Intelligent networked induction control platform for road traffic safety facilities

Abstract

The present application relates to the field of traffic control, specifically to a kind of road traffic safety facilities intelligent network connection response control platform.It includes traffic data acquisition module: the data stream of traffic sensor is collected, through identification technology, output with traffic data source identification traffic data;Traffic data processing module: according to traffic data source identification adaptation analysis protocol, obtain standardized traffic parameter;And according to index system, assess data reliability index;Situation fusion module: form vehicle driving track through graph neural network, generate global traffic situation map using data reliability index, build traffic network digital twin model;Decision and control module: analyze traffic state from traffic network digital twin model, predict traffic event and collision risk, generate traffic control instruction.The platform breaks information silos through device fingerprint and protocol reverse technology, ensures the comprehensiveness and high reliability of data source, significantly improves road traffic safety and traffic efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of traffic control, in particular to a kind of road traffic safety facilities intelligent network connection response control platform. BACKGROUND

[0002] With the rapid development of China's transportation infrastructure, society has put forward higher requirements for road traffic safety and traffic efficiency. With the active promotion of national policy and the joint efforts of the industry, the digitalization, networking and intelligent transformation of the transportation system are accelerating. Among them, the "vehicle-road cloud integration" intelligent transportation system centered on cellular vehicle networking technology has become the industry-recognized mainstream direction of development with its wide coverage and low-latency communication advantages. The system not only is the key support for realizing high-level automatic driving, but also is an effective path to solve the current urban traffic congestion and frequent safety accidents. It aims to build a "perspective" beyond the perception ability of any single traffic participant through all-round, real-time information exchange and collaborative computing between vehicles, roadside facilities and cloud platforms, so as to predict risks, optimize signal control and guide paths, and make more accurate and efficient traffic decisions and controls.

[0003] However, in the transition to full intelligentization, there are still some common defects in existing technologies. The current large number of deployed road safety facilities, such as independent video surveillance, radar speedometers and various active warning devices, are still essentially single-function, passive and localized systems. These systems are each other, and the data collected cannot be effectively interchanged, forming a typical "information island". Even some safety systems with more comprehensive functions, their perception and control range are often limited to a specific point or a single scene, and they cannot obtain traffic situation information in a wider range, so their decisions are local and lack foresight, which leads to insufficient capacity in dealing with cross-regional traffic incidents, chain congestion, sudden crossing of pedestrians or non-motor vehicles and complex dynamic traffic scenes such as bad weather, and bottlenecks in safety protection and efficiency improvement.

[0004] In summary, it is now necessary to effectively network and intelligently integrate these functionally different and dispersedly deployed localized road safety sensing and warning facilities.

[0005] To this end, a road traffic safety facility intelligent network connection response control platform is proposed. SUMMARY

[0006] The purpose of the present application is to provide a road traffic safety facility intelligent network connection response control platform to solve the problems raised in the background art.

[0007] To achieve the above purpose, the present application provides the following technical solution: a road traffic safety facility intelligent network connection response control platform, comprising:

[0008] Traffic data collection module: collect data streams of multiple roadside traffic sensors, establish traffic data source identification by combining physical layer radio frequency characteristics and network communication behavior recognition technology, and output traffic data with the traffic data source identification;

[0009] Traffic data processing module: adaptively parse protocols according to the traffic data source identification, parse the traffic data into standardized traffic parameters, and evaluate data reliability indicators of the traffic parameters according to an index system covering timeliness, spatial accuracy and completeness;

[0010] Situation fusion module: associate traffic targets in the traffic parameters to form continuous vehicle driving trajectories through a graph neural network, perform data fusion using the data reliability indicators as dynamic weights, and generate a global traffic situation map; use the global traffic situation map to build and update a dynamic traffic road network digital twin model in real time;

[0011] Decision and control module: analyze traffic states from the traffic road network digital twin model, predict potential traffic incidents and collision risks, and generate traffic control instructions for traffic guidance and early warning to vehicle terminals.

[0012] Preferably, the specific implementation process of generating traffic data with the traffic data source identification includes:

[0013] Extracting physical layer characteristics containing signal damage introduced by hardware tolerance in the traffic sensor communication signal to generate a physical layer radio frequency fingerprint; extracting network behavior characteristics containing protocol headers, data packet sequences and communication timing patterns in the traffic sensor communication data stream to generate a network behavior fingerprint; taking the physical layer radio frequency fingerprint and the network behavior fingerprint as input features, fusing them through a weighted scoring mechanism, and calculating and assigning traffic data source identification for the traffic sensor.

[0014] Preferably, the specific implementation process of parsing into standardized traffic parameters includes:

[0015] Querying a device archive according to the traffic data source identification to determine a communication protocol type, calling a standard parser if the communication protocol type is known, and starting protocol reverse engineering if it is unknown or private, inferring data payloads containing message formats, field boundaries and data types through sequence comparison, statistical analysis and heuristic algorithms; analyzing the data payloads through a data pattern inference engine, identifying and labeling key business fields containing vehicle identity, speed, position and timestamp, and converting the traffic data into the standardized traffic parameters.

[0016] Preferably, the specific implementation process of evaluating data reliability indicators of the traffic parameters includes:

[0017] computing the time stamp of the traffic parameter and the delay of the current processing time, quantifying the timeliness; evaluating the spatial positioning error according to the inherent technical specifications of the traffic sensor of the traffic parameter and the current working environment factors, quantifying the spatial accuracy; checking whether the traffic parameter contains all the predefined key business fields, quantifying the integrity; performing weighted summation on the quantification results of the timeliness, spatial accuracy and integrity, generating the data reliability index.

[0018] Preferably, the specific implementation process of generating the global traffic situation map includes:

[0019] the traffic targets in the traffic parameters derived from different times and different traffic sensors are taken as nodes of a graph; edges representing potential associations are established between the nodes based on the spatio-temporal proximity and appearance feature similarity of the traffic targets, a multi-target association graph is constructed, the graph neural network is applied to infer the multi-target association graph, the association confidence represented by each edge is calculated, and based on the high-confidence association path, the nodes belonging to the same traffic target are connected to form the continuous vehicle driving trajectory; the continuous vehicle driving trajectory and the traffic parameter are weighted averaged through the corresponding data reliability index, and the global traffic situation map is generated.

[0020] Preferably, the specific implementation process of generating the global traffic situation map includes:

[0021] a traffic road network digital twin model containing road network topology, traffic facility static attributes and traffic rules is constructed; the vehicle driving trajectory and the traffic parameter in the global traffic situation map are taken as real-time data sources, and are instantiated as dynamic traffic participant objects with real-time positions, speeds and attributes in the traffic road network digital twin model; the updated global traffic situation map is continuously received, and the states and mutual relationships of the dynamic traffic participant objects are refreshed at high frequency.

[0022] Preferably, the specific implementation process of generating the traffic control instruction includes:

[0023] The traffic network digital twin model is abstracted into a dynamic space-time graph, intersections and road segments are nodes of the graph, connections between nodes are edges, and the characteristics of the nodes include real-time traffic flow parameters; the evolution of the dynamic space-time graph is analyzed by applying a space-time graph neural network to obtain prediction results including the probability, type and space-time influence range of traffic events occurring in a specific time window in the future; a hierarchical reinforcement learning framework is applied to make control decisions, a high-level meta-agent sets a macro traffic optimization strategy target according to the prediction results, and a bottom-level local agent learns and outputs the optimal tactical action according to the traffic optimization strategy target and the traffic state in the respective region; the tactical action is converted into specific instructions for traffic guidance and signal timing optimization, and vehicle terminal early warning messages containing predictive risk attributes are generated.

[0024] Compared with the prior art, the beneficial effects of the present application are:

[0025] 1. The breadth and efficiency of information fusion are improved, a unique identity file is established for various heterogeneous facilities by combining the hybrid identification technology of physical layer radio frequency fingerprint and network behavior fingerprint, and non-standard data is converted into unified standardized traffic parameters by adaptively triggering protocol reverse engineering, thereby improving the seamless integration capability and data interworking efficiency of global traffic facilities. Breakthrough the limitations of traditional systems relying on standard interfaces or manual configuration, realize the automatic and low-cost access of existing and new facilities, to a large extent, alleviate the "information island" problem, and lay a solid foundation for building a networked collaborative sensing and control system.

[0026] 2. The accuracy and reliability of situation awareness are ensured, a real-time multi-dimensional data quality evaluation system is established to evaluate the dynamic data reliability index of each traffic parameter, and in the subsequent fusion process, the index is used as a dynamic weight to intelligently trust high-reliability data sources and suppress the interference of low-quality data sources, thereby improving the accuracy and reliability of global traffic situation awareness. This quality-driven fusion mechanism ensures that the final generated situation map is a high-fidelity result after information refinement, rather than a simple aggregation of data, providing high-quality data input for subsequent accurate prediction and control.

[0027] 3. The initiative and foresight of traffic management are realized, the high-fidelity road network model is used for forward-looking risk prediction, the hierarchical reinforcement learning framework is applied for networked collaborative control, and the control instructions are verified in a closed loop through digital twinning, which realizes the paradigm shift from "passive response" to "active prediction" of traffic management, and improves traffic safety and efficiency. The platform can predict and avoid potential risks, implement cross-regional and long-chain collaborative optimization strategies, and ensure the reliable execution of control instructions, thereby preventing and resolving traffic risks at the source and maximizing road network operation efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 A structural diagram of a road traffic safety facility intelligent network connection induction control platform is provided for the embodiments of the application;

[0029] Figure 2 A collaborative perception flowchart is provided for the embodiments of the application;

[0030] Figure 3 A prediction and collaborative control closed-loop diagram is provided for the embodiments of the application. DETAILED DESCRIPTION

[0031] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only some of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the application.

[0032] Please refer to Figures 1-3 The application provides a road traffic safety facility intelligent network connection induction control platform, and the specific modules are as follows:

[0033] The traffic data collection module collects data streams of multiple roadside traffic sensors, establishes a traffic data source identifier by combining the identification technology of physical layer radio frequency characteristics and network communication behavior, and outputs traffic data with the traffic data source identifier.

[0034] The traffic data processing module: according to the traffic data source identifier, the protocol is adapted and parsed, the traffic data is parsed into standardized traffic parameters, and the data reliability index of the traffic parameters is evaluated according to the index system covering timeliness, spatial accuracy and integrity.

[0035] The situation fusion module: the traffic targets in the traffic parameters are associated to form continuous vehicle driving tracks by a graph neural network, the data fusion is performed by using the data reliability index as a dynamic weight, and a global traffic situation map is generated. The global traffic situation map is used to build and update a dynamic traffic road network digital twin model in real time.

[0036] The decision and control module: analyzes the traffic state from the traffic road network digital twin model, predicts potential traffic events and collision risks, and generates traffic control instructions for traffic guidance and early warning to vehicle terminals.

[0037] The technical solutions of the application will be described in further detail below with reference to specific embodiments.

[0038] Embodiment one

[0039] The embodiment of the application discloses a road traffic safety facility intelligent network connection induction control platform, at an A intersection, the platform is used for sensing the current traffic condition, referring to Figure 1 The specific modules of the application include: a traffic data collection module: collecting data streams of traffic sensors, outputting traffic data with traffic data source identification through identification technology; a traffic data processing module: parsing the traffic data into standardized traffic parameters according to the traffic data source identification; and evaluating the data reliability index of the traffic parameters according to a multi-dimensional index system; a situation fusion module: forming continuous vehicle driving trajectories through a graph neural network, generating a global traffic situation map using the data reliability index; constructing and real-time updating a dynamic traffic road network digital twin model using the global traffic situation map; a decision and control module: analyzing the traffic state from the traffic road network digital twin model, predicting potential traffic events and collision risks, and generating traffic control instructions for traffic guidance and early warning to vehicle terminals.

[0040] Further, the data streams of multiple roadside traffic sensors are collected, the traffic data source identification is established through the identification technology combining the physical layer radio frequency characteristics and network communication behavior, and the traffic data with the traffic data source identification is output; corresponding to the above traffic data collection module, referring to Figure 2 ; the specific implementation process includes:

[0041] The communication channels of the region where the A intersection is located are scanned and listened to through a multi-modal radio frequency front end and a network sniffer. For example, the 24GHz frequency band is continuously monitored to capture millimeter wave radar signals, the traffic of wired Ethernet is monitored through switch port mirroring, and the LoRaWAN gateway frequency band is scanned to discover Internet of Things devices. When any regular signal or data packet activity is detected, the module will lock the signal source and capture the original I / Q sample data, electrical signal waveform or network data packet. The collected original signal is processed through a signal analysis algorithm. For wireless devices such as millimeter wave radars, the signal needs to be subjected to fast Fourier transform (FFT) to calculate the stable offset of the carrier frequency relative to the nominal value, and the phase noise at a specific offset frequency is quantified by analyzing the power spectral density; for wired devices such as IP cameras, identification is performed in combination with network traffic analysis and device handshake features, mainly including MAC address patterns, network protocol stack fingerprints, data packet timing characteristics and the like. After these physical characteristics are quantified, a multi-dimensional physical fingerprint vector is formed.

[0042] In the observation window period of about 5 minutes, statistical analysis is carried out on all the captured data packets from the same signal source, and it is found that the millimeter wave radar sends UDP data packets with a length of 128 bytes in a 50 millisecond cycle, the IP camera sends video stream image groups with a length of 50 during RTSP streaming, and the average size ratio of I frames and P frames remains in a relatively stable interval of 7.5:1, and the geomagnetic detector sends 32-byte packets with a fixed interval of 2 seconds. These stable patterns in the protocol header, packet timing, payload size, and entropy value are finally combined into network behavior fingerprint features. With the aging and failure of device hardware, both the physical layer radio frequency fingerprint and the network behavior fingerprint will exhibit a certain degree of drift. In practice, many variable environmental factors will also cause short-term fluctuations in the fingerprint features. Therefore, when tracking the time series of each device fingerprint, an adaptive filter is used to preprocess the data, distinguishing between long-term trend drift caused by hardware aging and short-term noise caused by environmental factors. A long short-term memory network is used to build a long-term health drift model of the device fingerprint, which monitors whether it has an abnormal drift beyond the normal environmental fluctuation range and generates warning information and predictive maintenance work orders. In this way, the platform is upgraded from a simple traffic sensing system to an intelligent infrastructure management platform with asset health management and predictive operation and maintenance capabilities, thereby improving its predictive ability.

[0043] The extracted N physical layer radio frequency fingerprint features (such as carrier frequency offset CFO, I / Q imbalance, etc.) and M network behavior fingerprint features (such as average packet interval time, load entropy, etc.) are normalized using the Z-score standardization method, and a mixed feature vector V with a dimension of N+M is constructed. This vector is then input into a support vector machine classification model that has been pre-trained. The model is trained using fingerprint samples from 100 different models and a total of 5000 devices, and uses a radial basis function as the kernel function. The optimal hyperparameters (such as C equal to 10, gamma equal to 0.1, etc.) are determined using the grid search method to calculate the most likely identity of the device and output the corresponding confidence score. For example, the model determines that the current device is a "D company millimeter wave radar - model A" device with high confidence in the current radio frequency environment. The platform then assigns a universally unique identifier (UUID: 1a2b3c4d-5e6f-7a8b-9c0d-1e2f3a4b5c6d, etc.) as the traffic data source identifier and stores it in the device archive along with the complete fingerprint feature data.

[0044] By combining the physical layer radio frequency fingerprint and network behavior fingerprint, a digital identity is established for the traffic sensor, solving the problems of high cost, long cycle and manual configuration in traditional system integration. In this way, without relying on the information reported by the equipment, the identification and filing of the global heterogeneous facilities can be reliably completed, providing a solid and reliable identity authentication foundation for breaking the "information island" and realizing subsequent data analysis and collaborative control.

[0045] Further, the traffic data is parsed into standardized traffic parameters according to the traffic data source identification adapted parsing protocol; the corresponding traffic data processing module refers to Figure 2 ; the specific implementation process includes:

[0046] When receiving the data stream with the traffic data source identification, the identification is used to query the equipment archive, so as to determine the communication protocol type associated with the data source. For example, for the equipment recorded in the archive as "H company IP camera", the query result will show that its protocol is "ONVIF Profile T", so the subsequent data stream will be scheduled to the pre-set standard ONVIF parser for corresponding processing. If it is "D company millimeter wave radar", its archive may be marked as "protocol unknown" at the initial time, in which case the platform will route the data stream to the protocol reverse engineering unit for processing.

[0047] For non-encrypted unknown protocol data, protocol reverse engineering will be performed, for example, UDP packets sent by the radar of Company D, this module will capture hundreds of consecutive packets (for example, 500 packets, total duration 25 seconds) to form an analysis corpus, using sequence alignment based on the Needleman-Wunsch algorithm, it can be found that the first part of each 128-byte packet is a fixed 4-byte 0xDEADBEEF, from which it can be inferred that this is likely to be the message header. Then use the heuristic algorithm rule library to analyze the remaining payload, which covers not only the following aspects: first, the field entropy value rule, that is, by calculating the byte entropy in the 4-byte sliding window, the region with constant and low entropy value is identified as a fixed identifier or state word; second, the field value domain and change rule, if the value domain distribution of a 4-byte floating point field always falls within the intersection geographic fence coordinate range, and its change conforms to the vehicle dynamics model, then this field will be highly suspected to be the position coordinate field; if a 2-byte integer field presents a linear small incremental trend, then this field will be inferred to be the packet sequence number. According to the above rules, the 2-byte field following the packet sequence number is found to present a linear incremental trend, which is marked as "packet sequence number". After a series of operations such as statistical analysis based on byte frequency distribution and sliding window entropy value and inference relying on the heuristic rule library on the remaining payload, a region with variable length composed of multiple 20-byte structures is identified. Further analysis of the fields inside the 20-byte structure shows that the value domain distribution of some 4-byte floating point fields is highly consistent with the coordinates and speed of the vehicle in the intersection, so the complete format of the private protocol is inferred: [message header]-[sequence number]-[target number]-[target object 1]-[target object 2]..., where each target object contains ID, three-dimensional coordinates, speed, etc.

[0048] Whether from the standard parser or the protocol reverse engineering module, at this time it is still only structured raw fields (e.g. field_1, field_2...), in order to give business meaning, the data mode inference engine intervenes, which analyzes the numerical characteristics, variation rules and mutual relations of these fields based on the model trained in advance through a large amount of traffic data set. For example, the above radar data, the engine finds that one of the fields is a small integer and remains unchanged in a short time, which meets the characteristics of "target ID"; the values of the other two fields fall within the geographic fence of A intersection and are labeled as "position coordinates"; the values of the other two fields match the typical vehicle speed range and are labeled as "speed component". In this way, field_1 is automatically mapped to object_id, field_2 and field_3 are mapped to position_x and position_y, etc., completing the conversion from raw data to data with clear business semantic information.

[0049] Finally, all the data with semantic annotation are uniformly packaged into the platform's internal pre-defined JSON format. For example, the vehicle target detected by the above radar at a certain time, its object_id: 101, position_x: 12.5, position_y: -30.2, etc. information will be converted into the standardized traffic parameter object: {source_id: UUID: 1a2b..., timestamp: 2025-07-18T20:31:00.050Z, objects: [{object_id: 101, object_type: vehicle, position: {lat: 34.0522, lon: -118.2437}, velocity: {speed_kph: 45.5, heading: 182.3}}]}.

[0050] Through protocol reverse engineering and data mode inference, the platform understands and utilizes those devices that use unencrypted private or unknown protocols and cannot be accessed, activates the stock value of a large number of old or non-standard facilities deployed on the road, and avoids expensive device replacement. By converting raw bit streams from any source into unified, standardized traffic parameters with clear business semantics, the protocol barriers between devices are removed, providing a rich source of data for building a comprehensive perception network.

[0051] Further, according to the index system covering timeliness, spatial accuracy and completeness, the traffic parameter evaluation data reliability index is evaluated, refer to Figure 2 ; the specific implementation process includes:

[0052] The timestamp field in the traffic parameter object is extracted to quantitatively evaluate the timeliness, which is compared with the system clock of the current platform that is precisely synchronized by the Network Time Protocol (NTP). The difference between the two is the end-to-end data delay. For example, a traffic parameter from the millimeter wave radar of Company D has a timestamp of 2025-07-18T20:31:00.050Z, while the system time of the platform processing this data is 2025-07-18T20:31:00.100Z, and the delay is calculated to be 50 milliseconds. The platform will convert this delay into a timeliness score according to the preset delay-score mapping curve (for example, a delay of less than 100 milliseconds scores 0.95), and for the geomagnetic detector data transmitted through the low-power LoRaWAN network, the delay may reach 1.5 seconds, and the timeliness score will be reduced accordingly (for example, the score is 0.60).

[0053] The module quantitatively evaluates the spatial accuracy of the parameter, which is not simply reading the nominal accuracy on the device specification, but a dynamic and context-aware process. The module will first retrieve the inherent technical specifications of the sensor (such as the pixel resolution of the IP camera of Company H, the angle resolution of the radar of Company D) from the device library, combined with real-time environmental information from other data sources. For example, when the weather sensor reports that it is currently raining heavily, the system will call an environmental impact evaluation model that has been pre-trained through historical data. This model takes real-time environmental parameters (such as rainfall, illumination, temperature, etc.) as input, combines the inherent specifications of the sensor, and outputs a dynamic spatial accuracy attenuation coefficient. According to this basis, the spatial accuracy rating of the camera is adjusted. For the millimeter wave radar of Company D, its positioning error may be stable at ±0.1 meters in sunny weather, and the spatial accuracy score is 0.99; but in heavy rain, the positioning error of the camera may increase to ±2 meters, and the spatial accuracy score may decrease to 0.70.

[0054] Thirdly, the module quantitatively evaluates the integrity of the parameter. The platform will check whether the current received traffic parameter object contains all the necessary key business fields according to the predefined complete data pattern that this type of data should have. For example, a complete vehicle target object should contain object_id, position, velocity, etc. If the platform finds that the velocity field is missing in a certain data due to network packet loss or overload of the sensor internal processing unit, the integrity of this data will be judged to be defective. The integrity score will be calculated according to the importance and number of missing fields, for example, missing secondary fields scores 0.90, and missing core fields such as position scores may be directly reduced to 0.30.

[0055] Finally, the module performs the weighted generation of the data reliability index (DQS). It sums the quantitative scores of the above three dimensions (timeliness, spatial accuracy, and integrity) according to a configurable weighting formula to generate a final, single data reliability index (DQS). Moreover, the weights of each dimension are adjusted according to the application scenario, for example, in the application for collision warning, the requirements for spatial accuracy and timeliness are extremely high, and their weights may be set to 0.5 and 0.4; while in the application for macro traffic flow analysis, the requirement for integrity is higher, and its weight may be increased. Finally, each standardized traffic parameter is attached with a DQS value between 0 and 1 (for example, 0.92).

[0056] For each standardized traffic parameter entering the platform, a comprehensive data quality score (DQS) is dynamically assessed in real time, establishing a unified and quantitative credibility standard for the originally mixed data. Through the context-aware evaluation of the timeliness, spatial accuracy, and integrity of the data, the platform clearly understands the reliability of each piece of information, which provides crucial quality assurance for subsequent high-quality data fusion and high-risk, high-value traffic control decisions.

[0057] Further, the traffic targets in the traffic parameters are associated to form continuous vehicle driving trajectories through a graph neural network, and the data reliability index is used as a dynamic weight for data fusion to generate a global traffic situation map; for the corresponding situation fusion module, refer to Figure 2 ; the specific implementation process includes:

[0058] The platform takes all the traffic target detection results (e.g. point cloud targets from radar, bounding boxes from camera, vehicle presence signals from geomagnetic coil) generated by all sensors in the past short time window (e.g. 5 seconds) as nodes of a graph, and establishes edges between these nodes to represent potential associations based on physical plausibility and feature similarity. For example, A radar detects a target at position P1 at t=0, and B camera detects a target at position P2 at t=0.5 seconds, if the distance between P1 and P2 (e.g. 10 meters) is within the normal speed range of this road segment, an edge is established between these two nodes. If B camera can also provide color feature (e.g. red), and A radar can also provide radar cross section (RCS) data, the similarity of these features will also be used as initial basis for calculating edge weight. Through this process, a complex spatio-temporal graph containing all possible associations is constructed. The complex spatio-temporal graph is input into a pre-trained graph neural network (GNN) model composed of 2 layers of graph convolution network (GCN) and 1 layer of gated recurrent unit (GRU). Unlike traditional methods, GNN can learn global, high-order association patterns, not just pairwise comparisons. For example, when a car is completely blocked by a large truck for 3 seconds between two cameras, GNN can infer with high confidence that the target that reappears after the end of the block is the same entity as the target that disappeared before the block, by analyzing the flow patterns and context relationships of other vehicles in this area. GNN performs inference calculation on the association confidence of each edge in the graph, and the platform then only retains edges with a confidence higher than a certain threshold (e.g. 0.90), and along these high-confidence connection paths, the discrete nodes belonging to the same traffic target are "stitched" into smooth, complete, and unique ID long-term trajectories.

[0059] At this time, the platform has continuous trajectories of all targets, but the same target may still have data from multiple sensors describing it at the same time (e.g. radar and camera both give its position and speed), and at this moment the module will use the DQS value generated in step S4 as a dynamic weight to perform weighted fusion of these redundant data. For example, at a certain moment, D company's radar (whose current data DQS is 0.99) reports the position of a vehicle as (12.50, -30.20), and H company's camera reports its position as (12.65, -30.45) due to the influence of heavy rain (its current data DQS is only 0.70). The final fused position calculated by the platform will be (Position_Radar*0.99+Position_Camera*0.70) / (0.99+0.70), which will be extremely close to the high-reliability radar data. This weighted fusion process will be applied to all kinematic parameters that can be fused, such as position, speed, acceleration, orientation, etc.

[0060] All the high-reliability long-term trajectories after quality-aware fusion and other traffic parameters (such as average speed of road segment, traffic incident alert, etc.) that cannot be associated with trajectories are superimposed on the high-precision geographic information layer. This ultimately forms a unified, dynamic, and high-fidelity global traffic situation map.

[0061] By stitching the target trajectories across the visual domain through the graph neural network and using DQS as the dynamic weight for data fusion, a global traffic situation map is generated that is much more complete, accurate, and reliable than any single sensor perception result. This step not only solves the problem of target loss in complex scenarios (such as occlusion, congestion), but also intelligently tends to trust the most reliable data source at the moment. This allows the platform to obtain a "perspective" that is closest to the "ground truth", providing an unprecedented high-quality and high-fidelity data foundation for subsequent digital twin modeling and risk prediction.

[0062] Further, a dynamic traffic road network digital twin model is constructed and updated in real time using the global traffic situation map, as described in Figure 3 ; the specific implementation process includes:

[0063] The platform loads three types of key static data from its background database: first, high-precision map data, which contains the precise geometry, topological connection relationship, stop line position, and slope curvature of all lanes at A intersection and the surrounding area; second, traffic facility asset data, which records the precise three-dimensional coordinates, model, and inherent attributes of each traffic signal, variable message board, and camera; third, traffic rules data, which defines the legal speed limit, turning restrictions (such as "no left turn during peak hours"), and standard phase scheme of the signal light. These three types of data collectively form a static but semantically rich traffic road network digital twin model. The global traffic situation map is used as a real-time data source, and the dynamic elements in it are mapped to the traffic road network digital twin model, for example, a high-reliability long-term trajectory with ID Vehicle-123 after quality-aware fusion in the situation map will be instantiated as a vehicle object with the corresponding three-dimensional appearance in the reference road network model. The real-time position, speed, acceleration, orientation, and other kinematic parameters of the vehicle object will strictly follow its trajectory data in the situation map. At the same time, other traffic parameters such as the average speed of the road segment will be mapped as dynamic attributes of the road segment object. Through this process, the static map is "activated" into a scene containing fresh traffic participants.

[0064] To ensure the strict synchronization of digital twins with the physical world, the platform receives updated global traffic situation maps at a configurable frequency (e.g., 0.5-1 Hz for regular monitoring, 1-2 Hz for emergency response) based on the needs of different application scenarios and the conditions of computing resources and network bandwidth. The specific update frequency will be dynamically adjusted according to the current processing load. Each update triggers a state refresh of all dynamic traffic participant objects in the road network model. More importantly, this process also assumes the responsibility of closed-loop verification. For example, after the control instruction changes the 2nd signal light at intersection A to green, the module continuously monitors the actual traffic state through the 2nd light reflected by the fusion data in the digital twin. If the traffic does not start as expected within the preset time (e.g., 2 seconds), the system will determine that the instruction execution has failed or the physical device has failed, and will feed back this abnormal state to the decision module. Through the framework of generative adversarial networks: the generator continuously creates extremely dangerous traffic scenarios, and the discriminator responds to or solves these extreme scenarios. The relationship between the two is "adversarial": the generator tries to create scenarios that can "fail" the discriminator, while the discriminator continuously learns to improve its ability to respond to various unknown risks through this confrontation. Through this self-contradictory and self-game training process, the platform can automatically and systematically find weaknesses in its control algorithm and strengthen them accordingly, so that the control system has undergone a much more rigorous and comprehensive "stress test" than any historical data set before being deployed in the real world, improving its safety and robustness.

[0065] The traffic road network digital twin model is constructed to upgrade discrete perception data to a structured, interactive, and deductive digital twin environment. This model is not only an accurate mirror of the physical world but also a virtual test field that can be used to analyze the causal relationships of complex traffic phenomena and simulate and verify the effectiveness of control strategies. At the same time, its closed-loop verification capability provides a guarantee for the reliable operation of the entire system.

[0066] Further, the traffic state is analyzed from the traffic road network digital twin model to predict potential traffic incidents and collision risks, and traffic control instructions for traffic guidance and early warning to vehicle terminals are generated; corresponding decision and control modules; the specific implementation process includes:

[0067] The platform abstracts the dynamic traffic road network digital twin model as a dynamic space-time graph, where each intersection and key road segment is defined as a node of the graph, and the roads connecting these nodes are defined as edges. The module extracts rich features from the road network model in real time to fill the graph: the features of the nodes include the real-time traffic flow, average speed, lane occupancy rate, and vehicle type distribution (e.g., heavy truck proportion is 15%) at that location; the weights of the edges dynamically represent the real-time travel time or correlation strength between nodes.

[0068] The dynamic spatio-temporal graph is input into a spatio-temporal graph neural network (STGNN) model trained on massive historical data and simulation data. The model can predict the probability, type, and spatio-temporal impact range of traffic incidents within a specific time window in the future (e.g., the next 15 minutes). For example, the model predicts that "in the next 2-5 minutes, the risk level of queue overflow is high risk (based on preset risk assessment standards)" by analyzing the inflow rate of the ramp upstream of intersection A and the traffic density of the main road, combined with historical statistical data and the current traffic pattern.

[0069] Based on the regional risk prediction graph output by the STGNN, decisions are made in a hierarchical reinforcement learning framework constrained by safety rules, which includes two levels: the meta-agent at the high level is responsible for observing the macro risk situation and setting strategic goals, for example, for the above overflow prediction, the meta-agent may set the strategic goal as "reduce the ramp congestion index by 50% within 10 minutes without significantly affecting the efficiency of the main road". After receiving the goal, the multiple local agents at the bottom level, such as the agents controlling the ramp signal and the upstream intersection signal, will combine more detailed real-time data to learn and output the optimal tactical action that can most efficiently achieve the macro goal under the premise of passing deterministic safety checks and not violating any safety rules in the local decision space (such as adjusting signal timing schemes and phase differences). The state space S of the bottom-level local agents is defined as a feature vector containing [real-time queue length vector of each entrance, average waiting time vector of each entrance, upstream traffic on the main road], and the action space A is designed as a discrete set containing 5 pre-set signal timing schemes for different traffic flow patterns. The hierarchical reinforcement learning framework also introduces a more resilient decision-making mechanism, in which the relationship between the multiple local agents at the bottom level is modeled as a "cooperative game". Each agent has its own independent reward function, as well as a shared reward function based on the entire regional macro goal, allowing for "altruistic" collaboration between agents through this mechanism. For example, the agent at intersection A may find through calculation that if it reduces the green light time by 5 seconds, it can create a complete "green wave" for intersection B, thereby bringing a shared positive reward to the entire system. The regional traffic control strategy formed ultimately is much more flexible than traditional centralized commands, and is also more efficient, especially in handling unexpected situations, with significantly stronger adaptive capabilities.

[0070] Finally, the optimal tactical actions output by the framework (e.g., reduce the ramp signal green signal ratio by 15%, extend the main road green light of the upstream signal by 10 seconds) are converted into a series of specific, standardized control instructions, which are then issued to the corresponding traffic control devices. At the same time, the platform also dynamically generates content-enhanced, predictive V2X messages. For example, generate an e-DENM message containing fields such as predictedRiskProbability: 0.9, riskType: QUEUE_SPILLBACK, timeToImpact: 300s, and recommendedAction: MERGE_LANE_EARLY, and broadcast it to connected vehicles entering the area through roadside units, guiding drivers to take evasive action in advance.

[0071] Based on digital twinning, forward-looking risk propagation prediction is carried out, and a hierarchical reinforcement learning framework is applied to collaborative control decision-making, which changes the mode of passive response to a single problem in traditional traffic management. By predicting and intervening in potential congestion and accident risks in advance, and formulating a collaborative optimization strategy from a macroscopic perspective of the entire road network, traffic risks are prevented and resolved from the source, changing the concept of traffic management from "how to handle after the event" to "how to prevent the event from happening", thereby fundamentally improving the safety and efficiency of road traffic.

[0072] The platform disclosed in the present application accesses and analyzes new and old, heterogeneous roadside facilities through hybrid device fingerprinting and protocol reverse technology, and establishes a real-time data quality evaluation system, breaking down "information silos" and ensuring the comprehensiveness and high reliability of data sources. Secondly, using quality-driven fusion algorithms and graph neural networks, multi-source data is refined into a high-fidelity global traffic situation map, and an interactive and deductive dynamic digital twinning model is constructed, realizing the qualitative change from discrete perception to global situation understanding. Based on digital twinning, forward-looking risk propagation prediction is carried out, and a hierarchical reinforcement learning framework is applied to formulate a networked collaborative control strategy, changing the traffic management from "event response" to "risk intervention" new paradigm, avoiding congestion and accidents at the source, and significantly improving road traffic safety and efficiency.

[0073] Embodiment Two

[0074] At B intersection, a multi-phase complex intersection connecting the main road and the urban expressway ramp, it is now the evening peak period, the west-east direction traffic is saturated, and there is a queue of vehicles at the ramp entrance. The specific implementation of the traffic data collection module is as follows:

[0075] The induction and control process of the platform is activated in the field of B intersection. First, the data acquisition module listens to the communication channel of the intersection area in real time, capturing the signals of three main data sources: a L company LiDAR located in the northwest corner of the intersection, a C company panoramic camera located in the southeast corner of the intersection, and a set of W company wireless magnetometer detectors at the ramp entrance. For the W company wireless magnetometer that communicates through a private protocol in the 433MHz frequency band, the module extracts its physical layer radio frequency fingerprint by analyzing its radio frequency signal, which has a carrier frequency offset of +4.1kHz and a signal envelope rise time of 15 microseconds. The peak frequency is determined by applying fast Fourier transform (FFT) to the captured I / Q sample data, and the offset is obtained by comparing it with the nominal frequency. At the same time, the network analysis unit finds that it sends data packets of 18 bytes in length with an average period of about 1.5 seconds by differentiating the timestamps of consecutive data packets and analyzing the distribution histogram of data packet length, which constitutes its network behavior fingerprint. By standardizing the quantitative feature vectors of these two fingerprints through Z-score to eliminate the dimension effect, and then inputting them into the support vector machine (SVM) classification model with radial basis function (RBF) kernel, the platform determines its identity and assigns it a unique traffic data source identifier UUID: fabc4567-89de-f012-3456-789abcdef0. For LiDAR and camera, the platform also completes identity filing by analyzing features such as PHY chip clock frequency offset (+48ppm) and I-frame interval jitter (standard deviation 12ms) of RTSP stream media.

[0076] Example Three

[0077] In B intersection, the specific implementation of the traffic data processing module is as follows:

[0078] Subsequently, the platform receives an 18-byte raw data stream from UUID: fabc4567-89de-f012-3456-789abcdef0 (a magnetometer from Company W). Since the device repository shows that its protocol is unknown, the data stream is routed to the protocol reverse engineering unit. Using a sequence alignment technique based on the Needleman-Wunsch algorithm, treating the continuous data packet byte sequence as a DNA sequence in biology, and finding highly conserved regions through alignment operations, it is found that the first 4 bytes are a fixed device serial number, and by calculating the entropy and conditional probability of the subsequent bytes, it is found that the 5th byte switches between 0x00 (no car) and 0x01 (car present), which is consistent with the characteristics of the status field, and is inferred to be the status field. The 6th and 7th bytes are a 16-bit integer representing the duration of the vehicle's presence in milliseconds. The data pattern inference engine intervenes and labels the 5th byte field as occupancy_status and the 6th and 7th byte fields as occupancy_duration_ms. Finally, a raw binary data stream is converted into a standardized traffic parameter object: {source_id: UUID: fabc4567-89de-f012-3456-789abcdef0, timestamp: 2025-07-20T18:44:30.150Z, events: [{event_type: occupancy, status: 1, duration_ms: 2350}]}.

[0079] After the data is standardized, the platform performs real-time multi-dimensional data quality assessment on all sources of traffic parameters. The data from Company L's laser radar has a timestamp delay of only 20 milliseconds, and its three-dimensional point cloud has a positioning error of less than ±5 cm in the current sunny weather, and the data fields are complete and intact, and its data reliability index (DQS) is calculated by a weighted sum formula (where the timeliness weight is 0.4, the spatial accuracy weight is 0.5, and the integrity weight is 0.1), the result is 0.98. The data from Company C's panoramic camera, due to local glare caused by sunlight, has a decrease in target detection confidence in the west side of the intersection area, and the analysis module queries an internal lookup table that maps weather data (such as a sun elevation angle of 22° and an azimuth angle of 255°) to a precision attenuation coefficient (0.88) based on the sensor orientation (270°), resulting in a dynamic reduction in its spatial accuracy evaluation score, and its final DQS is rated as 0.85. The data from Company W's magnetometer has a data delay of 1.5 seconds, and can only provide presence information, not accurate location, and its DQS is rated as 0.55.

[0080] Example Four

[0081] In B intersection, the implementation of the situation fusion module is as follows:

[0082] Subsequently, the platform's situation fusion module begins processing a white SUV that is passing through the intersection from west to east. The graph attention network (GAT) model within the module has already stitched together discrete detection results (each detection result is represented as a node containing a 2D bounding box, a 128-dimensional Re-ID feature vector, a timestamp, and a sensor DQS) of the SUV over the past few seconds into a continuous long-term trajectory. The module performs quality-aware weighted fusion of the position and velocity from both data sources. The laser radar of company L (DQS = 0.98) reports its position as (x: 15.22, y: -5.10) and velocity as 45.1 km / h; while the camera of company C (DQS = 0.85) reports its position as (x: 15.45, y: -5.32) and velocity as 44.5 km / h due to the influence of perspective and glare. The final fused position multiplies the position coordinate vector of the laser radar by its quality score 0.98, adds the position coordinate vector of the camera multiplied by its quality score 0.85, and then divides each component of the resulting vector by the sum of the two quality scores (1.83). X coordinate: (15.22*0.98+15.45*0.85) / 1.83≈15.33; Y coordinate: (-5.10*0.98+-5.32*0.85) / 1.83≈-5.2. The final fused position is (x: 15.33, y: -5.20) and the velocity is 44.9 km / h, which is significantly biased towards the high-reliability laser radar data.

[0083] Based on the fused high-reliability trajectory, the platform updates the dynamic traffic network digital twin model of B intersection in real time, instantiating each fused trajectory as a renderable 3D vehicle object and binding its position and rotation attributes to the trajectory data. This binding ensures that the digital twin remains synchronized with the physical world at a frequency of 10 Hz. At this point, the model clearly shows that there are 22 vehicles instantiated as dynamic objects on the straight lane from west to east, and their average speed has dropped to 15 km / h. In contrast, the east-to-west direction is relatively smooth, with an average speed of 55 km / h. More importantly, the model reveals a potential risk: the right-turn lane leading to the expressway ramp has formed a long queue, with the tail extending to the straight lane of the main road, interweaving with the straight traffic flow.

[0084] Example Five

[0085] In B intersection, the implementation of the decision and control module is as follows:

[0086] Finally, the decision and control module of the platform receives the dynamic traffic network digital twin model and abstracts it into a dynamic spatio-temporal graph. A spatio-temporal graph neural network (STGNN) model composed of a graph convolutional network (GCN) and a gated recurrent unit (GRU) analyzes the evolution of the graph, where the GCN layer is used to capture the spatial dependency between B intersection and its adjacent intersections, and the GRU layer is used to learn the law of the evolution of traffic states (such as queue length, lane occupancy) over time. The model gets a high probability risk prediction through the softmax output layer: "In the next 3 minutes, the probability of rear-end or lateral collision caused by ramp vehicle queue overflow on the west side of B intersection is 85%". Based on this prediction, the high-level meta-agent using the "soft actor-critic" (SAC) algorithm sets a strategic goal, such as {target ramp queue length: <20 meters, target main road delay increment: <10 seconds}. After receiving the goal, the bottom-layer local agent using the "proximal policy optimization" (PPO) algorithm to control the signal lights of B intersection, in its action space containing 5 pre-set emergency signal schemes and 2 fine-tuning actions, after fast learning optimization, outputs a tactical action, which is converted into specific control instructions: the signal light controller of B intersection reduces the east-to-south left-turn green light time by 5 seconds in the current cycle, and increases the west-to-east straight green light by 5 seconds; and generates an e-DENM message containing predictedRiskProbability: 0.85, riskType: TRAFFIC_JAM_MERGING_CONFLICT, recommendedAction: REDUCE_SPEED_AND_WATCH_FOR_MERGING_VEHICLES and other fields, which is broadcast to all connected vehicles within 500 meters of B intersection through the roadside unit. After the platform issues the instructions, it continuously monitors the actual traffic changes at B intersection through the digital twin to verify the execution effect of the control instructions.

[0087] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A road traffic safety facility intelligent network connection induction control platform, characterized in that, Comprise: Traffic data collection module: collect data streams of multiple roadside traffic sensors, extract physical layer features of signal impairment caused by hardware tolerance in traffic sensor communication signals, generate physical layer radio frequency fingerprints; extract network behavior features including protocol header, data packet sequence and communication timing mode from traffic sensor communication data streams, generate network behavior fingerprints; combine physical layer radio frequency features and network communication behavior recognition techniques through a weighted scoring mechanism to establish traffic data source identification and output traffic data with the traffic data source identification; Traffic data processing module: adaptively parse protocols according to the traffic data source identification, parse the traffic data into standardized traffic parameters; calculate the delay of the timestamp of the traffic parameter and the current processing time, quantify the timeliness; evaluate the spatial positioning error according to the inherent technical specifications of the traffic sensor of the traffic parameter and the current working environment factors, quantify the spatial accuracy; check whether the traffic parameter contains all predefined key business fields, quantify the integrity; and according to the index system covering timeliness, spatial accuracy and integrity, weighted sum the quantification results of timeliness, spatial accuracy and integrity, and evaluate the data reliability index of the traffic parameter; Situation fusion module: take traffic targets in the traffic parameter derived from different times and different traffic sensors as nodes of a graph; based on the proximity and appearance feature similarity of the traffic targets in space and time, establish edges between the nodes representing potential associations, construct a multi-target association graph, infer the multi-target association graph through a graph neural network, calculate the association confidence represented by each edge, and based on the high-confidence association path, associate the traffic targets in the traffic parameter to form continuous vehicle driving trajectories, and use the data reliability index as a dynamic weight to fuse the continuous vehicle driving trajectories and traffic parameters, generating a global traffic situation graph; use the global traffic situation graph to construct and update a dynamic traffic road network digital twin model in real time; Decision and control module: analyze traffic status from the traffic road network digital twin model, predict potential traffic events and collision risks, and generate traffic control instructions for traffic guidance and early warning to vehicle terminals. 2.The intelligent networked induction control platform for road traffic safety facilities according to claim 1, characterized in that, The specific implementation process of parsing into standardized traffic parameters includes: According to the traffic data source identification, query the equipment archive to determine the communication protocol type, if the communication protocol type is known, call the standard parser, if it is unknown and private, start protocol reverse engineering for unknown protocols that conform to regular features, infer the data payload including message format, field boundary and data type through sequence comparison, statistical analysis and heuristic algorithm; analyze the data payload through the data mode inference engine, identify and label the key business fields including vehicle identity, speed, position and timestamp in the data payload, and convert the traffic data into the standardized traffic parameters. 3.The intelligent networked induction control platform for road traffic safety facilities according to claim 1, characterized in that, The specific generation process of constructing the traffic road network digital twin model includes: A traffic network digital twin model is constructed, which includes the topology of the road network, the static properties of the traffic facilities, and the traffic rules. The vehicle trajectories and traffic parameters in the global traffic situation map are taken as real-time data sources, and are instantiated as dynamic traffic participant objects with real-time positions, speeds, and properties in the traffic network digital twin model. Updated global traffic situation maps are continuously received, and the states and mutual relationships of the dynamic traffic participant objects are refreshed at high frequency. 4.The intelligent networked induction control platform for road traffic safety facilities according to claim 1, characterized in that, The specific implementation process of generating the traffic control instructions includes: The traffic network digital twin model is abstracted as a dynamic space-time graph, intersections and road segments are taken as nodes of the graph, connections between the nodes are taken as edges, and the features of the nodes include real-time traffic flow parameters. A space-time graph neural network is applied to analyze the evolution of the dynamic space-time graph, and prediction results including the probability, type, and space-time influence range of traffic events occurring in a specific time window in the future are obtained. A hierarchical reinforcement learning framework is applied to make control decisions. High-level meta-agents set macro traffic optimization strategy targets according to the prediction results, and low-level local agents learn and output optimal tactical actions according to the traffic optimization strategy targets and the traffic states in their respective regions. The tactical actions are converted into specific instructions for traffic guidance and signal timing optimization, and vehicle terminal early warning messages containing predictive risk attributes are generated.

Images