An abnormal condition detection system and method for highway traffic

Abstract

The present invention relates to the field of highway traffic management, and discloses an abnormal condition detection system and method for highway traffic, including extracting a road monitoring area through semantic segmentation in road monitoring video images to obtain a continuous sequence of traffic observation frames; performing cross-frame multi-object tracking on vehicle targets in the frame sequence to construct corresponding vehicle motion state time series vectors; based on the time series vectors, using a trajectory aggregation algorithm to extract traffic flow evolution features and identify abnormal behavior patterns; for the detected abnormal behavior areas, constructing a graph neural network model in combination with the spatio-temporal relationship between vehicles to analyze local mutation nodes of the traffic interaction structure; according to the output result of the graph neural network, determining whether there is a macroscopic traffic abnormal event; if so, generating a traffic abnormal distribution heat map on the map according to the type, duration and coverage of the identified traffic abnormal event. The present invention has the advantages of improving the intelligent level and reliability of the detection system.

Description

Technical Field

[0001] The present invention relates to the field of highway traffic management, and specifically to an abnormal condition detection system and method for highway traffic. Background Art

[0002] Currently, in the field of highway traffic management, the detection of traffic abnormal conditions is widely applied in multiple links such as traffic control, emergency response, and information release. Traditional abnormal detection methods usually judge based on fixed thresholds of traffic parameters. For example, whether there is congestion or other abnormal conditions is judged by whether the vehicle speed, flow rate, or density is lower than a certain set value. This method is simple to implement and has a small amount of calculation, and was widely used in traffic monitoring systems in the early stage. However, with the continuous expansion and complexity of the traffic system, the traffic operation characteristics of different road sections at different time periods show significant differences. The traditional fixed-threshold judgment method is difficult to dynamically adjust according to the specific traffic environment and has poor adaptability. Especially during the morning and evening rush hours or when emergencies occur, traffic parameters fluctuate frequently. If static thresholds are still used, problems such as false alarms or missed alarms are likely to occur, seriously affecting the accuracy and practicality of traffic abnormal detection. Therefore, it is very necessary to design an abnormal condition detection system and method for highway traffic that improves the intelligent level and reliability of the detection system. Summary of the Invention

[0003] (I) Technical Problems to be Solved

[0004] Aiming at the deficiencies of the prior art, the present invention provides an abnormal condition detection system and method for highway traffic, which has the advantage of being able to dynamically adjust detection parameters according to different road sections and times to adapt to traffic flow changes, and solves the problems in the above background art.

[0005] (II) Technical Solutions

[0006] To achieve the above purpose of being able to dynamically adjust detection parameters according to different road sections and times to adapt to traffic flow changes, the present invention provides the following technical solutions: An abnormal condition detection method for highway traffic, including the following steps:

[0007] Through semantic segmentation in the road monitoring video image, extract the road monitoring area to obtain a continuous traffic observation frame sequence;

[0008] Perform cross-frame multi-object tracking on vehicle targets in the frame sequence, and combine the motion trajectory, target size, and occlusion situation to construct a corresponding vehicle motion state time series vector;

[0009] Based on the time series vector, use the trajectory aggregation algorithm to extract traffic flow evolution features and identify abnormal behavior patterns;

[0010] For the detected abnormal behavior area, a graph neural network model is constructed in combination with the spatio-temporal relationship between vehicles to analyze the local mutation nodes of the traffic interaction structure;

[0011] According to the output result of the graph neural network and combined with the change trend of the global structure entropy in the traffic image, it is judged whether there is a macroscopic traffic abnormal event;

[0012] If it exists, according to the type, duration and coverage of the identified traffic abnormal event, a traffic abnormal distribution heat map is generated on the map, and event labels are output for traffic dispatching response.

[0013] Preferably, the process of constructing the corresponding vehicle motion state time series vector is as follows:

[0014] For each frame image after semantic segmentation processing, an object detection algorithm is applied to detect all vehicle objects in the image, the bounding box, confidence level, and class label of each vehicle are extracted to generate a preliminary vehicle detection result, multi-dimensional feature information is extracted for each detected vehicle object, and based on the above object features, a multi-object tracking algorithm is used to match the objects in the current frame and the previous frame, and consistent object IDs are assigned.

[0015] Preferably, for each vehicle object with a continuous ID, based on its state change in consecutive frames, the following motion state time series vector is constructed, and the formula is:

[0016] ;

[0017] In the formula, is the center coordinate of the vehicle in the t-th frame, is the speed estimated based on the position change of adjacent frames, is the acceleration estimated by the speed change rate, is the width and height of the vehicle object, is the occlusion state identifier.

[0018] Preferably, the process of identifying the abnormal behavior pattern is as follows:

[0019] Receive the motion state time series vectors of each vehicle in consecutive frames constructed in the previous stage. The vector set of multiple vehicles constitutes a description of the traffic flow state within a time window. A trajectory aggregation algorithm is used to extract and cluster the pattern of the time series trajectories of multiple vehicles.

[0020] Preferably, the process of analyzing the local mutation nodes of the traffic interaction structure is as follows:

[0021] In the abnormal behavior area detected in the trajectory aggregation analysis, the set of vehicle objects active within a specific time window in this area is extracted, denoted as: , each vehicle contains a time - series vector of its motion state and spatial location information. Taking vehicles as nodes of a graph, edge connections are established between every pair of vehicles with potential interaction relationships to construct a spatio - temporal interaction graph, and a graph neural network model is constructed to embed and analyze the traffic graph.

[0022] Preferably, the process of determining whether there is a macroscopic traffic anomaly event is as follows:

[0023] The entire road monitoring area is divided into several sub - areas, and the structural information of vehicle distribution in each frame of image is extracted. Based on the extracted image spatial structure features, the structural entropy value reflecting the orderliness and balance of the traffic state is calculated.

[0024] Preferably, the process of generating a traffic anomaly distribution heat map on the map is as follows:

[0025] After completing the macroscopic traffic anomaly recognition, the anomaly event information is structured and sorted, the anomaly event is located in the actual map coordinate system, an anomaly heat map is generated on the electronic map or traffic road network map, the identified anomaly event area is assigned a heat value, and different heat areas are represented using a color gradient.

[0026] An abnormal condition detection system for highway traffic includes:

[0027] A road area extraction module that automatically identifies the monitoring area in the road monitoring video through semantic segmentation and extracts a sequence of continuous traffic image frames;

[0028] A target tracking module that performs multi - target tracking on vehicle targets in the video and generates a time - series vector containing the motion trajectory and state;

[0029] An anomaly recognition module that extracts traffic flow evolution features based on vehicle motion time - series data and identifies potential abnormal behaviors;

[0030] A graph neural network analysis module that establishes a graph structure by combining the spatio - temporal relationships of vehicles and identifies local mutation nodes in traffic interactions;

[0031] A scheduling response module that generates a heat map and event labels based on the comprehensive anomaly analysis results to support traffic scheduling and emergency response.

[0032] (III) Beneficial effects

[0033] Compared with the prior art, the present invention provides an abnormal condition detection system and method for highway traffic, having the following beneficial effects:

[0034] By combining semantic segmentation, object tracking, and graph neural network models, the present invention realizes multi-level anomaly recognition from microscopic vehicle behaviors to macroscopic traffic structures, significantly improving the accuracy and stability of traffic anomaly detection. Through multi-object tracking and motion state modeling, it effectively addresses complex scenario problems such as occlusion and perspective changes, enhancing the robustness of the system in the actual road environment. Using the trajectory aggregation and structural entropy analysis method, it can dynamically extract traffic flow evolution features, identify behavioral patterns during abnormal changes, and meet the requirements of real-time response. By combining local behavior analysis and global structural change trends, it can not only identify individual anomalies but also macroscopic traffic anomaly phenomena such as group congestion and emergencies. Generating a heat map of traffic anomaly distribution and event labels can visually present the scope of anomaly impact, providing an efficient and operable basis for traffic management departments to schedule and guide. It can continuously adapt to different roads, time periods, and traffic states, reducing the dependence on artificially set thresholds and enhancing the intelligent level of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic diagram of the method of the present invention;

[0036] Figure 2 It is a schematic diagram of the structure of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0037] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0038] Embodiment 1

[0039] Please refer to Figure 1 As shown, a method for detecting abnormal conditions of highway traffic according to an embodiment of the present invention includes the following steps:

[0040] S1: Through semantic segmentation in the road monitoring video image, extract the road monitoring area to obtain a continuous sequence of traffic observation frames.

[0041] From the video stream collected by road monitoring cameras, video image frames are extracted at a fixed frame rate (such as 5 frames per second) to form an original image sequence, which serves as the basic data for subsequent processing. For each frame of the image, it is input into a pre-trained deep semantic segmentation model (such as DeepLabV3+, SegNet or BiSeNet) to classify pixel-level regions in the image and identify semantic categories such as "lane lines", "motor vehicle lanes", "non-motor vehicle lanes", "pedestrian areas", "background", etc. According to the pixel regions belonging to the "motor vehicle lane" or "lane line" categories in the semantic segmentation results, the effective monitoring regions actually used for traffic detection in this frame of the image are extracted, and the image parts irrelevant to traffic detection (such as sky, buildings, sidewalks, etc.) are removed. A road area mask (Mask) of the same size as the original image is generated, and it is pixel-level fused with the original image to retain the image content within the road area, and the remaining areas are set to black or transparent for subsequent object detection and tracking. Each processed frame of the image is numbered and timestamped, and organized into a continuous sequence of traffic observation frames in chronological order as the input data for traffic object detection and status analysis.

[0042] The road monitoring area is accurately extracted through semantic segmentation, effectively excluding the interference of non-traffic-related areas in the image, and significantly improving the accuracy of subsequent vehicle detection and tracking. The semantic segmentation technology is adopted to adapt to monitoring scenarios under different weather, lighting conditions and complex backgrounds, ensuring that the traffic monitoring area is always accurately identified and improving the stability and robustness of the system in various road environments. After semantic masking of the image, only the effective road area is processed subsequently, reducing the computational burden of irrelevant areas and enhancing the operating efficiency and real-time response ability of the entire system. The generation of the continuous frame sequence provides a complete and continuous time-dimensional input for modeling the motion states of subsequent vehicle targets, helping to accurately capture vehicle trajectory changes and identify abnormal behavior patterns. Introducing the semantic understanding ability of deep learning into traditional traffic monitoring realizes the upgrade from "pixel-level image perception" to "semantic-level traffic cognition", providing a higher-level and more refined perception ability for traffic anomaly detection.

[0043] S2: Perform cross-frame multi-object tracking on vehicle targets in the frame sequence, and combine the motion trajectories, target sizes and occlusion situations to construct corresponding temporal vectors of vehicle motion states.

[0044] For each frame of the image after semantic segmentation processing, apply an object detection algorithm to detect all vehicle objects in the image, extract information such as the bounding box, confidence level, and class label of each vehicle, and generate preliminary vehicle detection results. Extract multi-dimensional feature information for each detected vehicle object, including but not limited to: spatial position (center point coordinates, bounding box size), appearance features (color histogram, texture features), vehicle size and class, detection confidence level. These features are used as the representation of the object in the current frame for subsequent cross-frame association. Based on the above object features, use a multi-object tracking algorithm to match the objects in the current frame and the previous frame, assign consistent object IDs, and achieve continuous tracking of vehicles in the frame sequence. In the case of occlusion, separation, and reappearance, combine Kalman filter prediction and Re-ID re-identification module to improve the coherence and robustness of vehicle tracking. For each vehicle object with a continuous ID, based on its state changes in consecutive frames, construct the following motion state time series vector, the formula is:

[0045] ;

[0046] In the formula, is the center coordinate of the vehicle in the t-th frame, is the speed estimated based on the position change of adjacent frames, is the acceleration estimated by the rate of change of speed, is the width and height of the vehicle object, is the occlusion state flag.

[0047] Structurally store the state vectors of each vehicle object within the entire observation window for subsequent traffic behavior modeling, trajectory aggregation, and anomaly recognition analysis.

[0048] Through cross-frame multi-object tracking technology, it is possible to continuously track the identity consistency of each vehicle in a continuous image sequence, avoid tracking interruptions caused by object loss, occlusion, or re-identification errors between frames, and improve the tracking integrity and stability. By constructing the motion state time series vector of the vehicle in consecutive frames, the dynamic change characteristics such as its speed, acceleration, and trajectory curvature can be reflected in real time, providing a fine time series basis for subsequent identification of abnormal motion behaviors. Introduce the analysis of object size changes and occlusion states to effectively improve the tracking accuracy of the system in complex traffic scenarios and reduce the risk of mis-identification or missed-identification caused by occlusion. Using the time series vectors of multiple vehicles as input can be used to construct an evolutionary trajectory map of the overall traffic flow, realize group behavior modeling such as abnormal pattern recognition, mutation detection, and traffic flow analysis, and improve the perception ability of the overall traffic operation state. The constructed time series vector can be used as the feature input of the nodes in the subsequent graph neural network, which is convenient for mining complex traffic patterns from the interaction relationships between vehicles and promoting the leap of anomaly detection from the single-vehicle level to the group level and the structural level.

[0049] S3: Based on the time series vector, the trajectory aggregation algorithm is used to extract the traffic flow evolution characteristics and identify abnormal behavior patterns.

[0050] Receive the motion state time series vector of each vehicle in the continuous frames constructed in the previous stage. The vector set of multiple vehicles constitutes a description of the traffic flow state within a time window. The trajectory aggregation algorithm is used to extract and group the time series trajectories of multiple vehicles, including:

[0051] Density-based trajectory clustering: identifying subgroups that are densely packed but have different motion patterns;

[0052] Feature-based trajectory vector embedding and clustering: embed complex trajectory features into low-dimensional space and then cluster them;

[0053] Trajectory shape matching: used to compare the consistency of motion paths;

[0054] Autoencoder / Transformer trajectory compression modeling: extracting implicit traffic flow features.

[0055] These aggregated results can reflect the “typical motion patterns” within the local area.

[0056] Based on trajectory aggregation, calculate the evolution characteristic indicators of traffic flow, including: local trajectory density changes; number of vehicles in the same direction / different directions; average speed and acceleration fluctuations; trajectory characteristics such as sudden stops and reverse driving. Input the above evolution characteristics into anomaly detection models, such as rule bases, anomaly clustering detection, and time series anomaly detection models (such as LSTM-AE, AutoEncoder, Isolation Forest), to identify whether there are abnormal behavior patterns, such as sudden stops, vehicles driving in reverse, local abnormal aggregation / congestion, and irregular lane crossings.

[0057] Using trajectory aggregation algorithm to cluster the time series vectors of a large number of vehicles can effectively identify different types of traffic behavior patterns, such as uniform speed, slow speed, vehicle gathering, emergency stop or U-turn, etc., and improve the ability to understand the overall traffic dynamics. By analyzing the trajectory evolution characteristics (such as sudden changes in speed, abnormal movement direction, and sudden increase in trajectory density), abnormal traffic behaviors can be accurately identified, such as emergency stops caused by traffic accidents, vehicle reversal, and sudden congestion, reducing false alarms and missed reports. Aggregation analysis does not rely on the state of a single vehicle, but reflects the overall traffic trend through group trajectory changes, making the system more adaptable to nonlinear fluctuations in traffic characteristics in complex scenarios such as morning and evening rush hours, holidays or bad weather. Compared with the traditional static threshold judgment method, the trajectory aggregation method is based on dynamic behavior changes for analysis, which can perceive and respond to abnormal behaviors as soon as they occur, greatly improving the system's real-time detection capabilities for emergencies. The aggregated abnormal behavior patterns can not only be used to identify traffic anomalies, but also provide structured and high-value input data for subsequent graph neural network modeling, heat map visualization, and dispatch instruction formulation modules.

[0058] S4: For the detected abnormal behavior areas, a graph neural network model is constructed based on the spatiotemporal relationship between vehicles to analyze the local mutation nodes of the traffic interaction structure.

[0059] In the abnormal behavior area detected in the trajectory aggregation analysis, the vehicle target set active in the area within a specific time window is extracted and recorded as: Each vehicle contains a time series vector of motion state and spatial position information. The vehicles are used as nodes of the graph, and edges are established between each pair of vehicles with potential interaction relationships to construct a spatiotemporal interaction graph. , where node features include vehicle speed, acceleration, direction of movement, trajectory change trend, time series features, etc. Edge construction strategies include those based on spatiotemporal proximity: such as when the distance between vehicles is lower than the set threshold and the time windows overlap; based on relative speed: such as when there are behaviors such as approaching, following, and avoiding between two vehicles; and edge weights are introduced to reflect the intensity of interaction or potential conflict risks. Build graph neural network models, such as GCN, GAT, or ST-GCN, to embed and analyze traffic maps:

[0060] Node update mechanism: integrates the vehicle’s own characteristics with the information of neighboring vehicles to learn its dynamic role in the group traffic structure;

[0061] Edge update mechanism: reflects the evolution of interaction strength over time;

[0062] Temporal modeling: Incorporate the time dimension into models such as ST-GCN to characterize the evolution of nodes over time.

[0063] Perform structural variation analysis on the intermediate representation output by the graph neural network to identify local mutation nodes, such as: vehicle nodes with mutated motion states, such as a vehicle whose speed suddenly drops to zero; key nodes with abnormal interactions with multiple vehicles, such as a "bottleneck vehicle" that may cause chain congestion; local density mutation regions of the graph structure, such as the congestion core; use an anomaly scoring mechanism to label abnormal nodes or edges. Finally, output the following information for the dispatching system to respond or assist in judgment: the ID of the abnormal interaction node and its position in the graph; the type of anomaly, such as local stagnation, reverse conflict, potential collision; the scope of anomaly influence or propagation trend;

[0064] The anomaly level evaluation value.

[0065] By introducing a graph neural network (GNN) to model the spatio-temporal interaction relationships between vehicles, it is no longer limited to single-vehicle features or global statistical metrics, but can deeply explore potential structural anomalies among vehicle groups, enhancing the ability to identify complex anomaly patterns, such as chain congestion and intersection jams. GNN can effectively capture the mutual influence between vehicles, such as following behavior, lane-changing avoidance, and intersection conflicts, enabling reasoning from individual anomalies to group anomalies and endowing the system with the ability to explain and discover "synergistic abnormal behaviors". Through graph structure analysis, mutation nodes in the traffic interaction network can be quickly located, potential congestion sources or conflict focal points can be identified, avoiding global interference caused by the spread of local anomalies, and effectively enhancing the timeliness and foresight of traffic dispatching. Compared with analysis methods based on rules or traditional models, graph neural networks have stronger modeling flexibility and generalization ability, can adapt to different types of roads, such as multi-lane roads, highways, intersections, and behavior pattern changes under different traffic densities, and enhance the environmental adaptability of the model. The graph structure embeddings output by GNN can be used to support downstream tasks, such as traffic state prediction, path optimization, signal dispatching, etc., providing richer and more context-semantic input data.

[0066] S5: According to the output results of the graph neural network and combined with the change trend of the global structural entropy in the traffic image, judge whether there is a macroscopic traffic anomaly event.

[0067] The process of judging macroscopic traffic anomaly events by combining the change trend of the global structural entropy in the traffic image is as follows:

[0068] For each frame of road monitoring image, based on the vehicle detection and tracking results, extract the position distribution, speed direction, and density information of the vehicles in the current frame;

[0069] Construct a traffic state graph with vehicles as nodes and the relative distance and speed difference between vehicles as edge weights;

[0070] Calculate the information entropy of the traffic state graph to obtain the global structural entropy value corresponding to this frame;

[0071] The structural entropy is used to measure the orderliness of the vehicle flow state. The more uniform the vehicle distribution and the more consistent the movement direction, the higher the structural entropy; conversely, if the vehicles are dense, the movement is chaotic or stagnant, the structural entropy decreases;

[0072] The global structural entropy values extracted from several consecutive frames form a time series;

[0073] Perform a rate-of-change analysis on this entropy value sequence, and calculate the first-order rate of change and the second-order rate of change of the entropy value (i.e., the speed and acceleration of the change);

[0074] If the entropy value drops sharply at multiple consecutive time points and the entropy rate of change exceeds the normal fluctuation range, it indicates that there may be an abnormal traffic flow;

[0075] The process of setting and determining the abnormal judgment threshold is as follows:

[0076] Based on the historical entropy change data of the target section at different time periods, establish a benchmark curve of entropy change under normal traffic conditions;

[0077] Set the entropy drop amplitude threshold and the entropy change rate threshold as the abnormal judgment criteria;

[0078] If the current entropy change meets any of the following conditions, it is judged that there is a macroscopic traffic abnormal event:

[0079] The current entropy value is lower than a certain proportion of the historical average entropy value (such as lower than 85%);

[0080] The entropy value drop rate exceeds the maximum rate of normal fluctuation;

[0081] The entropy change trend continues to be abnormal for more than the set time threshold (for example, it drops continuously for 5 minutes);

[0082] The process of comprehensive judgment combined with the results of the graph neural network is as follows:

[0083] The abnormal nodes in the local traffic area are identified by the graph neural network and fused with the global structural entropy change results;

[0084] When the number of local graph mutation nodes increases significantly and the global entropy change trend shows abnormalities synchronously, it is finally determined that there is a macroscopic traffic abnormality.

[0085] When judging whether there is a macroscopic traffic abnormal event, if there is no macroscopic traffic abnormal event, it meets one or more of the following situations:

[0086] The global structural entropy change is within the normal fluctuation range: the global structural entropy values of the extracted consecutive frame traffic images change smoothly, and the rate of change does not exceed the normal fluctuation threshold set based on historical data; even if the entropy value drops or rises, it is still within the acceptable range of normal traffic flow, without abnormal violent fluctuations.

[0087] There are no mutation nodes in the local traffic interaction structure: In the traffic interaction graph obtained by graph neural network analysis, the spatio-temporal relationship changes between nodes conform to the normal evolution pattern, and no significant local aggregation, breakage, or abnormal path reconstruction phenomena are detected; parameters such as the relative movement and spacing changes between vehicles remain continuous and stable, without large-scale mutations.

[0088] Local anomalies do not reach the macroscopic level: Even if individual vehicle anomalies (such as sudden stops, U-turns, etc.) are detected, their influence scope is limited to a very small local area and does not cause significant disturbances to the overall traffic flow structure; the change in global structure entropy is minimally affected by local disturbances, and the overall traffic system still maintains a stable operating state.

[0089] The time of abnormal change is not sufficient to constitute a macroscopic event: The duration of the entropy value decrease or local structure mutation is short (for example, less than the set minimum abnormal duration threshold, such as within 2 minutes), which belongs to instantaneous disturbance or self-recovery phenomena; no persistent congestion, traffic interruption, or long-term abnormal traffic flow phenomena are formed.

[0090] The entropy change trend and local mutations do not occur synchronously: If the global entropy change and local node mutations do not occur within the same time window, that is, the graph neural network outputs an anomaly but the entropy is stable, or the entropy is abnormal but the graph does not show mutations, it is judged that there is no macroscopic traffic anomaly event.

[0091] In the previous graph neural network model analysis, the abnormal node information and graph structure embedding features within the local area have been output, including the number, distribution, and abnormal scores of abnormal nodes; the characteristics of structural mutations in sub-regions of the graph; and the temporal change trends, such as the continuous growth of abnormal nodes. Based on the traffic observation frame sequence, the entire road monitoring area is divided into several sub-regions, and the structural information of vehicle distribution in each frame of image is extracted, such as: vehicle density and arrangement uniformity within each sub-region; connectivity and breakage of traffic flow between road sections; local graph clustering degree and vehicle aggregation degree. Based on the extracted image spatial structure features, the structure entropy values reflecting the orderliness and balance of traffic states are calculated, including:

[0092] Shannon entropy: Measures the distribution uncertainty of vehicles between different sub-regions;

[0093] Clustering entropy: Reflects whether vehicles are concentrated in a certain local area;

[0094] Entropy change rate: Reflects the speed and degree of structural mutation, and the formula is:

[0095] ;

[0096] In the formula, is the probability that a vehicle appears in the i-th sub-region.

[0097] By continuously calculating the structural entropy value for multiple time slices to form an entropy change sequence and analyzing the change trend: a sharp drop in the structural entropy indicates a sudden increase in the degree of vehicle aggregation, which may be a severe congestion or an emergency; a violent fluctuation in the structural entropy may be due to random vehicle driving and disordered distribution, reflecting an unstable traffic state; a high local entropy and a low overall entropy indicate the existence of an abnormal traffic bottleneck in a specific section. The local structural mutation information identified by the graph neural network is fused and analyzed with the structural entropy change trend. If a sudden node aggregation appears in a certain area and is accompanied by a sharp drop in the global entropy, it is determined as a macroscopic congestion event; if the abnormal amplitude of the entropy change exceeds the set threshold and is combined with the structural abnormal pattern of the graph neural network, it is determined as a macroscopic traffic anomaly. The determination result is converted into structured abnormal information, including the type of abnormal event, such as large-scale congestion, distribution disorder, traffic break; the covered area and the central section; the start and end times, the duration; the severity score, such as based on the entropy drop rate and the graph anomaly density; and output to the traffic dispatching system or the visualization interface for subsequent response mechanisms to use.

[0098] By combining the high-dimensional expression of local structural anomalies by the graph neural network with the global structural entropy change trend, an effective transition from microscopic abnormal features to macroscopic traffic states is achieved, enabling more accurate identification of large-scale traffic abnormal events such as congestion and blockage. As an indicator to measure the order degree of the traffic system, the structural entropy can reflect the real-time change in the spatial distribution of vehicles in the road network. Combining with the node anomaly analysis of the graph model helps to quickly discover early signals such as systematic disturbances and local anomaly diffusion, improving the timeliness of response. Compared with traditional methods based on fixed rules or single-parameter judgment, the dynamic analysis combined with the structural entropy can adapt to traffic scenarios with different densities and layouts and still maintain stable detection performance in complex structures such as intersections, main and auxiliary roads, and elevated and ground transitions. The change trend of the entropy value can reflect the evolution process of the traffic state from order to disorder or from disorder to stability. Combining with the dynamic characteristics of the graph structure helps to analyze the starting position, diffusion path, and development trend of traffic anomalies, providing a decision-making basis for emergency dispatching and intervention. This process realizes the collaborative modeling of the graph model (structural layer) and the image content (pixel layer), promotes the transformation of image semantic understanding to structured knowledge expression, and lays a foundation for building a higher-level traffic intelligent perception system.

[0099] S6: If it exists, generate a traffic anomaly distribution heat map on the map according to the type, duration, and coverage of the identified traffic abnormal event, and output event labels for traffic dispatching response.

[0100] After completing the identification of macroscopic traffic anomalies, the anomaly event information is structured and organized, including the following fields: anomaly type, such as severe congestion, traffic interruption, suspected accident, traffic disorder, etc.; start time and end time; the area covered by the anomaly event, including the starting and ending road segments, geographical coordinate boundaries, polygon areas or grid cells; anomaly intensity level, based on comprehensive index scores such as graph anomaly density and entropy change amplitude; traffic impact indicators, such as average speed reduction, vehicle detention time, and degree of reduction in regional traffic capacity. The anomaly event is located in the actual map coordinate system. The process includes mapping the pixel-level position information output by the surveillance video or graph neural network into longitude and latitude coordinates; performing spatial projection on the anomaly impact area, such as converting it into the form of GIS grid or polygon; if multi-source data is used, coordinate alignment processing is performed on the spatial benchmarks of different sources. Based on the above spatial information, an anomaly heat map is generated on the electronic map or traffic road network map, and the identified anomaly event area is assigned a heat value, and color gradients are used to represent different heat areas, such as red for severe anomalies and yellow for moderate anomalies. The system automatically generates event tags according to the characteristics of the event and outputs them to the traffic dispatching system. The tags include but are not limited to:

[0101] Tag name: such as severe congestion on the main road, suspected accident of multiple vehicle rear-end collisions, signal control anomaly at intersections, etc.;

[0102] Location: Display location marking points and coverage boundaries through the map interface;

[0103] Level: Divided into mild, moderate, and severe;

[0104] Emergency suggestions: such as suggesting traffic flow restriction and detour, signal priority control, releasing guidance information, etc.;

[0105] Unique event ID: Used for subsequent tracking, correlation analysis, or response records.

[0106] Push the above tags and heat map information to the traffic management platform; trigger response mechanisms, such as releasing information on the induction screen, switching signal control strategies, dispatching police resources, pushing early warning information to the user side, etc.; and support comparison with historical events to achieve event archiving and response efficiency evaluation.

[0107] By generating a heat map of traffic anomaly distribution on the map, traffic managers can intuitively grasp the distribution, severity and evolution trend of abnormal events in the geographical space, improving the perception efficiency and understanding depth of complex traffic situations. Combining the type, duration and coverage of abnormal events to output label information can quickly trigger matching dispatching strategies or emergency mechanisms, such as induced diversion, signal timing adjustment, on-site disposal plans, etc., to achieve rapid intervention and control of traffic anomalies. Different types of traffic anomalies are classified and layer-managed, which helps to accurately analyze events, prioritize them and schedule resources, improving the systematic and collaborative efficiency of anomaly management. The label information and spatial distribution results of abnormal events can be stored for a long time for subsequent event comparison, pattern recognition and traffic policy evaluation, providing data support for intelligent traffic planning and big data decision-making. Pushing the generated heat map and label results to the public information service platform can realize more targeted travel suggestions and route recommendations, optimizing the overall traffic operation efficiency and user experience.

[0108] Embodiment 2

[0109] Please refer to Figure 2 As shown, an abnormal condition detection system for highway traffic according to an embodiment of the present invention includes:

[0110] An abnormal condition detection system for highway traffic includes:

[0111] A road area extraction module automatically identifies the monitoring area in the road monitoring video through semantic segmentation and extracts a continuous sequence of traffic image frames;

[0112] A target tracking module performs multi-target tracking on vehicle targets in the video and generates a time series vector containing motion trajectories and states;

[0113] An anomaly recognition module extracts traffic flow evolution features based on vehicle motion time series data and identifies potential abnormal behaviors;

[0114] A graph neural network analysis module establishes a graph structure by combining the spatio-temporal relationships of vehicles and identifies local mutation nodes in traffic interactions;

[0115] A dispatching response module generates a heat map and event labels based on the comprehensive anomaly analysis results to support traffic dispatching and emergency response.

[0116] It should be noted that, in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0117] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting abnormal conditions in road traffic, characterized in that, It includes the following steps: Through semantic segmentation of road monitoring video images, extract the road monitoring area to obtain a continuous sequence of traffic observation frames; Perform cross-frame multi-object tracking on vehicle targets in the frame sequence, and combine the motion trajectory, target size, and occlusion situation to construct the corresponding time-series vector of vehicle motion states; Based on the time-series vector, use the trajectory aggregation algorithm to extract traffic flow evolution features and identify abnormal behavior patterns; For the detected abnormal behavior area, construct a graph neural network model in combination with the spatio-temporal relationship between vehicles, and analyze the local mutation nodes of the traffic interaction structure; According to the output result of the graph neural network and combined with the change trend of the global structure entropy in the traffic image, determine whether there is a macroscopic traffic abnormal event; Among them, the method for determining whether there is a macroscopic traffic abnormal event is: Construct a traffic state graph and calculate the corresponding global structure entropy; if it is detected that the entropy values drop sharply and the change rate is abnormal at multiple consecutive time points, it is determined that a macroscopic traffic abnormal event has occurred; If it exists, according to the type, duration, and coverage of the identified traffic abnormal event, generate a traffic abnormal distribution heat map on the map and output event labels for traffic dispatching response.

2. The abnormal condition detection method for highway traffic according to claim 1, characterized in that, The process of constructing the corresponding time-series vector of vehicle motion states is: For each frame image after semantic segmentation, apply an object detection algorithm to detect all vehicle targets in the image, extract the bounding box, confidence level, and class label of each vehicle to generate a preliminary vehicle detection result, extract multi-dimensional feature information for each detected vehicle target, and based on the above target features, use a multi-object tracking algorithm to match the targets in the current frame and the previous frame, and assign consistent target IDs.

3. The abnormal condition detection method for road traffic according to claim 2, wherein For each vehicle target with a continuous ID, based on its state change in consecutive frames, construct the following motion state time-series vector, and the formula is: ； In the formula, is the center coordinate of the vehicle at the t-th frame, is the speed estimated based on the position change of adjacent frames, is the acceleration estimated by the speed change rate, is the width and height of the vehicle target, is the occlusion status flag.

4. The abnormal condition detection method for road traffic according to claim 3, wherein The process of identifying abnormal behavior patterns is: Receive the time-series vector of the motion state of each vehicle in consecutive frames constructed in the previous stage. The vector set of multiple vehicles constitutes a description of the traffic flow state within a time window. Use the trajectory aggregation algorithm to perform pattern extraction and clustering on the time-series trajectories of multiple vehicles.

5. The abnormal condition detection method for highway traffic according to claim 4, wherein, The process of analyzing the local mutation nodes of the traffic interaction structure is: In the abnormal behavior area detected in the trajectory aggregation analysis, extract the set of vehicle targets that are active within a specific time window in this area, denoted as: , each vehicle contains a time series vector of motion state and spatial position information. Take the vehicles as the nodes of the graph, establish edge connections between every pair of vehicles with potential interaction relationships, construct a spatio-temporal interaction graph, construct a graph neural network model, and perform embedding and analysis on the traffic graph.

6. The abnormal condition detection method for highway traffic according to claim 5, characterized in that, The process of determining whether there is a macroscopic traffic abnormal event is: Divide the entire road monitoring area into several sub-regions, and extract the structural information of the vehicle distribution in each frame of image. Based on the extracted image spatial structure features, calculate the structure entropy value reflecting the order and balance of the traffic state.

7. The abnormal condition detection method for highway traffic according to claim 6, wherein, The process of generating a traffic abnormal distribution heat map on the map is: After completing the identification of macroscopic traffic anomalies, organize the abnormal event information in a structured manner, locate the abnormal event in the actual map coordinate system, generate an abnormal heat map on the electronic map or traffic road network map, assign heat values to the identified abnormal event areas, and use color gradients to represent different heat areas.

8. An abnormal condition detection system for road traffic, which is applied to an abnormal condition detection method for road traffic according to any one of claims 1-7, characterized in that, It includes: A road area extraction module that automatically identifies the monitoring area in the road monitoring video through semantic segmentation and extracts a continuous sequence of traffic image frames; An object tracking module that performs multi-object tracking on vehicle targets in the video and generates a time-series vector containing motion trajectories and states; Anomaly recognition module, which extracts traffic flow evolution features based on vehicle movement time-series data and identifies potential abnormal behaviors; Graph neural network analysis module, which constructs a graph structure by combining the spatio-temporal relationships of vehicles and identifies local mutation nodes in traffic interactions; Scheduling response module, which generates a heat map and event labels based on the comprehensive anomaly analysis results to support traffic scheduling and emergency response.

Images