the vehicles passing through the left-turn lane during the green light period when a dedicated left-turn signal is set up; This indicates the duration of the green light for the left-turn phase at that entrance lane; Indicates the distance between the front of the car when turning left; When no dedicated left-turn signal is set up: (7) in, This indicates vehicles passing through the left-turn lane during the green light period when there is no dedicated left-turn signal; This indicates the left-turn traffic flow reduction factor. Due to interference from through traffic, the left-turn traffic flow will be less than the flow when a dedicated left-turn signal is set up. Step 2.3.2.3, Formula for calculating vehicles passing through the right-turn lane during the green light period: (8) in, This indicates vehicles passing through the right-turn lane during the green light period; This indicates the duration of the green light for the right turn phase at that entrance lane; Indicates the distance to the front of the car when turning right; other parameters are the same as before; Step 2.3.2.4: Vehicles passing through the shared straight / left lane during the green light period. Calculation formula: When the left turn ratio When, the calculation formula is: (9) When the left turn ratio When, the calculation formula is: (10) When the left turn ratio When, the calculation formula is: (11) Therefore, it can be concluded that left-turning vehicles passing through the combined straight and left-turn lanes... Calculation formula: (12) Formula for calculating the number of vehicles going straight through a shared straight-left lane: (13) in, This indicates vehicles proceeding straight through the shared straight / left-hand lane during the green light period; This indicates the duration of the green light for the straight-ahead and left-turn phases at this approach lane; Indicates the headway of vehicles in the combined straight and left lane; This indicates the proportion of left-turning traffic in a shared straight-left lane; Indicates the reduction factor; Step 2.3.2.5: Vehicles passing through the shared straight / right lane during the green light period. Calculation formula: When the right turn ratio When, the calculation formula is: (14) When the right turn ratio When, the calculation formula is: (15) When the right turn ratio When, the calculation formula is: (16) Therefore, the formula for calculating the number of straight-going vehicles passing through the shared straight-and-right lane is as follows: (17) in, This indicates vehicles going straight through the shared straight / right lane during the green light period; This indicates the duration of the green light for the straight-ahead and right-turn phases at this approach lane; Indicates the headway of vehicles in the shared straight-and-right lane; This indicates the proportion of turning traffic in a shared straight-to-right lane; Indicates the reduction factor; Step 2.3.3: Estimate the number of vehicles in the queue; based on the queue length data on the map, calculate the number of vehicles queuing at each turn of the entrance lane. The calculation formula is as follows: (18) in, This indicates the queue of vehicles turning at each direction on the entrance lane; Indicates the queue length; Indicates the distance between vehicles; Indicates the length of a standard car; Step 2.3.4: Estimate the flow rate for a single signal cycle; based on the capacity of each turn. Combined with the number of vehicles in the queue The flow rate for each turning point in a single cycle is calculated using the following formula: when The flow rate calculation formula is as follows: (19) in, These represent left turn, straight ahead, and right turn, respectively. This represents the flow rate of a single signal cycle; This is the reduction factor; when The flow rate calculation formula is as follows: (20) Step 2.3.5: Calculate the total flow rate during the study period; based on the single-cycle flow rate, calculate the total flow rate at each turning point during the study period using the following formula: (21) in, Indicates the total flow during the research period; Indicates the total time of the research period; This indicates the signal cycle duration at the intersection where the approach lane is located; Step 2.3.6: Calculate the turning ratio of each approach lane; calculate the turning ratio of each approach lane at the intersection, that is, the ratio of left-turn, straight-through, and right-turn traffic flow. Step 3, traffic signal optimization, as detailed below: Step 3.1: Construct a traffic simulation model: Based on the signal timing scheme of the study area and the traffic state data obtained in step 2, a traffic simulation model of the study area is constructed using microscopic traffic simulation software to simulate the real traffic flow process. Step 3.2: Construct a traffic flow prediction model: The regional traffic flow operation status is simulated based on the macro traffic flow model in order to predict the traffic flow status of road segments after the current signal timing scheme is implemented; Step 3.3: Construct a regional traffic signal optimization model: Assume the set of road segments is I The set of intersections is J intersection j The phase set is The cycle duration is The current control time step of the model predictive controller is k The time step is Control time domain is N Road section i Controlling time steps k The traffic is intersection j phase p The green light time is The minimum green light time is The maximum green light time is The green light interval is The cycle duration is The control input (i.e., the decision variable) is the green light duration of all phases at each time step within the control time domain. In each control step, define and solve the following problem: (22) (23) In formula (23) Let represent the objective function, which aims to minimize the sum of the remaining traffic flow of the road segment at all time steps within the control time domain; Equation (23) represents the constraints. Given the known initial network state constraints, each control time step is obtained from the traffic simulation model constructed in step 3.1; intersection phase The green light time is The green light interval is ; , These are the upper and lower bound constraints for the decision variables, namely the minimum green light duration and the maximum green light duration; The cycle duration constraint means that the sum of the green light duration and the green light interval time for each phase at each intersection is equal to the cycle duration of that intersection. Step 3.4: Solve the regional signal optimization model: The trust-region Bayesian optimization algorithm is used to solve the regional signal optimization model constructed in step 3.3 to obtain the current control time step. k Optimal green light duration for each phase; Step 3.5: Determine whether to terminate the optimization; For the maximum control time step, if the current control time step If the signal optimization fails, terminate the signal optimization; otherwise, update the initial state of the road network and proceed to the next time step, letting... Return to step 3.1 and input the signal timing scheme obtained in step 3.4 into the traffic simulation model in step 3.

1.

2. The traffic signal optimization method based on model predictive control and trust region Bayesian optimization according to claim 1, characterized in that, Step 3.4 is as follows: Step 3.4.1: Construct the prior dataset Random selection was performed using the Latin hypercube sampling method. n Group signal timing scheme as initial sample Formula (22) is used to evaluate the initial sampling points and obtain the corresponding objective function values. and the initial prior dataset ; The initial training sample dataset is ; Step 3.4.2: Set algorithm parameters; the initial trust region's base radius length is... The maximum radius of the trust region is The minimum radius is ; Optimize the variable dimension to The batch sample size is The trust domain update success counter is The failure counter is The maximum number of iterations is ; Step 3.4.3: Initialize the trust domain space The initial trust region has a base radius length of . The center point of the initial trust region is ,in Let the number of iterations be... ; Step 3.4.4: Fit the objective function using local GP; based on the sample dataset. Local generalized physics (GP) is used to fit the nonlinear relationship between green light duration and the sum of traffic flow in the predicted time domain, thus obtaining the next batch of sampling points. The posterior probability distribution; Steps 3.4.5: In the trust domain Thompson sampling is performed; first, in the trust region space of the current iteration. In, generate a size of random samples Then, through Repeat the screening to select the next batch q The candidate points are used to predict the green light duration at each time step in the time domain. ,in, , ; Indicates the first The objective function values ​​corresponding to the candidate green light durations follow a Gaussian process. ) distribution, in which For the current control time step The One green light duration candidate scheme It is local The fitted average predicted value, It is about predicting uncertainty; Step 3.4.6: Evaluate candidate points; based on the traffic flow prediction model and objective function formula (22) in step 3.2, calculate the corresponding objective function value to obtain the next batch of sampling dataset. ; Step 3.4.6: Determine if the iteration has terminated; if If so, proceed to step 3.4.9; otherwise, continue to the next step. Step 3.4.7, Trust Region Update; First, determine if a better solution can be found in the current iteration. If so, let , ,in, Record the number of times a better solution is found consecutively. Record the number of consecutive times no better solution is found; otherwise, let , Then, determine whether the update condition for the trust region radius is met. Then let , ;if Then let , ,in, The threshold for the number of successful attempts. Define the search range based on the current base radius of the trust region. Find the maximum base radius of the trust region; finally, update the trust region. That is, the range of candidate points for the optimal green light duration; It is the central location of the trust domain, in which, , For the green light duration scheme of the trust domain center point, This is the initial green light duration sample set. This will serve as a candidate set for the next batch of green light durations. Step 3.4.8: Update the dataset with the next sample set, let Return to step 3.4.4; Step 3.4.9, obtain Optimal green light duration for each time step Choose the first k The result of the step is used as the optimal signal timing scheme for the current control time step.

Images