A multi-target tracking and space-time prediction regional intelligent transportation system

Abstract

This invention discloses a regional intelligent transportation system with multi-target tracking and spatiotemporal prediction, comprising a core framework module, an external interface module, an AI module, a prediction module, a node management module, a map interaction module, and a data management module. The data management module acquires real-time and historical data on traffic parameters, vehicle travel directions, and vehicle information. Based on the deep integration of the YOLO target detection framework and the ByteTrack multi-target tracking algorithm, it fuses spatiotemporal trajectories with real-time traffic parameters, monitors collision events, sharp turns, and sudden braking events in real time, outputs anomaly indices and event types, and predicts future congestion levels. This invention solves the problems of data fragmentation, response lag, and disconnect between prediction and decision-making in existing systems, achieving closed-loop management of "perception-prediction-decision," and improving the real-time performance of traffic situation awareness, prediction accuracy, and the intelligence level of decision-making.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, specifically to a regional intelligent transportation monitoring system with multi-target tracking and spatiotemporal prediction. Background Technology

[0002] With the acceleration of urbanization, intelligent transportation systems are gradually becoming a core tool for improving road efficiency and safety. Existing systems typically focus on single functional modules, such as vehicle detection, traffic parameter calculation, or congestion prediction, achieving localized optimization through independent technologies and providing basic support for traffic management. However, these systems currently suffer from the following significant shortcomings:

[0003] 1. Data fragmentation problem: Detection, tracking, and prediction modules operate independently, failing to form a closed-loop collaborative mechanism. For example, after identifying abnormal vehicle behavior, the system only records the data but cannot link it with the congestion prediction module, requiring manual intervention for decision-making.

[0004] 2. Response lag: Traditional target detection algorithms are not robust enough in complex scenarios and are easily affected by lighting and occlusion, which can lead to vehicle trajectory breaks and thus affect the real-time performance and accuracy of traffic parameter calculation.

[0005] 3. Disconnect between prediction and decision-making: Congestion prediction relies on static historical data and does not integrate real-time traffic parameters and the status of surrounding nodes, such as vehicle speed and traffic flow, causing the prediction results to deviate from the actual road conditions and failing to support dynamic scheduling.

[0006] 4. Insufficient intelligent interaction: The operation relies on a professional interface and lacks natural language interaction capabilities. Users need to have knowledge in this field to analyze complex traffic data and find it difficult to quickly obtain decision support.

[0007] 5. Small Sample Size Challenge: While ground-based monitoring equipment plays a crucial role in traffic flow monitoring, vehicle speed calculation, and abnormal behavior detection within current intelligent transportation systems, its inherent limitations lead to problems such as monitoring blind spots and insufficient data coverage. Ground-based monitoring equipment suffers from limitations including limited detection range, poor environmental adaptability, and delayed data updates.

[0008] Therefore, there is an urgent need to build a regional intelligent transportation system that integrates multi-modal collaboration, real-time tracking of multiple targets, spatiotemporal map prediction, and AI interaction to achieve closed-loop management of "perception-decision-response". Summary of the Invention

[0009] The purpose of this invention is to provide a regional intelligent transportation system with multi-target tracking and spatiotemporal prediction, which aims to solve the problems of data fragmentation, response lag, disconnect between prediction and decision-making, insufficient interactive intelligence, and small sample size in existing intelligent transportation systems.

[0010] The objective of this invention is achieved through the following technical solution:

[0011] A regional intelligent transportation system with multi-target tracking and spatiotemporal prediction includes a core framework module, an external interface module, an AI module, a prediction module, a node management module, a map interaction module, and a data management module.

[0012] The core framework module is connected to the data management module, the external interface module, the AI ​​module, and the map interaction module, respectively. The node management module and the map interaction module are connected to the prediction module, respectively. The data management module is connected to the AI ​​module, the visualization module, the prediction module, and the map interaction module, respectively, and is used to provide data support for the AI ​​module, the visualization module, the prediction module, and the map interaction module.

[0013] The core framework module is used for system scheduling, encapsulation of inter-module communication protocols, processing of request and response, and access to the database.

[0014] The external interface module is used to connect to meteorological data APIs, traffic information platforms, and third-party map services;

[0015] The AI ​​module integrates a natural language understanding engine and uses a bidirectional Transformer architecture to parse user-input traffic semantics to generate structured query instructions. Based on the query instructions, it performs multi-source data joint reasoning based on the data from the data management module, knowledge graph, and rule engine, and outputs root cause analysis conclusions or traffic optimization suggestions as decision results.

[0016] The data management module is used to collect traffic video streams and meteorological sensor data. It obtains real-time and historical data of traffic parameters, vehicle driving direction, vehicle information, and equipment status data from the traffic video streams. Based on the deep integration of the YOLO target detection framework and the ByteTrack multi-target tracking algorithm, it constructs a full-link processing capability covering vehicle perception, parameter calculation, and behavior analysis. At the same time, it generates disturbance-resistant traffic parameters based on a dynamic calibration algorithm and integrates a multi-level anomaly detection model to output a structured data stream with spatiotemporal tags. By integrating spatiotemporal trajectories and real-time traffic parameters, it identifies and monitors collision events, sharp turning behaviors, and emergency braking events through the intersection-exchange-comparison (IoC) algorithm, and outputs anomaly indices and event types.

[0017] The map interaction module is used to obtain system pixel coordinates, provide a node relationship configuration interface, respond to map click events to perform node creation, drag-and-drop positioning and deletion operations, construct a weighted connection matrix with spatiotemporal constraints, define traffic direction rules and dynamic weight coefficients, generate topology difference data packets in real time and push them to the prediction module, and a region division tool.

[0018] The node management module is used to dynamically configure each node according to the area calibration data of the map interaction module, set cameras and detection parameters at each node, establish a spatial relationship model between each node, dynamically store road connection directions, node device connection status and traffic capacity weights, maintain the topological relationship of camera nodes and send it to the prediction module, and diagnose the online status of the devices in real time.

[0019] The prediction module is used to perform node-level congestion calculation based on spatiotemporal graph neural network based on historical weather, traffic parameters, real-time weather, time characteristics, and map topology relationship data provided by the data management module and the map interaction module, and to predict the future congestion level.

[0020] Beneficial effects of this invention:

[0021] This invention dynamically maintains the camera topology through a node management module and achieves road network grid modeling by combining the regional calibration unit of the map interaction module, providing a topological foundation for spatiotemporal prediction. A data acquisition module integrates video streams and meteorological sensor data to provide data support for the data processing module. The prediction module, based on a spatiotemporal graph neural network, utilizes a weighted directed edge model generated by a graph construction unit, integrates real-time traffic parameters, meteorological data, and historical congestion characteristics, and outputs a future congestion probability value at a certain time node. The AI ​​module parses user commands through a natural language understanding unit, links with a decision-making reasoning unit to generate optimization suggestions, and drives a report generation unit to automatically integrate and generate graphic reports. The core framework module achieves millisecond-level communication through the gRPC protocol and, combined with a state coordinator unit, completes fault circuit interruption, forming a closed loop of "data acquisition-processing-prediction-decision." The data processing module, based on the data obtained from the data acquisition module, processes the data in real time using YOLOv11-ByteTrack technology to achieve vehicle identification and trajectory tracking, outputting a pixel-level structured data stream with vehicle coordinates, corresponding vehicle category labels, and timestamps, based on which traffic parameters are calculated and abnormal behavior is detected. This invention constructs a smart transportation system covering the entire chain of perception, prediction, and decision-making through a multi-level collaborative architecture and a spatiotemporal data fusion mechanism.

[0022] Multi-level technology collaboration enables a leap in the perception-decision closed loop: The data processing module improves vehicle detection rate in rainy and foggy environments through YOLOv11's C2PSA spatial attention mechanism, and reduces trajectory breakage rate by combining ByteTrack's low-confidence detection box reuse strategy; the spatiotemporal convolutional unit integrates real-time vehicle speed, abnormal events, and meteorological factors from surrounding nodes to reduce congestion prediction errors; the dynamic topology modeling of the map interaction module and the real-time correction of traffic weights by the node management module form a road network optimization closed loop; the AI ​​module generates structured query instructions through the Transformer architecture, reducing user interaction steps; and the gRPC protocol of the core framework, combined with the fault circuit interruption mechanism of the state coordinator unit, ensures system availability.

[0023] An integrated platform suitable for real-time perception and storage of traffic data, congestion prediction, multimodal decision support, and regional collaborative management. Attached Figure Description

[0024] Figure 1 This is a block diagram of a specific embodiment 1 of the present invention;

[0025] Figure 2 This is a block diagram of the core framework module of a specific embodiment 1 of the present invention;

[0026] Figure 3 This is a block diagram of the user login module in a specific embodiment 1 of the present invention;

[0027] Figure 4 This is a block diagram of the external interface module in a specific embodiment 1 of the present invention;

[0028] Figure 5 This is a block diagram of the AI ​​module in a specific embodiment 1 of the present invention;

[0029] Figure 6 This is a block diagram of the visualization module in a specific embodiment 1 of the present invention;

[0030] Figure 7 This is a block diagram of the prediction module in a specific embodiment 1 of the present invention;

[0031] Figure 8 This is a block diagram of the node management module in a specific embodiment of the present invention;

[0032] Figure 9 This is a block diagram of the map interaction module in specific embodiment 1 of the present invention;

[0033] Figure 10 This is a block diagram of the data acquisition module in a specific embodiment 1 of the present invention;

[0034] Figure 11 This is a block diagram of the data storage module in a specific embodiment 1 of the present invention;

[0035] Figure 12 This is a block diagram of the data processing module in a specific embodiment 1 of the present invention.

[0036] In the diagram: 1-Core Framework Module, 2-User Login Module, 3-External Interface Module, 4-AI Module, 5-Visualization Module, 6-Prediction Module, 7-Node Management Module, 8-Map Interaction Module, 9-Data Acquisition Module, 10-Data Storage Module, 11-Data Processing Module, 101-Communication Protocol Unit, 102-Request Distribution Unit, 103-Database Access Unit, 104-State Coordinator Unit, 201-Multi-Factor Authentication Unit, 202-Session Auditing Unit, 301-Weather Access Unit, 302-Map Service Unit, 401-Natural Language Understanding Unit, 402-Decision Reasoning Unit, 403-Report Generation Unit, 404-Personalized Learning Unit, 501-Prediction Display Units include: 502 - Parameter Visualization Unit, 503 - Comparison Chart Unit, 601 - Graph Construction Unit, 602 - Spatiotemporal Convolution Unit, 603 - Prediction Engine Unit, 701 - Node Configuration Unit, 702 - Topology Maintenance Unit, 703 - Status Monitoring Unit, 801 - Coordinate Mapping Unit, 802 - Node Editing Unit, 803 - Topology Engine Unit, 901 - Video Access Unit, 902 - Sensor Integration Unit, 903 - Protocol Conversion Unit, 1001 - Relational Database Unit, 1002 - Object Storage Unit, 1003 - Time Series Database Unit, 1101 - Video Analysis Unit, 1102 - Real-time Computing Unit, 1103 - Anomaly Detection Unit, and 1104 - Message Queue Unit. Specific Implementation Example 1:

[0038] See Figures 1 to 12 As shown, this invention provides a regional intelligent transportation system with multi-target tracking and spatiotemporal prediction, including a core framework module 1, a user login module 2, an external interface module 3, an AI module 4, a visualization module 5, a prediction module 6, a node management module 7, and a map interaction module 8.

[0039] The user login module 2, the external interface module 3, the AI ​​module 4, and the map interaction module 8 are respectively connected to the core framework module 1, which serves as the center. The data management module includes a data acquisition module 9, a data storage module 10, and a data processing module 11 that are interconnected. The data management module provides data support for the core framework module 1, the AI ​​module 4, the visualization module 5, the prediction module 6, and the map interaction module 8. The node management module 7 and the map interaction module 8 are respectively connected to the prediction module 6, and the prediction module 6 is connected to the visualization module 5.

[0040] The core framework module 1 is used for intelligent transportation system scheduling, inter-module communication protocol encapsulation, request and response processing, and connecting to data modules to access the database; specifically, it includes:

[0041] The core framework module 1 includes a communication protocol unit 101, a request distribution unit 102, a database access unit 103, and a state coordinator unit 104 connected in sequence.

[0042] The communication protocol unit 101 is used to encapsulate the gRPC communication protocol to realize millisecond-level data exchange between modules. In this specific implementation, the communication protocol unit 101 encapsulates the millisecond-level data exchange between modules using the gRPC protocol, and establishes a bidirectional streaming communication channel by defining a structured data interface through Protocol Buffers.

[0043] The request distribution unit 102 is used to dynamically route user requests to various modules through a load balancing strategy. In this specific example, the request distribution unit 102 dynamically routes user requests based on the data transmission rules defined by the communication protocol unit 101 through a load balancing strategy, and accurately distributes the requests to each functional module.

[0044] The database access unit 103 is used to establish a distributed connection pool with the data storage module, provide a query interface, and implement read-write separation optimization. In this specific embodiment, the database access unit 103 provides a multi-mode query interface and performs intelligent data access operations based on the routing instructions of the request distribution unit 102 and through a dynamically managed distributed connection pool system.

[0045] The state coordinator unit 104 is used to monitor the operating status of each module and achieve automatic isolation of faulty modules and service circuit breaking through heartbeat detection. In this specific embodiment: the state coordinator unit 104 uses a heartbeat detection mechanism to poll the health status of each module, and transmits the circuit breaking command to the request distribution unit 102 based on the circuit breaker mode, automatically isolating abnormal nodes and triggering service degradation strategies.

[0046] The core framework module ensures data transmission efficiency through the communication protocol unit, responds to high-concurrency scenarios with dynamic load balancing through the request distribution unit, achieves query availability through the database access unit, and ensures that the system maintains core services in the event of a single point of failure through the circuit breaker mechanism. This four-fold linkage enables the system to maintain a low average response latency.

[0047] The user login module 2 is used to assign operation permissions to users and verify user identities; the user management module 2 includes a multi-factor authentication unit 201 and a session auditing unit 202 connected in sequence.

[0048] The multi-factor authentication unit 201 is used to enhance login security through a two-factor authentication mechanism that combines password verification with dynamic tokens. In this specific implementation, the multi-factor authentication unit 201 first performs combined verification on the password and dynamic token submitted by the user, and generates an encrypted session token with permission identifier after the verification is successful.

[0049] The session auditing unit 202 is used to capture and encrypt user operation commands in real time, record user operation trajectories, and generate an immutable operation behavior-system state change traceability chain. In this specific implementation, the session auditing unit 202 captures user operation events in real time based on the session token, synchronously records operation commands, parameters, and system state change data, and generates a traceability evidence chain with timestamps.

[0050] The external interface module 3 is used to connect to the meteorological data API, provide geocoding services, and obtain third-party map services; the external interface module 3 includes a weather access unit 301 and a map service unit 302 connected in sequence.

[0051] The weather access unit 301 is used to access multi-source meteorological data API through a dynamic key polling mechanism, analyze precipitation intensity, visibility and road surface temperature data in real time, and convert them into internal meteorological data models of the system. In this specific implementation, the weather access unit 301 calls the meteorological API through dynamic key pool polling, and outputs standardized meteorological objects after data cleaning.

[0052] The map service unit 302 is used to provide geocoding services, realize bidirectional address-coordinate conversion, and dynamically load and display third-party map base maps through a map rendering engine. In this specific embodiment: the map service unit 302 realizes bidirectional address-coordinate conversion through a geocoder, and loads street-level vector map base maps by calling a third-party map API through a map rendering engine, providing a map base map foundation for the map interaction module 8.

[0053] The weather access unit of external interface module 3 acquires meteorological data and improves its availability based on key polling and dynamic parsing. The map service unit supports millisecond-level coordinate transformation of geocoding services and effectively provides the required map base map.

[0054] The AI ​​module 4 integrates a natural language understanding engine, a traffic decision reasoning engine, and a report generator; the AI ​​module 4 includes a natural language understanding unit 401, a decision reasoning unit 402, a report generation unit 403, and a personalized learning unit 404 connected in sequence.

[0055] The natural language understanding unit 401 is used to parse the traffic semantics input by the user through a bidirectional Transformer architecture, such as: Is the traffic flow heavy at the next intersection? It integrates node topology relationships, real-time traffic parameters, and prediction result context to generate structured query instructions. In this specific embodiment, the natural language understanding unit 401 parses the user's natural language instructions in real time through a bidirectional Transformer architecture, integrates the node topology relationships of the map interaction module 8, the real-time traffic parameters of the data acquisition module 9, and the congestion prediction results of the prediction module 6 to generate structured query instructions.

[0056] The decision reasoning unit 402 is used to perform joint reasoning based on knowledge graph and rule engine to output root cause analysis conclusions or traffic optimization suggestions, such as: because the traffic flow on the next road is too large, it is recommended that vehicles detour on the next road; in this specific embodiment, after receiving the instruction, the decision reasoning unit 402 performs joint reasoning based on traffic knowledge graph and rule engine Drools to output optimization results.

[0057] The report generation unit 403 is used to dynamically generate graphic reports based on the decision results, and automatically integrate image data such as heat maps, prediction curves and abnormal event snapshots through a template engine; in this specific embodiment: the report generation unit 403 generates reports in formats such as PDF that contain data visualization elements based on the heat map of the prediction module 6 and the traffic parameter charts of the visualization module 5.

[0058] The personalized learning unit 404 records the user's language habits and satisfaction with the results, and continuously optimizes the suggestion strategy through machine learning algorithms to achieve adaptive evolution. In this specific embodiment: the personalized learning unit 404 continuously optimizes the decision strategy based on the user's satisfaction feedback on the report and historical interaction logs, thereby improving the accuracy of semantic understanding.

[0059] The AI ​​module's natural language understanding unit improves the accuracy of traffic semantic parsing through a context-aware Transformer architecture; the decision reasoning unit enhances root cause analysis efficiency by relying on a multi-source data fusion reasoning engine; the report generation unit generates visual reports based on existing data; and the personalized learning unit continuously optimizes decision-making strategies to effectively improve suggestion adoption rates. This four-layer architecture drives the system to achieve a leap in intelligent traffic management decision-making.

[0060] The visualization module 5 is used to display prediction results, acquire data from the database and visualize traffic parameters, and generate charts for multi-dimensional analysis. The visualization module 5 includes a prediction display unit 501, a parameter visualization unit 502 and a comparison chart unit 503 connected in sequence.

[0061] The prediction display unit 501 is used to overlay and display the prediction results of node-level congestion on the map interface, supporting both heat map and numerical labeling formats. In this specific implementation, the prediction display unit 501 receives node-level congestion prediction data from the prediction module 6 in real time via the gRPC protocol, and combines it with the road grid topology provided by the map interaction module 8. It uses the open-source front-end map framework OpenLayers geographic rendering engine to implement two visualization formats: heat map mode and numerical labeling mode.

[0062] The parameter visualization unit 502 is used to visualize real-time traffic parameters and provide alerts for node equipment anomalies. The visualized data includes dynamic traffic flow maps, speed distribution maps, abnormal event markers, and equipment anomaly alerts. In this specific embodiment, the parameter visualization unit 502 pulls real-time data streams from the data storage module 10 through a distributed connection pool. The data streams include various visualization components, such as traffic flow statistics maps, vehicle speed statistics maps, and vehicle type statistics maps. The unit also dynamically generates node equipment anomaly alert maps by receiving equipment anomaly alerts from the node management unit 7.

[0063] The comparison chart unit 503 is used to generate multi-dimensional analysis charts, including time-period comparison line charts and prediction comparison charts. In this specific example, the comparison chart unit 503 generates various analysis charts based on existing prediction data sequences and historical traffic parameters, such as time-period comparison line charts and prediction comparison charts.

[0064] The prediction display unit uses dual-mode rendering to present congestion predictions intuitively. The parameter visualization unit visualizes parameters based on real-time and historical data and alerts users to abnormal device connections. The comparison chart unit generates a decision support view through spatiotemporal data fusion. These three units form a panoramic visualization chain, enabling managers to quickly understand complex traffic situations.

[0065] The prediction module 6 is used to perform node-level congestion calculation based on spatiotemporal graph neural networks based on data provided by the data management module, including historical weather, traffic parameters, real-time weather, traffic parameters, and map topology relationships between the node management module and the map interaction module; the node management module and the map interaction module are respectively connected to the prediction module.

[0066] The prediction module 6 includes a graph construction unit 601, a spatiotemporal convolution unit 602, and a prediction engine unit 603 connected in sequence.

[0067] The graph construction unit 601 is used to abstract nodes as graph vertices and model road connection relationships as weighted directed edges. In this embodiment, the graph construction unit 601 dynamically generates a weighted directed graph model based on the area calibration results of the map interaction module 8, maps user-configured nodes as graph vertices, and defines the travel direction and travel capacity weights between nodes as directed edge weights.

[0068] The spatiotemporal convolutional unit 602 is used to extract road network topology features based on data such as real-time traffic parameters of nodes, real-time traffic parameters of surrounding areas, historical congestion levels, weather conditions, road time, time features, event information, and node static attributes, and combines this with GRU to capture the dynamic evolution of the time dimension. In this specific embodiment, the spatiotemporal convolutional unit 602 is based on the weighted graph structure output by the graph construction unit 601, and integrates real-time traffic parameters, weather data, road time, time features, event information, and node static attributes provided by the data acquisition module 9, as well as historical congestion features in the data storage module 10. It achieves hierarchical extraction of spatiotemporal features through Chebyshev polynomial decomposition and introduces a gated cyclic unit (GRU) to capture the dynamic evolution of the time dimension.

[0069] The prediction engine unit 603 is used to output the probability value of congestion level within a preset time period. In this specific embodiment, the nodes of the next 15, 30, and 60 minutes are set as the probability values ​​of congestion level within each time node. The prediction engine unit 603 performs multi-scale time window aggregation on the feature tensor extracted by the spatiotemporal convolution unit 602, and outputs the node-level congestion probability value of the next 15 / 30 / 60 minutes through a fully connected layer and a Sigmoid activation function. This value is between 0 and 1, and the congestion level threshold can be dynamically configured.

[0070] The node management module sends the complete and full basic road network structure to the prediction module. When the system starts, the prediction module initializes or the topology data in the node management module is restored and updated in batches by the external system, i.e., the database, and then sent.

[0071] The map interaction module sends real-time dynamic differences to the prediction module. After each drag, click, or delete operation on the map interface, the module pushes the changes in topology in real time and at the millisecond level.

[0072] The graph construction unit of the prediction module realizes real-time modeling of road network topology through dynamic weight mapping. The spatiotemporal convolution unit adopts a hybrid neural network architecture based on Chebyshev and GRU to jointly optimize the spatiotemporal feature representation. The prediction engine unit improves the long-term prediction robustness through multi-scale time window fusion technology. The cascaded collaboration of these three units effectively improves the prediction accuracy and effectively reduces the error rate under rain and fog weather conditions.

[0073] The node management module 7 is used to maintain the topological relationship of camera nodes, configure detection parameters, and monitor their status; the static topological structure of the road network and the basic configuration information of the nodes are transmitted from the node management module to the prediction module 6.

[0074] The node management module 7 includes a node configuration unit 701, a topology maintenance unit 702, and a status monitoring unit 703 connected in sequence.

[0075] The node configuration unit 701 is used to set the camera and camera parameters at each node, such as resolution, frame rate, and detection sensitivity threshold. In this embodiment, the node configuration unit 701 receives the area calibration data from the map interaction module 8 through the geographic coordinate mapping interface and dynamically configures the camera parameters of each node, such as resolution, frame rate, and detection sensitivity threshold.

[0076] The topology maintenance unit 702 is used to establish a spatial relationship model between nodes and dynamically store road connection directions, node device connection status and traffic capacity weights. In this specific embodiment, the topology maintenance unit 702 constructs a weighted directed graph model based on configuration data and stores the traffic direction between nodes and real-time traffic capacity coefficients through an adjacency matrix.

[0077] The status monitoring unit 703 is used to diagnose the online status of the device in real time. It polls the device status periodically over a certain time period and issues an alarm promptly if the device status is abnormal, ensuring the device's usability. In this specific embodiment, the status monitoring unit 703 polls the device status every 10 seconds. When a node goes offline, it automatically triggers a topology weight degradation mechanism and pushes an alarm to the visualization module 5.

[0078] The node management module achieves millisecond-level node management at a multi-node scale through a three-level linkage mechanism of dynamic parameter configuration, topology modeling, and status feedback.

[0079] The map interaction module 8 provides a node relationship configuration interface, a topology relationship construction interface, and a region division tool, and synchronizes the topology difference data package to the prediction module. The map interaction module 8 includes a coordinate mapping unit 801, a node editing unit 802, and a topology engine unit 803 connected in sequence.

[0080] The coordinate mapping unit 801 is used to integrate third-party map services and realize sub-pixel level conversion from WGS84 coordinates to pixel coordinates through dynamic Mercator projection. In this embodiment, the coordinate mapping unit 801 converts the third-party map coordinates into system pixel coordinates in real time through the dynamic Mercator projection algorithm, effectively controlling the coordinate conversion error.

[0081] The node editing unit 802 is used to respond to map click events and perform node creation, drag-and-drop positioning, and deletion operations. Based on the R-tree spatial index engine, it processes user operation commands, automatically avoids node position conflicts, and positions the node to the nearest legal location. The legal location refers to the available point closest to the user's clicked or dragged position, provided it does not conflict with other nodes. In this specific embodiment, the node editing unit 802 implements 50-meter buffer zone conflict detection and automatic avoidance based on the R-tree spatial index.

[0082] The topology engine unit 803 is used to construct a spatiotemporally constrained weighted connection matrix, define traffic direction rules and dynamic weight coefficients, and generate topology difference data packets in real time, which are then pushed to the prediction module. In this specific embodiment: the topology engine unit 803 adopts a streaming graph computing architecture, completes the adjacency matrix reconstruction in a short time, reduces the size of the topology data packets by using binary differential compression technology, and synchronizes them to the prediction module 6. The map interaction module 8 sends the topology difference data packets to the prediction module 6, ensuring that the topology used for prediction is updated immediately.

[0083] The data acquisition module 9 is used to acquire data in real time, such as video streams and weather sensors, and simultaneously collects and records traffic parameters, vehicle driving direction, vehicle information, and equipment status data from the video stream, and completes the unification of communication protocols; it includes a video access unit 901, a sensor integration unit 902, and a protocol conversion unit 903 connected in sequence.

[0084] The video access unit 901 is used to connect to camera equipment or video streams to match videos with nodes. In this embodiment, the video access unit 901 actively matches or responds to the instructions of the node management module 7 by using a preset node identifier, and accesses the RTSP video stream or local video file of the specified monitoring node in real time, dynamically parses the video source metadata, and establishes a stable transmission channel.

[0085] The sensor integration unit 902 is used to access meteorological sensors, acquire third-party meteorological data, and receive meteorological data. In this specific embodiment, the sensor integration unit 902 polls or subscribes to meteorological sensor data through a third-party protocol, receives key meteorological parameters such as precipitation intensity, visibility, road surface temperature, and wind speed and direction in real time, and performs preliminary validity verification and timestamp synchronization.

[0086] The protocol conversion unit 903 is used to unify the communication protocols of heterogeneous devices. In this specific embodiment: the protocol conversion unit 903 receives raw video stream element information from the video access unit 901 and raw meteorological data packets from the sensor integration unit 902, and converts the heterogeneous data into a unified internal standard data model through a built-in protocol parsing engine, and encapsulates context information such as timestamps and node IDs, and outputs a structured and standardized real-time sensing data stream to the data processing module 11 and the data storage module 10.

[0087] The protocol conversion unit 903 of the data acquisition module 9 eliminates device heterogeneity through normalization processing; the video access unit 901 ensures the timeliness of traffic event capture through low latency; the sensor integration unit 902 achieves second-level synchronization of multi-source meteorological data; and the protocol conversion unit 903 eliminates device heterogeneity through normalization processing. The cascaded collaboration of these three units ensures the data acquisition integrity rate of the system.

[0088] The data storage module 10 is used for distributed storage of user information, historical weather and traffic parameters, historical forecast records, node operation logs, video and chart reports, model parameter data, and other data. The data storage module 10 includes a relational database unit 1001, an object storage unit 1002, and a time-series database unit 1003 connected in sequence.

[0089] The relational database unit 1001 is used to store structured data such as node configuration and user information, and supports ACID transactions. In this embodiment, the relational database unit 1001 receives structured data from the core framework module 1, ensures data consistency through ACID transactions, and generates metadata indexes to be passed to the object storage unit 1002.

[0090] The object storage unit 1002 is used for distributed storage of historical videos and visualizations of some data. In this specific embodiment, the object storage unit 1002 associates and stores historical video clips and visualization charts according to the metadata index, generates a unique access URL, and returns it to the visualization module 5.

[0091] The time-series database unit 1003 is used to write high-frequency traffic parameters and provide time-range aggregated queries. In this specific embodiment: the time-series database unit 1003 receives real-time traffic parameter streams and real-time weather data from the data processing module 11, stores them in timestamp-segmented chunks, and optimizes long-term data query efficiency through downsampling technology.

[0092] The data processing module 11, based on the deep integration of the YOLOv11 target detection framework and the ByteTrack multi-target tracking algorithm, constructs a full-link processing capability covering vehicle perception, parameter calculation, and behavior analysis. It is used to achieve continuous all-weather vehicle tracking through YOLOv11-ByteTrack technology, generate disturbance-resistant traffic parameters based on a dynamic calibration algorithm, and integrate a multi-level anomaly detection model to output a structured data stream with spatiotemporal tags.

[0093] The data processing module 11 includes a video analysis unit 1101, a real-time computing unit 1102, an anomaly detection unit 1103, and a message queue unit 1104 connected in sequence.

[0094] The video analysis unit 1101 employs an improved YOLOv11 model as its core detection engine, enhancing its small target feature extraction capabilities by introducing a C2PSA module. It also integrates the ByteTrack multi-target tracking algorithm and innovatively adopts a low-confidence detection box reuse strategy. Simultaneously, it processes video stream data in real time, achieving accurate vehicle identification and continuous trajectory tracking, outputting a pixel-level structured data stream with unique tracking IDs, vehicle coordinates, category labels, and timestamps. In this embodiment, the video analysis unit 1101 enhances vehicle feature extraction capabilities in rainy and foggy environments through YOLOv11's C2PSA spatial attention mechanism, and combines this with ByteTrack's low-confidence detection box reuse strategy to handle occluded scenes, outputting a structured data stream with tracking IDs to the real-time computing unit.

[0095] The real-time computing unit 1102 is used to process the original detection data based on the pixel-level data output by the video analysis unit through a dynamic calibration algorithm to obtain traffic parameters that are less susceptible to interference, more robust, and more accurate. The dynamic calibration algorithm includes: dynamic pixel-to-distance conversion and exponential moving average smoothing.

[0096] First, pixel displacement is converted into physical distance using a dynamic pixel-to-distance conversion formula to facilitate subsequent calculations. Simultaneously, traffic parameters such as vehicle speed, vehicle acceleration, traffic flow, and driving direction vector are calculated. During unit operation, the system also employs an exponential moving average algorithm to smooth vehicle speed fluctuations. In this specific embodiment, the real-time calculation unit 1102 maps vehicle coordinates to physical positions based on the dynamic pixel-to-distance conversion formula, uses an exponential moving average algorithm to smooth vehicle speed fluctuations, and simultaneously calculates a series of traffic parameters including acceleration, traffic flow, direction vector, vehicle density, and headway. The processing results are then transmitted to the anomaly detection unit.

[0097] The anomaly detection unit 1103 is used to fuse spatiotemporal trajectories and traffic parameters to achieve collision detection, sharp turn identification, and emergency braking determination. Specifically, it detects vehicle collision events in real time using the intersection-union ratio (IUGR) algorithm, identifies sharp turn events using the motion vector angle formula, determines whether emergency braking has occurred based on the rate of change of speed, and dynamically adjusts the determination threshold based on road type. Users can also set their own determination thresholds to minimize the false alarm rate. In this specific embodiment: the anomaly detection unit 1103 fuses spatiotemporal trajectories and traffic parameters, monitors collision events in real time using the IUGR algorithm, identifies sharp turn events using the motion vector angle formula, determines emergency braking events based on the rate of change of speed threshold, and finally outputs an anomaly index and event type.

[0098] The message queue unit 1104 is used to build a Kafka-based distributed asynchronous processing framework to achieve highly reliable access and processing of data from multiple heterogeneous devices, ensuring the transmission of traffic data. In this specific embodiment: the message queue unit 1104 uses the Kafka distributed architecture to achieve asynchronous buffering of device data, ensuring the high-concurrency data processing capabilities of the video analysis unit, real-time computing unit, and anomaly detection unit.

[0099] This data processing module innovatively achieves multi-technology collaborative optimization: YOLOv11's C2PSA module enhances small target detection capabilities through a cross-stage spatial attention mechanism, improving vehicle detection rates in rainy and foggy environments; ByteTrack's low-confidence detection box reuse strategy significantly reduces trajectory breakage rates, substantially improving tracking continuity in complex scenes; the dynamic pixel-distance conversion formula combined with a vehicle category adaptive calibration mechanism effectively controls distance mapping errors; the multi-method fusion strategy for anomaly detection, coupled with dynamic threshold adjustment for road types, effectively reduces false alarm rates; and the Kafka-based asynchronous processing framework supports high message throughput, providing technical support for relatively large-scale road network monitoring.

[0100] In this embodiment of the invention, the map interaction module 8 realizes road network grid modeling and dynamic topology construction, and enables visualized configuration of node relationships through the map interaction module. It supports 50-meter buffer zone conflict detection and dynamic correction of traffic weights, providing a real-time road network topology foundation for spatiotemporal prediction. The node management module 7 maintains the topological relationships of camera nodes and corrects traffic weights in real time. The data acquisition module 9 integrates video streams and meteorological sensor data. The data processing module 11 realizes accurate vehicle identification, continuous trajectory tracking, traffic parameter calculation, and abnormal behavior detection, outputting a structured data stream with spatiotemporal tags. Based on multi-source data fusion analysis, it detects abnormal behaviors such as vehicle collisions, sharp turns, and sudden braking in real time, automatically marks them through the visualization module, and pushes alarm information with a response latency of less than 50 milliseconds, realizing proactive anomaly warning. At the same time, it adopts the YOLOv11-C2PSA rain and fog enhancement algorithm and the ByteTrack low-confidence box reuse strategy to realize continuous vehicle tracking and abnormal behavior detection in complex environments. The data storage module 10 distributes and stores multi-source heterogeneous traffic data. The prediction module 6, based on a spatiotemporal graph neural network, fuses real-time traffic parameters, meteorological data, historical features, and topological relationships to output congestion probability values ​​at future time nodes, achieving multi-scale congestion prediction. The visualization module 5 provides a multimodal dynamic display of prediction results and traffic parameters. The AI ​​module 4 uses a bidirectional Transformer architecture to parse natural language commands, linking with the decision-making inference engine to generate root cause analysis and optimization suggestions, and outputting graphic reports as needed. The core framework module 1 achieves millisecond-level communication via the gRPC protocol and completes fault circuit interruption and load balancing. The external interface module 3 dynamically accesses multi-source meteorological APIs and third-party map services; the user login module 2 implements multi-factor authentication and operation traceability auditing. Overall, this invention, through a multi-level collaborative architecture and spatiotemporal data fusion mechanism, constructs a regional intelligent transportation system covering the entire link of perception, prediction, and decision-making, overcoming challenges such as response delays, data silos, data fragmentation, and small sample size issues in existing intelligent transportation systems, effectively improving the real-time performance of traffic situation awareness, the collaborative nature of prediction and decision-making, system robustness, and the level of intelligent management.

[0101] For example, during the evening rush hour from 5:30 PM to 7:00 PM on July 23rd, the traffic management center of a certain city used this system to manage the traffic situation in a specific area. This area included 12 key intersection nodes numbered N1-N12, equipped with high-definition cameras and weather sensors. Administrators used the system to predict congestion, respond to abnormal events, and optimize decision-making.

[0102] The user login module is responsible for user login and permission activation: the administrator enters a password and verifies a dynamic token to obtain operation permissions. The session audit unit records the login time, such as 17:30 for the administrator in this example. It also records the permission scope, such as: the ability to modify node configurations and view prediction reports.

[0103] The map interaction module is responsible for map topology construction: After the administrator logs in, it automatically calls a third-party map API. After the administrator manually selects or drags to define the boundary of the monitoring area, the WGS84 coordinates are converted into pixel coordinates. Then, the administrator drags and configures the node positions N1-N12, and the system automatically triggers a 50-meter buffer collision detection to avoid node overlap. After the administrator sets the nodes, the topology engine unit automatically constructs a weighted directed graph, where vertices = intersections and edges = road connections, defining the unidirectional traffic weights from N5 to N8. If there are any changes, a topology differential data packet will be generated in real time and pushed to the prediction module.

[0104] The data acquisition module is responsible for acquiring multi-source data: This module accesses the nodes' 4K video streams via the RTSP protocol for subsequent data processing. It also acquires real-time meteorological data, such as: N1 node: visibility 3km, road surface temperature 32°C, precipitation intensity 0mm; N2 node: visibility 2.8km, affected by localized dust, road surface temperature 35°C, precipitation intensity 0mm. After acquiring the data, the camera and meteorological sensor data are unified, and a standardized data stream is output.

[0105] The data processing module is responsible for real-time data processing: it uses YOLOv11-ByteTrack technology to identify and track vehicles, assigning each vehicle a unique ID, such as Car#0723057. After vehicle identification, it outputs pixel-level structured data with ID, coordinates, category, and timestamp. Taking Car#0723057 as an example, after system startup, the first time the system detects this vehicle, it obtains information indicating a car appearing near node N4 at 17:32:15. Based on dynamic pixel displacement calibration, Car#0723057 moves 120 pixels within 2 seconds, resulting in a vehicle speed of 32.4 km / h. The data processing module can also obtain other traffic parameters in real time, such as the current traffic flow at node N3 being 85 vehicles / minute, smoothed using an exponential moving average. At 17:33:32, the anomaly detection unit detected an SUV at node N7, ID#0723102, with a speed change rate of 45%, changing from 60km / h to 28km / h within 2 seconds, which was determined to be "emergency braking" with an anomaly index of 0.92.

[0106] The prediction module is responsible for congestion prediction: it abstracts N1-N12 as vertices and models road connections as weighted edges. It integrates historical and real-time data, such as: the average real-time vehicle speed for N4 is 22 km / h, the historical congestion probability for N6 is 0.75, and the current visibility for N10 is 3 km based on meteorological data. After obtaining the data, the module extracts spatiotemporal features using Chebyshev multinomials, captures evolutionary patterns using GRU, and then outputs the future congestion probability. Taking N8 as an example, the probability is 0.68 for 15 minutes, 0.83 for 30 minutes, and 0.91 for 60 minutes. A threshold > 0.8 will trigger a red alert.

[0107] The visualization module is responsible for data visualization and alerts: Based on the data obtained from the prediction module, node N8 will be displayed in dark red in the prediction view, with a congestion probability of 0.83. Clicking on the corresponding node will query the required traffic parameter data graphs. For example, the traffic flow bar chart for node N3 shows an increasing trend, from 85 to 112 vehicles / minute. In addition, administrators can also view the node congestion prediction comparison curve, which shows that the predicted value is larger than the historical average.

[0108] The AI ​​module provides AI decision support: When an administrator inputs the question, "Why is N8 congested?", the Transformer architecture parses the semantics, correlates topology data, and identifies one-way traffic between N5 and N8, with a real-time traffic flow of 120 vehicles per minute at N8. Combined with knowledge graph analysis, it finds that an accident at the upstream N5 node is one of the causes of traffic congestion. After analysis, the AI ​​outputs an optimization suggestion: "It is recommended to temporarily open the N5→N8 opposite lane, with an estimated relief time of 12 minutes." If needed, the administrator can also request the AI ​​to generate a report. The module will automatically integrate heatmaps, abnormal event data, and optimization suggestions to generate a PDF report titled "Suggested Handling Plan for N8 Node during Evening Rush Hour on July 23." The personalized learning unit records whether the administrator continues to ask follow-up questions. If the administrator adopts the suggestion, similar decisions with high adoption rates will be prioritized when encountering similar questions in the future.

[0109] Through the above process, the modules of this invention construct a regional intelligent transportation system covering the entire chain of perception, prediction, and decision-making, providing high-precision, low-latency, and full-chain collaborative technical support for city-level intelligent transportation.

[0110] The above-disclosed embodiments are merely preferred embodiments of a regional intelligent transportation system based on YOLO-ByteTrack multi-target tracking and spatiotemporal prediction according to the present invention. Of course, they should not be construed as limiting the scope of the present invention. Those skilled in the art can understand that all or part of the processes of the above embodiments can be implemented, and equivalent changes made in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A regional intelligent transportation system with multi-target tracking and spatiotemporal prediction, characterized in that: It includes a core framework module, an external interface module, an AI module, a prediction module, a node management module, a map interaction module, and a data management module. The core framework module is used for system scheduling, encapsulation of inter-module communication protocols, processing of request and response, and access to the database. The external interface module is used to connect to meteorological data APIs, traffic information platforms, and third-party map services; The AI ​​module integrates a natural language understanding engine and uses a bidirectional Transformer architecture to parse user-input traffic semantics to generate structured query instructions. Based on the query instructions, it performs multi-source data joint reasoning based on the data from the data management module, knowledge graph, and rule engine, and outputs root cause analysis conclusions or traffic optimization suggestions as decision results. The data management module is used to collect traffic video streams and meteorological sensor data, and to obtain real-time and historical data of traffic parameters, vehicle driving direction, vehicle information and equipment status data from the traffic video streams. Based on the deep integration of the YOLO target detection framework and the ByteTrack multi-target tracking algorithm, it builds a full-link processing capability covering vehicle perception, parameter calculation and behavior analysis. At the same time, it generates anti-disturbance traffic parameters based on dynamic calibration algorithm, and integrates multi-level anomaly detection model to output a structured data stream with spatiotemporal tags. By integrating spatiotemporal trajectories with real-time traffic parameters, it identifies and monitors collision events, sharp turns, and sudden braking events, and outputs anomaly indices and event types. The map interaction module is used to obtain system pixel coordinates, provide a node relationship configuration interface, respond to map click events to perform node creation, drag-and-drop positioning and deletion operations, construct a weighted connection matrix with spatiotemporal constraints, define traffic direction rules and dynamic weight coefficients, generate topology difference data packets in real time and push them to the prediction module, and a region division tool. The node management module is used to dynamically configure each node according to the area calibration data of the map interaction module, set cameras and detection parameters at each node, establish a spatial relationship model between each node, dynamically store road connection directions, node device connection status and traffic capacity weights, maintain the topological relationship of camera nodes and send it to the prediction module, and diagnose the online status of the devices in real time. The prediction module is used to perform node-level congestion calculation based on spatiotemporal graph neural network based on historical weather, traffic parameters, real-time weather, time characteristics, and map topology relationship data provided by the data management module and the map interaction module, and to predict the future congestion level.

2. The regional intelligent transportation system with multi-target tracking and spatiotemporal prediction according to claim 1, characterized in that: It also includes a visualization module, which is connected to both the data management module and the prediction module. The visualization module is used to overlay the node-level congestion prediction results and node device anomalies of the prediction module onto the map interface, as well as the road grid topology display and real-time traffic parameter visualization prompts provided by the data management module.

3. The regional intelligent transportation system with multi-target tracking and spatiotemporal prediction according to claim 2, characterized in that: The visualization module includes a prediction display unit, a parameter visualization unit, and a comparison chart unit connected in sequence. The prediction display unit is used to overlay and display the prediction results of node-level congestion on the map interface, supporting both heat map and numerical annotation formats. The parameter visualization unit is used to visualize real-time traffic parameters, including dynamic flow diagrams, speed distribution diagrams, abnormal event markers, and equipment abnormality prompts. The comparison chart unit is used to generate multi-dimensional analysis charts, including time period comparison line charts and prediction comparison charts.

4. The regional intelligent transportation system with multi-target tracking and spatiotemporal prediction according to claim 1 or 2, characterized in that: It also includes a user login module, which is connected to the core framework module. The user login module is used to allocate operation permissions and authenticate identities.

5. The regional intelligent transportation system with multi-target tracking and spatiotemporal prediction according to claim 4, characterized in that: The user login module includes a multi-factor authentication unit and a session audit unit connected in sequence; The multi-factor authentication unit is used to integrate password login and dynamic token verification to enhance login security. The session auditing unit is used to record user operation trajectories and generate an operation behavior-system state change traceability chain.

6. The regional intelligent transportation system with multi-target tracking and spatiotemporal prediction according to claim 1, characterized in that: The data management module includes an interconnected data acquisition module, a data processing module, and a data storage module. The data acquisition module is used to acquire video streams and meteorological sensor data in real time, and to collect and record traffic parameters, vehicle driving direction, vehicle information and equipment status data in the video stream. The data storage module is used for distributed storage of user information, historical weather and traffic parameters, historical forecast records, node operation logs, video and chart reports and model parameter data; The data processing module is used to achieve continuous all-weather vehicle tracking based on real-time data collected by the data acquisition module and historical data stored by the data storage module, using YOLO-ByteTrack technology. It generates anti-disturbance traffic parameters based on a dynamic calibration algorithm and integrates a multi-level anomaly detection model to output a structured data stream with spatiotemporal tags. It integrates spatiotemporal trajectories and traffic parameters, monitors collision events in real time using the intersection-exchange-comparison algorithm, identifies sharp turning behavior by combining the angle of motion vectors, determines emergency braking events based on the speed change rate threshold, and finally outputs anomaly index and event type.

7. The regional intelligent transportation system with multi-target tracking and spatiotemporal prediction according to claim 6, characterized in that: The data processing module includes a video analysis unit, a real-time computing unit, an anomaly detection unit, and a message queue unit connected in sequence. The video analysis unit is used to identify vehicle targets in the video in real time and preprocess the target data through the YOLO target detection model and ByteTrack multi-target tracking algorithm, so as to achieve accurate vehicle identification and continuous trajectory tracking, and generate pixel-level structured data with unique tracking ID, vehicle coordinates, category labels and timestamps. The real-time computing unit is used to accurately map image coordinates to physical locations based on pixel-level data output by the video analysis unit through a dynamic pixel-distance conversion formula, and simultaneously calculate a series of traffic parameters such as vehicle speed, acceleration, driving direction, traffic flow, vehicle density, and headway, and uses an exponential moving average algorithm to control vehicle speed fluctuation errors. The anomaly detection unit is used to fuse spatiotemporal trajectories and traffic parameters, monitor collision events in real time through the intersection-union algorithm, identify sharp turning behavior by combining the angle of motion vectors, determine emergency braking events based on the speed change rate threshold, and output anomaly index and event type. The message queue unit is used to implement asynchronous processing of device data based on Kafka, ensuring the high-concurrency data processing capabilities of the video analysis unit, the real-time computing unit, and the anomaly detection unit.

8. The regional intelligent transportation system with multi-target tracking and spatiotemporal prediction according to claim 1, characterized in that: The AI ​​module is also used to dynamically generate graphic reports and perform personalized learning based on the decision results.

9. The regional intelligent transportation system with multi-target tracking and spatiotemporal prediction according to claim 1, characterized in that: The prediction module includes a graph construction unit, a spatiotemporal convolution unit, and a prediction engine unit connected in sequence. The graph construction unit is used to abstract the nodes sent by the node management module as graph vertices and model the road connection relationship as weighted directed edges; The spatiotemporal convolutional unit is used to extract road network topology features based on real-time traffic parameters of nodes, real-time traffic parameters of surrounding areas, historical congestion levels, weather conditions, road time, time characteristics, event information, and static attribute data of nodes, using Chebyshev polynomials. The prediction engine unit is used to output the node congestion level value within a preset time period through a fully connected layer and a Sigmoid activation function.

10. The regional intelligent transportation system with multi-target tracking and spatiotemporal prediction according to claim 1, characterized in that: The map interaction module includes a coordinate mapping unit, a node editing unit, and a topology engine unit connected in sequence. The coordinate mapping unit is used to integrate third-party map services to realize the conversion from WGS84 coordinates to pixel coordinates; The node editing unit is used to respond to map click events and perform node creation, drag-and-drop positioning and deletion operations, while pre-setting a certain range of buffer zone conflict detection and automatic avoidance; The topology engine unit is used to construct a weighted connection matrix with spatiotemporal constraints, define traffic direction rules and dynamic weight coefficients, and generate and transmit topology differential data packets in real time.

Images