Congestion alarm method based on visual large model secondary verification

Abstract

The invention relates to the technical field of urban intelligent traffic, in particular to a congestion alarm method based on visual large model secondary verification, which comprises the following steps: periodically acquiring traffic parameters of each road section of a target road and a congestion alarm state of a previous period; according to the traffic parameters and the congestion alarm state of the previous period, performing preliminary congestion state identification, and generating a congestion alarm alternative set; based on historical false alarm information and historical disposal information of each road section of the target road, screening and priority ranking are carried out on the congestion alarm alternative set, and a congestion alarm to-be-analyzed list after priority ranking is generated; acquiring a real-time video stream corresponding to each road section in the congestion alarm to-be-analyzed list, inputting the real-time video streams into a preset visual large model, and acquiring a false alarm set and a final congestion alarm set; and performing congestion alarm according to the final congestion alarm set and a preset alarm rule. According to the invention, the accuracy and reliability of congestion alarm are obviously improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of urban intelligent transportation, and in particular to a congestion alarm method based on secondary verification of a visual large model. BACKGROUND

[0002] With the continuous acceleration of urbanization and the continuous growth of the number of motor vehicles, traffic congestion has become an important problem affecting the efficiency of urban operation, the travel experience of residents and the ecological environment. Accurate and timely identification of traffic congestion state not only plays a key role in optimizing traffic organization and improving road capacity, but also is the basic support for realizing intelligent traffic management and modernization of urban governance. Therefore, building an efficient, accurate and robust traffic congestion identification system has become an urgent need and core task of the development of modern urban intelligent transportation.

[0003] In current urban traffic management, congestion identification methods based on speed, flow, queue and other basic data have been widely used in various intelligent traffic systems and have become an important technical means to realize traffic state perception and dispatching decision. However, with the increasing complexity of urban road network structure and the growing demand for traffic, relying on single or limited dimensional data for congestion judgment has gradually exposed its limitations. Influenced by sensor device failure, communication delay and environmental interference, the traffic data obtained may fluctuate abnormally, leading to frequent false alarms and unnecessary dispatch response and resource waste.

[0004] Therefore, there is an urgent need for a congestion alarm method based on secondary verification of a visual large model. SUMMARY

[0005] (I) Technical problems to be solved In view of the above-mentioned shortcomings and deficiencies of the prior art, the present application provides a congestion alarm method based on secondary verification of a visual large model, which solves the problem of high false alarm rate caused by environmental interference in the prior art traffic congestion alarm method based on data analysis, and the problem of response delay, waste of computing power and environmental adaptability caused by the limitation of the visual large model full-time verification scheme.

[0006] (II) Technical solutions In order to achieve the above-mentioned purposes, the main technical solutions adopted by the present application include: The congestion alarm method based on secondary verification of a visual large model provided by the embodiments of the present application comprises: S100, periodically acquiring traffic parameters of each road section of a target road and congestion alarm state of the last period; performing preliminary congestion state identification according to the traffic parameters and the congestion alarm state of the last period to generate a congestion alarm candidate set; The traffic parameters include speed, flow and traffic capacity. S200, screening and priority sorting the congestion alarm candidate set based on historical false alarm information and historical handling information of each road section of the target road, to generate a congestion alarm to-be-analyzed list after priority sorting; S300, acquiring real-time video streams corresponding to each road section in the congestion alarm to-be-analyzed list, inputting the real-time video streams into a visual large model set in advance, and acquiring a false alarm set and a final congestion alarm set; S400, performing congestion alarm according to the final congestion alarm set and an alarm rule set in advance.

[0007] Optionally, in some embodiments of the present application, the S100 comprises: According to the traffic parameter, a traffic saturation of each road section of the target road is acquired; the traffic saturation is a ratio of flow per unit time to traffic capacity; If the congestion alarm state of the last period is no alarm, when the speed is less than a preset speed threshold and the saturation is greater than a preset saturation threshold, it is determined as suspected congestion; if the congestion alarm state of the last period is alarm, when the speed is less than a preset speed threshold, it is determined as suspected congestion; Basic information of each road section determined as suspected congestion is acquired to generate a congestion alarm candidate set; the basic information of each road section comprises road section identification, intersection identification, import identification, suspected congestion start time and corresponding traffic parameters.

[0008] Optionally, in some embodiments of the present application, the preset speed threshold is acquired according to the following steps: Historical traffic flow data of the target road is acquired, traffic state clustering analysis is performed on each road section under different road levels, free flow speed and critical congestion density of each road section are acquired; According to the free flow speed and the critical congestion density of each road section, a critical speed from free flow to congestion transition is acquired through a Greenshields model; The preset speed threshold is a times of the critical speed, and a is acquired according to different road levels; the road levels comprise expressway, express road, main road and secondary road.

[0009] Optionally, in some embodiments of the present application, the S200 comprises: S210, acquiring historical false alarm information and historical handling information of each road section of the target road; The historical false alarm information comprises false alarm records and false alarm time in a historical period, and the historical handling information comprises handling records and handling time in a historical period; S220, screen out, as a first congestion set, a road section with a false alarm record within a preset false alarm time window according to the historical false alarm information and historical handling information, screen out, as a second congestion set, a road section that has completed handling within a preset handling time window, and remove the first congestion set and the second congestion set from the congestion alarm candidate set to obtain a primary screening set; S230, obtain a real alarm set of each road section of a target road in a previous period, compare the primary screening set with the real alarm set of each road section of the target road in the previous period, screen the primary screening set, and generate a congestion alarm analysis list; S240, sort the congestion alarm analysis list according to a multi-dimensional hierarchical sorting rule set in advance to obtain a congestion alarm analysis list after priority sorting.

[0010] Optionally, in some embodiments of the present application, the S230 includes: The primary screening set is compared with the real alarm set of each road section of the target road in the previous period to identify continuous alarms, new alarms, and suspected disappeared alarms. The real alarm set of each road section of the target road in the previous period includes a road section identifier, an intersection identifier, an import identifier, a congestion start time, a congestion end time, a duration, and an alarm state. The continuous alarm is an alarm of a road section in the primary screening set that exists in the real alarm set of each road section of the target road in the previous period. The suspected disappeared alarm is an alarm of a road section that exists in the real alarm set of each road section of the target road in the previous period but does not exist in the primary screening set. The new alarm is an alarm of a road section that does not exist in the real alarm set of each road section of the target road in the previous period but exists in the primary screening set. According to the continuous alarm and the suspected disappeared alarm, a set of continuous alarms with a duration less than a preset duration threshold and a set of suspected disappeared alarms are obtained. The set of continuous alarms with a duration less than the preset duration threshold is removed from the primary screening set, the set of suspected disappeared alarms is added to the primary screening set, and the new alarm is retained in the primary screening set to obtain a congestion alarm analysis list.

[0011] Optionally, in some embodiments of the present application, the S240 specifically includes: Obtain a suspected congestion duration of each road section in the congestion alarm analysis list. Sort the congestion alarm analysis list according to three dimensions of the suspected duration, traffic saturation, and speed to obtain a congestion alarm analysis list after priority sorting. Specifically comprising: sorting each road segment in the congestion alarm analysis list by length to short according to the suspected duration; if the suspected duration is the same, sorting each road segment by traffic saturation from large to small; if the traffic saturation is the same, sorting each road segment by speed from low to high.

[0012] Optionally, in some embodiments of the present application, the S300 comprises: S310, obtaining the alarm type of each road segment in the congestion alarm analysis list after priority sorting; the alarm type comprises a new alarm, a continuous alarm and a suspected disappearance alarm; When the alarm type is a new alarm, obtaining the complete real-time video stream of the corresponding road segment; When the alarm type is a continuous alarm or a suspected disappearance alarm, extracting key frames of the real-time video stream of the corresponding road segment at a preset fixed interval to obtain a key frame sequence; S320, inputting the complete real-time video stream or the key frame sequence into the pre-set visual large model to obtain a false alarm set and a final congestion alarm set.

[0013] Optionally, in some embodiments of the present application, the visual large model is a deep learning model that fuses time sequence and spatial perception, comprising: An input preprocessing layer is used to decode the received real-time video stream into an image frame sequence, or directly receive the key frame sequence extracted by the upstream link; the sequence is subjected to size normalization, frame rate adjustment and tensor formatting processing to generate batch data meeting the model input requirements; A space-time feature encoding layer is used to extract a joint feature representation that fuses space-time dimensions from the batch data to obtain fused space-time features; the space-time feature encoding layer comprises a spatial feature extraction node and a time sequence feature modeling node; A scene understanding and decision layer is used to perform high-level semantic reasoning and state judgment based on the fused space-time features, comprising a traffic state classification node and a confidence evaluation node; the traffic state classification node adopts a fully connected neural network combined with a softmax activation function to output a traffic state classification result of each road segment, and the confidence evaluation node is used to output a confidence score corresponding to the classification result; The classification result comprises: free flow, slow traffic, congestion and severe congestion; An output layer is used to generate a false alarm set and a final congestion alarm set according to a predetermined decision rule.

[0014] Optionally, in some embodiments of the present application, the output layer specifically comprises: The road segment with a classification result of congestion or severe congestion and a confidence score higher than a preset high confidence threshold is included in the final congestion alarm set; Road segments classified as free-flowing or slow-moving, or those that are congested but have a confidence score below a preset low confidence threshold, are included in the false alarm set.

[0015] Optionally, in some embodiments of this application, the method further includes: S500. Based on the false alarm set, update the historical false alarm information, and based on the updated historical false alarm information, periodically filter the data in the false alarm set as the training set to perform incremental training and optimization on the visual large model.

[0016] (III) Beneficial Effects The congestion alarm method based on secondary verification using a large visual model provided in this application integrates data analysis with a secondary verification mechanism using a large visual model. After initial congestion alarm identification, the large visual model is introduced to effectively filter false alarms, improving the authenticity and reliability of traffic congestion identification. Furthermore, this application only calls the large visual model for verification on the priority-ranked list of congestion alarms to be analyzed, avoiding the enormous computational consumption of full-network, all-time visual processing and improving overall operational efficiency. Moreover, this application combines multi-source traffic data with visual feature analysis, making congestion identification more adaptable to different scenarios and able to cope with interference factors such as complex weather and special traffic behaviors, thus improving the stability of the identification. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a congestion alarm method based on secondary verification of a large visual model according to an embodiment of this application. Figure 2 This is a flowchart illustrating the process of generating a priority-sorted list of congestion alarms to be analyzed in step S200 of a congestion alarm method based on visual large model secondary verification according to an embodiment of this application. Detailed Implementation

[0018] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] Currently, in the congestion alarm technology, the basic data collection mainly relies on three means: one is the fixed detector, which is used to obtain the real-time speed, flow and the like of the road section; the other is the floating car GPS data, which can be used to infer the road section passing state through the vehicle trajectory; and the third is the video monitoring equipment, which is only used for post-tracing and does not participate in real-time alarm judgment. In the existing technology, the preliminary congestion identification mostly adopts the threshold judgment method, that is, the preset speed, flow and the like are used to determine whether the road section is in a congestion state. However, with the increasing complexity of the traffic scene, such as temporary construction, pedestrian crossing, and vehicle flow tide change, the traditional alarm technology gradually exposes many defects and cannot meet the fine management needs.

[0020] In order to improve the accuracy and reliability of the congestion alarm, it is necessary to introduce a more intelligent, diversified and multi-modal judgment mechanism to make up for the shortcomings of the traditional scheme in the complex traffic environment.

[0021] Therefore, in order to solve the problem of high false alarm rate caused by environmental interference in the traffic congestion alarm method based on data analysis in the prior art, and the problem of response delay, waste of computing power and environmental adaptability caused by the response delay and waste of computing power of the visual large model full-time verification scheme, the application provides a congestion alarm method based on secondary verification of a visual large model.

[0022] The congestion alarm method based on secondary verification of a visual large model of the embodiment first constructs a preliminary alarm mechanism based on multi-source traffic data, performs real-time monitoring and screening on the whole road network, identifies road sections that may exist congestion, and generates a candidate list to be verified; then, a visual large model is called to perform high-precision semantic analysis on real-time video and perform secondary confirmation on the congestion state; finally, under the premise of ensuring the authenticity and effectiveness of the alarm event, a formal congestion alarm is pushed to the traffic management system, and false alarm samples are recorded for system adaptive optimization. The method can not only ensure the alarm accuracy, but also effectively reduce the calling frequency of the visual large model, avoid the waste of computing resources and the problem of response delay caused by the full-time visual verification, provide a new technical path for the efficient deployment and precise management of the urban traffic intelligent perception system, and has good application prospect and popularization value.

[0023] In order to better understand the above technical solutions, the exemplary embodiments of the application will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the application are shown in the drawings, it should be understood that the application can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a clearer, more thorough understanding of the application and to convey the complete scope of the application to those skilled in the art.

[0024] Embodiment 1

[0025] Figure 1A flowchart of a congestion alarm method based on visual large model secondary verification according to an embodiment of the present application is shown in FIG. 1. The congestion alarm method includes the following steps. Figure 1 Step S100, periodically acquire traffic parameters of each road segment of the target road and the congestion alarm state of the last period; according to the traffic parameters and the congestion alarm state of the last period, perform preliminary congestion state identification to generate a congestion alarm candidate set; The traffic parameters include speed, flow, and capacity. The capacity refers to the maximum number of vehicles that can pass through a lane or a road segment within a certain time under specific traffic and control conditions.

[0026] Specifically, the above step S100 includes the following steps. According to the traffic parameters, acquire the traffic saturation of each road segment of the target road; the traffic saturation is the ratio of the flow per unit time to the capacity. If the congestion alarm state of the last period is not alarmed, when the speed is less than a preset speed threshold and the saturation is greater than a preset saturation threshold, it is determined as suspected congestion; if the congestion alarm state of the last period is already alarmed, when the speed is less than a preset speed threshold, it is determined as suspected congestion. Acquire the basic information of each road segment determined as suspected congestion to generate a congestion alarm candidate set; the basic information of each road segment includes the road segment identifier, the intersection identifier, the import identifier, the suspected congestion start time, and the corresponding traffic parameters, for example, as shown in Table 1: Table 1: Example of road segment basic information

[0027] The preset speed threshold is obtained according to the following steps: Obtain the historical traffic flow data of the target road, perform traffic state clustering analysis on each road segment under different road levels, and obtain the free flow speed and critical jam density of each road segment. According to the free flow speed and critical jam density of each road segment, obtain the critical speed from free flow to congestion transition through the Greenshields model. The preset speed threshold is a times the critical speed, and a is obtained according to different road levels; the road levels include expressway, expressway, main road, and secondary road.

[0028] In order to obtain an accurate preset speed threshold, first, the historical traffic flow data of the target road needs to be obtained, which covers a complete period of 12 months, including morning and evening peak, flat peak, low valley, and special periods of holidays, and covers different weather scenes such as sunny, rainy, snowy, and foggy, to ensure the spatiotemporal completeness of the samples.

[0029] ​After obtaining sufficient historical data, the "speed, flow and capacity" in the historical traffic flow data are taken as the three-dimensional feature vector. First, the three-dimensional feature vector is reduced to two-dimensional space by principal component analysis, and then the Gaussian mixture model is used to cluster the historical traffic flow data.

[0030] It should be noted that, due to the significant differences in design standards, speed limits, traffic composition and service objects of different types of roads, their traffic operation characteristics are also completely different, so it is necessary to model by grade, that is, to classify all observation points according to the functional grade of the road-highway, expressway, main road and secondary road. The goal of clustering is to identify typical traffic operation modes, such as "free flow", "steady flow", "transition zone" and "congestion flow" state clusters. By statistical analysis of the center value and distribution range of each cluster, the typical average speed of each road section in the "free flow" state, i.e. the free flow speed, can be accurately extracted, and the key nodes where the traffic flow changes from order to chaos can be identified, and the critical jam density, i.e. the spatial density of vehicles when the traffic begins to slow down and queue to sprout, can be determined.

[0031] Further, the classic Greenshields linear model is introduced to establish the linear relationship between speed and capacity, i.e. density, as shown in the following formula: V=V f (1-k / k j ); where k j is the jam density, i.e. the maximum density when vehicles are completely stopped, and k is the critical jam density, V f is the free flow speed, and V is the critical speed.

[0032] Finally, in order to enhance the robustness of congestion determination and reserve a reasonable buffer interval, the preset speed threshold actually used for congestion identification is set to α times the critical speed, and the adjustment coefficient α is dynamically adjusted according to the road grade, reflecting the sensitivity difference of different roads to congestion response. For example, the speed of highway is fast and the traffic continuity is strong, which is very sensitive to speed drop, so the value of α is small, taking 0.75; the expressway has entrance and exit control but has ramp interlacing, so α takes 0.8; the main road has frequent signal control and large traffic fluctuation, so α takes 0.85; the secondary road is disturbed by pedestrians and non-motor vehicles, so the natural speed fluctuation is large, and α takes a higher value of 0.90.

[0033] The preset speed threshold determined by the above method not only has a solid theoretical basis, but also can be adjusted adaptively according to the actual operation characteristics of different roads, thereby significantly improving the accuracy, adaptability and practicality of the congestion identification algorithm.

[0034] In the above step S100, the periodic monitoring and preliminary congestion identification of the traffic state of each road section of the target road are intended to be achieved. The process first collects the real-time traffic parameters of each road section in the target road at a fixed time period, such as every 5 minutes or 15 minutes, including the average speed of vehicles, the traffic flow per unit time, and the capacity of the road, while obtaining the congestion alarm state of the road section in the last period. Based on these data, the system will make a preliminary congestion state judgment and generate a "congestion alarm candidate set" for further analysis and confirmation.

[0035] Specifically, the "traffic saturation" in the traffic parameters is an important indicator for judging traffic pressure, which is defined as the ratio of the actual traffic flow per unit time to the designed or measured capacity of the road, reflecting the utilization degree of road resources. When the saturation is close to or exceeds 1, it indicates that the road is close to or exceeds the carrying limit. On this basis, the congestion determination logic adopts a differentiated strategy according to the alarm state of the last period: if the congestion alarm is not triggered in the last period, two conditions must be met to determine "suspected congestion" - the current average speed of vehicles is lower than the preset speed threshold, and the traffic saturation is higher than the preset saturation threshold. However, if the last period is in a congestion alarm state, the determination condition is more relaxed: only the current speed lower than the preset speed threshold can continue to determine "suspected congestion", which reflects the persistence and inertia characteristics of the congestion state, preventing frequent state jumps caused by short-term traffic fluctuations.

[0036] Further, the calculation of the preset saturation threshold also introduces a dynamic correction mechanism, which decreases according to the road grade and automatically scales according to the time period.

[0037] Through the above double determination criteria, false positives caused by temporary speed reduction or short-term traffic fluctuations can be effectively avoided, thereby improving the accuracy and adaptability of congestion identification.

[0038] Step S200, based on the historical false alarm information and historical disposal information of each road section of the target road, screening and priority sorting the congestion alarm candidate set to generate a priority-sorted congestion alarm analysis list; Specifically, referring to Figure 2 , step S200 includes: Step S210, obtaining the historical false alarm information and historical disposal information of each road section of the target road; The historical false alarm information includes false alarm records and false alarm times in the historical period, and the historical disposal information includes disposal records and disposal times in the historical period; For example, the historical false alarm information is shown in Table 2: Table 2: Example of historical false alarm information

[0039] The historical treatment information is as shown in Table 3: Table 3: Example of historical treatment information

[0040] In the specific implementation process, the acquisition of the historical false alarm information and the historical treatment information needs to establish a technical system of "multi-source data linkage + standardized storage + time effectiveness filtering", to ensure the completeness, accuracy and usability of the data. In terms of data source, the historical false alarm information is mainly taken from the past alarm verification log: on the one hand, the manual correction record of the traffic management platform is associated, and on the other hand, the false alarm record marked after the secondary verification of the visual large model is integrated; the historical treatment information is connected with the work order management system of the traffic command center and the mobile law enforcement terminal of the traffic police, and the core data includes the unique work order number of the treatment event, the corresponding congestion intersection identifier, the treatment start time, the treatment completion time, the treatment type and the treatment result feedback, and at the same time, the traffic parameter change curve during the treatment needs to be synchronously associated, to provide data support for subsequent screening.

[0041] After obtaining the historical false alarm information and the historical treatment information, the information needs to be preprocessed, and three layers of quality control mechanisms are adopted to ensure the effectiveness of the data: first, the field completeness check is adopted to eliminate invalid records such as "missing intersection identifier" and "abnormal timestamp", to ensure the completeness of the data; second, the duplicate data is removed, and for multiple false alarm or treatment records of the same road section in the same period, the "most recent time + most complete information" principle is adopted to retain the optimal data; third, the time effectiveness filtering is adopted, considering the dynamic changes of the traffic scene, the historical data of the last 3 months is retained, and special scene data such as holidays and extreme weather is marked separately, to avoid the interference of the data on the regular screening rules. Finally, all the data is stored in a time sequence database, to support the quick query according to the time window, the intersection identifier and other dimensions.

[0042] In the above step S210, through the "double-source false alarm data integration + three-layer quality control", the completeness and accuracy of the historical information are ensured, and combined with the "time sequence database classification storage + time effectiveness filtering", the data query efficiency is improved, to provide a reliable data basis for subsequent screening.

[0043] In step S220, according to the historical false alarm information and the historical treatment information, the road sections with false alarm records in the preset false alarm time window are screened out as a first congestion set, and the road sections with completed treatment in the preset treatment time window are screened out as a second congestion set, the first congestion set and the second congestion set are removed from the congestion alarm candidate set, to obtain a primary screening set. Specifically, the preset false alarm time window and the preset treatment event window in step S220 are both dynamic time windows. First, for the dynamic setting of the preset time window, different road levels are configured differently. For example, the expressway is set to 30 minutes because of strong traffic flow and short false alarm impact period, the fast road and the main road are set to 45 minutes because of signal control and more pedestrian interference, and the secondary road is set to 60 minutes because of high mixed traffic proportion and repeated false alarms. The treatment time window is adjusted according to the treatment type. For example, the time window is set to 20 minutes for fast treatment types such as “police diversion” and “temporary signal adjustment”, and the time window is extended to 30 minutes for long-time treatment types such as “accident cleanup” and “fault vehicle towing”. Meanwhile, dynamic optimization according to historical treatment efficiency is supported. For example, if multiple accident cleanups on a road segment are completed within 25 minutes, the time window is automatically reduced to 25 minutes.

[0044] Optionally, in the collection judgment and elimination link, the screening logic needs to be executed in two steps: the first step is the judgment of the first congestion set. In addition to the basic condition that “there is a false alarm record within the preset false alarm time window”, the false alarm reason needs to be further refined. For example, if the false alarm reason is device fluctuation and it has been restored to normal according to the period, the road segment is not included in the first congestion set. If the false alarm reason is “temporary event”, such as temporary parking on the roadside, and there is still a temporary event warning in the current period, it is preferred to be included in the first congestion set to avoid repeated false alarms. The second step is the judgment of the second congestion set, which needs to meet the conditions that “the treatment order status is “case closed” within the preset treatment time window” and “the traffic parameters meet the standards after treatment”. If only the treatment order is closed but the traffic parameters do not meet the standards, the road segment is not included in the second congestion set to avoid false elimination of effective congestion.

[0045] Further, in the elimination execution and result verification link, the “first judgment-then review” process is adopted: first, according to the above rules, the road segments of the first congestion set and the second congestion set are batch eliminated from the congestion alarm candidate set to generate the initial primary screening set; then, the elimination result is sampled and reviewed to correct the screening rule deviation by comparing the real-time video stream of the current period; at the same time, if the size of the primary screening set is too small, such as the remaining number is less than 30% of the original candidate set after elimination, the time window is automatically adjusted, and the screening is re-executed to ensure that enough road segments are retained after screening to avoid missing effective congestion due to excessive screening. The final primary screening set needs to record the “screening state” of each road segment to provide a basis for subsequent priority sorting and tracing.

[0046] The above-mentioned double filtering mechanism based on historical false alarms and treatment feedback not only enhances the self-learning and anti-interference ability of the scheme, but also realizes the closed-loop linkage of alarm logic and actual management process, laying a solid foundation for subsequent priority sorting and final alarm confirmation.

[0047] Step S230, obtaining the real alarm set of each road section of the target road in the last period, comparing the primary screening set with the real alarm set of each road section of the target road in the last period, screening the primary screening set, and generating a congestion alarm analysis list; The real alarm set of each road section of the target road in the last period includes the road section identifier, the intersection identifier, the import identifier, the congestion start time, the congestion end time, the duration, and the alarm state. Step S230 includes: Comparing the primary screening set with the real alarm set of each road section of the target road in the last period, identifying the continuous alarm, the new alarm, and the suspected disappearance alarm. The continuous alarm is the alarm of the road section in the primary screening set existing in the real alarm set of each road section of the target road in the last period. The suspected disappearance alarm is the alarm of the road section existing in the real alarm set of each road section of the target road in the last period but not existing in the primary screening set. The new alarm is the alarm of the road section not existing in the real alarm set of each road section of the target road in the last period but existing in the primary screening set. According to the continuous alarm and the suspected disappearance alarm, a set of the continuous alarm with a duration less than a preset duration threshold and a set of the suspected disappearance alarm are obtained, the set of the continuous alarm with a duration less than the preset duration threshold is removed from the primary screening set, the set of the suspected disappearance alarm is added to the primary screening set, and the new alarm is reserved in the primary screening set, to obtain the congestion alarm analysis list.

[0048] Specifically, the embodiment matches the "primary screening set" of the current period with the "real alarm set" confirmed in the last period item by item. Each record in the real alarm set contains complete space-time and state information: the road section identifier, the intersection identifier, the import direction identifier, the congestion start time, the congestion end time (if released), the actual duration, and the final confirmed alarm state (such as "in progress" or "released"). Through accurate matching based on the road section-intersection-import three-tuple, the following three types of key alarm state change types can be identified: Continuous alarm: refers to the road section alarm that has been confirmed as a real congestion in the last period and appears again in the primary screening set of the current period. This type of alarm indicates that the congestion state is still continuing, and belongs to a high-priority attention object.

[0049] New alarm: refers to a suspected congestion event that does not exist in the real alarm set of the last period but first appears in the primary screening set of the current period. This type of alarm represents a newly occurring traffic anomaly, which may be caused by a sudden accident, temporary regulation, or a surge in traffic demand.

[0050] Suspected disappearance alarm: refers to the alarm record that exists in the real alarm set of the last period, but is not re-identified in the primary screening set of the current period. This kind of situation may mean: (a) the congestion has naturally dissipated or has been effectively disposed of; (b) the traffic parameter temporarily rises, causing the system not to trigger an alarm; (c) the detection equipment is abnormal or data is missing. In order to prevent false judgment that the congestion has been lifted due to short-time data fluctuation, it is marked as "suspected disappearance" instead of directly closing the alarm.

[0051] After the above state recognition is completed, the primary screening set is dynamically reconstructed according to the preset logical rules to generate the final "congestion alarm to be analyzed list", which specifically includes: For the sustained alarm, it is further judged whether the accumulated duration is less than a preset duration threshold. If the duration is too short, it means that the congestion is still in the embryonic stage, and its stability and severity need to be observed. In order to avoid over-response to temporary fluctuations, the alarm with "insufficient duration" is temporarily excluded from the primary screening set. After exclusion, the sustained alarm with insufficient duration is directly included in the final congestion alarm set, waiting for the final alarm.

[0052] For the suspected disappearance alarm, it is added back to the primary screening set to ensure the continuous tracking of the occurred congestion event and avoid false judgment of complete recovery due to normal data.

[0053] For the new alarm, it is retained in the primary screening set as a new event that needs to be checked.

[0054] Finally, after the above "exclusion of temporary sustained items, retention of new items, and backfilling of suspected disappearance items", the congestion alarm to be analyzed list is generated. The list not only contains the current suspected congestion points, but also integrates the continuation information and dissipation trend of historical alarms, providing a comprehensive, dynamic and logically consistent data basis for subsequent priority sorting, multi-source data fusion verification and command and dispatch decision, significantly improving the intelligent level and response efficiency of traffic event management.

[0055] In the embodiment, step S230 further introduces a dynamic comparison mechanism in the time dimension on the basis of the primary screening set, and identifies the evolution trend of the traffic congestion state by fine comparison with the "real alarm set" generated in the last period, so as to optimize the alarm candidate list of the current period. This process not only focuses on "whether it is congested now", but also focuses on "how the congestion changes", realizes the tracking and logical correction of the congestion time life cycle, and effectively improves the continuity, stability and decision support ability of the congestion alarm.

[0056] Step S240 sorts the congestion alarm to be analyzed list according to the multi-dimensional hierarchical sorting rule set in advance to obtain the congestion alarm to be analyzed list after priority sorting.

[0057] For example, the basic information of a congestion road section in the congestion alarm to be analyzed list is shown in Table 4: Table 4 Basic information of a congestion road section

[0058] Step S240 specifically includes: obtaining the suspected congestion duration of each road section in the congestion alarm to be analyzed list; Specifically, the method for obtaining the suspected congestion duration is as follows: New alarm duration calculation: taking the "time when the suspected congestion condition is first met in the current period" as the starting point, if the current period has not ended, the duration is calculated as "current time-starting time"; if the current period has been completely run and the suspected congestion condition is met throughout the period, the duration is calculated as the period, and it is also necessary to check whether there is a "cross-period precursor" for the road section. If the speed is lower than the threshold value for 3 minutes at the end of the last period, the duration of this part is added to avoid missing the congestion accumulation at the junction of the periods. Continuous alarm duration calculation: starting from the "starting time of the congestion in the last period when the real alarm is triggered", the complete duration of the current period is added, and it is also necessary to exclude the "interruption period of the congestion in the last period", and only the actual congestion period is accumulated; for example, the duration of the last period is 4 minutes, and the current period has been running for 4 minutes, so the total duration is 4+4=8 minutes.

[0059] Suspected disappearance alarm duration calculation: taking the "starting time of the congestion in the last period when the real alarm is triggered" as the starting point, the total duration to the "starting time of the current period" is calculated.

[0060] If there is a missing duration data, such as device failure leading to incomplete starting time record, the historical analogy method is used to complete the missing data.

[0061] According to the suspected duration, traffic saturation and speed, the congestion alarm to be analyzed list is sorted to obtain the congestion alarm to be analyzed list after priority sorting; Specifically, each road section in the congestion alarm to be analyzed list is sorted according to the suspected duration from long to short; if the suspected duration is the same, each road section is sorted according to the traffic saturation from large to small; if the traffic saturation is the same, each road section is sorted according to the speed from low to high.

[0062] Specifically, a three-level progressive sorting logic is used for priority sorting: First, take the "suspected duration" as the highest priority dimension, arrange all road segments in the list in descending order according to their duration from large to small. This ensures that those long-term congestion points that have not been alleviated are always at the forefront of the alarm queue, avoiding being ignored due to transient index fluctuations, and reflecting the focus on "stubborn congestion" or "persistent congestion caused by major events".

[0063] When the suspected duration of multiple road segments is equal, the second level of judgment criteria-traffic saturation, that is, the ratio of unit time flow to road capacity, is introduced. Saturation directly reflects the overload of road resources: saturation greater than 1 indicates that demand exceeds supply, and queuing has occurred; the closer or exceeds the threshold value of 1.2, 1.5, etc., the greater the traffic pressure and the higher the risk of collapse. Therefore, in the case of the same duration, the road segment with higher saturation means more serious traffic load and should have higher priority for disposal.

[0064] When the duration and saturation of two or more road segments are the same, further introduce the average speed of vehicles as the final winning dimension, and arrange them in ascending order according to the speed value from low to high. Speed is the most intuitive reflection of traffic running state, and extremely low speed usually means that vehicles are in a stop-and-go or even stagnant state, with poor user experience, significant increase in emissions and safety hazards.

[0065] Through the above three-level hierarchical sorting strategy of "duration → saturation → speed", the congestion event is finely graded. This sorting method takes into account the cumulative effect of the time dimension, and combines the utilization intensity and running quality of the space resource, avoiding misjudgment caused by a single indicator. For example, a road segment with a duration of 5 minutes, a saturation of 1.6, and an average speed of only 8 km / h will be responded to in priority over a road segment with a duration of 3 minutes, a saturation of 1.4, and a speed of 15 km / h.

[0066] Step S300, acquire real-time video stream corresponding to each road segment in the congestion alarm list to be analyzed, input the real-time video stream into the visual large model set in advance, and acquire a false alarm set and a final congestion alarm set; Step S300 includes: Step S310, acquire the alarm type of each road segment in the congestion alarm list to be analyzed after priority sorting; the alarm type includes new alarm, continuous alarm, and suspected disappearance alarm; When the alarm type is a new alarm, acquire the complete real-time video stream corresponding to the road segment; When the alarm type is a continuous alarm or a suspected disappearance alarm, extract key frames of the real-time video stream of the corresponding road segment at a preset fixed interval to acquire a key frame sequence; Optionally, when extracting the key frames of the real-time video stream of the corresponding road segment, the following method is adopted: To improve the accuracy and scene adaptation ability of key frame extraction, and break through the limitations of traditional fixed interval sampling. Through the training of LSTM time series prediction model based on historical traffic flow data, the model has the ability to learn the periodicity of traffic state and the propagation characteristics of sudden disturbance, and can predict the congestion development trend in the next 10 seconds in real time, identify the key state turning points that will occur, such as congestion intensification or congestion beginning to ease, etc. Actively encrypt sampling 3 seconds before, at and after the predicted turning point, forming a "front-middle-back" three-frame sequence, accurately capturing the state transition moment, for example, in continuous alarm, it is predicted that the vehicle flow will disperse after the traffic light turns green, so the key frames before and after the green light are extracted, avoiding missing the core changes of fixed interval.

[0067] Further, to deal with unpredictable sudden events such as traffic accidents, vehicle breakdowns, and pedestrian intrusions, a multi-source heterogeneous data fusion trigger mechanism is introduced to deeply bind key frame extraction with external sensor events. By real-time access and analysis of multi-dimensional information streams from microwave radar, floating car GPS data, and signal light controllers, a real-time event monitoring layer is constructed. When any sensor detects the following high-value events, "event anchor point forced frame extraction" is triggered immediately: Radar detects that the acceleration of multiple vehicles in a lane has suddenly decreased; Floating car data shows that the average speed of a road segment has decreased by more than 60% in 10 seconds; The precise switching time of the signal light state from red to green or green to red, etc.

[0068] Once triggered, the video frame at the time of the event is immediately extracted, and the previous 2 seconds and the next 2 seconds before the event are automatically backtracked and extracted, forming a 5-frame short sequence, and giving 3 times the weight priority to retain, ensuring that the key frame is accurately bound to the traffic event, solving the problem of missing sudden events in traditional sampling.

[0069] The above key frame extraction method upgrades the key frame from "passive interval extraction" to "active capture of core features", providing highly structured and semantically rich inputs for visual large model verification, greatly improving the timeliness and reliability of congestion state recognition.

[0070] Step S320, input the complete real-time video stream or key frame sequence into the visual large model set in advance, and obtain the false alarm set and the final congestion alarm set.

[0071] The visual large model is a deep learning model that integrates time series and spatial perception, including: The input preprocessing layer is used to decode the received real-time video stream into an image frame sequence, or directly receive the key frame sequence extracted by the upstream link; the sequence is normalized in size, adjusted in frame rate, and formatted in tensor, generating batch data that meets the model input requirements; a spatio-temporal feature encoding layer configured to extract a joint feature representation of fused spatio-temporal dimensions from the batch data and obtain fused spatio-temporal features; the layer includes a spatial feature extraction node and a time series feature modeling node; the spatial feature extraction node is based on a deep convolutional neural network or a visual Transformer architecture, and is configured to extract spatial features in the image frame by frame; the time series feature modeling node is based on a time series convolutional network or a recurrent neural network architecture, and is configured to model a feature sequence of consecutive frames and capture dynamic change features of the traffic flow; a scene understanding and decision layer configured to perform high-level semantic reasoning and state determination based on the fused spatio-temporal features, including a traffic state classification node and a confidence score evaluation node; the traffic state classification node adopts a fully connected neural network combined with a softmax activation function, and outputs a traffic state classification result of each road segment; the confidence score evaluation node is configured to output a confidence score corresponding to the classification result; the classification result includes free flow, slow traffic, congestion, and severe congestion; an output layer configured to generate a false alarm set and a final congestion alarm set according to a predetermined decision rule.

[0072] the output layer specifically includes: a road segment with a classification result of congestion or severe congestion and a confidence score higher than a preset high confidence threshold is classified into the final congestion alarm set; a road segment with a classification result of free flow or slow traffic, or a classification result of congestion but a confidence score lower than a preset low confidence threshold is classified into the false alarm set.

[0073] In summary, step S320 realizes a leap from "data perception" to "semantic understanding" through deep fusion of artificial intelligence technologies, greatly improves the accuracy and intelligence level of congestion identification, and provides core technical support for building an efficient, reliable, and sustainable urban traffic management system.

[0074] Step S400: According to the final congestion alarm set and a preset alarm rule, congestion alarm is performed.

[0075] The congestion alarm method based on the secondary verification of the visual large model of the embodiment further includes: Step S500: According to the false alarm set, historical false alarm information is updated, and based on the updated historical false alarm information, data in the false alarm set is periodically screened as a training set to perform incremental training and optimization on the visual large model.

[0076] In addition, after obtaining the false alarm set, the false alarm set is analyzed, high-frequency false alarm road segments are identified, suspected data anomalies are reminded, and manual data and equipment checking is suggested, specifically including: The false positive data in the false positive set is analyzed, all false positive events are preliminarily classified according to false positive reasons, including data fluctuation type false positives (such as false positives caused by instantaneous flow drop or abnormal vehicle speed), device abnormality type false positives (such as sensor failure, network delay, etc.), scene interference type false positives (such as temporary construction, non-motor vehicle gathering, etc.), algorithm misjudgment type false positives (such as improper threshold setting of data analysis model, etc.). Based on the above classification, a time series statistical model of false positive events is established, and the false positive frequency, false positive periodicity characteristics and their distribution law in different time periods of each road section are calculated.

[0077] The false positive frequency of each road section is ranked by using statistical methods (such as chi-square test, sliding window mean comparison, etc.), and "high frequency false positive road sections" with false positive rate significantly higher than the average level are identified. For such road sections, warning information is generated and pushed to operation and maintenance personnel to remind suspected data anomalies and suggest manual data and device checking.

[0078] For each false positive event, the associated visual data is extracted, and the false positive reason is labeled by manual or semi-automatic method combined with the semantic analysis result of the visual large model (such as "no obvious congestion", "non-motor vehicle gathering", "pedestrian crossing influence", etc.). Subsequently, these labeled false positive cases are added to the training set of the visual model to form a multi-class sample library containing real congestion scenes and false positive scenes for subsequent model incremental learning and optimization.

[0079] The congestion alarm method based on the visual large model secondary verification of the embodiment first introduces high-precision semantic analysis based on the preliminary identification of congestion alarm through the fusion of data analysis and visual large model secondary verification mechanism, effectively filters false positive events, and improves the authenticity and reliability of traffic congestion identification. Secondly, the visual large model is called only for the suspected congestion areas preliminarily screened, avoiding the huge algorithm consumption caused by full road network and full period visual processing, and improving the overall operation efficiency of the system. Further, the congestion identification has better scene adaptability by combining multi-source traffic data and visual feature analysis, which can cope with complex weather, special traffic behavior and other interference factors, and improves the system stability. In addition, the application records, statistics and analysis of the false positive set, identifies high frequency false positive road sections and reminds manual checking, and puts the false positive set into the visual large model training set to continuously optimize the model performance, forming a complete closed loop of "identification-verification-feedback-optimization". The congestion alarm method based on the visual large model secondary verification of the application considers real-time, accuracy and resource utilization, and is suitable for intelligent monitoring deployment in large-scale urban road network, and provides strong support for traffic scheduling, emergency management and infrastructure optimization.

[0080] In the description of the application, it should be understood that the terms "first", "second" are used only for descriptive purposes and are not to be construed as indicating or implying relative importance or an indicated number of technical features. Therefore, the features defined as "first", "second" can explicitly or implicitly include one or more of the features. In the description of the application, the meaning of "a plurality of" is two or more, unless otherwise explicitly specified and limited.

[0081] In the present application, unless otherwise explicitly specified and limited, the terms "mounting", "connecting", "connecting", "fixing" and the like should be understood broadly, for example, it can be fixed connection, or detachable connection, or integrated; it can be mechanical connection, or electrical connection; it can be directly connected, or indirectly connected through intermediate medium; it can be the internal communication of two elements or the interaction relationship between two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0082] In the present application, unless otherwise explicitly specified and limited, the first feature is "on" or "under" the second feature, which can be direct contact between the first and second features, or indirect contact between the first and second features through intermediate medium. Moreover, the first feature "above", "over" and "on" the second feature, can be the first feature directly above or obliquely above the second feature, or only indicate that the first feature is higher than the second feature in horizontal height. The first feature "below", "under" and "under" the second feature, can be the first feature directly below or obliquely below the second feature, or only indicate that the first feature is lower than the second feature in horizontal height.

[0083] In the description of the present application, the description of the terms "one embodiment", "some embodiments", "embodiment", "example", "specific example" or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In the present application, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine the different embodiments or examples described in the present application and the features of the different embodiments or examples, without contradiction.

[0084] Although the embodiments of the present application have been shown and described above, it should be understood that the above embodiments are exemplary and cannot be construed as limiting the present application, and those skilled in the art can modify, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1.A congestion alarm method based on visual large model secondary verification, characterized in that, The method comprises the following steps: S100, acquiring traffic parameters of each road section of a target road and congestion alarm states of a previous period; According to the traffic parameters and the congestion alarm states of the previous period, a preliminary congestion state identification is performed to generate a congestion alarm candidate set; The traffic parameters include speed, flow and traffic capacity; S200, based on historical false alarm information and historical disposal information of each road section of the target road, the congestion alarm candidate set is screened and prioritized to generate a congestion alarm to-be-analyzed list after priority sorting; S300, real-time video streams corresponding to each road section in the congestion alarm to-be-analyzed list are acquired, the real-time video streams are input into a preset visual large model, a false alarm set and a final congestion alarm set are acquired; S400, according to the final congestion alarm set and a preset alarm rule, a congestion alarm is performed. 2.The congestion alarm method based on visual large model secondary verification according to claim 1, wherein, The S100 comprises the following steps: According to the traffic parameters, traffic saturation degrees of each road section of the target road are acquired; the traffic saturation degree is a ratio of flow per unit time to traffic capacity; If the congestion alarm state of the previous period is no alarm, when the speed is less than a preset speed threshold and the saturation degree is greater than a preset saturation degree threshold, it is determined that there is suspected congestion; if the congestion alarm state of the previous period is alarm, when the speed is less than the preset speed threshold, it is determined that there is suspected congestion; Basic information of each road section determined to have suspected congestion is acquired to generate a congestion alarm candidate set; the basic information of each road section includes road section identification, intersection identification, import identification, suspected congestion start time and corresponding traffic parameters. 3.The congestion alarm method based on visual large model secondary verification according to claim 2, characterized in that, The preset speed threshold is acquired according to the following steps: Historical traffic flow data of the target road are acquired, traffic state clustering analysis is performed on each road section under different road levels, free flow speeds and critical jam densities of each road section are acquired; According to the free flow speeds and the critical jam densities of each road section, a critical speed from free flow to congestion transition is acquired through a Greenshields model; The preset speed threshold is α times of the critical speed, and α is acquired according to different road levels; the road levels include expressway, express road, main road and secondary road. 4.The congestion alarm method based on visual large model secondary verification according to claim 1, wherein, The S200 comprises the following steps: S210, historical false alarm information and historical disposal information of each road section of the target road are acquired; The historical false alarm information includes false alarm records and false alarm times in a historical period, and the historical disposal information includes disposal records and disposal times in a historical period; S220, according to the historical false alarm information and the historical disposal information, a road section having a false alarm record in a preset false alarm time window is screened out as a first congestion set, and a road section having completed disposal in a preset disposal time window is screened out as a second congestion set, the first congestion set and the second congestion set are removed from the congestion alarm candidate set to acquire a primary screening set; S230, a real alarm set of each road section of the target road in the previous period is acquired, the primary screening set is compared with the real alarm set of each road section of the target road in the previous period, the primary screening set is screened to generate a congestion alarm to-be-analyzed list; S240. According to the multi-dimensional hierarchical sorting rule set in advance, the congestion alarm to-be-analyzed list is sorted to obtain a congestion alarm to-be-analyzed list after priority sorting. 5.The congestion alarm method based on visual large model secondary verification according to claim 4, characterized in that, The S230 includes: The primary screening set is compared with a real alarm set of each road section of the target road in the last period to identify a continuous alarm, a new alarm, and a suspected disappearance alarm; the real alarm set of each road section of the target road in the last period includes a road section identifier, an intersection identifier, an import identifier, a congestion start time, a congestion end time, a duration, and an alarm state; The continuous alarm is an alarm of a road section in the primary screening set that exists in the real alarm set of each road section of the target road in the last period; the suspected disappearance alarm is an alarm of a road section that exists in the real alarm set of each road section of the target road in the last period but does not exist in the primary screening set; and the new alarm is an alarm of a road section that does not exist in the real alarm set of each road section of the target road in the last period but exists in the primary screening set; According to the continuous alarm and the suspected disappearance alarm, a set of the continuous alarms with a duration less than a preset duration threshold and a set of the suspected disappearance alarms are obtained, the set of the continuous alarms with a duration less than the preset duration threshold is removed from the primary screening set, the set of the suspected disappearance alarms is added to the primary screening set, and the new alarm is retained in the primary screening set to obtain a congestion alarm to-be-analyzed list. 6.The congestion alarm method based on visual large model secondary verification according to claim 4, characterized in that, The S240 specifically includes: A suspected congestion duration of each road section in the congestion alarm to-be-analyzed list is obtained. The congestion alarm to-be-analyzed list is sorted according to the suspected duration, traffic saturation, and speed in three dimensions to obtain a congestion alarm to-be-analyzed list after priority sorting. Specifically, each road section in the congestion alarm to-be-analyzed list is sorted according to the suspected duration from long to short; if the suspected duration is the same, each road section is sorted according to the traffic saturation from large to small; and if the traffic saturation is the same, each road section is sorted according to the speed from low to high. 7.The congestion alarm method based on visual large model secondary verification according to claim 5, wherein, The S300 includes: S310. An alarm type of each road section in the congestion alarm to-be-analyzed list after priority sorting is obtained; the alarm type includes a new alarm, a continuous alarm, and a suspected disappearance alarm. When the alarm type is a new alarm, a complete real-time video stream of the corresponding road section is obtained. When the alarm type is a continuous alarm or a suspected disappearance alarm, a key frame sequence is obtained by extracting key frames of the real-time video stream of the corresponding road section at a preset fixed interval. S320. The complete real-time video stream or the key frame sequence is input into the visual large model set in advance to obtain a false alarm set and a final congestion alarm set. 8.The congestion alarm method based on visual large model secondary verification according to claim 7, wherein, The visual large model is a deep learning model that fuses time sequence and spatial perception and includes: An input preprocessing layer is configured to decode a received real-time video stream into a sequence of image frames, or directly receive a sequence of key frames extracted by an upstream link; the sequence is subjected to size normalization, frame rate adjustment and tensor formatting processing to generate batch data meeting the input requirements of the model; A spatio-temporal feature encoding layer is configured to extract a joint feature representation integrating spatial and temporal dimensions from the batch data, and obtain fused spatio-temporal features; the spatio-temporal feature encoding layer includes a spatial feature extraction node and a time series feature modeling node; A scene understanding and decision-making layer is configured to perform high-level semantic reasoning and state determination based on the fused spatio-temporal features, including a traffic state classification node and a confidence evaluation node; the traffic state classification node adopts a fully connected neural network combined with a softmax activation function to output traffic state classification results for each road segment, and the confidence evaluation node is configured to output confidence scores corresponding to the classification results; The classification results include: free flow, slow traffic, congestion, and severe congestion; An output layer is configured to generate a false alarm set and a final congestion alarm set according to predetermined decision rules. 9.The congestion alarm method based on visual large model secondary verification according to claim 8, wherein, The output layer specifically includes: Road segments with classification results of congestion or severe congestion and confidence scores higher than a preset high confidence threshold are included in the final congestion alarm set; Road segments with classification results of free flow or slow traffic, or with classification results of congestion but with confidence scores lower than a preset low confidence threshold are included in the false alarm set. 10.The congestion alarm method based on visual large model secondary verification according to claim 1, wherein, The method further includes: S500, updating the historical false alarm information according to the false alarm set, and periodically screening data in the false alarm set as a training set based on the updated historical false alarm information to perform incremental training and optimization on the visual large model.

Images