Multi-source sensing traffic abnormal event detection method and system

Abstract

The invention discloses a multi-source sensing traffic abnormal event detection method and system, and relates to the technical field of traffic abnormal detection.The method comprises the steps that synchronous data acquisition is executed at a multi-mode sensing array of a target intersection, and a multi-mode spatio-temporal data stream is obtained; carrying out traffic abnormity identification, and outputting a traffic abnormity event; traffic behavior backtracking is carried out, and a historical trajectory flow is obtained; performing event evolution trend prediction, and outputting a space-time influence vector; carrying out traffic influence range framing; generating a real-time traffic regulation and control intervention strategy by taking the traffic influence range as a dynamic constraint; and performing signal control real-time optimization and regional road network collaborative scheduling by adopting a real-time traffic regulation and control intervention strategy. According to the method and the device, the technical problems of insufficient traffic abnormal event detection accuracy and traffic regulation intervention lag in the prior art are solved, and the technical effects of realizing accurate detection of traffic abnormal events and real-time traffic regulation and control and improving intelligence and high efficiency of traffic management are achieved.

Description

Technical Field

[0001] This invention relates to the field of traffic anomaly detection technology, specifically to a multi-source sensing method and system for detecting traffic anomalies. Background Technology

[0002] With the acceleration of urbanization and the surge in motor vehicle ownership, urban road network traffic flow continues to climb, leading to frequent traffic anomalies such as traffic violations, accidents, and congestion. These anomalies not only severely reduce road network efficiency and increase travel time costs but also easily trigger secondary accidents and other chain reactions, placing stringent demands on the real-time, accurate, and coordinated nature of traffic management. Existing traffic anomaly detection technologies mostly rely on single-sensor data collection, which suffers from drawbacks such as data asynchrony, limited detection dimensions, and low anomaly identification accuracy. Furthermore, they lack the ability to effectively predict the evolution trend of anomalies, resulting in delayed traffic control interventions, insufficient regional road network coordinated scheduling, and difficulty in quickly responding to and mitigating the traffic impact of anomalies. Consequently, they fail to meet the demands of modern traffic management for intelligent and efficient handling.

[0003] Existing technologies suffer from insufficient accuracy in detecting traffic anomalies and delayed traffic control interventions. Summary of the Invention

[0004] This application provides a multi-source sensing method and system for detecting traffic anomalies, which addresses the technical problems of insufficient accuracy in detecting traffic anomalies and lagging traffic control intervention in existing technologies.

[0005] In view of the above problems, this application provides a method and system for detecting abnormal traffic events using multi-source sensing.

[0006] The first aspect of this application provides a method for detecting traffic anomalies using multi-source sensing, the method comprising:

[0007] A multimodal sensor array deployed at the target intersection is driven by the NTP protocol to perform synchronous data acquisition, obtaining a multimodal spatiotemporal data stream. Based on this multimodal spatiotemporal data stream, traffic anomaly identification is performed, outputting traffic anomaly events. These events include traffic event type, traffic anomaly subject, traffic anomaly behavior, and anomaly event coordinates. Traffic behavior of the traffic anomaly subject is traced back to obtain a historical trajectory stream. Based on the historical trajectory stream, traffic event type, and traffic anomaly behavior, event evolution trend prediction is performed, outputting a spatiotemporal impact vector. The traffic impact range is defined based on the spatiotemporal impact vector and anomaly event coordinates. Using the traffic impact range as a dynamic constraint, a real-time traffic control and intervention strategy is generated. This real-time traffic control and intervention strategy is then used for real-time optimization of signal control and coordinated scheduling of the regional road network.

[0008] A second aspect of this application provides a multi-source sensing traffic anomaly detection system, the system comprising:

[0009] A spatiotemporal data stream acquisition module is used to drive a multimodal sensor array deployed at the target intersection to perform synchronous data acquisition via the NTP protocol, thereby obtaining a multimodal spatiotemporal data stream. A traffic anomaly event output module is used to identify traffic anomalies based on the multimodal spatiotemporal data stream and output traffic anomaly events, wherein the traffic anomaly events include traffic event type, traffic anomaly subject, traffic anomaly behavior, and anomaly event coordinates. A historical trajectory stream acquisition module is used to backtrack the traffic behavior of the traffic anomaly subject to obtain a historical trajectory stream. A spatiotemporal influence vector output module is used to predict the event evolution trend based on the historical trajectory stream, traffic event type, and traffic anomaly behavior, and output a spatiotemporal influence vector. A traffic influence range definition module is used to define the traffic influence range based on the spatiotemporal influence vector and anomaly event coordinates. A traffic control intervention strategy generation module is used to generate a real-time traffic control intervention strategy using the traffic influence range as a dynamic constraint. A collaborative scheduling module is used to perform real-time signal control optimization and regional road network collaborative scheduling using the real-time traffic control intervention strategy. One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0010] A multimodal sensor array deployed at the target intersection is driven by the NTP protocol to perform synchronous data acquisition, obtaining a multimodal spatiotemporal data stream. Based on this data stream, traffic anomaly identification is performed, outputting traffic anomaly events. Each traffic anomaly event includes a traffic event type, a traffic anomaly subject, a traffic anomaly behavior, and anomaly event coordinates. Traffic behavior of the traffic anomaly subject is traced back to obtain a historical trajectory stream. Based on the historical trajectory stream, traffic event type, and traffic anomaly behavior, event evolution trend prediction is performed, outputting a spatiotemporal impact vector. The traffic impact range is defined based on the spatiotemporal impact vector and anomaly event coordinates. Using this impact range as a dynamic constraint, a real-time traffic control intervention strategy is generated. This strategy is then used for real-time optimization of signal control and coordinated scheduling of the regional road network. This achieves the technical effect of accurate detection and real-time traffic control of traffic anomaly events, improving the intelligence and efficiency of traffic management. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1A schematic flowchart of a multi-source sensing traffic anomaly detection method provided in an embodiment of this application;

[0013] Figure 2 This is a schematic diagram of the structure of a multi-source sensing traffic anomaly detection system provided in an embodiment of this application.

[0014] Figure labeling: 10 Spatiotemporal data stream acquisition module, 20 Traffic anomaly event output module, 30 Historical trajectory stream acquisition module, 40 Spatiotemporal influence vector output module, 50 Traffic influence range delineation module, 60 Traffic control intervention strategy generation module, 70 Coordinated scheduling module. Detailed Implementation

[0015] This application provides a multi-source sensing method and system for detecting abnormal traffic events, which addresses the technical problems of insufficient accuracy in detecting abnormal traffic events and lagging traffic control intervention in existing technologies.

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0017] Example 1, as Figure 1 As shown, this application provides a multi-source sensing method for detecting abnormal traffic events, the method comprising:

[0018] Step S100: Drive the multimodal sensor array deployed at the target intersection to perform synchronous data acquisition via the NTP protocol to obtain a multimodal spatiotemporal data stream.

[0019] Specifically, at the target intersection, based on the effective monitoring range of each sensor, a multimodal sensor array covering the intersection is deployed, including a video surveillance array, a millimeter-wave radar array, a geomagnetic detection array, and a lidar array. The multimodal sensor array is driven by the NTP protocol to perform synchronous data acquisition at the target intersection, acquiring array-type multi-view video streams, array-type target trajectory point clouds, array-type lane magnetic field disturbance sequences, and array-type high-precision spatial sampling streams. Then, according to the overlapping topology of the field of view of each sensor array, the above-mentioned data streams are spatiotemporally aligned and stitched together to finally obtain the spatiotemporal trajectory stream of the moving target, the lane-level physical field stream, the dynamic road network structure stream, and the holographic traffic scene stream, which together constitute a multimodal spatiotemporal data stream.

[0020] Step S200: Based on the multimodal spatiotemporal data stream, traffic anomaly identification is performed, and traffic anomaly events are output. The traffic anomaly events include traffic event type, traffic anomaly subject, traffic anomaly behavior, and anomaly event coordinates.

[0021] Specifically, firstly, a multi-target tracking algorithm is used to segment multiple vehicle-level spatiotemporal trajectory sub-streams from the moving target spatiotemporal trajectory stream in the multimodal spatiotemporal data stream. Then, based on these vehicle-level spatiotemporal trajectory sub-streams, lane-level physical field streams, dynamic road network structure streams, and holographic traffic scene streams are mapped and segmented to obtain multiple vehicle magnetic field fingerprints, multiple vehicle traffic permission domains, and multiple vehicle risk fields. Subsequently, a first vehicle coordinate system is established with the center of the front bumper of the first traveling vehicle as the spatial origin and the direction of movement as the positive X-axis. This coordinate system is used to unify the spatiotemporal references of the aforementioned vehicle-level spatiotemporal trajectory sub-streams, vehicle magnetic field fingerprints, vehicle traffic permission domains, and vehicle risk fields. A sliding segmentation method using a preset time window is used to construct a system containing motion state vectors, physical field vectors, and road... Multiple state vector slices of weight vector and risk vector; based on pre-constructed sample state vectors of various sample abnormal event types, the matching confidence of the state vector slices and sample state vectors is calculated using a pre-set confidence threshold rule. Abnormal state slices are obtained by filtering based on deviation from extreme value features. If the abnormal state slice is a non-empty set, the dominant abnormal event type is traversed and statistically analyzed through a preset continuous time window as the traffic event type. The vehicle fingerprint identifier of the first driving vehicle is called as the traffic abnormal subject. The peak slice is retrieved and aggregated according to the event type and behavioral parameter distribution mapping to generate traffic abnormal behavior. The GIS coordinates of the peak potential energy point of the vehicle's risk field are called as the abnormal event coordinates. Finally, the traffic abnormal event containing the above four pieces of information is output.

[0022] Step S300: Perform traffic behavior backtracking on the traffic anomaly subject to obtain historical trajectory flow.

[0023] Specifically, based on the vehicle fingerprint of the traffic anomaly subject identified in the identified traffic anomaly event, i.e. the target abnormal vehicle, the historical collection records corresponding to the vehicle fingerprint are retrieved from the multimodal spatiotemporal data stream that was synchronously collected by the multimodal sensor array and spatiotemporally aligned and stitched together. The relevant data such as the spatiotemporal trajectory substream of the abnormal vehicle before the traffic anomaly event occurred, the vehicle magnetic field fingerprint, the vehicle passage permission domain and the vehicle risk field are extracted and integrated after spatiotemporal benchmark unification to form a complete historical trajectory stream that reflects its past driving status and path.

[0024] Step S400: Based on the historical trajectory flow, traffic event type, and abnormal traffic behavior, predict the event evolution trend and output the spatiotemporal impact vector.

[0025] Specifically, the core input data consists of traffic event types and abnormal traffic behaviors identified by traffic anomaly identification, as well as historical trajectory streams obtained by tracing back the subjects of traffic anomalies. The data is integrated with relevant information from the historical trajectory stream, such as the spatiotemporal trajectory sub-streams of abnormal vehicle past movements, vehicle magnetic field fingerprints, vehicle access permission domains, and vehicle risk fields, after being unified by a spatiotemporal reference. Combined with the abnormal characteristics corresponding to the traffic event type and the parameter distribution of abnormal traffic behaviors, a trend prediction model is used to analyze the continuous diffusion pattern of the traffic anomaly event in the time dimension and the degree of radiation impact on surrounding lanes and road network nodes in the spatial dimension. Finally, a spatiotemporal impact vector containing key parameters such as the time span of impact, the spatial radius of impact, and the probability of road network congestion propagation is quantitatively output.

[0026] Step S500: Define the traffic impact range based on the spatiotemporal impact vector and the coordinates of the abnormal event.

[0027] Specifically, the GIS coordinates of the abnormal event output by the traffic anomaly, i.e. the peak point of the risk field potential energy of the target abnormal vehicle, are used as the core spatial benchmark. Combined with key parameters such as the time span of the impact, the spatial radius of the impact, and the probability of road network congestion propagation contained in the spatiotemporal impact vector output by the event evolution trend prediction, and superimposed with the dynamic road network structure flow of the target intersection and the topological relationship of the surrounding road network, the continuous impact range of the traffic anomaly in the time dimension and the radiation coverage of the surrounding lanes, intersections and related road network nodes in the spatial dimension are accurately defined through spatial range modeling and impact boundary calculation, forming a clear definition of the traffic impact range.

[0028] Step S600: Using the traffic impact range as a dynamic constraint, generate a real-time traffic control and intervention strategy.

[0029] Specifically, the defined traffic impact range, including the time duration of impact, spatial radiation coverage, and the impact boundaries on surrounding lanes, intersections, and related road network nodes, serves as dynamic constraints. Combined with the dynamic road network structure flow, lane-level physical field flow, and regional road network topology in the multimodal spatiotemporal data stream, and considering key parameters such as the type of traffic anomaly, the degree of impact of abnormal traffic behavior, and the probability of road network congestion propagation in the spatiotemporal impact vector, targeted real-time traffic control intervention strategies are formulated. These strategies include adjusting lane access permissions around the anomaly area, optimizing signal timing at related intersections, planning diversion paths, and allocating road network traffic priorities. This ensures that the strategies accurately match the impact scenarios of traffic anomalies, providing a clear basis for subsequent signal control optimization and regional road network collaborative scheduling.

[0030] Step S700: Using the real-time traffic control intervention strategy, perform real-time optimization of signal control and coordinated scheduling of regional road network.

[0031] Specifically, based on the generated real-time traffic control and intervention strategy, and combined with the dynamic road network structure flow, lane-level physical field flow, and regional road network topology in the multimodal spatiotemporal data stream, the signal timing scheme is dynamically adjusted for related intersections within the traffic impact range. This includes optimizing green light duration and adjusting phase sequence to achieve real-time optimization of signal control. Simultaneously, based on the diversion path planning and road network traffic priority allocation rules defined in the strategy, the system links with traffic control nodes in the surrounding regional road network to coordinate and schedule the traffic flow of different road segments. Through the coordinated cooperation of various intersections and road segments within the region, the system disperses traffic pressure in the abnormal event area, prevents congestion from spreading, achieves coordinated scheduling of the regional road network, and ensures overall traffic efficiency.

[0032] In one possible implementation, step S100 further includes:

[0033] Step S110: Based on the effective monitoring range of the sensors, deploy a multimodal sensor array at the target intersection, wherein the multimodal sensor array includes a video surveillance array, a millimeter-wave radar array, a geomagnetic detection array, and a lidar array.

[0034] Step S120: Drive the multimodal sensor array through the NTP protocol to perform synchronous data acquisition on the target intersection, and obtain array-type multi-view video stream, array-type target trajectory point cloud, array-type lane magnetic field disturbance sequence and array-type high-precision spatial sampling stream.

[0035] Step S130: Based on the overlapping topological relationship of the field of view of the video surveillance array, millimeter-wave radar array, geomagnetic detection array and lidar array, the array-type multi-view video stream, array-type target trajectory point cloud, array-type lane magnetic field disturbance sequence and array-type high-precision spatial sampling stream are spatiotemporally aligned and stitched together to obtain the moving target spatiotemporal trajectory stream, lane-level physical field stream, dynamic road network structure stream and holographic traffic scene stream, thus forming the multimodal spatiotemporal data stream.

[0036] Specifically, during the planning and deployment phase of the target intersection, the core performance parameters of the video surveillance array, millimeter-wave radar array, geomagnetic detection array, and lidar array, such as their effective monitoring distance, detection angle, and data acquisition accuracy, are fully considered. Combined with the number of lanes at the intersection, the layout of the intersection area, the road width, and the distribution of key traffic nodes in the surrounding area, a multi-modal sensor array deployment design is implemented for full coverage. By rationally planning the installation positions and angles of each sensor array, it is ensured that the video surveillance array can comprehensively capture visual images of vehicles, pedestrians, and other targets at the intersection; the millimeter-wave radar array accurately detects the spatial position and movement trajectory of targets; the geomagnetic detection array effectively senses the magnetic field disturbances caused by vehicles passing through each lane; and the lidar array achieves high-precision sampling of the intersection's spatial environment and road network details. Simultaneously, the monitoring ranges of the four types of sensor arrays are interconnected and complementary, with no blind spots, ultimately achieving comprehensive and seamless perception coverage of the traffic scene at the target intersection.

[0037] First, based on the time synchronization characteristics of the NTP protocol, unified time calibration and drive control are performed on the multimodal sensor arrays that have been deployed and covered, including video surveillance arrays, millimeter-wave radar arrays, geomagnetic detection arrays, and lidar arrays. This ensures that the four types of sensor arrays work collaboratively under the same time reference, avoiding time differences in data acquisition. Then, a synchronous data acquisition process is initiated. The video surveillance array captures dynamic visual images of traffic participants such as vehicles and pedestrians within the target intersection through a multi-view layout, generating an array-style multi-view video stream. The millimeter-wave radar array accurately captures the spatial position, speed, and trajectory information of the target through electromagnetic wave detection technology, forming an array-style target trajectory point cloud. The geomagnetic detection array is deployed in each lane to sense changes in magnetic field disturbances caused by vehicle passage, outputting an array-style lane magnetic field disturbance sequence. The lidar array achieves high-precision three-dimensional sampling of the intersection spatial environment, road network structure, and targets through laser scanning, acquiring an array-style high-precision spatial sampling stream. Finally, the above four types of targeted raw sensor data streams are collected synchronously.

[0038] First, the overlapping topological relationships of the field of view of the video surveillance array, millimeter-wave radar array, geomagnetic detection array, and lidar array after deployment at the target intersection are analyzed to clarify the intersection and complementary range of the monitoring areas of each sensor array. Based on this, a spatial correlation mapping of multi-source data is established. Then, a spatiotemporal alignment algorithm is adopted, using a unified timestamp and spatial coordinate system as a reference, to perform time synchronization calibration and spatial position matching on the collected array-type multi-view video stream, array-type target trajectory point cloud, array-type lane magnetic field disturbance sequence, and array-type high-precision spatial sampling stream, thereby eliminating the spatiotemporal deviation between data from different sensors. Through data stitching and fusion processing, the visual information of traffic targets in multi-view video streams is associated with the spatial motion information of trajectory point clouds. Combined with the traffic flow change characteristics reflected by the lane magnetic field disturbance sequence, and superimposed with high-precision road network spatial data collected by lidar, these data streams are integrated to form a spatiotemporal trajectory stream of moving targets that can accurately depict the motion state of traffic participants, a lane-level physical field stream that reflects the real-time traffic status of lanes, a dynamically updated road network structure stream, and a holographic traffic scene stream that fully restores the traffic scene at the intersection. These four types of data streams together constitute a complete multimodal spatiotemporal data stream.

[0039] In one possible implementation, step S200 further includes:

[0040] Step S210: Divide the spatiotemporal trajectory stream of the moving target into multiple vehicle-level spatiotemporal trajectory sub-streams using a multi-target tracking algorithm.

[0041] Step S220: Based on the multiple vehicle-level spatiotemporal trajectory sub-stream mappings, segment the lane-level physical field flow, dynamic road network structure flow, and holographic traffic scene flow to obtain multiple vehicle magnetic field fingerprints, multiple vehicle access permission domains, and multiple vehicle risk fields.

[0042] Step S230: After spatiotemporally aligning the multiple vehicle-level spatiotemporal trajectory sub-streams, multiple vehicle magnetic field fingerprints, multiple vehicle access permission domains, and multiple vehicle risk fields, perform single-vehicle behavior flow anomaly detection to locate the traffic anomaly event of the target abnormal vehicle.

[0043] Specifically, for the spatiotemporal trajectory stream of moving targets that has been integrated in the multimodal spatiotemporal data stream, a multi-target tracking algorithm is used to perform fine segmentation processing. This algorithm identifies the motion characteristics of different traffic targets in the trajectory stream, such as driving speed, direction, trajectory curvature, spatial position differences and trajectory continuity, and establishes a target association matching model. The trajectory stream, which originally contained comprehensive motion information of multiple vehicles, is accurately separated into independent data units corresponding to each driving vehicle, i.e., multiple vehicle-level spatiotemporal trajectory sub-streams.

[0044] Using multiple vehicle-level spatiotemporal trajectory substreams segmented from the spatiotemporal trajectory stream of moving targets through a multi-target tracking algorithm as the core mapping basis, and leveraging the precise spatiotemporal coordinate information contained in each vehicle-level spatiotemporal trajectory substream, targeted mapping and segmentation are performed on the lane-level physical field stream, dynamic road network structure stream, and holographic traffic scene stream in the multimodal spatiotemporal data stream. Specifically, for the lane-level physical field stream, magnetic field disturbance feature data corresponding to the driving trajectory of each vehicle is extracted to form multiple vehicle magnetic field fingerprints that can uniquely identify the magnetic field characteristics of a single vehicle; for the dynamic road network structure stream, the legal passage boundaries and ranges of each vehicle in the current road network environment are delineated by combining the road network region corresponding to the vehicle trajectory, resulting in multiple vehicle passage permission domains; for the holographic traffic scene stream, multiple vehicle risk fields that can reflect the distribution of traffic risks around a single vehicle are constructed based on scene information such as the distribution of traffic participants and road conditions around the vehicle trajectory. Ultimately, the precise association between the lane-level physical field stream, dynamic road network structure stream, and holographic traffic scene stream and the vehicle-level spatiotemporal trajectory substream is achieved, outputting multi-dimensional feature data matching a single vehicle.

[0045] First, a first vehicle coordinate system is established with the center of the front bumper of the first vehicle as the spatial origin and the direction of motion as the positive X-axis. This coordinate system is used to unify the spatiotemporal references of multiple vehicle-level spatiotemporal trajectory substreams, multiple vehicle magnetic field fingerprints, multiple vehicle access permission domains, and multiple vehicle risk fields, eliminating spatiotemporal deviations between data of different dimensions. Then, a preset time window is used to perform sliding segmentation on the unified single-vehicle related data, including vehicle-level spatiotemporal trajectory substreams, vehicle magnetic field fingerprints, vehicle access permission domains, and vehicle risk fields, constructing multiple state vector slices containing motion state vectors, physical field vectors, right-of-way vectors, and risk vectors. Based on the pre-constructed sample state of multiple sample anomaly event types... The state vector is used to calculate the matching confidence of each state vector slice with the sample state vector using a pre-set confidence threshold rule. Abnormal state slices are obtained by filtering based on deviation from extreme value features. If the abnormal state slice is a non-empty set, the dominant abnormal event type is traversed and statistically analyzed through a preset continuous time window as the traffic event type. The vehicle fingerprint identifier of the vehicle is called as the traffic abnormal subject. The peak slice is retrieved and aggregated to generate traffic abnormal behavior based on the event type and the distribution mapping of behavioral parameters of the abnormal state slice. The GIS coordinates of the peak potential energy point of the vehicle's risk field are called as the abnormal event coordinates to complete the detection of single-vehicle behavior flow anomalies and locate the target abnormal vehicle. The traffic abnormal event containing the above four core information items is output.

[0046] In one possible implementation, step S230 further includes:

[0047] Step S231: Establish the first vehicle coordinate system with the center of the front bumper of the first vehicle as the spatial origin and the direction of movement as the positive X-axis.

[0048] Step S232: Use the first vehicle coordinate system to unify the spatiotemporal reference of the first vehicle-level spatiotemporal trajectory sub-stream, the first vehicle magnetic field fingerprint, the first vehicle access permission domain, and the first vehicle risk field.

[0049] Step S233: The first vehicle-level spatiotemporal trajectory sub-stream, the first vehicle magnetic field fingerprint, the first vehicle access permission domain, and the first vehicle risk field are slidably segmented using a preset time window to construct multiple state vector slices.

[0050] Step S234: Perform anomaly detection based on the multiple state vector slices to obtain multiple abnormal state slices.

[0051] Step S235: If the multiple abnormal state slices are a non-empty set, then perform abnormal event verification on the multiple abnormal state slices and output the traffic abnormal event.

[0052] Specifically, for the first vehicle to be subjected to single-vehicle behavior flow anomaly detection, in order to achieve accurate spatiotemporal benchmark unification of its multi-dimensional data, the center of the front bumper of the vehicle is taken as the spatial origin, the current driving direction of the vehicle is defined as the positive X-axis, and the Y-axis perpendicular to the driving direction and the Z-axis perpendicular to the road surface are defined in combination with the vehicle's own structure and driving characteristics to construct a unique first vehicle coordinate system.

[0053] Using the established first vehicle coordinate system as a unified spatiotemporal reference standard, spatiotemporal benchmark calibration is performed on the first vehicle-level spatiotemporal trajectory sub-stream, the first vehicle magnetic field fingerprint, the first vehicle access permission domain, and the first vehicle risk field corresponding to the first vehicle. Specifically, for the first vehicle-level spatiotemporal trajectory sub-stream, its original spatial coordinates are converted to relative coordinates under the first vehicle coordinate system, and the trajectory timing is corrected based on a unified timestamp. For the first vehicle magnetic field fingerprint, the spatial orientation and distribution correlation of the magnetic field disturbance data are calibrated according to the spatial orientation of the vehicle coordinate system. For the first vehicle access permission domain, the spatial parameter definition of the access boundary is adjusted according to the dimensional specifications of the vehicle coordinate system. For the first vehicle risk field, the spatial distribution model and temporal correlation logic of the risk potential energy are reconstructed based on the vehicle coordinate system. Ultimately, the benchmarks of the four types of data are unified in both spatial and temporal dimensions, eliminating spatiotemporal deviations caused by differences in the original acquisition reference system and ensuring that all types of data accurately correspond to the actual driving state of the first vehicle.

[0054] After unifying the spatiotemporal references of the first vehicle-level spatiotemporal trajectory substream, the first vehicle magnetic field fingerprint, the first vehicle access permission domain, and the first vehicle risk field, a time window of preset duration is used to perform sliding segmentation processing on the above four types of unified continuous data. According to the stepping rules of the time window, relevant data segments within the corresponding time period are extracted sequentially. The corresponding data of the first vehicle-level spatiotemporal trajectory substream, the first vehicle magnetic field fingerprint feature data, the first vehicle access permission domain parameter data, and the first vehicle risk field distribution data within each time period are integrated to construct multiple state vector slices containing motion state vectors, physical field vectors, right-of-way vectors, and risk vectors. Each slice fully represents the comprehensive state of the first vehicle in the dimensions of motion, physical field, right-of-way, and risk within that time period, providing structured data support for subsequent anomaly detection.

[0055] Multiple sample state vectors are pre-constructed, covering various sample abnormal event types, such as illegal lane changing, speeding, running red lights, and driving in the wrong direction. Each sample state vector contains standard motion state vector, physical field vector, right-of-way vector, and risk vector features under the corresponding abnormal event type. Then, for the constructed multiple state vector slices, a pre-set confidence threshold rule is used to calculate the matching confidence between each state vector slice, i.e., the first state vector slice, and the sample state vectors corresponding to various sample abnormal event types, quantifying the degree of feature fit between the two. Finally, by analyzing the deviation extreme value features of multiple matching confidence, the matching results that exceed the preset threshold for the feature matching degree of the sample abnormal event type and deviate from the normal driving state features are selected. The corresponding sample abnormal event type is output as the first abnormal state slice. Finally, multiple abnormal state slices are integrated to complete the anomaly screening of each state vector slice.

[0056] If multiple abnormal state slices obtained through anomaly detection are non-empty sets, indicating that the first vehicle is exhibiting abnormal behavior, the abnormal event confirmation process is initiated: First, all abnormal state slices are traversed using a preset continuous time window, and the dominant abnormal event type with the highest frequency and most significant impact is statistically analyzed, identified as the traffic event type, and output. Next, the unique first vehicle fingerprint identifier of the first vehicle is used as the unique corresponding traffic anomaly subject. Then, based on the identified traffic event type, and combined with the distribution characteristics of behavioral parameters such as motion state vectors, physical field vectors, right-of-way vectors, and risk vectors contained in multiple abnormal state slices, slices that can centrally represent the peak of abnormal behavior are mapped and retrieved from the abnormal state slices, and specific traffic anomaly behavior is generated through data aggregation processing. Finally, the GIS coordinates corresponding to the peak point of the risk field potential energy of the first vehicle are used as the abnormal event coordinates to accurately locate the location of the anomaly. Ultimately, the four core pieces of information—traffic event type, traffic anomaly subject, traffic anomaly behavior, and abnormal event coordinates—are integrated to output the complete traffic anomaly event.

[0057] In one possible implementation, step S233 further includes:

[0058] The state vector slice includes motion state vector, physical field vector, path weight vector, and risk vector.

[0059] Specifically, the state vector slice is a structured data unit that integrates the multi-dimensional operational characteristics of the first vehicle. Its core comprises four types of vectors: motion state vector, which mainly covers parameters reflecting the vehicle's dynamic motion attributes such as speed, acceleration, direction, and trajectory curvature, accurately depicting the vehicle's motion state within a corresponding time period; physical field vector, constructed based on the magnetic field disturbance characteristics caused by vehicle passage, including key information such as magnetic field disturbance intensity, distribution range, and frequency of change, reflecting the interaction between the vehicle and its surrounding physical environment; right-of-way vector, clarifying the vehicle's legal passage boundaries, priority, and related right-of-way restrictions in the current road network environment, providing a basis for judging whether the vehicle has engaged in unauthorized driving behavior; and risk vector, which comprehensively considers factors such as the distribution of traffic participants around the vehicle, road obstacles, and traffic flow density, quantitatively characterizing the traffic risk level, risk distribution range, and high-risk area location around the vehicle, comprehensively reflecting the safety situation during vehicle operation. These four types of vectors work together to form a complete state vector slice.

[0060] In one possible implementation, step S234 further includes:

[0061] Preconstruct multiple sample state vectors for various sample anomaly event types.

[0062] Using a pre-set confidence threshold rule, multiple matching confidence scores are calculated for the first state vector slice and the multiple sample state vectors.

[0063] Based on the deviation extreme value characteristics of the multiple matching confidence scores, the first sample abnormal event type is selected and output as the first abnormal state slice.

[0064] Specifically, a sample abnormal event type library is pre-built to cover a variety of common traffic abnormal scenarios, including but not limited to illegal lane changing, speeding, running red lights, driving in the wrong direction, illegal parking, and low-speed lane occupation. For each sample abnormal event type, its typical features are extracted and multiple corresponding sample state vectors are constructed. Each sample state vector consists of a motion state vector, a physical field vector, a right-of-way vector, and a risk vector, forming a complete sample feature library.

[0065] Using pre-set confidence threshold rules, the weights of motion state vector, physical field vector, path right-of-way vector, and risk vector are defined as 0.4, 0.3, 0.2, and 0.1, respectively. Based on this, the matching confidence of the first state vector slice with multiple sample state vectors corresponding to multiple sample abnormal event types is calculated. First, the feature matching degree of the motion state vector, physical field vector, path right-of-way vector, and risk vector in the first state vector slice with the corresponding dimension vectors of each sample state vector is calculated. Then, the matching degrees of the four dimensions are weighted and summed according to the set weight ratio to obtain the comprehensive matching confidence of the first state vector slice with each sample state vector. Finally, multiple matching confidence results corresponding one-to-one with the sample state vectors are generated.

[0066] After obtaining multiple matching confidence scores of the first state vector slice and multiple sample state vectors, all matching confidence score data are first normalized to clarify the degree of deviation of each confidence score value from the baseline confidence score under normal driving conditions. Then, the extreme value characteristics of these matching confidence scores are analyzed in detail, and extreme matching confidence scores with deviations exceeding the preset normal range and values ​​reaching or exceeding the confidence threshold are selected, i.e., peak results that are significantly higher than other confidence scores. The sample abnormal event type corresponding to the extreme matching confidence score is determined as the first sample abnormal event type. At the same time, the motion state vector, physical field vector, right-of-way vector and risk vector features corresponding to the sample abnormal event type are integrated to form the first abnormal state slice and output it.

[0067] In one possible implementation, step S235 further includes:

[0068] The multiple abnormal state slices are traversed using a preset continuous time window to perform statistics on the dominant abnormal event types, and the traffic event type is output.

[0069] The first vehicle fingerprint identifier of the first driving vehicle is used as the subject of the traffic anomaly.

[0070] Based on the traffic event type and the distribution of behavioral parameters of multiple abnormal state slices, peak slices are retrieved from the multiple abnormal state slices and aggregated to generate the abnormal traffic behavior.

[0071] The GIS coordinates of the peak potential energy point of the first driving vehicle in the risk field are used as the coordinates of the abnormal event.

[0072] Specifically, after multiple abnormal state slices are non-empty sets, the abnormal event confirmation process is initiated to output the complete traffic abnormal event: First, a continuous time window of preset duration is used to traverse and analyze all abnormal state slices, and the frequency, duration and impact intensity of the abnormal event type corresponding to each abnormal state slice are counted. The abnormal event type with the highest frequency, longest duration and most significant impact is determined as the dominant abnormal event type, and this is used as the output traffic event type.

[0073] Next, the fingerprint identifier of the first vehicle, which is uniquely bound to the first vehicle, is invoked. This identifier can accurately link the vehicle's basic information and driving records, thus identifying it as the subject of the traffic anomaly.

[0074] Subsequently, based on the output traffic event types, we conducted an in-depth analysis of the distribution characteristics of behavioral parameters such as motion state vectors, physical field vectors, right-of-way vectors, and risk vectors contained in multiple abnormal state slices. We located the slices corresponding to the peak data that can centrally reflect the intensity and typical characteristics of abnormal behavior among each parameter, namely peak slices. Through feature extraction and data aggregation processing of peak slices, we generated specific and accurate descriptions of abnormal traffic behavior.

[0075] Finally, the GIS coordinates corresponding to the peak point with the highest potential energy in the risk field of the first vehicle are retrieved. These coordinates can accurately locate the specific spatial location where the abnormal event occurred. They are used as the coordinates of the abnormal event. Finally, the traffic event type, the subject of the traffic abnormality, the traffic abnormal behavior, and the coordinates of the abnormal event are integrated to complete the output of the traffic abnormal event.

[0076] Example 2, based on the same inventive concept as the multi-source sensing traffic anomaly detection method in the foregoing examples, such as... Figure 2 As shown, this application provides a multi-source sensing traffic anomaly detection system. The system and method embodiments in this application are based on the same inventive concept. The system includes:

[0077] The spatiotemporal data stream acquisition module 10 is used to drive the multimodal sensor array deployed at the target intersection to perform synchronous data acquisition through the NTP protocol, thereby obtaining a multimodal spatiotemporal data stream.

[0078] The traffic anomaly event output module 20 is used to identify traffic anomalies based on the multimodal spatiotemporal data stream and output traffic anomaly events, wherein the traffic anomaly events include traffic event type, traffic anomaly subject, traffic anomaly behavior, and anomaly event coordinates.

[0079] The historical trajectory stream acquisition module 30 is used to trace the traffic behavior of the traffic anomaly subject and obtain the historical trajectory stream.

[0080] The spatiotemporal influence vector output module 40 is used to predict the event evolution trend based on the historical trajectory flow, traffic event type and abnormal traffic behavior, and output the spatiotemporal influence vector.

[0081] The traffic impact range definition module 50 is used to define the traffic impact range based on the spatiotemporal impact vector and the coordinates of the abnormal event.

[0082] The traffic control intervention strategy generation module 60 is used to generate real-time traffic control intervention strategies with the traffic impact range as a dynamic constraint.

[0083] The collaborative scheduling module 70 is used to perform real-time optimization of signal control and collaborative scheduling of regional road networks using the real-time traffic control intervention strategy.

[0084] Furthermore, the system is also used to implement the following functions:

[0085] At the target intersection, a multimodal sensor array is deployed based on the effective monitoring range of the sensors. This multimodal sensor array includes a video surveillance array, a millimeter-wave radar array, a geomagnetic detection array, and a lidar array. The multimodal sensor array is driven by the NTP protocol to perform synchronous data acquisition at the target intersection, resulting in an array-based multi-view video stream, an array-based target trajectory point cloud, an array-based lane magnetic field disturbance sequence, and an array-based high-precision spatial sampling stream. Based on the overlapping topology of the field of view of the video surveillance array, millimeter-wave radar array, geomagnetic detection array, and lidar array, the array-based multi-view video stream, array-based target trajectory point cloud, array-based lane magnetic field disturbance sequence, and array-based high-precision spatial sampling stream are spatiotemporally aligned and stitched together to obtain a moving target spatiotemporal trajectory stream, a lane-level physical field stream, a dynamic road network structure stream, and a holographic traffic scene stream, thus constituting the multimodal spatiotemporal data stream.

[0086] Furthermore, the system is also used to implement the following functions:

[0087] Multiple vehicle-level spatiotemporal trajectory sub-streams are segmented from the spatiotemporal trajectory stream of the moving target using a multi-target tracking algorithm. The lane-level physical field stream, dynamic road network structure stream, and holographic traffic scene stream are then segmented based on the multiple vehicle-level spatiotemporal trajectory sub-streams to obtain multiple vehicle magnetic field fingerprints, multiple vehicle access permission domains, and multiple vehicle risk fields. After spatiotemporally aligning the multiple vehicle-level spatiotemporal trajectory sub-streams, multiple vehicle magnetic field fingerprints, multiple vehicle access permission domains, and multiple vehicle risk fields, single-vehicle behavior flow anomaly detection is performed to locate the traffic anomaly event of the target abnormal vehicle.

[0088] Furthermore, the system is also used to implement the following functions:

[0089] A first vehicle coordinate system is established with the center of the front bumper of the first vehicle as the spatial origin and the direction of motion as the positive X-axis. The first vehicle coordinate system is used to unify the spatiotemporal references of the first vehicle-level spatiotemporal trajectory sub-stream, the first vehicle magnetic field fingerprint, the first vehicle traffic permission domain, and the first vehicle risk field. A preset time window is used to slide and segment the first vehicle-level spatiotemporal trajectory sub-stream, the first vehicle magnetic field fingerprint, the first vehicle traffic permission domain, and the first vehicle risk field, constructing multiple state vector slices. Anomaly detection is performed based on the multiple state vector slices to obtain multiple abnormal state slices. If the multiple abnormal state slices are non-empty sets, abnormal events are confirmed on the multiple abnormal state slices, and the traffic anomaly event is output.

[0090] Furthermore, the system is also used to implement the following functions:

[0091] The state vector slice includes motion state vector, physical field vector, path weight vector, and risk vector.

[0092] Furthermore, the system is also used to implement the following functions:

[0093] Multiple sample state vectors of various sample abnormal event types are pre-constructed; multiple matching confidence scores of the first state vector slice and the multiple sample state vectors are calculated using a pre-set confidence threshold rule; based on the deviation extreme value characteristics of the multiple matching confidence scores, the first sample abnormal event type is selected as the first abnormal state slice output.

[0094] Furthermore, the system is also used to implement the following functions:

[0095] A preset continuous time window is used to traverse the multiple abnormal state slices to perform statistics on the dominant abnormal event type and output the traffic event type; the first vehicle fingerprint identifier of the first driving vehicle is called as the main body of the traffic anomaly; based on the traffic event type and the distribution of behavioral parameters of the multiple abnormal state slices, peak slices are retrieved from the multiple abnormal state slices and aggregated to generate the traffic anomaly behavior; the GIS coordinates of the risk field potential energy peak point of the first driving vehicle are called as the coordinates of the abnormal event.

[0096] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0097] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0098] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A method for detecting traffic anomaly events from multi-source perception, characterized in that, The method comprises: Synchronous data acquisition is performed on a multi-modal sensor array deployed at a target intersection through an NTP protocol to obtain a multi-modal spatio-temporal data stream; Traffic anomaly identification is performed based on the multi-modal spatio-temporal data stream to output a traffic anomaly event, wherein the traffic anomaly event comprises a traffic event type, a traffic anomaly subject, a traffic anomaly behavior and an anomaly event coordinate; Traffic behavior backtracking is performed on the traffic anomaly subject to obtain a historical trajectory stream; Event evolution trend prediction is performed based on the historical trajectory stream, the traffic event type and the traffic anomaly behavior to output a spatio-temporal influence vector; A traffic influence range is framed according to the spatio-temporal influence vector and the anomaly event coordinate; A real-time traffic control intervention strategy is generated with the traffic influence range as a dynamic constraint; Real-time optimization of signal control and regional road network collaborative scheduling are performed by using the real-time traffic control intervention strategy. 2.The multi-source perception traffic anomaly event detection method of claim 1, wherein, Synchronous data acquisition is performed on a multi-modal sensor array deployed at a target intersection through an NTP protocol to obtain a multi-modal spatio-temporal data stream, and the method comprises: Multi-modal sensor array coverage deployment is performed at the target intersection based on a sensor effective monitoring range, wherein the multi-modal sensor array comprises a video monitoring array, a millimeter wave radar array, a geomagnetic detection array and a laser radar array; Synchronous data acquisition is performed on the target intersection by the multi-modal sensor array through an NTP protocol to obtain an array multi-view video stream, an array target trajectory point cloud, an array lane magnetic field disturbance sequence and an array high-precision space sampling stream; According to the field angle overlapping topological relationship of the video monitoring array, the millimeter wave radar array, the geomagnetic detection array and the laser radar array, the array multi-view video stream, the array target trajectory point cloud, the array lane magnetic field disturbance sequence and the array high-precision space sampling stream are spatio-temporally aligned and spliced to obtain a moving target spatio-temporal trajectory stream, a lane-level physical field stream, a dynamic road network structure stream and a holographic traffic scene stream, which constitute the multi-modal spatio-temporal data stream. 3.The multi-source perception traffic anomaly event detection method of claim 2, wherein, Traffic anomaly identification is performed based on the multi-modal spatio-temporal data stream to output a traffic anomaly event, and the method comprises: A plurality of vehicle-level spatio-temporal trajectory sub-streams are segmented from the moving target spatio-temporal trajectory stream through a multi-target tracking algorithm; The lane-level physical field stream, the dynamic road network structure stream and the holographic traffic scene stream are segmented according to the plurality of vehicle-level spatio-temporal trajectory sub-streams to obtain a plurality of vehicle magnetic field fingerprints, a plurality of vehicle passage permission domains and a plurality of vehicle risk fields; After spatio-temporal alignment of the plurality of vehicle-level spatio-temporal trajectory sub-streams, the plurality of vehicle magnetic field fingerprints, the plurality of vehicle passage permission domains and the plurality of vehicle risk fields, single-vehicle behavior stream anomaly detection is performed to locate the traffic anomaly event of a target abnormal vehicle. 4.The multi-source perception traffic anomaly event detection method of claim 3, wherein, After spatio-temporal alignment of the plurality of vehicle-level spatio-temporal trajectory sub-streams, the plurality of vehicle magnetic field fingerprints, the plurality of vehicle passage permission domains and the plurality of vehicle risk fields, single-vehicle behavior stream anomaly detection is performed to locate the traffic anomaly event of a target abnormal vehicle, and the method comprises: A first vehicle body coordinate system is established with the center of the front bumper of a first driving vehicle as a spatial origin and the moving direction as the positive direction of the X axis. Adopting the first vehicle body coordinate system to unify the space-time reference of the first vehicle level space-time trajectory sub-flow, the first vehicle magnetic field fingerprint, the first vehicle passing permission domain and the first vehicle risk field; Adopting a preset time window sliding segmentation to segment the first vehicle level space-time trajectory sub-flow, the first vehicle magnetic field fingerprint, the first vehicle passing permission domain and the first vehicle risk field, and constructing multiple state vector slices; Based on the multiple state vector slices, performing anomaly detection to obtain multiple anomaly state slices; If the multiple anomaly state slices are a non-empty set, performing anomaly event verification on the multiple anomaly state slices, and outputting the traffic anomaly event. 5.The multi-source perception traffic anomaly event detection method of claim 4, wherein, The state vector slice includes a motion state vector, a physical field vector, a right vector and a risk vector.

6. The multi-source perception traffic anomaly event detection method of claim 5, wherein, Based on the multiple state vector slices, performing anomaly detection to obtain multiple anomaly state slices, the method comprising: Pre-constructing multiple sample state vectors of multiple sample anomaly event types; Adopting a preset confidence threshold rule to calculate multiple matching confidence of the first state vector slice and the multiple sample state vectors; According to the deviation extreme value characteristics of the multiple matching confidence, filtering a first sample anomaly event type as a first anomaly state slice output.

7. The multi-source perception traffic anomaly event detection method of claim 4, wherein, If the multiple anomaly state slices are a non-empty set, performing anomaly event verification on the multiple anomaly state slices, and outputting the traffic anomaly event, the method comprising: Adopting a preset continuous time window to traverse the multiple anomaly state slices to perform dominant anomaly event type statistics, and outputting the traffic event type; Calling the first vehicle fingerprint identifier of the first driving vehicle as the traffic anomaly subject; According to the traffic event type and the behavior parameter distribution of the multiple anomaly state slices, mapping and calling the peak slice from the multiple anomaly state slices, and aggregating to generate the traffic anomaly behavior; Calling the risk field potential peak point GIS coordinate of the first driving vehicle as the anomaly event coordinate.

8. A multi-source perception based traffic anomaly event detection system, characterized by, The system is used to implement the multi-source perception traffic anomaly event detection method of any one of claims 1-7, and the system comprises: A space-time data flow acquisition module for performing synchronous data acquisition by a multi-modal sensing array deployed at a target intersection through an NTP protocol driving, to obtain a multi-modal space-time data flow; A traffic anomaly event output module for identifying a traffic anomaly based on the multi-modal space-time data flow, and outputting a traffic anomaly event, wherein the traffic anomaly event comprises a traffic event type, a traffic anomaly subject, a traffic anomaly behavior and an anomaly event coordinate; A historical trajectory flow acquisition module for performing traffic behavior backtracking on the traffic anomaly subject to obtain a historical trajectory flow; A space-time influence vector output module for performing event evolution trend prediction based on the historical trajectory flow, the traffic event type and the traffic anomaly behavior, and outputting a space-time influence vector; A traffic influence range framing module for framing a traffic influence range according to the space-time influence vector and the anomaly event coordinate; A traffic control intervention strategy generation module for generating a real-time traffic control intervention strategy with the traffic influence range as a dynamic constraint; The cooperative scheduling module is configured to adopt the real-time traffic control intervention strategy to perform real-time optimization of signal control and cooperative scheduling of regional road networks.

Images