Traffic scene abnormal event rapid detection method and system

Abstract

The invention provides a traffic scene abnormal event rapid detection method and system, and relates to the technical field of abnormal event detection, and the method comprises the steps: carrying out the key region recognition of a target traffic scene region, obtaining N key traffic regions, and collecting N region multi-source traffic operation data streams through a multi-sensor monitoring module; collecting a traffic anomaly event data set, and building a lightweight traffic anomaly recognition model; carrying out anomaly identification on the multi-source traffic operation data streams of the N regions to obtain traffic anomaly feature sets of the N regions; and uploading the abnormal feature set to a cloud processing center to carry out abnormal event detection, determining a target traffic scene abnormal event, and carrying out abnormal response decision through the target traffic scene abnormal event. According to the invention, the technical problems of low detection precision and slow response speed of abnormal event detection in the prior art can be solved, and the technical effect of effectively ensuring smooth traffic operation and travel safety is achieved.

Description

Technical Field

[0001] This application relates to the field of abnormal event detection technology, and in particular to a rapid detection method and system for abnormal events in traffic scenarios. Background Technology

[0002] With the acceleration of urbanization and the continuous growth of traffic flow, traffic anomalies such as traffic accidents, road congestion, and pedestrian violations are frequent, seriously affecting traffic efficiency and posing a significant threat to travel safety. Existing traffic anomaly detection technologies lack scientific identification of key areas in traffic scenarios, leading to excessive consumption of monitoring resources in non-core areas. Most solutions rely on data from single sensors, making them susceptible to adverse environmental conditions such as heavy rain, fog, and insufficient nighttime lighting, resulting in large data deviations and low accuracy in anomaly identification. This often leads to missed or false detections, further delaying the handling of anomalies and exacerbating traffic congestion or safety risks.

[0003] In summary, existing technologies for anomaly detection suffer from low detection accuracy and slow response speed. Summary of the Invention

[0004] The purpose of this application is to provide a rapid detection method and system for abnormal events in traffic scenarios, in order to solve the technical problems of low detection accuracy and slow response speed in the existing abnormal event detection technology.

[0005] In view of the above problems, this application provides a method and system for rapid detection of abnormal events in traffic scenarios.

[0006] Firstly, this application provides a method for rapid detection of abnormal events in traffic scenes. This method is implemented through a rapid detection system for abnormal events in traffic scenes, and includes the following components:

[0007] The system identifies N key traffic areas within a target traffic scenario and deploys multi-sensor monitoring modules in these areas to collect multi-source traffic operation data streams. It also collects a traffic anomaly event dataset and builds a lightweight traffic anomaly identification model based on the multi-sensor monitoring modules and the dataset. The model is then used to identify anomalies in the multi-source traffic operation data streams of the N areas, resulting in N sets of traffic anomaly features. These features are then uploaded to a cloud processing center for anomaly event detection, identifying anomalies in the target traffic scenario and making anomaly response decisions based on these events.

[0008] Secondly, this application also provides a rapid detection system for abnormal events in traffic scenes, used to execute the rapid detection method for abnormal events in traffic scenes as described in the first aspect, wherein the rapid detection system for abnormal events in traffic scenes includes:

[0009] The system comprises the following modules: a region identification module, which identifies key regions of the target traffic scenario to obtain N key traffic regions, and deploys multi-sensor monitoring modules on these N key traffic regions to collect multi-source traffic operation data streams; a data acquisition module, which collects a traffic anomaly event dataset and builds a lightweight traffic anomaly identification model based on the multi-sensor monitoring modules and the traffic anomaly event dataset; an anomaly identification module, which identifies anomalies in the multi-source traffic operation data streams of the N regions based on the lightweight traffic anomaly identification model to obtain N region traffic anomaly feature sets; and an event detection module, which uploads the N region traffic anomaly feature sets to a cloud processing center for anomaly event detection, identifies anomalies in the target traffic scenario, and makes anomaly response decisions based on the anomalies in the target traffic scenario.

[0010] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0011] By accurately identifying key traffic areas based on factors such as traffic flow characteristics and accident risks, and optimizing sensor deployment, the monitoring accuracy of core areas is improved while reducing resource consumption. Furthermore, a lightweight anomaly identification model is built based on multi-source data collected from multiple sensors, avoiding the bias of data from a single sensor and enhancing real-time processing capabilities. Finally, a cloud processing center maps multi-regional anomaly features onto a traffic map for visualization and global correlation analysis, reducing missed and false positives. Ultimately, this improves the efficiency and accuracy of anomaly event detection in traffic scenarios, shortens anomaly identification and response time, and effectively ensures smooth traffic flow and travel safety.

[0012] The above description is merely an overview of the technical solution of this application. To enable a clearer understanding of the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating the rapid detection method for abnormal events in traffic scenarios proposed in this application.

[0015] Figure 2 This is a schematic diagram of the structure of the rapid detection system for abnormal events in traffic scenarios in this application.

[0016] Explanation of reference numerals in the attached diagram: Area identification module 11, data acquisition module 12, anomaly identification module 13, event detection module 14. Detailed Implementation

[0017] This application provides a rapid detection method and system for abnormal events in traffic scenarios, which solves the technical problems of low detection accuracy and slow response speed in the existing abnormal event detection technology.

[0018] The technical solutions of this application will now be clearly and completely described 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. It should be understood that this application is not limited to the exemplary embodiments described herein. 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. It should also be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not all of them.

[0019] Example 1, please refer to the appendix. Figure 1 This application provides a method for rapid detection of abnormal events in traffic scenes, wherein the method is applied to a rapid detection system for abnormal events in traffic scenes, and the method specifically includes the following steps:

[0020] S100: Identify key areas in the target traffic scenario area to obtain N key traffic areas, and deploy multi-sensor monitoring modules in the N key traffic areas to collect multi-source traffic operation data streams from the N areas through the multi-sensor monitoring modules.

[0021] Specifically, the identification of key areas in the target traffic scenario involves first acquiring a set of key factors for traffic areas, including flow characteristics, accident risk, regional functions, and topological impacts. Then, based on this factor set, a systematic criticality assessment is conducted on each area of ​​the target region, calculating the traffic scenario area criticality coefficient for each area. Finally, based on the coefficient levels, N key traffic areas that significantly impact the overall traffic operation are selected. This key area identification refers to accurately locating areas requiring focused monitoring by quantifying the core influencing factors of traffic areas. Subsequently, for the identified N key traffic areas, video sensors, millimeter-wave radar, and traffic flow detectors are deployed. A multi-sensor monitoring module, composed of various devices such as environmental sensors, can complement each other's functions. Deploying a multi-sensor monitoring module means building a collaborative monitoring unit with multiple devices in key areas. Through the collaborative work between devices, traffic status information that is difficult for a single sensor to capture comprehensively is covered. Finally, using the deployed multi-sensor monitoring module, multi-source traffic operation data streams are collected in real time in each key area. These data streams cover multi-dimensional information such as traffic flow data, speed data, event data, and environmental data. Multi-source traffic operation data streams refer to a comprehensive set of data collected by different types of sensors that can fully reflect the real-time operation status of traffic scenarios.

[0022] S200: Collect a traffic anomaly event dataset, and build a lightweight traffic anomaly recognition model based on the multi-sensor monitoring module and the traffic anomaly event dataset.

[0023] Specifically, the first step is to collect a traffic anomaly event dataset, which includes event types such as traffic accidents, traffic congestion, and traffic violations. The traffic anomaly event dataset integrates the characteristic information of various traffic anomalies with the corresponding multimodal monitoring data. After completing the dataset collection, a lightweight traffic anomaly recognition model was built based on the multi-sensor monitoring module and the dataset. According to the equipment configuration of the multi-sensor monitoring module, the multimodal traffic data types that the model needs to process were identified. Correlational anomaly features were extracted from the traffic anomaly event dataset, and key features corresponding to various anomaly events were selected. These features were then categorized and integrated according to the multimodal data types to form a multimodal traffic anomaly feature set. Deep neural networks were used to perform supervised training on the multimodal traffic anomaly feature set to obtain a multimodal traffic anomaly recognition model set capable of processing single-modal data. The reliability of the model set was evaluated to determine the reliability coefficient set of the multimodal recognition model. Weighted decision fusion was performed on each model according to the coefficients to obtain the basic traffic anomaly recognition model. Performance verification and parameter optimization were carried out on the basic model, ultimately constructing a lightweight traffic anomaly recognition model that balances high recognition accuracy and fast operating efficiency, adapts to the real-time detection needs of traffic scenarios, and has low hardware resource consumption.

[0024] S300: Based on the lightweight traffic anomaly identification model, perform anomaly identification on the multi-source traffic operation data streams of the N regions to obtain the traffic anomaly feature sets of the N regions.

[0025] Specifically, the first step is to preprocess the multi-source traffic operation data streams collected in N key traffic areas through multi-sensor monitoring modules. This includes eliminating data noise and invalid data generated by temporary sensor failures, unifying data formats, and adjusting the timestamps and data units output by different sensors to a consistent standard and time dimension to ensure the correspondence of different modal data at the same time point. This ensures the validity and consistency of multi-source data, including video image data, radar speed data, traffic statistics data, environmental parameter data, etc. Subsequently, the lightweight traffic anomaly identification model that has been built is invoked. Because this model has been trained and optimized with multimodal data and weighted decision fusion, it can maintain high identification accuracy and has efficient data processing capabilities. The model first distributes the preprocessed multi-source traffic operation data stream to the corresponding identification sub-modules according to modal type, namely, video image data input image feature identification sub-module, vehicle speed and position data captured by radar input time sequence feature identification sub-module, and traffic volume data collected by traffic flow detector input statistical feature identification sub-module. Each sub-module performs anomaly identification based on the normal traffic pattern feature library and anomaly feature template learned in the training phase, through deviation analysis of real-time data features and normal features, and calculation of anomaly feature matching degree, etc., that is, it uses model algorithms to accurately determine whether there is information in real-time traffic data that deviates from the normal operation state and locate potential traffic anomalies. After completing the independent anomaly identification of each modality, the model performs cross-modal fusion verification on the anomaly information output by each sub-module, eliminates misidentified results, and supplements feature details not captured by a single modality. Then, for each key area, it extracts regional anomaly features containing information such as anomaly type, precise time and geographical location of anomaly occurrence, and parameters of anomaly impact range. These features of N key areas are sorted and integrated to form a traffic anomaly feature set, which is a set that provides a structured and standardized description of traffic anomalies in each key traffic area. This set presents the specific attributes and quantitative characteristics of anomalies in each area.

[0026] S400: Upload the N regional traffic anomaly feature sets to the cloud processing center for anomaly event detection, determine the target traffic scene anomaly events, and make anomaly response decisions based on the target traffic scene anomaly events.

[0027] Specifically, the N regional traffic anomaly feature sets acquired in the early stages are uploaded to the cloud processing center via an encrypted transmission protocol. The cloud processing center is a platform with massive data storage, traffic scene map retrieval, multi-dimensional data visualization, and global analysis capabilities, capable of integrating anomaly features from scattered key areas. After uploading, the cloud processing center first retrieves a high-precision traffic scene area map of the target traffic scene region. Based on the geographical location of the key areas corresponding to each anomaly feature set, it maps them onto the map and visualizes them using differentiated colors and symbols, forming a clear distribution area of ​​traffic anomaly features. Next, based on this distribution area, anomaly event detection is performed. Anomaly event detection uses a cloud-based global analysis algorithm to assess the correlation between anomaly features in each key area, extracts and classifies the global anomaly feature set of the traffic scene, and finally identifies the target traffic scene anomaly events. These target traffic scene anomaly events are specific traffic anomalies that have been globally verified and clearly classified, including key information such as anomaly type, vehicle collision, road congestion, pedestrian jaywalking, precise location, affected area, number of affected key areas, degree of interference with traffic flow, and time of occurrence. Once the target abnormal event is identified, the cloud processing center combines the preset response rule base with the real-time traffic situation to generate an abnormal response decision. The abnormal response decision is a targeted response measure formulated according to the severity of the abnormality. For minor accidents, it triggers the adjustment of the timing of surrounding traffic lights and pushes navigation detour prompts. For large-scale congestion, it initiates regional traffic control and notifies law enforcement personnel to guide traffic on-site. For dangerous violations, it preserves evidence and pushes it to the law enforcement system.

[0028] Furthermore, N key transportation areas were identified, including:

[0029] Obtain a set of key factors for a traffic area, including traffic flow characteristics, accident risk, regional function, and topological impact; conduct a criticality assessment of the target traffic scenario area according to the set of key factors to obtain a set of key coefficients for the traffic scenario area; identify key areas of the target traffic scenario area based on the set of key coefficients to obtain the N key traffic areas.

[0030] Specifically, the first step is to obtain a set of key factors for traffic areas, which serves as a systematic indicator set to measure the traffic importance of each area within the target traffic scenario. This set includes four key dimensions: traffic flow characteristics, accident risk, regional function, and topological impact. Traffic flow characteristics refer to the area's traffic flow carrying capacity and actual traffic volume per unit time, the difference in traffic flow between off-peak and peak hours, and vehicle density per unit area, reflecting the area's traffic load. Accident risk refers to the probability and impact of historical traffic events in the area, typically requiring statistics on the frequency, type, duration of traffic disruptions, and casualties of traffic accidents in the area over the past 1-3 years, directly related to the safety of the traffic scenario. Regional function refers to the area's service attributes within the overall transportation network, whether it is a city transportation hub (e.g., near subway stations, long-distance bus stations, densely populated functional areas, or areas residents must pass through daily commutes), determining the regularity and intensity of regional traffic demand. Topological impact refers to the transmission effect of the area's location in the road network structure on surrounding traffic flow, whether it is a road network node, bottleneck section, or key passage connecting multiple core areas; changes in the traffic status of such areas can rapidly affect the traffic efficiency of a large surrounding road network.

[0031] After obtaining the set of key factors for traffic areas, a criticality assessment of the target traffic scenario area is conducted according to this set. Through standardized quantitative analysis, the qualitative characteristics of each area are transformed into comparable values. Specifically, mathematical models such as the analytic hierarchy process (AHP) and weighted summation can be used. First, based on the management priorities of the target traffic scenario, weight coefficients are assigned to four dimensions: traffic flow characteristics, accident risk, regional function, and topological impact. Then, the indicators of each dimension for each area are quantified, and peak-hour vehicle traffic volume is categorized as below 500 vehicles / hour (1 point), 500-1000 vehicles / hour (2 points), and above 1000 vehicles / hour (3 points). The risk of accidents is assigned a graded score, with 0-1 accidents per month (1 point), 2-3 accidents per month (2 points), and 4 accidents or more per month (3 points). Finally, a weighted summation formula is used to calculate the quantitative result of the criticality of each region, which is: Preliminary regional criticality score = Traffic characteristic score × Traffic weight + Accident risk score × Accident weight + Regional function score × Function weight + Topological impact score × Topological weight. This forms a set of regional criticality coefficients for traffic scenarios. This set of coefficients is a quantitative representation of the traffic criticality of each region. Each region corresponds to a unique coefficient value. The higher the coefficient, the greater the impact of the region on the overall traffic scenario operation and the more important it is to monitor.

[0032] After completing the criticality assessment, key areas are identified in the target traffic scenario region based on the criticality coefficient set of the traffic scenario region. The criticality coefficient threshold is set in combination with the scale of monitoring resources and traffic management priority of the target scenario. Then, through threshold screening or coefficient descending sorting, areas that meet the requirements of criticality coefficient are extracted from all administrative regions / road network units of the target scenario, and finally N key traffic areas are obtained. These areas are the core units in the target traffic scenario that play a decisive role in traffic smoothness and safety.

[0033] Furthermore, a lightweight traffic anomaly identification model was developed, including:

[0034] Based on the multi-sensor monitoring module, the multimodal traffic data type is determined; the traffic anomaly event dataset is classified and trained according to the multimodal traffic data type to obtain a multimodal traffic anomaly recognition model set; the multimodal traffic anomaly recognition model set is weighted and fused to build a lightweight traffic anomaly recognition model.

[0035] Specifically, based on the previously deployed multi-sensor monitoring module, which refers to a collaborative monitoring unit deployed in N key traffic areas and composed of devices such as video sensors, millimeter-wave radar, traffic flow detectors, and environmental sensors, its function is to collect traffic scene data from different dimensions, providing multi-source input for subsequent model training and detection. Based on the module's equipment composition and data acquisition capabilities, the types of multimodal traffic data can be clearly defined, namely, the data categories with differentiated forms and characteristics output by different sensors: for example, image / video data generated by video sensors are used to capture visual anomalies such as vehicle collisions and pedestrian crossings; point cloud data output by millimeter-wave radar is used to accurately acquire dynamic anomalies such as sudden changes in vehicle speed and positional shifts; numerical traffic flow data recorded by traffic flow detectors, such as sudden increases / decreases in traffic volume per unit time; and environmental parameter data collected by environmental sensors, such as environmental anomalies affecting traffic like rainfall and fog. These different modalities of data collectively constitute the objects processed by the model.

[0036] Subsequently, based on the determined multimodal traffic data types, classification training was conducted on the previously collected traffic anomaly event dataset. First, the dataset was split into corresponding subsets according to data type, namely image anomaly subsets, point cloud anomaly subsets, and numerical traffic flow anomaly subsets. Then, suitable deep neural networks were selected for supervised training based on the features of each subset. For image data, convolutional neural networks were used to learn visual anomaly features; for time-series traffic flow data, long short-term memory networks were used to capture abnormal patterns in traffic flow fluctuations; and for radar point cloud data, point cloud convolutional networks were used to identify abnormal vehicle movement trajectories. By independently designing training processes and optimizing model parameters for each modality of data, each modality of data corresponds to a model capable of accurately identifying traffic anomalies under that modality. This ultimately forms a multimodal traffic anomaly recognition model set. This model set contains multiple single-modal anomaly recognition models, each focusing solely on processing anomaly recognition tasks for a specific type of data and outputting anomaly judgment results for the corresponding modality.

[0037] After constructing the model set, a weighted decision fusion is performed. First, the reliability of each model in the multimodal traffic anomaly recognition model set is evaluated using a test set. Based on metrics such as accuracy, recall, and F1 score, a set of reliability coefficients for the multimodal recognition models is obtained, which serves as the weighting basis for each model. Then, the output results of each model are integrated according to their weights to reduce the risk of misjudgment by a single model. Its functional expression can be defined as:

[0038]

[0039] Where: Y represents the final anomaly probability value after weighted decision fusion, with a value range of [0,1]. If Y is greater than the preset anomaly judgment threshold, it is judged that an anomaly exists; k represents the number of models in the multimodal traffic anomaly recognition model set; w i This represents the weight of the i-th single-modal model, i.e., the reliability coefficient, satisfying 0 ≤ w i ≤1, the higher the reliability of the model, the larger the weight; P i This represents the anomaly probability value output by the i-th single-modal model. By weightedly integrating the outputs of each model through this function, a basic traffic anomaly recognition model can be obtained. Its performance is then validated by testing real-time data processing speed, accuracy in complex scenarios, and parameter optimization. This involves reducing redundant network layers and quantizing parameters to lower computational complexity. Finally, a lightweight traffic anomaly recognition model is built. This model balances high recognition accuracy with low resource consumption, significantly reducing computational latency and hardware resource usage while ensuring anomaly recognition accuracy, thus adapting to the real-time detection needs of traffic scenarios.

[0040] Furthermore, a multimodal traffic anomaly identification model set was obtained, including:

[0041] The traffic anomaly event dataset is subjected to associated anomaly feature extraction to obtain a traffic anomaly associated feature set; the traffic anomaly associated feature set is classified and integrated according to the multimodal traffic data type to obtain a multimodal traffic anomaly feature set; a deep neural network is used to perform supervised training on the multimodal traffic anomaly feature set to obtain the multimodal traffic anomaly recognition model set.

[0042] Specifically, the first step is to extract associated anomaly features from the previously collected traffic anomaly event dataset. Through correlation analysis and feature importance ranking algorithms, key features directly related to traffic anomaly events, such as vehicle collisions, road congestion, and pedestrian violations, are screened from the massive amount of raw data, including vehicle speed, image frames, traffic volume, and environmental parameters. The intrinsic relationships between different features are then explored to form an anomaly association feature set. This set is a structured data collection that integrates multi-dimensional association features of various anomaly events. For example, the anomaly association feature set of a rear-end collision will simultaneously include the vehicle speed change curve at the time of the accident, the visual features of the vehicle collision in the video, the real-time traffic data of the accident section, and the environmental visibility parameters, fully reflecting the overall characteristics of the anomaly event.

[0043] After completing the extraction of associated anomaly features, the anomaly associated feature set needs to be classified and integrated based on the multimodal traffic data types previously determined by the multi-sensor monitoring module. That is, according to the data modality attribute to which the features belong, the various features in the anomaly associated feature set are assigned to the corresponding modality category. Vehicle collision visual features and pedestrian illegal crossing image features are classified into the image modality feature category, vehicle speed change features and vehicle position shift features are classified into the radar point cloud modality feature category, and sudden drop in traffic volume per unit time and surge in traffic density features are classified into the numerical traffic flow modality feature category. Through this process, a multimodal traffic anomaly feature set is obtained. This set is divided into multiple independent sub-feature sets based on modality. Each sub-feature set contains only features related to the anomaly event under the corresponding modality, effectively avoiding training interference caused by the mixing of features from different modalities.

[0044] Subsequently, deep neural networks adapted to different modal data features were employed. For image modalities, convolutional neural networks were used to extract visual spatial features; for temporal traffic modalities, long short-term memory networks were used to capture traffic change patterns; and for radar point cloud modalities, point cloud convolutional networks were used to process 3D spatial motion features. Supervised training was conducted on each sub-feature set of the multimodal traffic anomaly feature set. This training process used sub-feature sets labeled with anomaly types—rear-end collisions, road congestion, and pedestrian red-light running—as training samples. After the feature data was input into the corresponding deep neural network, the network weights and bias parameters were continuously adjusted through backpropagation, allowing the model to gradually learn the difference threshold between normal and abnormal traffic features in that modality. Ultimately, the model possessed the ability to accurately determine the existence and type of anomalies based on the input features. Once the training of all modal sub-feature sets reached the preset accuracy standard, a model capable of independently identifying traffic anomalies in that modality was generated for each modality. These models collectively constituted a multimodal traffic anomaly identification model set, encompassing the anomaly identification capabilities of each individual modality.

[0045] Furthermore, a weighted decision fusion is performed on the multimodal traffic anomaly identification model set to build a lightweight traffic anomaly identification model, including:

[0046] The reliability of the multimodal traffic anomaly identification model set is evaluated to obtain a set of multimodal identification model reliability coefficients; the multimodal traffic anomaly identification model set is weighted and fused according to the set of multimodal identification model reliability coefficients to obtain a basic traffic anomaly identification model; the basic traffic anomaly identification model is validated and its parameters are optimized to build the lightweight traffic anomaly identification model.

[0047] Specifically, a reliability evaluation is performed on each model in the multimodal traffic anomaly recognition model set—that is, using an independent test dataset, i.e., containing anomaly samples not used in training, to measure the recognition accuracy (Acc). i Correctly identify the proportion of abnormal samples and the recall rate Rec i The model successfully captured indicators such as the proportion of abnormal samples to quantify its performance, and then calculated the reliability coefficient w. i Its functional expression can be defined as:

[0048] w i =α×Acc i +β×Rec i ;

[0049] Where α and β are weighting coefficients, satisfying α + β = 1, w i The value ranges from [0,1], and the higher the value, the stronger the reliability of the model.

[0050] Based on the reliability coefficient, a weighted decision fusion is performed on a multimodal traffic anomaly identification model set. By integrating the output results of different modal models, the risk of misjudgment by a single model is reduced. Its functional expression can be defined as:

[0051]

[0052] Where: Y represents the final anomaly probability value after weighted decision fusion, with a value range of [0,1]. If Y is greater than the preset anomaly judgment threshold, it is judged that an anomaly exists; k represents the number of models in the multimodal traffic anomaly recognition model set; w i This represents the weight of the i-th single-modal model, i.e., the reliability coefficient, satisfying 0 ≤ w i ≤1, the higher the reliability of the model, the larger the weight; P i This represents the anomaly probability value output by the i-th single-modal model.

[0053] The performance of the basic traffic anomaly recognition model was verified by testing its recognition accuracy and response latency in different scenarios. If there were problems with insufficient accuracy or excessive latency, parameter optimization was carried out, including model pruning, removing redundant network layers to reduce computation, parameter quantization, and converting 32-bit floating-point numbers to 16-bit integers to reduce resource consumption. Finally, a lightweight traffic anomaly recognition model was built. This model balances high recognition accuracy and low resource consumption and can efficiently process real-time anomaly recognition of multi-source traffic data streams.

[0054] Furthermore, identify abnormal events in the target traffic scenario, including:

[0055] The cloud processing center calls the traffic scene area map of the target traffic scene region; the N regional traffic anomaly feature sets are mapped onto the traffic scene area map for visualization and identification, resulting in the distribution area of ​​traffic anomaly features; based on the distribution area of ​​traffic anomaly features, abnormal events are detected in the N regional traffic anomaly feature sets to determine the abnormal events in the target traffic scene.

[0056] Specifically, the first step is to access the traffic scene area map of the target traffic scene area through the cloud processing center. The traffic scene area map contains a high-precision digital map of the road network structure of the target area, road directions, number of lanes, intersection distribution, specific geographical coordinates of N key traffic areas identified in the early stage, traffic facilities, traffic lights, traffic signs and other information. This provides a clear geographical reference framework for the subsequent spatial positioning of abnormal features, ensuring that the abnormal information can accurately correspond to the actual geographical location. Next, the N regional traffic anomaly feature sets obtained from the lightweight traffic anomaly identification model are mapped onto the traffic scene area map. This involves matching the anomaly occurrence location, specific road segment number, and lane position recorded in each regional traffic anomaly feature set with the latitude and longitude or road network coordinates on the traffic scene area map, ensuring that the anomaly features of each key area are anchored to specific spatial points on the map. Subsequently, the mapped anomaly features are visualized, using differentiated visual symbols and color systems to distinguish anomaly types. Vehicle collisions are marked with red triangles, traffic congestion is covered with yellow gradient blocks, and pedestrians illegally crossing the road are marked with blue circles. Key information such as the anomaly occurrence time, impact range, length of congested road segments, and number of lanes occupied by the accident is also attached next to the markers. Through this intuitive visual presentation, the final traffic anomaly feature distribution area is obtained. This distribution area transforms the abstract anomaly features scattered across N key areas into a globally visible map, clearly showing the spatial location, type differences, and initial impact range of each anomaly, effectively avoiding the isolation and fragmentation of anomaly information in a single area. Finally, based on the distribution areas of traffic anomalies, abnormal events are detected in N regional traffic anomaly feature sets. The cloud processing center uses spatial correlation analysis algorithms to determine the correlation between the distribution areas of each anomaly feature, and to determine whether the congestion features of two adjacent key areas are a chain reaction caused by the same traffic flow backlog. At the same time, combined with the quantitative parameters of each anomaly feature set, such as the duration of the anomaly and the decrease in the average vehicle speed, false anomalies caused by temporary sensor failures are eliminated, and effective anomaly information with actual traffic impact is integrated. After the above global verification and correlation analysis, the target traffic scene anomaly event is finally determined. The target traffic scene anomaly event refers to a complete description of anomalies whose specific anomaly type and severity have been clarified after spatial positioning, visualization verification, and global correlation analysis.

[0057] Furthermore, based on the traffic anomaly feature distribution area, abnormal event detection is performed on the traffic anomaly feature sets of the N regions to determine the abnormal events of the target traffic scene, including:

[0058] Based on the distribution area of ​​the traffic anomaly features, global anomaly detection is performed on the N regional traffic anomaly feature sets to obtain a global traffic scene anomaly feature set; the global traffic scene anomaly feature set is then classified into anomaly events to determine the target traffic scene anomaly events.

[0059] Specifically, based on the spatial visualization of traffic anomaly feature distribution areas, global anomaly detection is carried out on N regional traffic anomaly feature sets. Global anomaly detection starts from the overall perspective of the target traffic scene, combined with the spatial correlation of traffic anomaly feature distribution areas, whether there is a transmission of anomalies in adjacent key areas, such as whether there is a causal relationship between a sudden drop in vehicle speed in area A and traffic congestion in area B, and temporal correlation, and changes in anomaly features of the same key area at different time nodes, such as a sudden drop in traffic flow followed by vehicle stagnation in area C. The scattered regional anomaly features are integrated and analyzed across regions and across time, while local false anomaly features caused by temporary sensor errors are eliminated. For example, a single false pedestrian violation in a certain area is judged as false if it does not form a correlation feature in adjacent areas. Finally, the anomaly information with global correlation and temporal coherence is integrated to form a global anomaly feature set of the traffic scene. This set contains a global anomaly information set including the spatial spread range of the abnormal event, the temporal development process, the number of key areas affected, and the core impact parameters, fully presenting the overall picture of the anomaly in the entire traffic scene.

[0060] After constructing the global anomaly feature set for traffic scenarios, anomaly events are classified. Based on the preset traffic anomaly event classification standards, the anomalies are divided into traffic accidents, traffic congestion, and traffic violations according to their nature, and into local single-point anomalies and regional spread anomalies according to their impact range. The core information in the global anomaly feature set is matched and classified. If the global feature set contains information such as a sudden drop in vehicle speed, vehicle collision patterns captured in video images, and the impact on traffic flow in two adjacent key areas, it is classified as a traffic accident, specifically a rear-end collision. If it contains information such as a surge in traffic flow, an average vehicle speed dropping below 10 km / h, and the impact spreading to three key areas, it is classified as a traffic congestion, specifically a regional one. During the classification process, the specific attributes of each abnormal event are clearly defined, including the type of abnormal event, the precise geographical location, and the location of the specific road segment or lane by combining the traffic scene area map, the time of occurrence, the scope of impact, and the severity level. Finally, the complete abnormal events that need to be focused on and dealt with are selected from the classification results, namely the target traffic scene abnormal events. This event is a concrete result that has been verified by global anomaly detection and clearly classified as an abnormal event, such as: At the intersection of XX Road and XX Road, a rear-end collision occurred in the key area XX at XX:XX, occupying X lanes, causing congestion in the downstream key areas XX and XX, and the average vehicle speed decreased by XX%.

[0061] In summary, the rapid detection method for abnormal events in traffic scenarios provided in this application has the following technical effects:

[0062] By accurately identifying key traffic areas based on factors such as traffic flow characteristics and accident risks, and optimizing sensor deployment, the monitoring accuracy of core areas is improved while reducing resource consumption. Furthermore, a lightweight anomaly identification model is built based on multi-source data collected from multiple sensors, avoiding the bias of data from a single sensor and enhancing real-time processing capabilities. Finally, a cloud processing center maps multi-regional anomaly features onto a traffic map for visualization and global correlation analysis, reducing missed and false positives. Ultimately, this improves the efficiency and accuracy of anomaly event detection in traffic scenarios, shortens anomaly identification and response time, and effectively ensures smooth traffic flow and travel safety.

[0063] Example 2: Based on the same inventive concept as the rapid detection method for abnormal events in traffic scenes in Example 1, this application also provides a rapid detection system for abnormal events in traffic scenes. Please refer to the appendix. Figure 2 The system includes:

[0064] The region identification module 11 identifies key regions in the target traffic scene area to obtain N key traffic areas, and deploys multi-sensor monitoring modules in the N key traffic areas to collect multi-source traffic operation data streams in the N areas through the multi-sensor monitoring modules; the data acquisition module 12 collects a traffic anomaly event dataset, and builds a lightweight traffic anomaly identification model based on the multi-sensor monitoring modules and the traffic anomaly event dataset; the anomaly identification module 13 performs anomaly identification on the multi-source traffic operation data streams in the N areas based on the lightweight traffic anomaly identification model to obtain N area traffic anomaly feature sets; the event detection module 14 uploads the N area traffic anomaly feature sets to the cloud processing center for anomaly event detection, determines the target traffic scene anomaly events, and makes anomaly response decisions based on the target traffic scene anomaly events.

[0065] Furthermore, the area identification module 11 in the rapid detection system for abnormal events in traffic scenarios is used to: acquire a set of key factors for traffic areas, which includes traffic flow characteristics, accident risk, regional functions, and topological influence; perform a criticality assessment on the target traffic scenario area according to the set of key factors for traffic areas to obtain a set of key coefficients for traffic scenario areas; and identify key areas of the target traffic scenario area based on the set of key coefficients for traffic scenario areas to obtain the N key traffic areas.

[0066] Furthermore, the data acquisition module 12 in the rapid detection system for abnormal traffic events is used to: determine the multimodal traffic data type according to the multi-sensor monitoring module; classify and train the traffic abnormal event dataset according to the multimodal traffic data type to obtain a multimodal traffic abnormality recognition model set; and perform weighted decision fusion on the multimodal traffic abnormality recognition model set to build a lightweight traffic abnormality recognition model.

[0067] The data acquisition module 12 is further configured to: extract associated abnormal features from the traffic abnormal event dataset to obtain a traffic abnormal associated feature set; classify and integrate the traffic abnormal associated feature set according to the multimodal traffic data type to obtain a multimodal traffic abnormal feature set; and use a deep neural network to perform supervised training on the multimodal traffic abnormal feature set to obtain the multimodal traffic abnormal recognition model set.

[0068] The data acquisition module 12 is further configured to: perform reliability assessment on the multimodal traffic anomaly identification model set to obtain a multimodal identification model reliability coefficient set; perform weighted decision fusion on the multimodal traffic anomaly identification model set according to the multimodal identification model reliability coefficient set to obtain a basic traffic anomaly identification model; and perform performance verification and parameter optimization on the basic traffic anomaly identification model to build the lightweight traffic anomaly identification model.

[0069] Furthermore, the event detection module 14 in the rapid detection system for abnormal traffic events is used to: call the traffic scene area map of the target traffic scene area through the cloud processing center; map the N regional traffic anomaly feature sets onto the traffic scene area map for visualization and identification, thereby obtaining the traffic anomaly feature distribution area; and perform abnormal event detection on the N regional traffic anomaly feature sets based on the traffic anomaly feature distribution area to determine the abnormal events in the target traffic scene.

[0070] The event detection module 14 is further configured to: perform global anomaly detection on the N regional traffic anomaly feature sets based on the traffic anomaly feature distribution area to obtain a global traffic scene anomaly feature set; classify the anomaly events in the global traffic scene anomaly feature set to determine the target traffic scene anomaly event.

[0071] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Figure 1The rapid detection method and specific examples for traffic scene anomaly events in Embodiment 1 are also applicable to the rapid detection system for traffic scene anomaly events in this embodiment. Through the foregoing detailed description of the rapid detection method for traffic scene anomaly events, those skilled in the art can clearly understand the rapid detection system for traffic scene anomaly events in this embodiment; therefore, for the sake of brevity, it will not be described in detail here. As for the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant details can be found in the method section.

[0072] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0073] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application also intends to include such modifications and variations.

Claims

1. A rapid detection method for abnormal events in traffic scenarios, characterized in that, The method includes: Key areas are identified in the target traffic scenario area to obtain N key traffic areas, and multi-sensor monitoring modules are deployed in the N key traffic areas to collect multi-source traffic operation data streams from the N areas through the multi-sensor monitoring modules; Collect a traffic anomaly event dataset, and build a lightweight traffic anomaly recognition model based on the multi-sensor monitoring module and the traffic anomaly event dataset; Based on the lightweight traffic anomaly identification model, anomalies are identified in the multi-source traffic operation data streams of the N regions to obtain traffic anomaly feature sets of the N regions. The N regional traffic anomaly feature sets are uploaded to the cloud processing center for anomaly event detection, to determine the anomaly events in the target traffic scene, and to make anomaly response decisions based on the anomaly events in the target traffic scene.

2. The rapid detection method for abnormal events in traffic scenarios as described in claim 1, characterized in that, We have identified N key traffic areas, including: Obtain a set of key factors for a traffic area, which includes traffic flow characteristics, accident risk, regional functions, and topological impacts. The key factors of the traffic area are used to conduct a key assessment of the target traffic scenario area to obtain a key coefficient set of the traffic scenario area. Based on the set of key coefficients for the traffic scene area, key areas are identified for the target traffic scene area to obtain the N key traffic areas.

3. The rapid detection method for abnormal events in traffic scenarios as described in claim 1, characterized in that, Build a lightweight traffic anomaly detection model, including: The multimodal traffic data type is determined based on the multi-sensor monitoring module. The traffic anomaly event dataset is classified and trained according to the multimodal traffic data type to obtain a multimodal traffic anomaly recognition model set; A lightweight traffic anomaly identification model is built by weighted decision fusion of the multimodal traffic anomaly identification model set.

4. The rapid detection method for abnormal events in traffic scenarios as described in claim 3, characterized in that, Obtain a multimodal traffic anomaly identification model set, including: The traffic anomaly event dataset is subjected to associated anomaly feature extraction to obtain a traffic anomaly associated feature set; The traffic anomaly associated feature set is classified and integrated according to the multimodal traffic data type to obtain a multimodal traffic anomaly feature set; The multimodal traffic anomaly recognition model set is obtained by using a deep neural network to perform supervised training on the feature set of the multimodal traffic anomalies.

5. The rapid detection method for abnormal events in traffic scenarios as described in claim 3, characterized in that, A lightweight traffic anomaly identification model is constructed by weighted decision fusion of the multimodal traffic anomaly identification model set, including: The reliability of the multimodal traffic anomaly identification model set is evaluated to obtain a set of multimodal identification model reliability coefficients; The multimodal traffic anomaly identification model set is weighted and fused according to the reliability coefficient set of the multimodal identification model to obtain the basic traffic anomaly identification model; The performance of the basic traffic anomaly identification model was verified and the parameters were optimized to build the lightweight traffic anomaly identification model.

6. The rapid detection method for abnormal events in traffic scenarios as described in claim 1, characterized in that, Identify abnormal events in the target traffic scenario, including: The cloud processing center retrieves the traffic scene area map of the target traffic scene region. The N regional traffic anomaly feature sets are mapped onto the traffic scene area map for visualization and identification, thereby obtaining the traffic anomaly feature distribution area; Based on the distribution area of ​​the traffic anomaly features, abnormal events are detected in the N regional traffic anomaly feature sets to determine the abnormal events in the target traffic scene.

7. The rapid detection method for abnormal events in traffic scenarios as described in claim 6, characterized in that, Based on the traffic anomaly feature distribution area, abnormal event detection is performed on the traffic anomaly feature sets of the N regions to determine the abnormal events of the target traffic scene, including: Based on the traffic anomaly feature distribution area, global anomaly detection is performed on the N regional traffic anomaly feature sets to obtain a global anomaly feature set for the traffic scene. The abnormal events in the global abnormal feature set of the traffic scene are classified to determine the abnormal events in the target traffic scene.

8. A rapid detection system for abnormal events in traffic scenarios, characterized in that, The system comprises: steps for implementing the method according to any one of claims 1 to 7, wherein the system includes: The area identification module identifies key areas in the target traffic scene area to obtain N key traffic areas, and deploys a multi-sensor monitoring module on the N key traffic areas to collect multi-source traffic operation data streams from the N areas through the multi-sensor monitoring module. The data acquisition module collects a traffic anomaly event dataset and builds a lightweight traffic anomaly identification model based on the multi-sensor monitoring module and the traffic anomaly event dataset. The anomaly identification module identifies anomalies in the multi-source traffic operation data streams of the N regions based on the lightweight traffic anomaly identification model, and obtains traffic anomaly feature sets for the N regions. The event detection module uploads the N regional traffic anomaly feature sets to the cloud processing center for anomaly event detection, identifies anomaly events in the target traffic scene, and makes anomaly response decisions based on the anomaly events in the target traffic scene.

Images