Machine Vision-Based High-Speed ​​Traffic Flow Prediction and Traffic Control Methods

Abstract

This invention discloses a machine vision-based method for high-speed traffic flow prediction and traffic control, belonging to the field of data analysis technology. The method includes: analyzing historical traffic flow trends on various roads, quantifying the traffic status of each road, constructing a traffic scene recognition model for monitoring the high-speed traffic network, selecting attribution datasets for traffic scenes on various roads in the high-speed traffic network, generating probability distribution vectors for traffic scenes on various roads using Bayesian priors, substituting these vectors into a high-speed traffic space-time convolutional topology network, and predicting traffic flow for short-term monitoring of traffic scenes on various roads in the high-speed traffic network. Based on several known sets of traffic control actions for high-speed roads, an adaptive traffic control function for high-speed roads is constructed to generate the optimal control strategy for each road in the high-speed traffic network. The beneficial effects of this invention are: comprehensively improving the operational efficiency of the road network while ensuring the practical feasibility and economy of the control strategy.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, specifically to a method for high-speed traffic flow prediction and traffic control based on machine vision. Background Technology

[0002] Existing machine vision-based methods for high-speed traffic flow prediction and traffic control mostly remain at the stage of extracting surface features from visual data, lacking a deep semantic understanding of the evolution of traffic conditions. Prediction models are mostly pure data-driven black-box systems, which are difficult to integrate prior knowledge of traffic physics and lack interpretability of the decision-making process. Control strategies often rely on empirical rules or local optimization, failing to establish an end-to-end intelligent closed loop from perception to decision-making. This results in insufficient adaptability of the system in complex dynamic scenarios, low coordination efficiency between control units, and difficulty in dealing with sudden traffic anomalies. Summary of the Invention

[0003] To address the aforementioned technical problems, this technical solution provides a machine vision-based method for high-speed traffic flow prediction and traffic control, thus resolving the aforementioned issues.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] Machine vision-based methods for high-speed traffic flow prediction and traffic control include:

[0006] S1. Obtain historical traffic flow data for each road in the monitored high-speed traffic network, analyze the historical traffic flow change trend of each road, quantify the traffic status of each road, construct a traffic scene identification model for the monitored high-speed traffic network, and screen the traffic scene attribution dataset for each road in the monitored high-speed traffic network.

[0007] S2. Based on the attention attribution dataset of various road traffic scenarios in the high-speed traffic network, Bayesian prior is used to generate the probability distribution vector of each road traffic scenario. This vector is then substituted into the high-speed traffic space-time convolutional topology network to predict the traffic flow of each road traffic scenario in the short-term monitoring of the high-speed traffic network.

[0008] S3. Based on a known set of traffic control actions on highways, quantify the scheduling flow vector of each known traffic control action and monitor the traffic flow of each road in the highway network in a short time. Construct an adaptive traffic control function for highways and generate the optimal control strategy for each road in the highway network.

[0009] Preferably, step S1 specifically includes:

[0010] Data cleaning and alignment processes were performed on historical traffic flow data for various roads in the high-speed transportation network.

[0011] Based on the historical traffic flow data of each road in the monitoring high-speed traffic network, the trend term, seasonal term and residual term of the historical traffic flow of each road are extracted according to the STL decomposition method, and the time series of historical traffic flow of each road in the monitoring high-speed traffic network are constructed.

[0012] Predefine several traffic state clusters for high-speed traffic;

[0013] Based on the historical traffic flow time series of each road in the high-speed traffic network, the traffic volume, average speed and time occupancy characteristics of the historical traffic flow of each road are extracted. Using K-Means clustering, according to the centroid of several predefined communication state clusters of high-speed traffic, the minimum Euclidean distance of the traffic volume, average speed and time occupancy characteristics of the historical traffic flow of each road is calculated for iterative division, and an enhanced dataset of historical traffic flow status labels for each road in the high-speed traffic network is constructed.

[0014] Preferably, step S1 further includes:

[0015] Several scenarios for the evolution of traffic conditions on high-speed roads are predefined;

[0016] Based on recurrent neural networks (RNNs), this method takes the historical traffic flow status label augmentation dataset of each road in the high-speed traffic network as input. According to several defined traffic flow status evolution scenarios of high-speed traffic, it captures the hidden state of the last time step of the historical traffic flow status label augmentation data of each road within a unit time window. Substitutes the data into a fully connected layer and a Softmax layer to generate the probability of the traffic flow status evolution scenario to which the historical traffic flow status label augmentation data of each road belongs. This model is then used to train and construct a traffic scenario recognition model for monitoring high-speed traffic networks.

[0017] Based on the traffic scene identification model of the high-speed traffic network, a feature interpreter is established using SHAP value theory. The maximum contribution value of the historical traffic state label enhancement data of each road per unit time to several traffic state evolution scenarios that define the high-speed traffic is selected in the probability of the traffic state evolution scenario. This maximum contribution value is denoted as the attention attribution dataset of each road traffic scene in the high-speed traffic network.

[0018] Preferably, step S2 specifically includes:

[0019] Based on the monitoring of the attribution dataset of various road traffic scenarios in the high-speed traffic network, the SHAP contribution value of the attribution data of various road traffic scenarios within a unit time is normalized, and weights are assigned to the attribution data of various road traffic scenarios and several predefined traffic state evolution scenarios of high-speed traffic. A weighted Euclidean distance is established to calculate the spatial distance between the attribution data of various road traffic scenarios and the traffic state evolution scenarios within a unit time.

[0020] Using the Gaussian kernel function, the spatial distance between the attribution data of each road traffic scenario and the evolution scenario of each traffic state within a unit of time is mapped to the matching score between the attribution data of each road traffic scenario and the evolution scenario of each traffic state within a unit of time according to a monotonically decreasing function.

[0021] Based on the matching scores between the attribution data of each road traffic scenario within a unit of time and the evolution scenarios of each traffic state, the Softmax function is used to generate the probability distribution vector of each road traffic scenario within a unit of time.

[0022] Preferably, step S2 further includes:

[0023] Based on STGNN graph neural network, each road in the monitored high-speed traffic network is taken as a node, and the adjacency matrix of the road spatial topology of the high-speed traffic network is taken as the connection relationship. Chebyshev graph convolution is used to extract the aggregation features of the adjacency nodes of each road, and capture the spatial dependence of the spatial interaction influence relationship between adjacent roads in the monitored high-speed traffic network.

[0024] Based on STGNN graph neural network, using historical traffic flow data of each road in the high-speed traffic network as input, and using TCN temporal convolutional network, we can capture the time dependence of the traffic flow of independent roads in the high-speed traffic network as the traffic status changes over a unit of time.

[0025] Based on the spatial dependence of the spatial interaction influence relationship between adjacent roads in the monitored high-speed traffic network and the temporal dependence of the traffic flow of independent roads in the monitored high-speed traffic network as the traffic status changes with unit time, a high-speed traffic spatial-temporal convolutional topology network is constructed to extract the spatial semantic embedding vector and the temporal evolution scenario embedding vector between adjacent roads in the monitored high-speed traffic network.

[0026] The probability distribution vector of each road traffic scenario within a unit time is concatenated with the spatial semantic embedding vector and temporal evolution scenario embedding vector between adjacent roads in the monitored high-speed traffic network. This concatenation is then fed into a fully connected layer to predict traffic flow for each road traffic scenario in the short-term monitored high-speed traffic network.

[0027] Preferably, step S3 specifically includes:

[0028] Determine a known set of traffic control actions for a highway.

[0029] Based on the adjacency matrix of the road spatial topology of the high-speed transportation network, the traffic flow of each road traffic scenario in the high-speed transportation network is monitored in a short time as the road attribute in the adjacency matrix of the road network spatial topology. Using traffic simulation open-source software, the changes in traffic flow of each road traffic scenario in the high-speed transportation network under given known traffic control actions are verified, and a road control action-effect mapping table of the high-speed transportation network is constructed.

[0030] Based on the road control action-effect mapping table of the high-speed traffic network, a multivariate linear regression model is established, with road control actions as independent variables as inputs and known traffic control action scheduling flow vectors as dependent variables as outputs.

[0031] Preferably, step S3 further includes:

[0032] Obtain the maximum traffic flow capacity of each road in the high-speed transportation network;

[0033] Based on NLP nonlinear programming, the objective function is to minimize the overall travel time of the highway network and maximize the traffic flow. The traffic constraints are: the traffic flow after knowing several traffic control actions on the highway is within the maximum traffic flow capacity of each road in the highway network; the mutual exclusion constraint is that the known traffic control actions on the highway cannot be executed simultaneously; and the cost constraint is that the cost of implementing control actions on each road in the highway network cannot exceed the budget cost. An adaptive traffic control function for highways is constructed.

[0034] Standardization processing is performed on the scheduling flow vectors of known traffic control actions and the traffic flow of various road traffic scenarios in the short-term monitoring of the highway traffic network.

[0035] Based on the adaptive traffic control function for highways, the optimal control strategy for each road in the highway traffic network is generated by taking the known scheduling flow vectors of each traffic control action and the traffic flow of each road in the short-term monitoring of the highway traffic network as inputs and iteratively solving the problem according to the metaheuristic genetic algorithm.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] This invention proposes a high-speed traffic flow prediction and traffic control scheme based on machine vision. By constructing a full-process intelligent closed loop of perception-prediction-optimization-control, it achieves deep integration of machine vision and explainable AI. The system can not only accurately identify the current traffic status, but also proactively predict the trend of congestion evolution. By combining Bayesian priors guided by physical knowledge with spatiotemporal graph neural networks, the accuracy, robustness, and interpretability of the prediction results are significantly improved. Finally, by using constraint optimization decision-making based on system simulation, a globally collaborative optimal control strategy is generated, enabling traffic management to shift from passive response to proactive intervention. While comprehensively improving the efficiency of road network operation, it ensures the practical feasibility and economy of the control strategy, forming an adaptive intelligent traffic control system. Attached Figure Description

[0038] Figure 1 This is a flowchart of a machine vision-based method for high-speed traffic flow prediction and traffic control. Detailed Implementation

[0039] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0040] Reference Figure 1 As shown, a machine vision-based method for high-speed traffic flow prediction and traffic control includes:

[0041] S1. Obtain historical traffic flow data for each road in the monitored high-speed traffic network, analyze the historical traffic flow change trend of each road, quantify the traffic status of each road, construct a traffic scene identification model for the monitored high-speed traffic network, and screen the traffic scene attribution dataset for each road in the monitored high-speed traffic network.

[0042] Step S1 specifically includes:

[0043] Data cleaning and alignment processes were performed on historical traffic flow data for various roads in the high-speed transportation network.

[0044] Based on the historical traffic flow data of each road in the monitoring high-speed traffic network, the trend term, seasonal term and residual term of the historical traffic flow of each road are extracted according to the STL decomposition method, and the time series of historical traffic flow of each road in the monitoring high-speed traffic network are constructed.

[0045] Predefine several traffic state clusters for high-speed traffic;

[0046] As a further development, several communication state clusters for high-speed traffic are predefined, including: a smooth traffic state label cluster, a slow-moving traffic state label cluster, and a congested traffic state label cluster. The physical meanings of each cluster are as follows: Smooth traffic state label cluster: high speed, medium to low flow, low occupancy; Slow-moving traffic state label cluster: medium speed, medium flow, medium occupancy; Congested traffic state label cluster: low speed, low flow, high occupancy. The values ​​of the predefined clusters depend on the settings of the implementers and will not be described here.

[0047] Based on the monitoring of historical traffic flow time series of various roads in the high-speed traffic network, the traffic volume, average speed and time occupancy characteristics of historical traffic flow of each road are extracted. Using K-Means clustering, according to the centroid of several predefined communication state clusters of high-speed traffic, the minimum Euclidean distance of traffic volume, average speed and time occupancy characteristics of historical traffic flow of each road is calculated for iterative division, and an enhanced dataset of traffic status labels for historical traffic flow of various roads in the high-speed traffic network is constructed.

[0048] Step S1 also includes:

[0049] Several scenarios for the evolution of traffic conditions on high-speed roads are predefined;

[0050] As a further development, several traffic state evolution scenarios for high-speed traffic are predefined, including: stable and smooth traffic scenario, congestion formation scenario, congestion dissipation scenario, and intermittent congestion scenario. The physical meanings of each scenario are as follows: Stable and smooth traffic scenario: the traffic flow characteristic value is smooth for multiple consecutive units of time; Congestion formation scenario: the traffic flow characteristic value changes from smooth to slow to congested within a unit of time; Congestion dissipation scenario: the traffic flow characteristic value changes from congested to slow to smooth within a unit of time; Intermittent congestion scenario: the traffic flow characteristic value frequently alternates between smooth and slow within a unit of time.

[0051] Based on recurrent neural networks (RNNs), this method takes the historical traffic flow status label augmentation dataset of each road in the high-speed traffic network as input. According to several defined traffic flow status evolution scenarios of high-speed traffic, it captures the hidden state of the last time step of the historical traffic flow status label augmentation data of each road within a unit time window. Substitutes the data into a fully connected layer and a Softmax layer to generate the probability of the traffic flow status evolution scenario to which the historical traffic flow status label augmentation data of each road belongs. This model is then used to train and construct a traffic scenario recognition model for monitoring high-speed traffic networks.

[0052] Based on the traffic scene identification model of the high-speed traffic network, a feature interpreter is established using the SHAP value theory. The maximum contribution value of the historical traffic flow status label enhancement data of each road under the unit time in the probability of traffic state evolution of the traffic state is selected as the focus attribution dataset of each road traffic scene in the high-speed traffic network.

[0053] When using it, please refer to the steps outlined above:

[0054] As a further step, a progressive data mining process elevates raw traffic flow data into interpretable management knowledge. STL decomposition and K-Means clustering are used to objectively quantify traffic conditions. Then, an RNN model is used to identify dynamically evolving scenarios, and key causal features under different scenarios are extracted based on SHAP interpretability technology. This achieves the transformation from low-level data to high-level semantic knowledge. It enables the forward-looking and interpretable understanding of traffic conditions: it can not only identify what is happening now, but also predict what will happen next (such as congestion formation), and clearly reveal the decision-making basis (such as a sudden increase in upstream occupancy), thus laying a solid and reliable data foundation for subsequent accurate prediction and adaptive management.

[0055] Furthermore, as an embodiment of step S1 above:

[0056] Project Name: Intelligent Traffic Sensing System for Shanghai Section of Shanghai-Nanjing Expressway;

[0057] Objective: To achieve accurate perception and scene recognition of traffic conditions on this road segment.

[0058] Implementation process:

[0059] Data preparation and status quantification:

[0060] Input: Get 5-minute granular data (traffic flow, speed, occupancy) of all cameras and loop detectors on this road segment over the past year.

[0061] Processing: Data cleaning and STL decomposition were performed to identify significant daily and weekly cycles.

[0062] Clustering: K-Means clustering (K=3) was performed on the traffic-speed-occupancy data for the entire year. Results show:

[0063] Cluster 1 Center: (Speed ​​= 85km / h, Flow = 1200 vehicles / hour, Occupancy = 8%) -> Smooth flow;

[0064] Cluster 2 Center: (Speed ​​= 45km / h, Flow = 1800 vehicles / hour, Occupancy = 18%) -> Slow-moving state;

[0065] Cluster 3 Center: (Speed ​​= 15km / h, Flow = 800 vehicles / hour, Occupancy = 35%) -> Congested State;

[0066] Output: A dataset of state labels for each 5-minute time slice, named Dhistory′Dhistory′, was generated.

[0067] Scene recognition model training:

[0068] Sample construction: Construct sequence samples using a 30-minute (6 time steps) time window. For example, the sequence [smooth, smooth, slow, slow, congested, congested] is labeled as congestion formation.

[0069] Model training: A single-layer GRU network was used for training. The input consisted of state labels at 6 time steps (converted into vectors through an embedding layer), and the output consisted of probabilities for 4 scenes. The model achieved an accuracy of 92% on the test set.

[0070] Attribution analysis and knowledge base construction:

[0071] SHAP analysis: Perform SHAP analysis on the trained GRU model.

[0072] Finding 1: For congestion scenarios, the most significant contributing feature is the sudden increase in occupancy rates at the current time t-1 and t-2, while the contribution of speed decrease is relatively delayed.

[0073] Finding 2: For intermittent congestion scenarios, the most contributing feature is the frequent and rapid alternation between smooth and slow-moving states in the sequence.

[0074] Finding 3: Identifying congestion requires focusing on data from the past 20-30 minutes (i.e., the critical time window TCkTCk), while identifying stable and smooth traffic only requires the past 5-10 minutes.

[0075] Output attribution dataset:

[0076] A_Congestion Formation = {Core Characteristics: [Occupancy Rate], Key Time Window: 30 minutes, Spatial Impact: [This Road Section and Upstream Road Sections]};

[0077] A_Intermittent Congestion = {Core Characteristics: [State Switching Frequency], Key Time Window: 15 minutes, Spatial Impact: [This Road Section]}.

[0078] S2. Based on the attention attribution dataset of various road traffic scenarios in the high-speed traffic network, Bayesian prior is used to generate the probability distribution vector of each road traffic scenario. This vector is then substituted into the high-speed traffic space-time convolutional topology network to predict the traffic flow of each road traffic scenario in the short-term monitoring of the high-speed traffic network.

[0079] Step S2 specifically includes:

[0080] Based on the monitoring of the attribution dataset of various road traffic scenarios in the high-speed traffic network, the SHAP contribution value of the attribution data of various road traffic scenarios within a unit time is normalized, and weights are assigned to the attribution data of various road traffic scenarios and several predefined traffic state evolution scenarios of high-speed traffic. A weighted Euclidean distance is established to calculate the spatial distance between the attribution data of various road traffic scenarios and the traffic state evolution scenarios within a unit time.

[0081] Using the Gaussian kernel function, the spatial distance between the attribution data of each road traffic scenario and the evolution scenario of each traffic state within a unit of time is mapped to the matching score between the attribution data of each road traffic scenario and the evolution scenario of each traffic state within a unit of time according to a monotonically decreasing function.

[0082] Based on the matching scores between the attribution data of each road traffic scenario within a unit of time and the evolution scenarios of each traffic status, the Softmax function is substituted to generate the probability distribution vector of each road traffic scenario within a unit of time.

[0083] Step S2 also includes:

[0084] Based on STGNN graph neural network, each road in the monitored high-speed traffic network is taken as a node, and the adjacency matrix of the road spatial topology of the high-speed traffic network is taken as the connection relationship. Chebyshev graph convolution is used to extract the aggregation features of the adjacency nodes of each road, and capture the spatial dependence of the spatial interaction influence relationship between adjacent roads in the monitored high-speed traffic network.

[0085] Based on STGNN graph neural network, using historical traffic flow data of each road in the high-speed traffic network as input, and using TCN temporal convolutional network, we can capture the time dependence of the traffic flow of independent roads in the high-speed traffic network as the traffic status changes over a unit of time.

[0086] Based on the spatial dependence of the spatial interaction influence relationship between adjacent roads in the monitored high-speed traffic network and the temporal dependence of the traffic flow of independent roads in the monitored high-speed traffic network as the traffic status changes with unit time, a high-speed traffic spatial-temporal convolutional topology network is constructed to extract the spatial semantic embedding vector and the temporal evolution scenario embedding vector between adjacent roads in the monitored high-speed traffic network.

[0087] The probability distribution vector of each road traffic scenario within a unit time is concatenated with the spatial semantic embedding vector and the temporal evolution scenario embedding vector between adjacent roads in the monitored high-speed traffic network. This concatenation is then fed into a fully connected layer to predict traffic flow for each road traffic scenario in the short-term monitored high-speed traffic network.

[0088] When using it, please refer to the steps outlined above:

[0089] As a further development, the physical attributions extracted in S1 are transformed into Bayesian prior probabilities using weighted Euclidean distance and the Softmax function. These prior probabilities are then deeply fused with the data-driven features extracted by the spatiotemporal graph network. This generates traffic flow predictions that are more physically consistent, forward-looking, and robust. When faced with data uncertainty, the system can maintain stable outputs based on prior knowledge and provides semantically rich and interpretable predictive data for subsequent management, fundamentally improving the reliability and practicality of the prediction results.

[0090] Furthermore, as an embodiment of step S2 above:

[0091] Project Name: Short-Term Traffic Flow Prediction System for Shanghai Section of Shanghai-Nanjing Expressway;

[0092] Objective: To achieve accurate prediction of traffic flow (speed, volume) at various points along this road segment within the next 15 minutes.

[0093] Implementation process:

[0094] Prior probabilities of generated scenes:

[0095] Input: Obtain the attribution dataset from S1, such as the known congestion formation scenario whose core feature is upstream occupancy.

[0096] calculate:

[0097] Get real-time occupancy data for the current moment and the past 30 minutes.

[0098] Calculate the weighted Euclidean distance between the current data and the standard feature vectors of scenarios such as congestion formation and stable smooth flow (the weight is the SHAP value obtained from S1, with the upstream occupancy rate having the highest weight).

[0099] Use the Gaussian kernel function to convert the distance into a matching score.

[0100] The prior probability vector is obtained by using the Softmax function: P(scenario) = [stable and smooth: 0.1, congestion formation: 0.7, congestion dissipation: 0.1, intermittent congestion: 0.1].

[0101] Output: The prior probability of congestion forming in the next time step for this road segment at the current time is 70%.

[0102] Spatiotemporal graph neural network prediction:

[0103] Graph construction: The Shanghai section of the Shanghai-Nanjing Expressway is divided into 100 road segment units as graph nodes, and an adjacency matrix is ​​constructed based on the road connection relationship.

[0104] Model input: Historical data on traffic flow, speed, and occupancy of all road segments over the past hour (12 five-minute intervals).

[0105] Model processing:

[0106] Spatial Convolution: STGNN uses graph convolutional layers to allow each node in a road segment to aggregate information from its upstream and downstream neighbors. For example, node A (downstream) aggregates the speed decrease information from nodes B and C (upstream).

[0107] Temporal convolution: TCN independently analyzes the velocity sequence of each node over the past hour to identify a continuous, slow, and decreasing temporal pattern.

[0108] Feature fusion: The spatiotemporal embedding vector [0.85, -0.32, ..., 1.22] (128 dimensions) extracted by STGNN is concatenated with the prior probability vector [0.1, 0.7, 0.1, 0.1] (4 dimensions) generated in step 1 to obtain a 132-dimensional fused feature vector.

[0109] Final prediction output:

[0110] The fused feature vector is fed into the fully connected layer, which outputs the predicted traffic flow value for the road segment in the next 15 minutes.

[0111] Example of prediction results: Downstream of road segment A, predicted average speed for the next 15 minutes: 28 km / h, traffic flow: 950 vehicles / hour.

[0112] System Interpretation: By incorporating 70% of the prior knowledge of congestion formation, the model, after integrating spatiotemporal data, provides a low-speed prediction. This result is not merely a number; it is supported by clear physical meaning: the increase in upstream occupancy (prior knowledge), the propagation of upstream speed decline (spatial dependence), and the continuous downward trend of its own speed (time dependence) all point to congestion forming.

[0113] S3. Based on a known set of traffic control actions on highways, quantify the scheduling flow vector of each known traffic control action and monitor the traffic flow of each road in the highway network in a short time. Construct an adaptive traffic control function for highways and generate the optimal control strategy for each road in the highway network.

[0114] Step S3 specifically includes:

[0115] Determine a known set of traffic control actions for a highway.

[0116] Based on the adjacency matrix of the road spatial topology of the high-speed transportation network, the traffic flow of each road traffic scenario in the high-speed transportation network is monitored in a short time as the road attribute in the adjacency matrix of the road network spatial topology. Using traffic simulation open-source software, the changes in traffic flow of each road traffic scenario in the high-speed transportation network under given known traffic control actions are verified, and a road control action-effect mapping table of the high-speed transportation network is constructed.

[0117] Based on the road control action-effect mapping table of the high-speed traffic network, a system is established using multiple linear regression, with road control actions as independent variables as inputs and known traffic control action scheduling flow vectors as dependent variables as outputs.

[0118] Step S3 also includes:

[0119] Obtain the maximum traffic flow capacity of each road in the high-speed transportation network;

[0120] Based on NLP nonlinear programming, the objective function is to minimize the overall travel time of the highway network and maximize the traffic flow. The traffic constraints are: the traffic flow after knowing several traffic control actions on the highway is within the maximum traffic flow capacity of each road in the highway network; the mutual exclusion constraint is that the known traffic control actions on the highway cannot be executed simultaneously; and the cost constraint is that the cost of implementing control actions on each road in the highway network cannot exceed the budget cost. An adaptive traffic control function for highways is constructed.

[0121] As a further point, the cost constraint is that the cost of implementing control measures on each road in the high-speed transportation network cannot exceed the budgeted cost. Without considering fixed machinery and equipment (gates and signal signs) on the highway, the deployment of police forces and vehicle matching on the highway are also cost constraints.

[0122] Standardization processing is performed on the scheduling flow vectors of known traffic control actions and the traffic flow of various road traffic scenarios in the short-term monitoring of the highway traffic network.

[0123] Based on the adaptive traffic control function for highways, the optimal control strategy for each road in the highway traffic network is generated by taking the known scheduling flow vectors of each traffic control action and the traffic flow of each road in the short-term monitoring of the highway traffic network as inputs and iteratively solving the problem according to the metaheuristic genetic algorithm.

[0124] When using it, please refer to the steps outlined above:

[0125] As a further step, the impact of various control actions on road network traffic is quantified through traffic simulation to form a scheduling flow vector. Then, with the objective function of minimizing the total travel time or maximizing the total traffic flow, multiple constraints such as road capacity, mutual exclusivity of control actions, and implementation costs are comprehensively considered. Finally, a metaheuristic algorithm is used for intelligent solution. This achieves a fundamental shift from experience-based local control to data-driven global optimization, automatically generating the optimal collaborative control strategy at the system level. This significantly improves the overall operational efficiency of the road network while ensuring the feasibility and economy of the strategy in actual deployment, ultimately constructing a complete closed loop of intelligent traffic management encompassing perception, prediction, optimization, and control.

[0126] Furthermore, as an embodiment of step S3 above:

[0127] Project Name: Dynamic Integration System for Ramp Coordination Control and Mainline Speed ​​Limit of Shanghai Section of Shanghai-Nanjing Expressway;

[0128] Objective: To automatically generate the optimal combination of control strategies during the evening rush hour congestion formation period in order to alleviate congestion.

[0129] Implementation process:

[0130] Effect Quantification and Modeling:

[0131] Control action set A: {a1: Ramp R1 flow restriction (60%), a2: Ramp R2 closure, a3: Mainline speed limit (80km / h), a4: Mainline speed limit (60km / h)};

[0132] Simulation Quantization: A model of this road segment was built in SUMO, and the above actions were simulated respectively. Measurements:

[0133] v1=[R1:-0.4, downstream main road L1:-0.1, downstream main road L2:-0.05,...] (vector elements represent relative flow rate changes)

[0134] v2=[R2:-1.0, L1:-0.3, L2:-0.2,...];

[0135] v3 = [L1: speed +5%, L2: speed +3%, ...];

[0136] Establish an optimization model:

[0137] Decision variables: z = [z1, z2, z3, z4], z_i ∈ {0, 1};

[0138] Objective function: MinJ(z) = Total travel time of the entire network (calculated based on the BPR function).

[0139] Constraints:

[0140] Traffic capacity: Predicted traffic flow + ∑(z_i*v_i) ≤ Road capacity;

[0141] Mutual exclusion constraint: z1+z2≤1 (two adjacent ramps cannot have one with flow restriction and the other closed, otherwise it will cause local chaos);

[0142] Cost constraint: 1*z1+3*z2+1*z3+1*z4≤4 (assuming that closing the ramp is more costly);

[0143] Solution and strategy generation:

[0144] Input: S2 predicts the traffic flow X_hat for the next 15 minutes, and the scheduling flow vectors v1, v2, v3, v4.

[0145] Solution using genetic algorithm:

[0146] Initialization: Randomly generate 100 policy populations, such as [1, 0, 1, 0], [0, 1, 0, 1]...;

[0147] Iterative evolution: Calculate the objective function value J(z) (total travel time) for each policy. The shorter the total time, the higher the fitness of the policy.

[0148] Selection, crossover, and mutation: Retaining excellent strategies, combining them to generate new strategies, and introducing random variations.

[0149] After 100 iterations, the algorithm converges to the optimal solution: z*=[1, 0, 1, 0];

[0150] Optimal strategy: Execute actions a1 and a3, namely: implement a 60% flow restriction on ramp R1, while simultaneously activating a variable speed limit of 80km / h on the main line.

[0151] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A machine vision-based method for high-speed traffic flow prediction and traffic control, characterized in that, include: S1. Obtain historical traffic flow data for each road in the monitored highway traffic network, analyze the historical traffic flow change trends of each road, quantify the traffic status of each road, construct a traffic scene recognition model for the monitored highway traffic network, and select the attribution dataset for traffic scenes of each road in the monitored highway traffic network. Step S1 specifically includes: Data cleaning and alignment processes were performed on historical traffic flow data for various roads in the high-speed transportation network. Based on the historical traffic flow data of each road in the monitoring high-speed traffic network, the trend term, seasonal term and residual term of the historical traffic flow of each road are extracted according to the STL decomposition method, and the time series of historical traffic flow of each road in the monitoring high-speed traffic network are constructed. Predefine several traffic state clusters for high-speed traffic; Based on the monitoring of historical traffic flow time series of various roads in the high-speed traffic network, the traffic volume, average speed and time occupancy characteristics of historical traffic flow of each road are extracted. Using K-Means clustering, according to the centroid of several predefined communication state clusters of high-speed traffic, the minimum Euclidean distance of traffic volume, average speed and time occupancy characteristics of historical traffic flow of each road is calculated for iterative division, and an enhanced dataset of traffic status labels for historical traffic flow of various roads in the high-speed traffic network is constructed. Several scenarios for the evolution of traffic conditions on high-speed roads are predefined; Based on recurrent neural networks (RNNs), this method takes the historical traffic flow status label augmentation dataset of each road in the high-speed traffic network as input. According to several defined traffic flow status evolution scenarios of high-speed traffic, it captures the hidden state of the last time step of the historical traffic flow status label augmentation data of each road within a unit time window. Substitutes the data into a fully connected layer and a Softmax layer to generate the probability of the traffic flow status evolution scenario to which the historical traffic flow status label augmentation data of each road belongs. This model is then used to train and construct a traffic scenario recognition model for monitoring high-speed traffic networks. Based on the traffic scene identification model of the high-speed traffic network, a feature interpreter is established using the SHAP value theory. The maximum contribution value of the historical traffic flow status label enhancement data of each road under the unit time in the probability of traffic state evolution of the traffic state is selected as the focus attribution dataset of each road traffic scene in the high-speed traffic network. S2. Based on the attention attribution dataset of various road traffic scenarios in the monitored high-speed traffic network, using Bayesian priors, generate probability distribution vectors for each road traffic scenario, and substitute them into the high-speed traffic space-time convolutional topology network to predict traffic flow for each road scenario in the short-term monitored high-speed traffic network. Step S2 specifically includes: Based on the monitoring of the attribution dataset of various road traffic scenarios in the high-speed traffic network, the SHAP contribution value of the attribution data of various road traffic scenarios within a unit time is normalized, and weights are assigned to the attribution data of various road traffic scenarios and several predefined traffic state evolution scenarios of high-speed traffic. A weighted Euclidean distance is established to calculate the spatial distance between the attribution data of various road traffic scenarios and the traffic state evolution scenarios within a unit time. Using the Gaussian kernel function, the spatial distance between the attribution data of each road traffic scenario and the evolution scenario of each traffic state within a unit of time is mapped to the matching score between the attribution data of each road traffic scenario and the evolution scenario of each traffic state within a unit of time according to a monotonically decreasing function. Based on the matching scores between the attribution data of each road traffic scenario within a unit of time and the evolution scenarios of each traffic status, the Softmax function is substituted to generate the probability distribution vector of each road traffic scenario within a unit of time. Based on STGNN graph neural network, each road in the monitored high-speed traffic network is taken as a node, and the adjacency matrix of the road spatial topology of the high-speed traffic network is taken as the connection relationship. Chebyshev graph convolution is used to extract the aggregation features of the adjacency nodes of each road, and capture the spatial dependence of the spatial interaction influence relationship between adjacent roads in the monitored high-speed traffic network. Based on STGNN graph neural network, using historical traffic flow data of each road in the high-speed traffic network as input, and using TCN temporal convolutional network, we can capture the time dependence of the traffic flow of independent roads in the high-speed traffic network as the traffic status changes over a unit of time. Based on the spatial dependence of the spatial interaction influence relationship between adjacent roads in the monitored high-speed traffic network and the temporal dependence of the traffic flow of independent roads in the monitored high-speed traffic network as the traffic status changes with unit time, a high-speed traffic spatial-temporal convolutional topology network is constructed to extract the spatial semantic embedding vector and the temporal evolution scenario embedding vector between adjacent roads in the monitored high-speed traffic network. The probability distribution vector of each road traffic scenario within a unit time is concatenated with the spatial semantic embedding vector and the temporal evolution scenario embedding vector between adjacent roads in the monitored high-speed traffic network. This concatenation is then fed into a fully connected layer to predict traffic flow for each road traffic scenario in the short-term monitored high-speed traffic network. S3. Based on a known set of traffic control actions on highways, quantify the scheduling flow vector of each known traffic control action and monitor the traffic flow of each road in the highway network in a short time. Construct an adaptive traffic control function for highways and generate the optimal control strategy for each road in the highway network.

2. The high-speed traffic flow prediction and traffic control method based on machine vision according to claim 1, characterized in that, Step S3 specifically includes: Determine a known set of traffic control actions for a highway. Based on the adjacency matrix of the road spatial topology of the high-speed transportation network, the traffic flow of each road traffic scenario in the high-speed transportation network is monitored in a short time as the road attribute in the adjacency matrix of the road network spatial topology. Using traffic simulation open-source software, the changes in traffic flow of each road traffic scenario in the high-speed transportation network under given known traffic control actions are verified, and a road control action-effect mapping table of the high-speed transportation network is constructed. Based on the road control action-effect mapping table of the high-speed traffic network, a multivariate linear regression model is established, with road control actions as independent variables as inputs and known traffic control action scheduling flow vectors as dependent variables as outputs.

3. The high-speed traffic flow prediction and traffic control method based on machine vision according to claim 2, characterized in that, Step S3 also includes: Obtain the maximum traffic flow capacity of each road in the high-speed transportation network; Based on NLP nonlinear programming, the objective function is to minimize the overall travel time of the highway network and maximize the traffic flow. The traffic constraints are: the traffic flow after knowing several traffic control actions on the highway is within the maximum traffic flow capacity of each road in the highway network; the mutual exclusion constraint is that the known traffic control actions on the highway cannot be executed simultaneously; and the cost constraint is that the cost of implementing control actions on each road in the highway network cannot exceed the budget cost. An adaptive traffic control function for highways is constructed. Standardization processing is performed on the scheduling flow vectors of known traffic control actions and the traffic flow of various road traffic scenarios in the short-term monitoring of the highway traffic network. Based on the adaptive traffic control function for highways, the optimal control strategy for each road in the highway traffic network is generated by taking the known scheduling flow vectors of each traffic control action and the traffic flow of each road in the short-term monitoring of the highway traffic network as inputs and iteratively solving the problem according to the metaheuristic genetic algorithm.

Images