Traffic dynamic cooperative control method and system based on multi-source heterogeneous data fusion

Abstract

The invention provides a traffic dynamic cooperative control method and system based on multi-source heterogeneous data fusion, and relates to the technical field of traffic control, and the method comprises the steps: obtaining a multi-source heterogeneous data flow of a management and control region through a ubiquitous perception terminal cluster, carrying out the fine-grained data fusion, and outputting a space-time fusion state matrix; performing blind area state completion to generate a panoramic traffic state map; performing dynamic demand prediction based on the space-time deep learning model, and outputting a pressure distribution matrix; dynamically delimiting a traffic control boundary; performing multi-agent collaborative decision, and outputting a collaborative strategy packet; and according to the real-time dissipation rate deviation, performing closed-loop updating of the traffic coordination strategy. The technical problems that traffic data in the prior art mainly depends on a single data source, data missing or coverage blind areas are prone to occurring, real traffic conditions cannot be comprehensively reflected, traffic flow prediction errors and bottleneck prediction errors are caused, and finally traffic congestion is aggravated are solved.

Description

Technical Field

[0001] This invention relates to the field of traffic control technology, specifically to a method and system for dynamic collaborative traffic control based on the fusion of multi-source heterogeneous data. Background Technology

[0002] Traffic congestion not only affects the daily travel of urban residents but also brings a series of problems such as environmental pollution and energy waste. Therefore, how to effectively improve the operational efficiency of urban transportation, reduce traffic congestion, and optimize traffic flow has become the core objective of modern urban traffic management. However, existing traffic data mainly relies on single data sources, such as single traffic sensors or cameras, which are prone to data gaps or blind spots, resulting in incomplete information that cannot fully reflect the real traffic situation. This leads to errors in traffic flow prediction and bottleneck prediction, which may ultimately exacerbate traffic congestion and reduce management efficiency. Summary of the Invention

[0003] This application provides a traffic dynamic collaborative control method and system based on the fusion of multi-source heterogeneous data. It aims to solve the technical problem that existing traffic data mainly rely on a single data source, which is prone to data loss or coverage blind spots, and cannot fully reflect the real traffic conditions, leading to errors in traffic flow prediction and bottleneck prediction, and ultimately exacerbating traffic congestion.

[0004] The first aspect disclosed in this application provides a method for dynamic collaborative traffic control based on multi-source heterogeneous data fusion. The method includes: obtaining multi-source heterogeneous data streams of a controlled area through a ubiquitous sensing terminal cluster; performing fine-grained data fusion based on local spatiotemporal registration to output a spatiotemporal fusion state matrix; completing blind spot states of the spatiotemporal fusion state matrix based on road network topology to generate a panoramic traffic state map; performing dynamic demand prediction of the panoramic traffic state map based on a spatiotemporal deep learning model to output a pressure distribution matrix, wherein the pressure distribution matrix includes vehicle flow pressure distribution, pedestrian flow pressure distribution, and road network bottleneck node distribution; dynamically delineating traffic control boundaries based on road network carrying capacity and the pressure distribution matrix; performing multi-agent collaborative decision-making at the traffic control boundaries based on the pressure distribution matrix to output a collaborative strategy package, wherein the collaborative strategy package includes a tiered release strategy and a multimodal resource scheduling strategy; and performing closed-loop updates of the traffic collaborative strategy based on the real-time dissipation rate deviation of the panoramic traffic state map during the execution of the collaborative strategy package.

[0005] The second aspect of this application discloses a traffic dynamic collaborative control system based on the fusion of multi-source heterogeneous data. This system is used in the aforementioned traffic dynamic collaborative control method based on the fusion of multi-source heterogeneous data. The system includes: a fine-grained data fusion module, used to obtain multi-source heterogeneous data streams of the control area through a ubiquitous sensing terminal cluster, and then perform fine-grained data fusion based on local spatiotemporal registration to output a spatiotemporal fusion state matrix; a blind spot state completion module, used to complete the blind spot states of the spatiotemporal fusion state matrix based on the road network topology to generate a panoramic traffic state map; and a dynamic demand prediction module, used to predict the dynamic demand of the panoramic traffic state map based on a spatiotemporal deep learning model. The system predicts and outputs a pressure distribution matrix, which includes vehicle pressure distribution, pedestrian pressure distribution, and road network bottleneck node distribution. A traffic control boundary delineation module dynamically delineates traffic control boundaries based on the road network carrying capacity and the pressure distribution matrix. A multi-agent collaborative decision-making module performs multi-agent collaborative decisions based on the pressure distribution matrix at the traffic control boundaries, outputting a collaborative strategy package, which includes a tiered release strategy and a multimodal resource scheduling strategy. A closed-loop update module updates the traffic collaborative strategy in a closed loop based on the real-time dissipation rate deviation of the panoramic traffic state map during the execution of the collaborative strategy package.

[0006] One or more technical solutions provided in this application have at least the following beneficial effects:

[0007] Multi-source heterogeneous data from various ubiquitous sensing terminals is acquired from the controlled area. Combined with local spatiotemporal registration technology, fine-grained data fusion is performed, and the output spatiotemporal fusion state matrix provides more accurate and detailed traffic state information for subsequent traffic analysis, better capturing important parameters such as traffic flow, density, and speed. By completing blind spots in the spatiotemporal fusion state matrix based on road network topology, a panoramic traffic state map is generated, representing the complete traffic network state. This map displays information such as traffic flow, bottleneck nodes, vehicle speed, and queue length, comprehensively reflecting the real-time situation in the traffic network. A spatiotemporal deep learning model is used to dynamically predict demand from the panoramic traffic state map, outputting a pressure distribution matrix, which shows the distribution of vehicle flow, pedestrian flow, and road network bottleneck nodes. This information helps predict future traffic demand and identify potential traffic bottlenecks. This system effectively provides decision-making support for traffic control and resource allocation. Based on the road network's carrying capacity and pressure distribution matrix, it dynamically adjusts traffic control boundaries, enabling real-time adjustments to the control scope for different traffic conditions and avoiding resource waste or blind spots caused by static boundary divisions. Multiple control agents collaborate on decisions based on the pressure distribution matrix, outputting collaborative strategy packages, including tiered release strategies and multimodal resource allocation strategies. These strategies help rationally allocate traffic resources among different control subdomains, optimizing release order and traffic signal control. During the execution of the collaborative strategy packages, changes in the panoramic traffic state map are monitored in real-time, and closed-loop updates of the collaborative strategies are performed based on dissipation rate deviations. This allows for continuous optimization of decision-making strategies based on real-time traffic changes, thereby improving the stability and safety of traffic flow.

[0008] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the traffic dynamic collaborative control method based on multi-source heterogeneous data fusion provided in the embodiments of this application.

[0010] Figure 2 A schematic diagram of the traffic dynamic collaborative control system structure for multi-source heterogeneous data fusion provided in this application embodiment.

[0011] Figure labeling: Fine-grained data fusion module 10, blind spot state completion module 20, dynamic demand prediction module 30, traffic control boundary delineation module 40, multi-agent collaborative decision-making module 50, closed-loop update module 60. Detailed Implementation

[0012] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0013] Example 1, as Figure 1 As shown in the embodiments of this application, a traffic dynamic cooperative control method based on multi-source heterogeneous data fusion is provided, the method comprising: A100: After obtaining multi-source heterogeneous data streams from the controlled area through a cluster of ubiquitous sensing terminals, it performs fine-grained data fusion based on local spatiotemporal registration and outputs a spatiotemporal fusion state matrix.

[0014] The ubiquitous sensing terminal cluster refers to various sensors and sensing devices distributed throughout the entire controlled area, capable of acquiring traffic data in real time. These devices include, but are not limited to, cellular base stations, edge computing cameras, ride-hailing platform API gateways, and intelligent roadside units. Different types of data streams are collected in parallel by these devices, including mobile phone signaling spatiotemporal trajectories, video stream target detection data, real-time ride-hailing demand data, and time-series data from roadside sensors. Since these data come from different sensor devices and different time and spatial coordinate systems, spatiotemporal registration is needed to unify these heterogeneous data into a standardized spatiotemporal framework. By fusing these spatiotemporally registered data, data from different sources are merged on a unified time and spatial grid, ultimately forming a spatiotemporal fusion state matrix. This matrix represents the traffic status of the traffic controlled area at different times and spatial locations.

[0015] A200: Based on the road network topology, complete the blind spot state of the spatiotemporal fusion state matrix to generate a panoramic traffic state map.

[0016] Road network topology refers to the structural model of a transportation network, including road segments (edges) and intersections (nodes). This topological information allows for the analysis of traffic condition propagation paths and affected areas. Since data collection inherently contains blind spots—such as missing or incomplete data in certain areas—road network topology is used to fill these gaps. Specific methods include: assigning weights to different road segments and intersections in the road network topology based on real-time traffic conditions, reflecting information such as traffic flow, vehicle speed, and congestion; performing spatiotemporal diffusion processing on the data blind spots based on topological relationships; and filling in missing data in the blind spots through interpolation or other methods using surrounding non-blind spot data. After blind spot completion, a panoramic traffic condition map covering the entire transportation network is obtained, including traffic conditions, flow, and speed information at every moment and location. This provides a comprehensive foundation for subsequent traffic prediction and management.

[0017] A300: Based on a spatiotemporal deep learning model, dynamic demand prediction is performed on the panoramic traffic state map, and a pressure distribution matrix is ​​output, wherein the pressure distribution matrix includes vehicle flow pressure distribution, pedestrian flow pressure distribution and road network bottleneck node distribution.

[0018] Spatiotemporal deep learning models combine spatial, temporal, and topological information to analyze the dynamic changes in traffic networks. These models include: a spatial encoder, which extracts spatial features from traffic state maps using a 3D convolutional algorithm to obtain the characteristics of traffic congestion propagation in local areas; a temporal encoder, which uses a bidirectional long short-term memory network to capture periodic trends in time series and predict changes in traffic flow; and a topological encoder, which uses a graph attention network to analyze the interaction relationships between nodes in the road network topology to obtain the traffic impact between different road segments.

[0019] After model processing, the output pressure distribution matrix describes the traffic pressure in different areas and road segments of the traffic network. Specifically, it includes: vehicle flow pressure distribution, reflecting the vehicle flow density and congestion level of different road segments; pedestrian flow pressure distribution, reflecting the distribution of pedestrian density on road segments; and road network bottleneck node distribution, reflecting the bottleneck nodes in the traffic network. These nodes have high traffic pressure and are key areas of traffic congestion.

[0020] A400: Dynamically delineate traffic control boundaries based on the road network carrying capacity and the pressure distribution matrix.

[0021] Road network carrying capacity refers to the maximum traffic flow that each road segment or intersection in a transportation network can handle per unit of time. It depends on factors such as road design, number of lanes, traffic signal control, and traffic management measures. The pressure distribution matrix describes the traffic pressure in various areas of the transportation network. Areas with high pressure can lead to traffic congestion, delays, and even the risk of traffic accidents. Based on the pressure distribution matrix, areas with high pressure in the transportation network are identified, i.e., potential traffic bottlenecks. These bottlenecks are where traffic flow exceeds the road network's carrying capacity. Based on these high-pressure areas and the road network's carrying capacity, traffic control boundaries are dynamically delineated. This means that the scope of the traffic control area will vary under different traffic conditions. In areas with high pressure, such as traffic bottlenecks, the control area will be expanded to reduce further congestion and traffic problems.

[0022] A500: Based on the pressure distribution matrix, the traffic control boundary performs multi-agent collaborative decision-making and outputs a collaborative strategy package, wherein the collaborative strategy package includes a tiered release strategy and a multimodal resource scheduling strategy.

[0023] Multi-agent systems refer to systems composed of multiple agents (independent decision-making units). In traffic management, these agents are responsible for managing traffic control systems in different areas or road segments. They can make collaborative decisions under shared goals and resources. Collaborative decision-making means that multiple agents achieve common traffic management goals, such as reducing traffic congestion and improving traffic efficiency, through information sharing and coordination. Each agent makes decisions based on its own control area and traffic pressure, and exchanges information with other agents to achieve global optimization.

[0024] Tiered release strategy is a strategy of tiered management in areas with high traffic volume after the traffic control boundary is delineated. Specifically, different levels of release will be implemented in different areas according to their traffic pressure. For example, high-pressure areas will release high-priority vehicles first to reduce congestion. Multimodal resource scheduling strategy refers to the rational allocation of traffic resources according to the needs of different traffic modes. This scheduling strategy can effectively balance the resource allocation between different modes of transportation and avoid excessive congestion of one mode of transportation causing delays in other modes.

[0025] A600: During the execution of the cooperative strategy package, the traffic cooperative strategy is updated in a closed loop based on the real-time dissipation rate deviation of the panoramic traffic state map.

[0026] A panoramic traffic status map can display information such as traffic status, traffic flow, vehicle speed, and traffic density in various areas of the traffic network. Real-time dissipation rate deviation refers to the deviation between the dissipation rate of traffic flow in the traffic network and the expected rate. Generally speaking, dissipation rate refers to the speed at which traffic flow changes from a certain state, such as high-density congestion, to a low-density flow state.

[0027] During the execution of the collaborative strategy package, traffic conditions change in real time. By monitoring and updating the panoramic traffic status map in real time, the effectiveness of the current strategy is assessed. If the traffic flow dissipation rate is lower than expected, meaning traffic is not effectively alleviated, the current strategy needs adjustment. Based on the real-time dissipation rate deviation, the traffic collaborative strategy is updated in a closed loop. This means that the strategy is corrected according to the actual situation, enabling the traffic management system to adaptively respond to different traffic demands and changes.

[0028] Furthermore, after obtaining multi-source heterogeneous data streams from the controlled area through a ubiquitous sensing terminal cluster, fine-grained data fusion is performed based on local spatiotemporal registration to output a spatiotemporal fusion state matrix. The method includes: A110: Parallel acquisition of mobile phone signaling spatiotemporal trajectory, video stream target detection data, ride-hailing real-time demand data, and roadside sensor time-series data is performed through cellular base stations, edge computing cameras, ride-hailing platform API gateways, and intelligent roadside units in the ubiquitous sensing terminal cluster to obtain the multi-source heterogeneous data stream; A120: Local spatiotemporal registration is performed on the multi-source heterogeneous data stream to obtain a spatiotemporal registration state group; A130: After mapping the spatiotemporal registration state group to a unified time grid, grid-level data fusion is performed based on the source confidence weight allocation mechanism to obtain a spatiotemporal fusion state matrix.

[0029] Cellular base stations utilize mobile phone signaling data to capture the location information of moving vehicles and pedestrians. By analyzing signal strength and timing information, user movement trajectories can be inferred. Edge computing cameras, installed in key locations, monitor video streams and perform target detection, capturing real-time visual information such as traffic flow, pedestrian flow, and accident occurrences. Ride-hailing platform API gateways obtain real-time ride-hailing demand data from the platform's API, including vehicle location and call status, helping to clarify ride-hailing distribution and demand density. Intelligent roadside units include roadside sensor devices such as geomagnetic sensors, lidar, and traffic flow detectors, which can monitor traffic flow, vehicle speed, and lane occupancy in real time. Through parallel data acquisition from these data sources, various types of heterogeneous data streams are obtained.

[0030] Spatiotemporal registration maps data from different data sources into a unified spatiotemporal framework for analysis. Since these data sources differ in timestamps, spatial coordinate systems, sampling frequencies, and precision, they must first be aligned to ensure comparison and fusion at the same time and spatial location. For example, cellular base station signals provide data over a large area, while video stream data provides images within the range of a fixed-location camera. Therefore, spatiotemporal registration combines this information into a common spatiotemporal network, such as geographic coordinates and time windows, for further analysis. After registration, the data is organized into spatiotemporal registration state groups, which contain the correspondence between the various data sources within the spatiotemporal framework.

[0031] Mapping spatiotemporal registration state groups to a unified time grid ensures that all data is analyzed at a unified time and spatial scale. A unified time grid is a way to divide the control area into multiple time and spatial units. For example, the control area can be divided by cell block, grid, or geographical region, and time can be divided by minutes, seconds, or other units. This method allows data from different data sources to be accurately mapped to the same spatiotemporal coordinate system.

[0032] In the process of multi-source data fusion, the reliability and accuracy of each data source vary. For example, video stream data has high accuracy but is limited by the camera's position and field of view; mobile signaling data has wide coverage but lower accuracy. The source confidence weighting mechanism assigns different weights based on the reliability, accuracy, and credibility of each data source. During data fusion, data sources with higher credibility have a greater impact on the fusion result and therefore receive higher weights. Based on the assigned weights, the spatiotemporal registration state groups are fused. Within each time grid cell, data from different sources are fused, and the final result is adjusted according to the weights. This process involves various data fusion methods, such as weighted average and Bayesian inference. The final spatiotemporal fusion state matrix contains the traffic status of the controlled area in each time period and each spatial grid.

[0033] Furthermore, the method involves performing local spatiotemporal registration on the multi-source heterogeneous data streams to obtain a spatiotemporal registration state group, the method comprising: A121: Perform spatiotemporal density clustering on the spatiotemporal trajectory of mobile phone signaling to obtain the pedestrian density distribution; A122: Output the vehicle queue length by performing queue topology aggregation based on lane geometry constraints on the target detection data of the video stream; A123: Based on the spatiotemporal interpolation to complete the real-time demand data of ride-hailing vehicles, perform speed field reconstruction to obtain the lane-level speed matrix; A124: Perform multimodal event fusion on the time-series data of the roadside sensors to obtain the demand hotspot coordinates; wherein, the pedestrian density distribution, vehicle queue length, lane-level speed matrix, and demand hotspot coordinates constitute the spatiotemporal registration state group.

[0034] Mobile phone signaling spatiotemporal trajectories are signal location data of mobile phone users, reflecting the user's location changes over a certain period of time. Analyzing these trajectories using clustering algorithms such as K-means and DBSCAN groups trajectories with similar time and spatial locations into spatiotemporal density clusters. The purpose of clustering is to identify high-density areas, i.e., places where people are concentrated. Spatiotemporal density refers to the density of people per unit area per unit time. By clustering the trajectory data, the distribution of people density in each area at different time periods can be obtained, thereby identifying congested areas or areas with concentrated crowds. After clustering, the distribution of people in different areas at a specific moment or time period is obtained. High-density areas require priority for crowd control or safety management.

[0035] Video stream target detection data is video data collected by surveillance cameras installed on traffic sections. Target detection technology can be used to identify vehicles in the video and obtain traffic flow information for each lane. During video stream analysis, aggregation is performed based on lane geometric features, such as lane width and vehicle queuing patterns. Queue topology aggregation aggregates and analyzes multiple vehicles on a lane according to their queuing position and lane geometry to estimate queue length. Lane geometric constraints refer to the physical properties of the lanes, such as lane width and lane boundaries, which affect the queuing pattern and number of vehicles. This geometric information allows for a more accurate analysis of queue length and queue formation. Finally, by aggregating the target detection data, the queue length on each lane is calculated. This information reflects the severity of traffic congestion and provides timely flow information for traffic management, determining whether to implement traffic control or traffic management measures.

[0036] Real-time ride-hailing demand data originates from the APIs of ride-hailing platforms and contains real-time demand information. Since this data may be incomplete or missing, spatiotemporal interpolation algorithms are used to fill in the gaps. Spatiotemporal interpolation fills in the missing demand data using surrounding known data, thus forming a complete ride-hailing demand dataset. After the real-time demand data is complete, the speed field of the entire road segment is reconstructed using this data; that is, the speed distribution in different lanes and at different time periods. The speed field is an indicator of traffic flow, indicating the traffic conditions on different road segments. By reconstructing the speed field, a lane-level speed matrix containing multiple lanes and different time points is obtained, representing the speed distribution in different lanes.

[0037] Roadside sensor time-series data includes information such as traffic flow, vehicle speed, and occupancy rate, and this data is time-series, meaning it changes over time. Different types of roadside sensors collect data with varying time-series characteristics and data types. Multimodal event fusion refers to the joint analysis and fusion of data from different sensors to identify traffic events under different temporal and spatial conditions, such as vehicle queues, accidents, and road closures. By fusing sensor data, high-demand hotspot areas can be identified. Demand hotspots are areas with high traffic flow, high vehicle concentration, or traffic bottlenecks. The specific coordinates of these hotspot areas are output using the fused demand hotspot information. These coordinates are used to optimize traffic management strategies, such as adding traffic signal control in high-demand areas and optimizing road resource allocation.

[0038] The above information is combined through spatiotemporal registration to form a complete spatiotemporal registration state group, providing comprehensive support for subsequent traffic analysis, prediction, and control decisions.

[0039] Furthermore, after mapping the spatiotemporal registration state group to a unified time grid, grid-level data fusion is performed based on the source confidence weight allocation mechanism to obtain a spatiotemporal fusion state matrix. The method includes: A131: After dividing the control area based on a preset grid scale, a spatiotemporal grid matrix is ​​constructed to obtain the unified time grid. Each grid cell in the unified time grid has a spatial coordinate code and a time window identifier, and the integrated container configuration of each grid cell includes a crowd density container, a queue length container, a speed vector container, and a demand intensity container. A132: After mapping the spatiotemporal registration state group to the unified time grid according to the spatiotemporal coordinates, the integrated container configuration performs grid-level state fusion calculation based on the source confidence weight allocation mechanism to obtain the grid fusion state. A133: The grid fusion state is matrix-encapsulated to output the spatiotemporal fusion state matrix.

[0040] The entire control area is divided by a preset grid scale. The grid scale is set according to the actual traffic control needs. It is based on the spatial division of geographical areas and the division of time windows. The grid scale setting should accurately reflect the spatiotemporal characteristics of traffic and support efficient data storage and calculation.

[0041] Based on a preset grid scale, the entire traffic control area is divided into multiple small units, each of which is a grid cell. Each grid cell has not only spatial coordinates but also a time window identifier, enabling each grid cell to represent the traffic status at a specific time and spatial location. Each grid cell carries a spatial coordinate code and a time window identifier to accurately identify the time and spatial range of that grid cell.

[0042] Each grid cell is also configured with an integrated container, which is used to store the fusion results of multi-source data, including: a pedestrian density container, which stores pedestrian density information within the grid cell; a queue length container, which stores the length of vehicle queues; a speed vector container, which stores lane-level speed vector information; and a demand intensity container, which stores data such as ride-hailing demand intensity.

[0043] Spatiotemporal registration state groups are mapped onto a pre-constructed unified time grid. Data from each state group is placed in its corresponding grid cell, ensuring that each grid cell can store traffic information for a specific time period and spatial location. During data fusion, each data source has a different weight. For example, queue length data from video streams is more accurate than mobile signaling data; therefore, data sources with higher weights will have a greater impact on the final result. Source confidence refers to the assessment of the reliability of a data source. Different weights are assigned based on factors such as accuracy, sampling frequency, and reliability. Data sources with higher confidence have greater weight and influence.

[0044] The data within each grid cell is weighted and fused to calculate the fused traffic state. The data fusion method for each grid cell is as follows: pedestrian density data is weighted and averaged; vehicle queue data is weighted and processed using the median, which effectively addresses the impact of outliers and provides more robust results; lane speed data is weighted and fused, with speed vector fusion performed along lane direction to accurately reflect the directionality of traffic flow; and ride-hailing demand data is weighted and synthesized, using a maximum value aggregation method to reflect the maximum intensity of demand at a given moment. After completing the weighted data fusion, each grid cell obtains a grid fusion state, which contains the fused results of all traffic information within that grid.

[0045] The grid fusion state is organized into a multi-dimensional matrix according to a spatiotemporal coordinate system. The purpose of matrix encapsulation is to store spatiotemporal data in a structured manner for subsequent analysis, decision-making, and prediction. Ultimately, the output is a four-dimensional matrix (X×Y×T×4), where X and Y represent the coordinates of the geographic grid, specifically the rows and columns of the grid, i.e., the spatial division of the control area; T represents the sequence of time windows, with T being the time dimension, indicating the traffic state at different times; and 4 represents the four dimensions of the traffic state, specifically including pedestrian density, vehicle queue length, vehicle speed, and demand intensity.

[0046] Furthermore, based on the road network topology, blind spot state completion is performed on the spatiotemporal fusion state matrix to generate a panoramic traffic state map. The method includes: A210: Abstract road intersections as topological nodes and road segments as topological edges, and construct a connection relationship matrix based on the road network topology of the controlled area; A220: Calculate the weights of the road network topological edges according to the real-time traffic status of the controlled area, and dynamically assign weights to the connection relationship matrix to obtain a dynamic topological relationship matrix; A230: Perform data missing detection on the spatiotemporal fusion state matrix to locate the coordinates of the hierarchical blind zone grids; A240: Perform spatiotemporal state diffusion of physical constraints on the coordinates of the hierarchical blind zone grids according to the dynamic topological relationship matrix to generate the panoramic traffic state map.

[0047] In transportation networks, intersections are where traffic flows converge and disperse, typically crucial points in the road network. For topological analysis, road intersections are abstracted as topological nodes, each representing a convergence or divergence point of traffic flow, controlling traffic flow in different directions. Road segments, connecting two intersections, are the main paths of traffic flow and are abstracted as topological edges to represent the connections between nodes. Based on these node and edge abstractions, a connection matrix is ​​constructed, reflecting the topological structure of nodes and edges within the controlled area.

[0048] Real-time traffic conditions include traffic flow, speed, and density, reflecting the actual traffic situation. Each topological edge has its weight calculated based on the real-time traffic conditions. For example, road segments with high traffic flow and slow speeds have higher weights, indicating greater traffic pressure and that they are traffic bottlenecks. By dynamically assigning weights to each topological edge in the connectivity matrix, a dynamic topological matrix is ​​obtained. This matrix reflects the dynamic changes in the road network within the controlled area, representing the real-time traffic load and carrying capacity of each road segment.

[0049] Data missing detection identifies areas with missing data by analyzing blank or outlier values ​​in the spatiotemporal fusion state matrix. Missing data can affect subsequent analysis and decision-making, thus requiring special handling. Blind spots refer to areas where accurate information cannot be obtained due to data missingness, equipment problems, or other reasons. In the spatiotemporal matrix, blind spots can be divided into three categories: complete blind spots, where these grid cells have absolutely no data and traffic status cannot be inferred by any means; partial blind spots, where data in these grid cells is partially missing, but some information can be inferred from neighboring areas; and low-confidence blind spots, where data in these grid cells is incomplete, but some information can be inferred from other data sources, although with low confidence.

[0050] For complete and partial blind spots, missing data is supplemented using spatiotemporal diffusion. Spatiotemporal state diffusion is based on traffic state information of adjacent grid cells, and uses physical constraints, such as traffic flow and vehicle speed, to predict and fill in the state of missing areas. Physical constraints refer to diffusion based on traffic flow patterns and road network topology, combined with data from adjacent grid cells. Specific methods can include weighted interpolation, topology diffusion, and confidence weighting. After spatiotemporal diffusion processing, a panoramic traffic state map is finally generated, which displays the comprehensive traffic state within the controlled area.

[0051] Furthermore, based on the dynamic topology matrix, a spatiotemporal state diffusion of physical constraints is performed on the hierarchical blind zone grid coordinates to generate the panoramic traffic state map. The method includes: A241: Extract the distributed fusion state of adjacent non-blind zone grids from the spatiotemporal fusion state matrix as the state source based on the hierarchical blind zone grid coordinates; A242: Use the pedestrian density attenuation rule and vehicle speed transmission rule as diffusion constraints, and perform differentiated state propagation processing on the hierarchical blind zone grid coordinates based on the distributed fusion state to obtain the blind zone completion state. Among them, topology weighted interpolation is performed on the complete blind zone, effective container retention and missing container topology filling are performed on the partial blind zone, and confidence weighted fusion is performed on the low confidence blind zone; A243: Project the blind zone completion state onto the dynamic topology relationship matrix for spatiotemporal joint optimization, and output the panoramic traffic state map.

[0052] To fill in data gaps, information from adjacent non-blind-spot grids is used, and the traffic status of each grid cell reflects the traffic conditions of the area. Distributed fusion status refers to traffic status obtained by fusing information from multiple data sources. For adjacent non-blind-spot grids, these fused statuses provide crucial data sources, helping to predict and fill in data gaps. By extracting the traffic status of these adjacent grids as a status source, a basis is provided for completing the blind-spot data.

[0053] In areas with high pedestrian density, such as busy intersections, changes in pedestrian density affect the surrounding area. However, this effect gradually diminishes with increasing distance. When filling blind spots, the pedestrian density attenuation rule means that density information from neighboring non-blind spots gradually weakens with distance, resulting in a decrease in influence. Vehicle speed is determined by various factors such as traffic conditions, road conditions, and traffic signals. In traffic flow, speed changes are not localized but propagate along road segments. For blind spots, the vehicle speed transmission rule is used to determine how vehicle speed within the blind spot can be transmitted from neighboring grid cells, especially when traffic flow propagates to adjacent road segments.

[0054] Based on these diffusion constraints, a differentiated propagation algorithm is used to fill blind spots, employing different completion strategies for each type of blind spot: Complete blind spots, lacking any data, are filled using topological weighted interpolation. This method combines the topological structure and relative distance of the surrounding grid, using weighted interpolation to infer the blind spot's state. Partial blind spots, with partially missing data, are filled by preserving effective containers and topologically filling missing containers. Effective containers refer to those that can provide reliable data, while missing containers are filled through topological diffusion. Low-confidence blind spots, with incomplete data, are filled using a confidence-weighted fusion method. This means fusing data from different sources using weights adjusted according to the confidence level of the data sources. Through these diffusion constraints and differentiated propagation processing, the completed blind spot states are obtained, reflecting the traffic conditions within the blind spots.

[0055] Mapping blind spot completion status onto a dynamic topology matrix, spatiotemporal joint optimization refers to optimizing the distribution of traffic states by combining temporal and spatial information and integrating factors such as road network topology, traffic flow, vehicle speed, and lane capacity. This process aims to ensure that the data after blind spot completion matches the actual situation and constraints of the entire traffic network. After spatiotemporal joint optimization, the final generated panoramic traffic state map contains the traffic state of each grid unit within the control area, covering multiple dimensions of information such as pedestrian flow, vehicle flow, vehicle speed, and demand, providing a comprehensive view of the entire traffic network.

[0056] Furthermore, the method involves dynamically predicting demand based on a spatiotemporal deep learning model of the panoramic traffic state map, outputting a pressure distribution matrix. A310: Based on the spatial encoder constructed using a 3D convolution algorithm, the temporal encoder constructed using a bidirectional long short-term memory network, and the topological encoder constructed using a graph attention network, the outputs of the spatial encoder, the temporal encoder, and the topological encoder are connected through a gated feature fusion unit to obtain a multimodal fusion model framework; A320: The traffic flow conservation equation and the pedestrian flow speed constraint term are injected into the loss function layer of the multimodal fusion model framework to complete the construction of the spatiotemporal deep learning model.

[0057] The role of a spatial encoder is to extract spatial features from traffic data, especially traffic conditions in different geographic areas. 3D convolution is an extended convolution operation that performs convolution calculations in both spatial and temporal dimensions, effectively extracting features from spatial data. Through 3D convolution, both local and global features of spatial data can be captured. In traffic management, this feature extraction can obtain traffic conditions at different geographical locations, such as traffic density and vehicle flow.

[0058] The role of a time-series encoder is to extract dynamic features from time-series data, particularly the patterns of traffic flow and pedestrian flow over time. Bidirectional Long Short-Term Memory (BSSM) networks are deep learning models capable of processing sequential data and capturing long-term dependencies. BSSM networks can simultaneously consider information from the past to the future (forward propagation) and from the future to the past (backward propagation), enabling the model to comprehensively understand the temporal patterns of traffic conditions, especially the time-varying trends of traffic volume and pedestrian density.

[0059] The role of a topology encoder is to extract the relationship features between nodes (such as intersections) and edges (such as roads) in a traffic network. Graph attention networks are a type of graph-based neural network that can effectively process graph-structured data such as traffic networks. Graph attention networks assign different weights to each node in the graph with its neighboring nodes through an attention mechanism. This allows them to adaptively focus on more important neighboring nodes when learning the relationships between nodes. For example, the traffic conditions on some road segments have a greater impact on the flow of traffic on other road segments.

[0060] The gated feature fusion unit combines the outputs of the spatial encoder, temporal encoder, and topology encoder. This unit uses a gating mechanism to control how features from different sources are fused, resulting in a more accurate global traffic state representation. This unit combines spatial, temporal, and topological information extracted from different encoders using a weighted fusion method. The gating mechanism effectively allows the model to focus on more important features while reducing reliance on irrelevant or unimportant information.

[0061] Finally, the outputs of the spatial, temporal, and topology encoders are connected through a gated feature fusion unit to obtain a multimodal fusion model framework. This framework can simultaneously process data from different modalities, including spatial data, time-series data, and topological data, and comprehensively analyze traffic conditions from multiple perspectives.

[0062] The loss function is a core component in deep learning model training. It measures the difference between the model's predictions and the actual values. In the task of traffic dynamic collaborative control, the loss function includes constraints related to traffic flow and pedestrian speed to ensure that the model not only accurately predicts traffic conditions but also conforms to the physical laws of traffic flow.

[0063] The traffic flow conservation equation refers to the conservation law of vehicle flow in the entire road network. Simply put, the input of traffic flow equals the output of traffic flow, and the changes in traffic flow should conform to this conservation relationship. By adding the traffic flow conservation equation to the loss function, it can be ensured that the model follows the physical law of traffic flow conservation when predicting traffic conditions.

[0064] The pedestrian flow speed constraint is a constraint in the model prediction that takes into account that the speed of pedestrian movement should be within a certain range and is related to traffic conditions. This constraint ensures that the pedestrian flow speed predicted by the model does not exceed a reasonable range and is consistent with factors such as traffic density and road conditions in the surrounding environment.

[0065] By injecting the traffic flow conservation equation and pedestrian flow speed constraint into the loss function, the model not only aims to minimize the prediction error during training, but also ensures that the model output conforms to the physical and behavioral laws of traffic flow. This makes the spatiotemporal deep learning model not only accurate in generating prediction results, but also consistent with the dynamic characteristics of actual traffic flow.

[0066] Furthermore, the method involves dynamically predicting demand based on a spatiotemporal deep learning model of the panoramic traffic state map, outputting a pressure distribution matrix. A330: Decouple the panoramic traffic state map to obtain a geographic grid tensor, a time window sequence vector, a normalized state matrix, and a topological relationship graph; S1: Use the normalized state matrix as the basic state observation value and the geographic grid tensor as input to the spatial encoder of the spatiotemporal deep learning model, and extract local congestion propagation features through 3D convolution; S2: Use the normalized state matrix as the basic state observation value and the time window sequence vector as input to the temporal encoder of the spatiotemporal deep learning model, and capture periodic trend features through bidirectional LSTM; S3: Use the normalized state matrix as the basic state observation value and the time window sequence vector as input to the temporal encoder of the spatiotemporal deep learning model, and capture periodic trend features through bidirectional LSTM; The state matrix serves as the basic state observation value, and the topology graph is input into the topology encoder of the spatiotemporal deep learning model. Node interaction features are modeled through a graph attention network. A340: After parallel execution of steps S1 to S3, the local congestion propagation features, periodic trend features, and node interaction features are integrated through the gating feature fusion unit to output future state predictions. A350: Based on the traffic flow conservation equation and pedestrian speed constraints, the future state predictions are reconstructed to obtain the traffic flow pressure distribution, pedestrian flow pressure distribution, and road network bottleneck node distribution, which constitute the pressure distribution matrix.

[0067] The panoramic traffic status map contains comprehensive traffic data for the controlled area. By decoupling this data, it is broken down into several key components to facilitate model processing. Among them, the geographic grid tensor divides the traffic area into multiple spatial grids. Each grid represents the traffic status of a specific geographical location. Each grid not only contains spatial coordinate information but also road network hierarchy. Road network hierarchy refers to the road type or complexity of a particular grid, such as arterial roads or secondary roads, which affects the size and speed of traffic flow.

[0068] The time window sequence vector contains timestamps and period markers for traffic data. The timestamps refer to the specific time of each data point, while the period markers reflect the periodic characteristics of traffic conditions, such as daytime peak and off-peak periods, or seasonal changes, such as holiday traffic patterns.

[0069] The normalized state matrix normalizes traffic state data, such as pedestrian density, queue length, vehicle speed, and ride-hailing demand. Normalization helps ensure that different types of data are processed on the same scale, making the model training process more stable.

[0070] A topology graph describes the structure of a traffic network. It is a graph structure in which the node connection matrix represents the connection relationship between the nodes in the network, and the edge weight represents the traffic status of the road segment. The weight can be dynamically adjusted according to information such as traffic flow and speed to reflect the congestion level or carrying capacity of each road segment.

[0071] Each element in the normalized state matrix represents the traffic state of a grid cell at a given time. The geographic grid tensor contains the spatial coordinates and road network hierarchy of each grid cell, which helps the spatial encoder understand the geographical location and hierarchical relationships of each region in the traffic network. The spatial encoder uses a 3D convolution algorithm to process spatial data. 3D convolution performs convolution calculations on 3D data to extract local traffic features within a region. Here, 3D convolution helps extract local congestion propagation features, which refer to how traffic congestion propagates over time and space within a specific region, indicating the spread of traffic flow from one road segment to another. This is used to identify traffic bottleneck areas or predict potential congestion areas.

[0072] The task of a time-series encoder is to extract time-series features from traffic data, especially the changing patterns of traffic flow, vehicle speed, and queue length over time. Traffic flow exhibits periodic characteristics, such as the changes in traffic density during morning and evening rush hours. Bidirectional LSTM (Long Short-Term Memory) is an improved LSTM that can simultaneously consider the relationships between the past and the future, and between the future and the past. Through bidirectional LSTM, it can capture the periodic trends of traffic flow, identify patterns of increasing and decreasing traffic volume, and predict future changes in traffic conditions. Periodic trend features refer to the regular patterns of change in traffic flow, vehicle speed, etc., over time. For example, traffic flow patterns differ between weekdays and weekends, and traffic flow fluctuations also exhibit periodicity during holidays. Bidirectional LSTM learns from time window sequences to capture these periodic changes, thereby improving the accuracy of traffic flow prediction.

[0073] The task of a topology encoder is to extract the relationship features between nodes and edges in a traffic network, especially the interaction between intersections (nodes) and road segments (edges). In traffic flow, the traffic status of certain road segments or intersections has a significant impact on the traffic conditions of surrounding road segments or areas.

[0074] Graph Attention Networks (GANs) are deep learning networks capable of processing graph-structured data. By assigning different attention weights to nodes, GANs enable the model to focus on more important parts of the graph. In this way, GANs can learn the importance of each node based on the relationships and topology of nodes in a road network and extract interaction features between nodes. In traffic networks, GANs capture the mutual influence between intersections; for example, the traffic conditions at one node may affect the traffic flow at adjacent nodes. Node interaction features refer to the interaction of traffic flows between intersections. For instance, when congestion occurs at an intersection, it can lead to increased traffic flow on adjacent road segments and even spread to more distant areas. Through GANs, these interactions can be learned based on the topology of the traffic network, optimizing traffic flow prediction.

[0075] The spatial encoder, temporal encoder, and topology encoder process the input data in parallel. This means that spatial, temporal, and topological features are encoded independently and then fused together for the final prediction. The gated feature fusion unit fuses the features extracted from the spatial encoder, temporal encoder, and topology encoder, including local congestion propagation features, periodic trend features, and node interaction features. Through the gating mechanism, features from these different sources can be flexibly fused, highlighting key features and suppressing irrelevant noise. Finally, after fusing all features, the model outputs a future state prediction, that is, a prediction of the traffic state at a future moment or time period. This includes predictions of vehicle flow, pedestrian flow, speed, demand, and other information, helping to predict traffic flow and the formation of traffic bottlenecks.

[0076] The traffic flow conservation equation reflects the fundamental physical laws of traffic flow, namely that the input of traffic flow equals the output. Pedestrian speed constraints ensure that the predicted pedestrian speed matches the actual walking speed range, especially in densely populated areas such as intersections or stations. After obtaining the future state prediction, the prediction results are reconstructed based on the traffic flow conservation equation and pedestrian speed constraints. This reconstruction process ensures that the traffic state prediction not only conforms to the physical laws of traffic flow but also accurately reflects the actual state of traffic flow. The resulting traffic pressure distribution, pedestrian pressure distribution, and road network bottleneck node distribution constitute a pressure distribution matrix, which details the pressure status of the entire traffic network in the future.

[0077] Furthermore, the method involves performing multi-agent collaborative decision-making based on the pressure distribution matrix at the traffic control boundary and outputting a collaborative strategy package, the method comprising: A510: After expanding the road network topology based on the traffic control boundary, the control area is divided into N control subdomains; A520: N control agents are allocated to the N control subdomains, wherein each control agent is configured with control range boundary coordinates, a list of associated bottleneck nodes, and an available resource pool; A530: A predefined pressure threshold response mechanism and multimodal resource coupling rules are injected into the N control agents as decision rules; A540: After inputting the pressure distribution matrix into the N control agents, based on the regional pressure extraction based on the N control range boundary coordinates, pressure gradient focusing analysis is performed based on the list of associated bottleneck nodes, and capacity constraint modeling is performed based on the available resource pool, outputting N distributed release strategies and N distributed scheduling strategies; A550: Based on the adjacent boundary connection relationship, the N control agents interact with the N distributed release strategies and N distributed scheduling strategies to perform conflict resolution iteration, outputting the hierarchical release strategy and multimodal resource scheduling strategy as the collaborative strategy package output.

[0078] Traffic control boundaries refer to the boundaries of traffic management areas. Road network topology refers to the structure of the road system, including road connections and intersections. By analyzing the existing road network topology, it can be determined which road segments and intersections need to be controlled within the expanded control area. Based on the expanded road network topology, the entire control area is divided into multiple smaller control sub-domains. Each control sub-domain represents a relatively independent traffic management area, which can be a specific street block, road segment, or intersection area. N control sub-domains mean that the traffic control area is divided into N smaller areas, facilitating more refined management and scheduling.

[0079] A traffic control agent is an intelligent system or algorithm unit responsible for managing traffic in a specific subdomain. Within each subdomain, a dedicated traffic control agent is assigned to perform traffic control and decision-making. Each agent has clearly defined boundary coordinates for its management area, i.e., the specific geographical region or traffic segment it manages. These coordinates help determine the agent's management area and ensure that the agent only responds to traffic conditions within its managed area. Each agent also has a list of associated bottleneck nodes, which are road network nodes with high traffic volume and prone to congestion or other traffic problems, such as intersections with traffic lights and major traffic hubs. Each agent possesses a pool of available resources, including traffic lights, guidance screens, traffic guidance displays, and other equipment that can alter traffic flow or pace.

[0080] The pressure threshold response mechanism refers to setting a set of traffic pressure indicators for each controlled sub-domain. When the traffic pressure reaches a certain threshold, the corresponding response mechanism is activated. For example, when the traffic flow in a certain area exceeds the road segment's capacity, or when the traffic density at a bottleneck node is too high, emergency measures are taken, such as adjusting traffic signals, guiding vehicles to divert traffic, or dispatching police forces. The pressure threshold is set based on historical data, real-time traffic flow, and road segment capacity. When the traffic pressure reaches the predetermined threshold, the response mechanism is automatically triggered to carry out traffic diversion or flow control.

[0081] Multimodal resource coupling rules refer to how to effectively combine and schedule different types of traffic resources. In traffic management, the rational allocation and scheduling of these resources is key to achieving efficient traffic flow. Through predefined coupling rules, the control agent can flexibly schedule various resources according to different traffic conditions. For example, during peak traffic hours, the frequency of traffic light scheduling can be increased, and traffic flow can be guided through guidance screens; in the event of a traffic accident, police forces can be dispatched to quickly clear the traffic.

[0082] These predefined pressure threshold response mechanisms and multimodal resource coupling rules are injected into the control agent as the basis for decision-making. In real-time traffic monitoring and scheduling, the control agent will use these rules to comprehensively consider factors such as traffic flow, congestion, and bottleneck nodes to perform optimal resource allocation and traffic scheduling.

[0083] The pressure distribution matrix is ​​input into N control agents. Each control agent is responsible for traffic scheduling in its assigned control subdomain. The pressure distribution matrix contains traffic pressure information for different areas within each control subdomain, helping the agents identify areas of concentrated traffic flow and areas that require special attention.

[0084] Each control agent extracts regional pressure based on the boundary coordinates of its controlled area. These boundary coordinates help the agent identify traffic pressure sources within its controlled area, thereby enabling reasonable resource scheduling and pressure management.

[0085] The list of associated bottleneck nodes comprises key traffic bottleneck nodes identified by each control agent within its control area. These bottleneck nodes are the primary sources of traffic congestion, and agents need to prioritize the traffic conditions of these nodes. Pressure gradient focusing analysis refers to analyzing traffic pressure gradients to determine the pressure transmission path from high-pressure areas to low-pressure areas. By focusing on these pressure gradients, agents can optimize flow in bottleneck areas and reduce traffic congestion in high-pressure areas.

[0086] During traffic management, capacity constraints refer to the maximum traffic flow that each traffic facility can handle within a specific timeframe. The management agent needs to model available resources, taking into account the limitations and constraints of each resource. The available resource pool includes resources such as traffic lights, police force, and traffic guidance screens. The capacity and usage of these resources determine the effectiveness of traffic management. Through capacity constraint modeling, the agent can rationally allocate resources based on current traffic pressure to alleviate traffic bottlenecks and improve traffic flow.

[0087] Based on the above analysis, distributed release strategies and distributed scheduling strategies are output. These strategies aim to reduce congestion and improve traffic flow efficiency within the region. The distributed release strategy refers to each control agent designing an appropriate release strategy based on the traffic pressure in its area. For example, in high-pressure areas, the agent might increase the green light duration or prioritize the release of traffic in a certain direction. The distributed scheduling strategy refers to scheduling strategies designed based on traffic pressure and resource availability. For example, this might involve dispatching traffic police, adjusting information on guidance screens, or changing traffic signals.

[0088] Each control agent not only manages traffic in its own region, but also needs to collaborate with control agents in neighboring regions. In this step, the adjacent boundary connection relationship is used to describe the boundaries and interactions between the control subdomains. For example, the traffic conditions in one region will affect the traffic flow in the neighboring regions.

[0089] When formulating distributed traffic release and distributed scheduling strategies, each control agent considers not only the situation in its own area but also the strategies of neighboring areas. Conflicts may arise between strategies from different areas; for example, two adjacent roads might both prioritize certain traffic flows, leading to traffic control conflicts. Conflict resolution iteration refers to gradually eliminating these conflicts through interaction between agents. For instance, agents can achieve optimal coordination by coordinating traffic light timings, adjusting police deployment, or sharing guidance information.

[0090] Tiered release strategies are optimized strategies derived from multiple conflict resolution processes. They rationally allocate release time and resources across different pressure zones. This means that in areas with high pressure, high-priority traffic flows, such as public transportation or emergency vehicles, are released first, while in areas with low pressure, regular traffic flow is released. Multimodal resource scheduling strategies refer to the coordinated scheduling of multiple resources to optimize their use. For example, in areas with high traffic pressure, traffic guidance screens are prioritized for traffic guidance, or police forces are deployed manually to manage traffic as needed.

[0091] Through the output of tiered release strategies and multimodal resource scheduling strategies, each control agent ultimately outputs a collaborative strategy package. The collaborative strategy package contains the optimal decision for the current traffic conditions, aiming to globally optimize traffic flow, reduce congestion, and improve efficiency.

[0092] Example 2, based on the same inventive concept as the traffic dynamic cooperative control method of multi-source heterogeneous data fusion in the foregoing examples, such as... Figure 2 As shown in the embodiment of this application, a traffic dynamic cooperative control system based on multi-source heterogeneous data fusion is provided, the system comprising: The fine-grained data fusion module 10 is used to obtain multi-source heterogeneous data streams from the control area through a ubiquitous sensing terminal cluster, and then perform fine-grained data fusion based on local spatiotemporal registration to output a spatiotemporal fusion state matrix. The blind spot state completion module 20 is used to complete the blind spot states of the spatiotemporal fusion state matrix based on the road network topology, generating a panoramic traffic state map. The dynamic demand prediction module 30 is used to predict the dynamic demand of the panoramic traffic state map based on a spatiotemporal deep learning model, outputting a pressure distribution matrix, wherein the pressure distribution matrix includes vehicle flow pressure distribution and pedestrian flow pressure distribution. The system includes: a pressure distribution and road network bottleneck node distribution module; a traffic control boundary delineation module 40, used to dynamically delineate the traffic control boundary based on the road network carrying capacity and the pressure distribution matrix; a multi-agent collaborative decision-making module 50, used to perform multi-agent collaborative decision-making based on the pressure distribution matrix at the traffic control boundary and output a collaborative strategy package, wherein the collaborative strategy package includes a graded release strategy and a multimodal resource scheduling strategy; and a closed-loop update module 60, used to perform closed-loop updates of the traffic collaborative strategy based on the real-time dissipation rate deviation of the panoramic traffic state map during the execution of the collaborative strategy package.

[0093] Furthermore, the fine-grained data fusion module 10 is used to perform the following operation steps: The multi-source heterogeneous data stream is obtained by parallel acquisition of mobile phone signaling spatiotemporal trajectory, video stream target detection data, real-time demand data of ride-hailing vehicles, and time-series data of roadside sensors through cellular base stations, edge computing cameras, ride-hailing platform API gateways, and intelligent roadside units in the ubiquitous sensing terminal cluster. The multi-source heterogeneous data stream is then locally spatiotemporally registered to obtain a spatiotemporal registration state group. After mapping the spatiotemporal registration state group to a unified time grid, grid-level data fusion is performed based on the source confidence weight allocation mechanism to obtain a spatiotemporal fusion state matrix.

[0094] Furthermore, the fine-grained data fusion module 10 is used to perform the following operation steps: Spatiotemporal density clustering is performed on the spatiotemporal trajectory of mobile phone signaling to obtain the pedestrian density distribution; queue topology aggregation based on lane geometry constraints is performed on the target detection data of video stream to output the vehicle queue length; based on the spatiotemporal interpolation to complete the real-time demand data of ride-hailing vehicles, speed field reconstruction is performed to obtain the lane-level speed matrix; multimodal event fusion is performed on the time series data of the roadside sensors to obtain the demand hotspot coordinates; wherein, the pedestrian density distribution, vehicle queue length, lane-level speed matrix and demand hotspot coordinates constitute the spatiotemporal registration state group.

[0095] Furthermore, the fine-grained data fusion module 10 is used to perform the following operation steps: After dividing the control area based on a preset grid scale, a spatiotemporal grid matrix is ​​constructed to obtain the unified time grid. Each grid cell in the unified time grid has a spatial coordinate code and a time window identifier, and the integrated container configuration of each grid cell includes a crowd density container, a queue length container, a velocity vector container, and a demand intensity container. After mapping the spatiotemporal registration state group to the unified time grid according to the spatiotemporal coordinates, the integrated container configuration performs grid-level state fusion calculation based on the source confidence weight allocation mechanism to obtain the grid fusion state. The grid fusion state is then matrix-encapsulated to output the spatiotemporal fusion state matrix.

[0096] Furthermore, the blind spot state completion module 20 is used to perform the following operation steps: Road intersections are abstracted as topological nodes, and road segments are abstracted as topological edges. A connection relationship matrix is ​​constructed based on the road network topology of the controlled area. According to the real-time traffic status of the controlled area, the weights of the road network topological edges are calculated, and dynamic weight assignment is performed on the connection relationship matrix to obtain a dynamic topological relationship matrix. Data missing detection is performed on the spatiotemporal fusion state matrix to locate the coordinates of the hierarchical blind zone grids. According to the dynamic topological relationship matrix, spatiotemporal state diffusion with physical constraints is performed on the coordinates of the hierarchical blind zone grids to generate the panoramic traffic state map.

[0097] Furthermore, the blind spot state completion module 20 is used to perform the following operation steps: Based on the hierarchical blind zone grid coordinates, the distributed fusion state of adjacent non-blind zone grids is extracted from the spatiotemporal fusion state matrix as the state source; using the pedestrian density attenuation rule and vehicle speed transmission rule as diffusion constraints, differentiated state propagation processing is performed on the hierarchical blind zone grid coordinates based on the distributed fusion state to obtain the blind zone completion state. Among them, topology weighted interpolation is performed on complete blind zones, effective container retention and missing container topology filling are performed on partial blind zones, and confidence weighted fusion is performed on low-confidence blind zones; the blind zone completion state is projected onto the dynamic topology relation matrix for spatiotemporal joint optimization, and the panoramic traffic state map is output.

[0098] Furthermore, the dynamic demand forecasting module 30 is used to perform the following operation steps: Based on the spatial encoder constructed using a 3D convolution algorithm, the temporal encoder constructed using a bidirectional long short-term memory network, and the topology encoder constructed using a graph attention network, a multimodal fusion model framework is obtained by connecting the outputs of the spatial encoder, the temporal encoder, and the topology encoder through a gated feature fusion unit. The vehicle flow conservation equation and the pedestrian flow speed constraint term are injected into the loss function layer of the multimodal fusion model framework to complete the construction of the spatiotemporal deep learning model.

[0099] Furthermore, the dynamic demand forecasting module 30 is used to perform the following operation steps: Decouple the panoramic traffic state map to obtain a geographic grid tensor, a time window sequence vector, a normalized state matrix, and a topology graph; S1: Input the normalized state matrix as the basic state observation value and the geographic grid tensor into the spatial encoder of the spatiotemporal deep learning model, and extract local congestion propagation features through 3D convolution; S2: Input the normalized state matrix as the basic state observation value and the time window sequence vector into the temporal encoder of the spatiotemporal deep learning model, and capture periodic trend features through bidirectional LSTM; S3: Input the normalized state matrix as the basic state observation value and the topology graph into the topology encoder of the spatiotemporal deep learning model, and model node interaction features through graph attention network; After executing steps S1 to S3 in parallel, integrate the local congestion propagation features, periodic trend features, and node interaction features through the gating feature fusion unit to output future state prediction; Reconstruct the future state prediction based on the traffic flow conservation equation and pedestrian flow speed constraints to obtain the traffic flow pressure distribution, pedestrian flow pressure distribution, and road network bottleneck node distribution, which constitute the pressure distribution matrix.

[0100] Furthermore, the multi-agent collaborative decision-making module 50 is used to perform the following operational steps: After expanding the road network topology based on the traffic control boundary, the control area is divided into N control subdomains. N control agents are assigned to these N subdomains, each agent configured with control range boundary coordinates, a list of associated bottleneck nodes, and an available resource pool. A predefined pressure threshold response mechanism and multimodal resource coupling rules are injected into the N control agents as decision rules. After inputting the pressure distribution matrix into the N control agents, based on regional pressure extraction using the N control range boundary coordinates, pressure gradient focusing analysis is performed based on the list of associated bottleneck nodes, and capacity constraint modeling is performed based on the available resource pool, outputting N distributed release strategies and N distributed scheduling strategies. Based on adjacent boundary connections, the N control agents interact with the N distributed release strategies and N distributed scheduling strategies to perform conflict resolution iterations, outputting the hierarchical release strategy and multimodal resource scheduling strategy as the collaborative strategy package.

[0101] Through the foregoing detailed description of the traffic dynamic cooperative control method based on multi-source heterogeneous data fusion, those skilled in the art can clearly understand the traffic dynamic cooperative control system based on multi-source heterogeneous data fusion in this embodiment. Since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and relevant parts can be referred to the method section.

[0102] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A traffic dynamic cooperative control method based on multi-source heterogeneous data fusion, characterized in that, The method includes: After obtaining multi-source heterogeneous data streams from the control area through a ubiquitous sensing terminal cluster, fine-grained data fusion is performed based on local spatiotemporal registration to output a spatiotemporal fusion state matrix. Based on the road network topology, the blind spot state of the spatiotemporal fusion state matrix is ​​filled in to generate a panoramic traffic state map. Dynamic demand prediction of the panoramic traffic state map is performed based on a spatiotemporal deep learning model, and a pressure distribution matrix is ​​output, wherein the pressure distribution matrix includes vehicle flow pressure distribution, pedestrian flow pressure distribution and road network bottleneck node distribution. Traffic control boundaries are dynamically delineated based on the road network carrying capacity and the pressure distribution matrix. Based on the pressure distribution matrix, a multi-agent collaborative decision is made at the traffic control boundary, and a collaborative strategy package is output, wherein the collaborative strategy package includes a tiered release strategy and a multimodal resource scheduling strategy. During the execution of the collaborative strategy package, the traffic collaborative strategy is updated in a closed loop based on the real-time dissipation rate deviation of the panoramic traffic state map. 2.The multi-source heterogeneous data fusion traffic dynamics cooperative control method of claim 1, wherein, After obtaining multi-source heterogeneous data streams from the controlled area through a ubiquitous sensing terminal cluster, fine-grained data fusion is performed based on local spatiotemporal registration to output a spatiotemporal fusion state matrix. The method includes: The multi-source heterogeneous data stream is obtained by parallel collection of mobile phone signaling spatiotemporal trajectory, video stream target detection data, real-time demand data of ride-hailing vehicles, and time-series data of roadside sensors through cellular base stations, edge computing cameras, ride-hailing platform API gateways, and intelligent roadside units in the ubiquitous sensing terminal cluster. Local spatiotemporal registration is performed on the multi-source heterogeneous data streams to obtain a spatiotemporal registration state group; After mapping the spatiotemporal registration state group to a unified time grid, grid-level data fusion is performed based on the source confidence weight allocation mechanism to obtain the spatiotemporal fusion state matrix. 3.The multi-source heterogeneous data fusion traffic dynamics cooperative control method of claim 2, wherein, The method involves performing local spatiotemporal registration on the multi-source heterogeneous data streams to obtain a spatiotemporal registration state group, the method comprising: Spatiotemporal density clustering of mobile phone signaling trajectories yields the distribution of pedestrian density. The vehicle queue length is output by performing queue topology aggregation based on lane geometry constraints on the target detection data in the video stream. Based on the spatiotemporal interpolation to complete the real-time demand data of ride-hailing vehicles, a speed field reconstruction is performed to obtain a lane-level speed matrix; Multimodal event fusion is performed on the time-series data from the roadside sensors to obtain the coordinates of the demand hotspots; The pedestrian density distribution, vehicle queue length, lane-level speed matrix, and demand hotspot coordinates constitute the spatiotemporal registration state group.

4. The traffic dynamic collaborative control method based on multi-source heterogeneous data fusion as described in claim 2, characterized in that, After mapping the spatiotemporal registration state group to a unified time grid, grid-level data fusion is performed based on the source confidence weight allocation mechanism to obtain the spatiotemporal fusion state matrix. The method includes: After dividing the control area based on a preset grid scale, a spatiotemporal grid matrix is ​​constructed to obtain the unified time grid. Each grid cell in the unified time grid has a spatial coordinate code and a time window identifier, and the integrated container configuration of each grid cell includes a crowd density container, a queue length container, a velocity vector container, and a demand intensity container. After mapping the spatiotemporal registration state group to the unified time grid according to the spatiotemporal coordinates, the integrated container configuration performs grid-level state fusion calculation based on the source confidence weight allocation mechanism to obtain the grid fusion state. The grid fusion state is matrix-encapsulated to output the spatiotemporal fusion state matrix.

5. The traffic dynamic collaborative control method based on multi-source heterogeneous data fusion as described in claim 1, characterized in that, The method involves completing the blind spot states of the spatiotemporal fusion state matrix based on the road network topology to generate a panoramic traffic state map. Road intersections are abstracted as topological nodes, and road segments are abstracted as topological edges. A connection matrix is ​​constructed based on the road network topology of the controlled area. Based on the real-time traffic status of the controlled area, the road network topology edge weights are calculated, and the dynamic weights of the connection relationship matrix are dynamically assigned to obtain the dynamic topology relationship matrix. Data missing detection is performed on the spatiotemporal fusion state matrix to locate the grid coordinates of hierarchical blind zones; Based on the dynamic topology matrix, the spatiotemporal state diffusion of physical constraints is performed on the hierarchical blind zone grid coordinates to generate the panoramic traffic state map.

6. The traffic dynamic collaborative control method based on multi-source heterogeneous data fusion as described in claim 5, characterized in that, Based on the dynamic topological relationship matrix, a spatiotemporal state diffusion of physical constraints is performed on the hierarchical blind zone grid coordinates to generate the panoramic traffic state map. The method includes: Based on the coordinates of the hierarchical blind zone grid, the distributed fusion state of the adjacent non-blind zone grid is extracted from the spatiotemporal fusion state matrix as the state source; Using the pedestrian density attenuation rule and the vehicle speed transmission rule as diffusion constraints, differential state propagation processing is performed on the hierarchical blind zone grid coordinates based on the distributed fusion state to obtain the blind zone completion state. Specifically, topology weighted interpolation is performed on the complete blind zone, effective container retention and missing container topology filling are performed on the partial blind zone, and confidence weighted fusion is performed on the low-confidence blind zone. The blind spot completion state is projected onto the dynamic topology matrix for spatiotemporal joint optimization, and the panoramic traffic state map is output.

7. The traffic dynamic collaborative control method based on multi-source heterogeneous data fusion as described in claim 1, characterized in that, The method involves dynamically predicting demand based on a spatiotemporal deep learning model of the panoramic traffic state map and outputting a pressure distribution matrix. Based on the spatial encoder constructed using a 3D convolution algorithm, the temporal encoder constructed using a bidirectional long short-term memory network, and the topology encoder constructed using a graph attention network, a multimodal fusion model framework is obtained by connecting the outputs of the spatial encoder, the temporal encoder, and the topology encoder through a gated feature fusion unit. The vehicle flow conservation equation and pedestrian flow speed constraint term are injected into the loss function layer of the multimodal fusion model framework to complete the construction of the spatiotemporal deep learning model.

8. The traffic dynamic collaborative control method based on multi-source heterogeneous data fusion as described in claim 7, characterized in that, The method involves dynamically predicting demand based on a spatiotemporal deep learning model of the panoramic traffic state map and outputting a pressure distribution matrix. Decouple the panoramic traffic state map to obtain the geographic grid tensor, time window sequence vector, normalized state matrix and topological relationship graph; S1: The normalized state matrix is ​​used as the basic state observation value, and the geographic grid tensor is input into the spatial encoder of the spatiotemporal deep learning model to extract local congestion propagation features through three-dimensional convolution. S2: The normalized state matrix is ​​used as the basic state observation value, and the time window sequence vector is input into the temporal encoder of the spatiotemporal deep learning model to capture periodic trend features through bidirectional LSTM; S3: The normalized state matrix is ​​used as the basic state observation value, and the topology graph is input into the topology encoder of the spatiotemporal deep learning model to model the node interaction features through a graph attention network; After executing steps S1 to S3 in parallel, the local congestion propagation features, periodic trend features, and node interaction features are integrated by the gating feature fusion unit to output a future state prediction. Based on the traffic flow conservation equation and pedestrian flow speed constraints, the future state prediction is reconstructed to obtain the traffic flow pressure distribution, pedestrian flow pressure distribution, and road network bottleneck node distribution, which constitute the pressure distribution matrix.

9. The traffic dynamic collaborative control method based on multi-source heterogeneous data fusion as described in claim 1, characterized in that, The method involves performing multi-agent collaborative decision-making based on the pressure distribution matrix at the traffic control boundary and outputting a collaborative strategy package, comprising: After expanding the road network topology based on the traffic control boundary, the control area is divided into N control subdomains; N control agents are allocated in the N control subdomains, wherein each control agent is configured with control range boundary coordinates, a list of associated bottleneck nodes, and an available resource pool. A predefined pressure threshold response mechanism and multimodal resource coupling rules are injected into the N management and control agents as decision rules; After inputting the pressure distribution matrix into the N control agents, based on the regional pressure extraction of the N control range boundary coordinates, pressure gradient focusing analysis is performed based on the list of associated bottleneck nodes, and capacity constraint modeling is performed based on the available resource pool, outputting N distributed release strategies and N distributed scheduling strategies. The N control agents interact based on adjacent boundary connections, execute conflict resolution iterations through the N distributed release strategies and N distributed scheduling strategies, and output the hierarchical release strategy and multimodal resource scheduling strategy as the collaborative strategy package.

10. A traffic dynamic cooperative control system based on multi-source heterogeneous data fusion, characterized in that, The system is used to implement the traffic dynamic collaborative control method based on multi-source heterogeneous data fusion according to any one of claims 1-9, the system comprising: The fine-grained data fusion module is used to obtain multi-source heterogeneous data streams from the control area through the ubiquitous sensing terminal cluster, and then perform fine-grained data fusion based on local spatiotemporal registration to output a spatiotemporal fusion state matrix. The blind spot state completion module is used to complete the blind spot state of the spatiotemporal fusion state matrix based on the road network topology, and generate a panoramic traffic state map. The dynamic demand prediction module is used to predict the dynamic demand of the panoramic traffic state map based on the spatiotemporal deep learning model and output a pressure distribution matrix, wherein the pressure distribution matrix includes vehicle flow pressure distribution, pedestrian flow pressure distribution and road network bottleneck node distribution. The traffic control boundary delineation module is used to dynamically delineate the traffic control boundary based on the road network carrying capacity and the pressure distribution matrix. A multi-agent collaborative decision-making module is used to make multi-agent collaborative decisions based on the pressure distribution matrix at the traffic control boundary and output a collaborative strategy package, wherein the collaborative strategy package includes a tiered release strategy and a multimodal resource scheduling strategy. The closed-loop update module is used to perform closed-loop updates of the traffic coordination strategy based on the real-time dissipation rate deviation of the panoramic traffic state map during the execution of the coordination strategy package.

Images