A highway incident data false positive processing system

Abstract

This invention relates to the field of traffic management technology, specifically a highway event data false alarm processing system, comprising: acquiring highway images; labeling events triggered by the highway images with event types; extracting data based on the distribution time period and duration of the triggered events; setting an event feature set; establishing time constraints for the same event type based on the semantic features and duration of the event feature set; verifying the time distribution of the triggered events through a time fitting method; obtaining target events from the verification results; comparing the event features of the target events; if the event features of the target events are consistent with the event type description, the compared data is regarded as highway event data; performing reverse matching on the highway event data; and determining the false alarm processing strategy for the current highway event data based on the false alarm events identified by the reverse matching; thus achieving both efficiency and accuracy in false alarm processing.

Description

Technical Field

[0001] This invention relates to the field of traffic management technology, specifically a highway incident data false alarm processing system. Background Technology

[0002] With the continuous expansion of the highway network and the sustained increase in vehicle ownership, highway traffic safety and emergency response have become important issues in the field of traffic management. Emergency events on highways, such as traffic accidents, road damage, and severe weather, are often accompanied by complex situations. If the causes of highway events are unclear during identification, it will hinder the rapid and effective resolution of the events, limit the efficiency of highway event handling, and seriously affect people's driving experience and road safety.

[0003] For example, Chinese Patent Publication No. CN118887803A discloses a method and apparatus for automatically generating 3D videos for historical reconstruction of highway traffic incidents. The method includes: acquiring static information of the highway road; constructing a static 3D model of the road based on the static information; generating 3D images of vehicles using a 3D database and 3D modeling technology; obtaining vehicle driving trajectories based on target detection and tracking results; optimizing car following models, straight-line models, and turning models based on the vehicle driving trajectories; analyzing the causes of vehicle accidents and the resulting road congestion based on vehicle behavior patterns and road factors detected by the target detection model; extracting the causes of accidents; calculating the course of the accident based on influencing factors; and overlaying the 3D images of vehicles with the pre-modeled static 3D images of the road to generate a 3D simulation video.

[0004] For example, Chinese Patent Publication No. CN119323887A discloses a method, system, terminal, and storage medium for handling highway incidents based on big data. The method includes collecting real-time data from highways, preprocessing the collected real-time data, analyzing the preprocessed real-time data to determine whether an emergency has occurred on the highway, and if so, inputting the collected real-time data into a pre-built highway emergency incident handling model to generate emergency incident handling measures.

[0005] Existing technologies identify vehicle accident occurrences by recognizing road facilities, equipment, and road surface information; and clarify the weight allocation for highway incident handling by making decisions based on the status of people, vehicles, roads, environment, and information. However, existing technologies neglect the handling of false alarms in highway incidents, leading to missed and incorrect cases in the statistical events, thus reducing the accuracy of false alarm handling in highway incidents. Summary of the Invention

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a highway event data false alarm processing system, comprising: an event acquisition module, used to acquire highway images, mark the events triggered by the highway images by event type, extract data according to the distribution time period and duration of the triggered events, and set event feature sets.

[0007] The target recognition module is used to establish time constraints for the same event type based on the semantic features and duration of the event feature set, and to verify whether the duration of multiple triggering events conforms to the time distribution of the event type through a time fitting method, and to obtain multiple target events from the verification results.

[0008] The feature comparison module is used to compare the event features of the target event with the relevant data of the target event in the corresponding environment. If the event features of the target event are consistent with the event type description, the compared data will be regarded as highway event data.

[0009] The decision matching module is used to perform reverse matching on highway event data and determine the false alarm handling strategy for the current highway event data based on the false alarm events identified in the reverse matching.

[0010] The beneficial effects of this invention are as follows: First, this invention extracts target samples with semantic features from each frame of image by using a method of target detection to obtain samples, regular expression matching to filter, and confidence ranking. It parses the type combination rules and calls the event type regular expression in the database to match the semantic features of the samples with the regular expression to achieve initial labeling of the event type. Then, it calculates the confidence of each sample and filters it according to the threshold, retaining the image frames corresponding to high-confidence samples. This reduces invalid data when reporting highway events, avoids subjective preliminary labeling errors, and ensures the initial credibility of highway event reporting.

[0011] Second, this invention groups the filtered samples by the same location and the same event type, merges consecutive image frames to generate the distribution time period of the triggering event, and records the duration synchronously; this avoids the same event being split or different events being confused, laying a data foundation for subsequent verification of time patterns; secondly, it associates the current event type with the identification of continuous historical cycles, statistically analyzes the periodic historical common features and intensity fields, and realizes that the extracted event feature set meets the characteristics of historical adaptation and full scene coverage, avoiding individual events being ignored due to low frequency.

[0012] Third, this invention employs a method of defining time constraints by grouping, quantifying probability values, and verifying through time distribution fitting. It groups events by event type and semantic features, combining duration constraints, time continuity constraints, and time interval constraints to form basic time constraints, and then adjusts the constraint parameters using semantic features. Under the premise that the constraints meet the standards, the ratio of the number of effective semantic feature frames to the total number of frames is used as the probability value within the event duration. After sorting events by probability value, duration data of events of the same type are extracted to fit a probability distribution model, verifying whether the current event falls within the fitted interval. This achieves scenario-based adaptation of time constraints, avoiding misjudgments of different event types using a single threshold. By quantifying the stability of event features through probability values ​​and verifying the rationality of time patterns through distribution fitting, it double-filters out false alarms caused by instantaneous feature jumps and abnormal time patterns, ensuring that the time dimension of the selected target events is truly effective.

[0013] Fourth, this invention compares the target event's own attributes and surrounding environmental attributes using feature quantification comparison and weighted scoring, avoiding mismatches and false alarms caused by partial feature similarity, thus further ensuring the accuracy of highway event data. Subsequently, the highway event data is divided into event subsets based on the false alarm events associated with each event. A list of false alarm events for each subset is output step-by-step according to the processing flow. Candidate relationships between false alarm events are determined through false alarm triggering conditions, and finally, a false alarm handling strategy is output. This achieves structured tracing of false alarm events, allowing for precise matching of processing strategies and improving the efficiency and accuracy of false alarm handling. Attached Figure Description

[0014] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0015] Figure 1 This is a system framework diagram of a highway incident data false alarm processing system.

[0016] Figure 2 This is a flowchart of the event acquisition module of a highway event data false alarm processing system.

[0017] Figure 3 This is a flowchart illustrating the target identification module of a highway incident data false alarm processing system.

[0018] Figure 4 This is a flowchart illustrating the feature comparison module of a highway incident data false alarm processing system.

[0019] Figure 5 This is a flowchart illustrating the decision matching module of a highway incident data false alarm processing system. Detailed Implementation

[0020] The embodiments of the present invention are described in detail below. The embodiments described below are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. Where specific techniques or conditions are not specified in the embodiments, they shall be performed in accordance with the techniques or conditions described in the literature in the art or in accordance with the product manual.

[0021] See Figure 1 A highway incident data false alarm processing system includes: an incident acquisition module, a target recognition module, a feature comparison module, and a decision matching module; wherein, the output of the incident acquisition module is connected to the target recognition module, the output of the target recognition module is connected to the feature comparison module, and the output of the feature comparison module is connected to the decision matching module.

[0022] The event acquisition module is used to acquire highway images, identify events triggered by the highway images by event type, extract data based on the distribution time period and duration of the triggered events, and set event feature sets.

[0023] The target recognition module is used to establish time constraints for the same event type based on the semantic features and duration of the event feature set, and to verify whether the duration of multiple triggering events conforms to the time distribution of the event type through a time fitting method, and to obtain multiple target events from the verification results.

[0024] The feature comparison module is used to compare the event features of the target event with the relevant data of the target event in the corresponding environment. If the event features of the target event are consistent with the event type description, the compared data will be regarded as highway event data.

[0025] The decision matching module is used to perform reverse matching on highway event data and determine the false alarm handling strategy for the current highway event data based on the false alarm events identified in the reverse matching.

[0026] The event acquisition module analyzes each frame of the acquired image to determine whether it has triggered a preset highway event, such as a car accident, an abnormal pedestrian, road repair, or debris. It then labels the image or video frame with the corresponding event type. For example, if a frame triggers a car accident, it is labeled as the event type: car accident.

[0027] When setting up an event feature set, the event feature set is a preliminary data processing method to identify highway events that trigger alarms, such as pedestrians or non-motorized vehicles, maintenance pedestrians, traffic cones and other signs in construction areas, abnormal parking, littering, and abnormal driving. When an event matching the corresponding description is identified in an image, it is described using an event type label, and the trigger time and duration of the corresponding event are recorded, which is then used as the event feature set for preliminary processing.

[0028] The time zone in which the event occurred needs to be examined, as well as the duration of the event within the marked area. This data is used for tracking and processing highway events. The time period can exclude isolated, non-continuously time-related false trigger data, such as a single frame misidentified as a pedestrian, or an image without any preceding or following time context. The duration can record the length of time from the occurrence to the end of the event, excluding instantaneous false triggers, such as scattered debris tags that disappear within 1 second due to wind blowing debris. In this case, the duration is more inclined to assess the lasting impact of events occurring on highways, as well as the process of eliminating the impact of each event, to enrich the form of description for each event type.

[0029] Highway image data can be acquired using devices such as fixed high-definition cameras, tunnel-specific cameras, drone inspection equipment, and vehicle-mounted mobile acquisition equipment. The acquired highway images are recorded at a rate of 10-15 frames per second. When a preset event such as a car accident or an abnormal pedestrian occurs in any frame of the image, the corresponding data is coarsely extracted, tagged, and then the time is recorded to determine the valid triggering event.

[0030] like Figure 2 As shown, the implementation of the event acquisition module includes: A1, based on object detection, at least one target sample is obtained from each frame of highway image, and each target sample corresponds to semantic features; the target sample represents a specific graphic in the image. These data will be compared with the labeled image set by loading a standard image set to extract positive samples such as stopped vehicles, pedestrians, traffic cones and scattered objects, as well as negative samples such as normally moving vehicles and shadows. These samples will correspond to an event type; the object detection algorithm can use YOLOv8 / YOLOv9 to load the target samples appearing in the image; then each target sample is described using semantic features, such as text like oil spill marks and traffic cones.

[0031] A2 performs regular expression matching to filter target samples and initially labels the event type of each target sample. For example, if multiple vehicle collision postures are detected and oil leak marks are found, it is labeled as a suspected car accident; if traffic cones are detected and people are found, it is labeled as a suspected road repair. The filtering method is to use some detected images and structures as the parts that need to be warned, and then use semantic labels to label them.

[0032] A3. Based on the event type after each target sample is labeled, determine the confidence level corresponding to each target sample, sort the target samples according to the confidence level, and retain the image frames corresponding to the target samples that meet the confidence level requirements.

[0033] During the screening process, it is also necessary to record the confidence level of the corresponding label. Target samples with a confidence level ≥ 0.7 should be retained to exclude false triggers such as flying birds and light and shadow, and sorted according to the confidence level value. At this time, the confidence level is calculated by the target detection algorithm by multiplying the probability of the target's existence by the probability of class matching.

[0034] Preferably, when performing regular expression matching to filter target samples, the implementation method further includes: parsing the type combination rules of the target samples, obtaining the regular expression of the event type from the database according to the type combination rules, matching the semantic features of the target samples with the regular expression, and obtaining the event type matched by the target samples in the same frame image.

[0035] Type combination rules represent the logical relationship between preset event types and semantic features, such as car accident = vehicle + collision deformation + oil leak, and road maintenance = traffic cone + personnel + emergency lane.

[0036] Regular expressions are used to structurally represent combinations of semantic features, describe the AND, OR, and NOT relationships of features using logical symbols, and store the corresponding data in a database, providing a basis for matching semantic features in tracking images.

[0037] When setting up the event feature set later, data processing along the time dimension is used to align each target sample with the timeline and extract valid event data that exhibits a certain time pattern.

[0038] A4 groups the target samples according to the same location and the same event type, merges consecutive image frames to form the distribution time period of the triggering event, and records the duration of the triggering event.

[0039] When identifying event feature sets in highway images, it is also necessary to statistically analyze the common features that appear in historical data, verify whether the current event conforms to its common features, and fill in the corresponding data to provide a more comprehensive basis for time fitting and comparison.

[0040] The implementation of the story event acquisition module also includes: obtaining the same event type as the current triggering event from historical data; if the event type is identified in consecutive periods; then statistically analyzing the target samples in each period, extracting the historical common features of the current triggering event, and filling them into the event feature set.

[0041] If the event type is not identified in multiple consecutive historical periods, the current event type is confirmed as an event with low intensity, and an intensity marker field is added to the event feature set.

[0042] The period for event types can be set by day, such as 1 day per period; the continuous period to be viewed can initially view relevant data for 3 consecutive days, statistically analyze the common characteristics of frequently identified highway events, and synchronize these common characteristics to each view to verify whether the current event is a false alarm.

[0043] Historical common characteristics can be illustrated by the following examples: common occurrence location: G15 Expressway K120-K130 section; average duration: 25 minutes; common environment: sunny day, morning rush hour; these characteristics represent historical patterns found from continuous periods. These historical patterns will be used to verify the currently reported expressway events in order to obtain a more detailed description of the relevant events.

[0044] As for events with added intensity marker fields, this part of the data needs to be carefully verified. In the subsequent comparison process, multiple data will be used for comparison, combining radar, vehicle attributes, surrounding environment attributes and other data to verify the specific description of the event and prevent the problem of incorrect description when reporting the event.

[0045] The output event feature set will include event ID, preliminary event type, location of occurrence, semantic features, and the corresponding time data of the event. This data will serve as the data set for the current preliminary identification.

[0046] In one embodiment of the present invention, when the target recognition module processes the event feature set, it sets time constraints according to the duration, temporal continuity and time interval of the event type in the event feature set, and comprehensively describes the image frame time that appears continuously in the image at each time to determine whether the time distribution of the current triggered event conforms to the verification result.

[0047] like Figure 3 As shown, the implementation of the target recognition module includes: B1, grouping the triggering events according to event type and semantic features, and defining time constraints for each group of data using duration and semantic features.

[0048] At this point, the event types identified by the event features are further grouped, such as Group 1: Event type: car accident, semantic features: driving lane + multi-vehicle collision; Group 2: Event type: car accident, semantic features: emergency lane + single vehicle failure.

[0049] Then, time constraints are defined for the data of each group to indicate whether each triggering event conforms to the normal time distribution. When defining time constraints, the implementation methods include: defining duration constraints based on the shortest effective duration of the triggering event, defining time continuity constraints based on the maximum allowable interruption interval of the time series, and setting time interval constraints based on the minimum time difference between two consecutive triggering events of the same event type.

[0050] These three constraints represent time constraints in terms of duration, interval, and continuity. The minimum effective duration can be set based on the shortest time in historical data to effectively identify the current event type, such as a car accident ≥ 30 seconds, debris ≥ 15 seconds. These times represent the processing time of the identified event, which can filter out fleeting false alarms. The maximum interruption interval is set by finding the maximum interval allowed for the same event type to be interrupted in highway images from the database. This represents the situation where the camera is briefly obstructed during identification, losing relevant image details. For example, if the camera is briefly obstructed by a truck for 3 seconds before detecting the relevant car accident, a large interval may indicate that the currently reported event has already ended or that the event reporting is delayed. In cases of delay; the minimum time difference setting is to avoid repeatedly judging the same event and not reporting the same event repeatedly; these three values ​​are extracted from the database, or confidence intervals are set according to the same event type in historical data. Confidence intervals with a 95% confidence level are set for the shortest effective duration, the maximum interruption interval, and the minimum time difference between two consecutive triggering events, respectively, in the form of the average ± 1.96 times the standard deviation. If the duration of the current triggering event is less than the shortest effective duration, greater than the maximum interruption interval, and less than the minimum time difference, it is considered an event with false alarms, and data that meets these three conditions is considered to have met the time constraint; at the same time, the selected data can be historical event data from the past 3 months.

[0051] By combining the duration constraint, time continuity constraint, and time interval constraint, we obtain the current basic time constraint.

[0052] Since triggering events of the same event type are affected by different semantic descriptions, the basic time constraints set will also change accordingly. At this time, it is necessary to adjust the set time constraints in accordance with their semantic characteristics.

[0053] The basic time constraints are adjusted using semantic features, and the adjusted basic time constraints are used as the output time constraints.

[0054] When adjusting using semantic features, the calculation is performed on data with the same semantic features from historical data, in the same way as the calculation of basic time constraints.

[0055] For example, if the semantic features include oil leak + multi-vehicle collision, the duration constraint is increased to ≥60 seconds; if it is only a minor scratch on a single vehicle, the duration constraint is decreased to ≥30 seconds. The adjusted semantic features represent the difficulty of handling the corresponding event on the highway.

[0056] B2, when the time constraint is met, determines the effective semantic features of the current triggering event, and uses the ratio of the number of frames corresponding to the effective semantic features to the total number of frames as the probability value of each triggering event within the corresponding duration.

[0057] Effective semantic features represent the core semantic information extracted from each image frame of a highway image that can directly and uniquely point to the authenticity of a certain type of event. This means that not all semantic features in an image are used to distinguish between real events and false alarms. For example, semantic features such as vehicle collision deformation, oil leak marks, multiple stationary vehicles superimposed, and hazard lights continuously on are effective semantic features that point to traffic accidents. Vehicle color, license plate number, etc., cannot directly point to traffic accidents and are only used as semantic features for auxiliary identification and are invalid semantic features.

[0058] At this point, it is necessary to record the number of valid semantic features in each image frame and determine whether the current event is real based on the ratio of the number of valid features to the total number of frames during the duration of the current triggering event. For example, 0.7 can generally be chosen as the judgment value. When the ratio is greater than 0.7, the current event is considered to exist stably during the duration. For example, collision deformation in a car accident can be detected in almost every frame of the image before the accident is dealt with, while oil leak marks caused by light and shadow misjudgment may only appear in a few frames of the image, making it impossible to directly determine the authenticity of the car accident.

[0059] B3. The triggering events are sorted by the probability value of each triggering event within the corresponding duration. The duration data of triggering events under the same event type are extracted, fitted into a probability distribution model, and the distribution of the current triggering event is verified to be within the fitting interval.

[0060] At this point, the triggering events are sorted according to their probability values ​​within the corresponding duration. This is done by filtering the triggering events based on their probability values ​​being greater than 0.7 and less than 0.7, with the portion having a probability value greater than 0.7 used as the subject for time fitting judgment. Then, the probability distribution model is fitted according to the duration. The reasonableness of the duration is determined by whether the duration falls within the duration fitting range for the same event type. For example, road repairs usually last 30-120 minutes, and debris usually lasts 15-60 minutes. The duration length is used to exclude some stable but abnormally long features.

[0061] It should be noted that the current processing method is biased towards historical pattern time fitting. The duration can be used as the length value in historical data, and a Gaussian distribution can be used as the current probability distribution model to obtain the mean and standard deviation of the duration under the normal distribution. The fitting interval is set in the form of mean ± 1.5 standard deviation.

[0062] B4. If both the current triggering event and the previous triggering event are within the fitting interval, then the corresponding triggering event will be used as the target event for output.

[0063] If both the current triggering event and the previous triggering event are within the fitting interval, the event is considered reasonable; otherwise, it is considered a duplicate false alarm or other form of false alarm. Data within the fitting interval will fit the false alarm screening scenario. Data that has undergone time constraint and time fitting processing will be used as the verification result, and the corresponding triggering events of these data will be output as the target events.

[0064] Preferably, when acquiring multiple target events, the implementation method further includes: determining whether the target event has been updated; if so, recording the target event at the time of the update; if not, outputting the current target event.

[0065] By performing update status determination on verified highway target events, when there are updates in dimensions such as features and status of the event, the updated event data is retained; when there are no updates to the event, the target event data fitted at the current time is output, ensuring the timeliness, uniqueness and traceability of the output data.

[0066] In one embodiment of the present invention, the feature comparison module compares the target event’s own attributes and the surrounding environment attributes, uses these attributes as event features, and compares whether these descriptions are consistent with the event type description. This thoroughly filters out latent false alarms that are stable in time but have incorrect features, and finally outputs highway event data with accurate types and real features, laying the foundation for subsequent handling and review of false alarms.

[0067] In the target recognition module, effective semantic features are used to sort the triggered events. Here, we will further verify the description of the effective semantic features on the target event's own attributes and the surrounding environment attributes, so that the initially identified event types can be more specific and accurate.

[0068] like Figure 4 As shown, the feature comparison module is implemented by using the target event's own attributes and surrounding environment attributes as event features, and quantifying the comparison results by using the matching score between the event features and the event type.

[0069] Each event feature is assigned a weight, and each event feature is scored according to the part that is consistent with the event type description. The weighted sum of these scores is used as the matching score. When the matching score is greater than the preset matching score, the event feature is considered to be consistent with the event type description.

[0070] At this point, the target event's attributes will be categorized into vehicle events and pedestrian or debris events based on the currently identified target event. Vehicle events will be checked for characteristics such as vehicle type, status (whether hazard lights / brake lights are continuously on, whether there is smoke / oil leakage), and posture (whether it is tilted / deformed due to a collision). Pedestrian or debris events will be checked for characteristics such as pedestrian clothing (whether it is high-speed work clothing, to distinguish between normal workers and abnormal pedestrians) and debris material (whether it is easily dropped goods, to distinguish between real debris and road stains).

[0071] The surrounding environment attributes can be divided into lane verification and surrounding associated target verification. Lane verification checks the lane where the target is located (driving lane / emergency lane) and whether the lane is the location described by the event type. Associated target verification checks the rationality of surrounding associated targets (e.g., there should be other vehicles' avoidance trajectories around a car accident event, and there should be traffic cones around road repairs; if there are no associated targets, an anomaly is marked).

[0072] For each labeled event feature, its weight value can be directly extracted from the database, or a correlation matrix can be constructed from the currently extracted event features. The weight of each event feature can be obtained from the correlation matrix. The correlation matrix is ​​a matrix of feature number × feature number. The value of the correlation matrix is ​​set by calculating the Pearson correlation coefficient. Then, the weight of each event feature label can be obtained by weight normalization.

[0073] As for scoring each event feature, the part that matches the event type label is given a value of 1, and the part that does not match is given a value of 0. The weighted sum of the event features is calculated to obtain the current matching degree score. The preset matching degree is generally set to 90%, but if the currently identified event is a high-weight event, it can be increased to 95% to adjust the accuracy of event type recognition.

[0074] In one embodiment of the present invention, in the decision matching module, based on the dual processing of the target recognition module and the feature comparison module, the output highway event data is data after removing false alarms, representing relatively accurate data after removing false alarms. This part of the data will be directly stored in the database or system cache for later direct viewing.

[0075] The purpose of reverse matching is to review the removed false alarm events, divide these false alarm events into multiple false alarm subsets according to their positions, determine the location where the false alarm events occurred, and unify the data of false alarms.

[0076] like Figure 5As shown, the implementation of the decision matching module includes: D1, based on the false alarm events associated with each event in the highway event data, dividing the highway event data into multiple event subsets; determining the false alarm event list for each event subset, the false alarm event list includes event ID, false alarm triggering conditions, and the basis for quantification processing. Then, by removing the false alarm highway event data and adding the removed false alarm data, it is divided into multiple false alarm events. The false alarm events are then statistically analyzed according to their location, such as highway segment station interval, lane, and the removed process.

[0077] D2 outputs the list of false alarm events for each event subset step by step according to the processing flow, and uses the false alarm triggering conditions to determine the candidate relationship between each false alarm event.

[0078] False alarm trigger conditions represent the basis for filtering corresponding false alarm data during the comparison and time fitting process. These conditions may include contradictions in the target event attributes and environment comparison, failure of time fitting verification, meeting time constraints but with extremely low matching degree, repeated triggering on the same road segment with consistent features, and failure to meet time constraints. These specific conditions triggering the filtering process are considered false alarm trigger conditions. Simultaneously, each false alarm event list will quantify specific processing criteria, such as quantitative indicators during filtering: matching degree 25% < threshold 60%, frame interval 6 seconds > threshold 5 seconds, etc., to verify the main conditions and types of false alarms. Finally, these data will be used to apply the false alarm handling strategy, completing the differentiation and archiving verification of false alarm events. The verified data can also be used to update the content currently stored in the database, improving the efficiency of subsequent false alarm event processing.

[0079] D3: For each false alarm event in the candidate relationship identification process, determine at least one false alarm handling strategy corresponding to the false alarm event, and filter the false alarm handling strategies according to the candidate relationship between multiple false alarm events to determine the current output false alarm handling strategy.

[0080] When screening false alarm handling strategies, the implementation methods include: based on at least one false alarm event associated with the false alarm triggering condition, the intersection of the false alarm triggering conditions corresponding to each false alarm event is regarded as a candidate relationship; if both false alarm events A and B fail the time fitting verification, then A and B are candidate relationships. The candidate relationships are automatically combined when the false alarm events are extracted according to the processing flow.

[0081] When filtering based on candidate relationships, at least one false alarm handling strategy form is obtained. It is ensured that the fields in the false alarm handling strategy form that apply the false alarm triggering conditions can match the false alarm triggering conditions in the false alarm event list. The false alarm handling strategy that the false alarm triggering conditions point to is taken as the output false alarm handling strategy, thereby filtering out the strategies that are applicable to the current false alarm event and also applicable to its candidate relationship false alarm events.

[0082] The part that the false alarm triggering condition points to at the same time means that multiple different false alarm triggering conditions can be solved by the same false alarm handling strategy. For example, if the time fitting fails, the frame interval exceeds the threshold, or the same road segment is repeatedly triggered, the false alarms triggered by these conditions can all be handled by selecting a strategy, avoiding repeatedly formulating and calling different strategies for the same type of false alarms triggered by different conditions, thus improving the efficiency and consistency of false alarm handling.

[0083] Suppose that false alarm triggering conditions A (time fitting < 0.7) and B (frame break interval > 5 seconds) both point to strategy X (camera transmission status verification + image encoding parameter optimization). This is because the false alarms under these three conditions are essentially caused by camera transmission lag / encoding abnormalities. Strategy X can solve the problem at its source. At this point, by examining the filtered data and corresponding conditions, the efficiency of the current system in handling false alarm events can be improved, thereby improving the timeliness and efficiency of the system's response to highway events.

[0084] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention, which are still covered within the protection scope of the present invention.

Claims

1. A highway incident data false positive processing system, characterized by, include: The event acquisition module is used to acquire highway images, identify the events triggered by the highway images by event type, extract data based on the distribution time period and duration of the triggered events, and set event feature sets. The target recognition module is used to establish time constraints for the same event type based on the semantic features and duration of the event feature set, and to verify whether the duration of multiple triggering events conforms to the time distribution of the event type through a time fitting method, and to obtain multiple target events from the verification results. The feature comparison module is used to compare the event features of the target event with the relevant data of the target event in the corresponding environment. If the event features of the target event are consistent with the event type description, the compared data will be regarded as highway event data. The decision matching module is used to perform reverse matching on highway event data and determine the false alarm handling strategy for the current highway event data based on the false alarm events matched in reverse. The event acquisition module can be implemented in the following ways: Based on object detection, at least one target sample is obtained from each frame of highway image, and each target sample corresponds to semantic features; The target samples are filtered using regular expression matching to initially label the event types of each target sample; Based on the event type after each target sample is labeled, the confidence level corresponding to each target sample is determined, the target samples are sorted according to the confidence level, and the image frames corresponding to the target samples that meet the confidence level requirements are retained. The target samples are grouped according to the same location and the same event type. Continuous image frames are merged to form the distribution time period of the triggering event, and the duration of the triggering event is recorded. The target recognition module can be implemented in the following ways: Triggered events are grouped according to event type and semantic features, and time constraints are defined for each group of data using duration and semantic features; When the time constraint is met, the effective semantic features of the current triggering event are determined, and the ratio of the number of frames corresponding to the effective semantic features to the total number of frames is used as the probability value of each triggering event within the corresponding duration. Triggering events are sorted by probability value of each event within the corresponding duration, duration data of triggering events of the same event type are extracted, fitted into a probability distribution model, and the distribution of the current triggering event is verified to be within the fitted interval. If both the current triggering event and the previous triggering event are within the fitting interval, then the corresponding triggering event will be used as the target event for output. When defining time constraints, the implementation methods include: The duration constraint is defined based on the shortest effective duration of the triggering event, the time continuity constraint is defined based on the maximum allowable interruption interval of the time series, and the time interval constraint is set based on the minimum time difference between two consecutive triggering events of the same event type. By combining the duration constraint, time continuity constraint, and time interval constraint, we obtain the current defined basic time constraint; The basic time constraints are adjusted using semantic features, and the adjusted basic time constraints are used as the output time constraints. When retrieving multiple target events, the implementation methods also include: Determine if the target event has been updated. If it has, record the target event at the time of the update. If it has not been updated, output the current target event.

2. A highway incident data false positive processing system as claimed in claim 1, wherein, When performing regular expression matching to filter target samples, the implementation methods also include: The target sample's type combination rules are parsed, and the event type regular expression is obtained from the database according to the type combination rules. The semantic features of the target sample are matched with the regular expression to obtain the event type matched by the target sample in the same frame image.

3. A highway incident data false positive processing system as defined in claim 1, wherein, The implementation methods of the event acquisition module also include: Obtain the event type that is the same as the current triggering event from historical data. If the event type is identified in consecutive periods, then count the target samples in each period, extract the historical common features of the current triggering event, and populate the event feature set. If the event type is not identified in multiple consecutive historical periods, the current event type is confirmed as an event with low intensity, and an intensity marker field is added to the event feature set.

4. A highway incident data false positive processing system as defined in claim 1, wherein, The feature comparison module can be implemented in the following ways: The target event's own attributes and surrounding environmental attributes are used as event features, and the matching score between the event features and the event type is used to quantify the comparison results. Each event feature is assigned a weight, and each event feature is scored according to the part that is consistent with the event type description. The weighted sum of these scores is used as the matching score. When the matching score is greater than the preset matching score, the event feature is considered to be consistent with the event type description.

5. A highway incident data false positive processing system as defined in claim 1, wherein, The decision matching module can be implemented in the following ways: Based on the false alarm events associated with each event in the highway event data, the highway event data is divided into multiple event subsets; The list of false alarm events for each subset of events is output step by step according to the processing flow, and the candidate relationship between each false alarm event is determined by using the false alarm triggering conditions; For each false alarm event in the candidate relationship identification process, at least one false alarm handling strategy corresponding to the false alarm event is determined. Based on the candidate relationship between multiple false alarm events, the false alarm handling strategies are filtered to determine the current output false alarm handling strategy.

6. A highway incident data false positive processing system as claimed in claim 5, wherein, When filtering false alarm handling strategies, the implementation methods include: Based on at least one false alarm event associated with the false alarm triggering condition, the intersection of the false alarm triggering conditions corresponding to each false alarm event is considered as a candidate relationship; When filtering based on candidate relationships, obtain at least one false alarm handling strategy form, ensure that the fields in the false alarm handling strategy form that apply the false alarm triggering conditions match the false alarm triggering conditions in the false alarm event list, and use the false alarm handling strategy that the false alarm triggering conditions point to simultaneously as the output false alarm handling strategy.

Images