Search

CN-121982643-A - Intelligent analysis method and system for intelligent security and protection AI image of park

CN121982643ACN 121982643 ACN121982643 ACN 121982643ACN-121982643-A

Abstract

The invention relates to a intelligent analysis method and system for intelligent security and protection AI images of a park, and belongs to the technical field of intelligent security and protection monitoring. The method comprises the steps of encoding behavior characteristics of video frames in a park security image based on a preset violation event-image characteristic relation network, distributing dynamic monitoring signals according to the area of an abnormal candidate area and the occurrence frequency of abnormal behaviors, constructing an abnormal intelligent judging model, judging the behavior types of abnormal objects through a multi-scale area matching algorithm, mapping the identified abnormal types and object space positions to a data twin cockpit for twin mirror image display, pushing abnormal monitoring alarms to a park security control center for secondary verification and risk scoring, synchronously updating mirror images in the data twin cockpit to generate a structured event report, constructing a judging enhancement sample training set, and dynamically adjusting the space attention weight of the abnormal intelligent judging model.

Inventors

  • SHENG ZHICHAO
  • SHI SHAOJUN
  • SHAO WENJUN

Assignees

  • 上海至盛信息技术股份有限公司

Dates

Publication Date
20260505
Application Date
20260128

Claims (12)

  1. 1. A park intelligent security AI image intelligent analysis method is characterized by comprising the following steps: S1, acquiring a park security image, encoding behavior characteristics of video frames in the park security image based on a preset violation event-image characteristic relation network, calculating pixel change rates of a current frame and a historical reference frame through an inter-frame difference method to obtain an abnormal candidate region, and distributing dynamic monitoring signals according to the area of the abnormal candidate region and the occurrence frequency of the abnormal behavior; S2, constructing an abnormal intelligent judgment model based on the dynamic analysis signal, inputting the park security image into a violation event-image characteristic relation network, judging the behavior type of an abnormal object through a multi-scale region matching algorithm, mapping the identified abnormal type and the object space position to a data twin cockpit, carrying out twin mirror image display on a real park scene and a digital park scene through a region mapping method, and outputting an abnormal monitoring alarm; S3, pushing the abnormal monitoring alarm to a park security control center for secondary verification and risk scoring, synchronously updating mirror image scenes in the data twin cockpit, and automatically generating a structured event report containing a time stamp and a time type; S4, synchronously counting high-frequency abnormal areas and types of various abnormal events in the park through the structured event report to obtain a judging enhancement sample training set, dynamically adjusting the space attention weight of the abnormal intelligent judging model according to the high-frequency abnormal areas, and carrying out self-adaptive matching on the actual risk distribution of the park and the space attention weight.
  2. 2. The method of claim 1, wherein the method for obtaining the campus security image comprises the steps of performing time synchronization and coordinate alignment on video frames, encoding an image stream according to edge computing nodes, and capturing thermal signal characteristics in a low-light environment by combining an infrared thermal imaging technology to obtain the campus security image.
  3. 3. The method of claim 1, wherein the method for encoding the behavior features in the park security image by the illegal event-image feature relation network is characterized in that a correlation map among illegal event types is established through a correlation rule mining algorithm based on historical security data, the behavior features in the park security image are used as nodes to be embedded into the illegal event-image feature relation network, and an integrated attention mechanism is used for weighting and integrating the dependency intensity among the features to generate an abnormal type dynamic relation table.
  4. 4. The method according to claim 1, wherein the method for distributing the dynamic monitoring signal is: extracting environment factors of a park based on the area of the abnormal candidate region and the occurrence frequency of abnormal behaviors, calculating contribution degrees of abnormal sensitivity parameters in different regions to overall risk, performing correlation fitting on the abnormal sensitivity parameters and the environment factors of the park, and screening out the abnormal candidate region by an interframe difference method; Counting the occurrence frequency of the abnormality in the continuous time window in the abnormality candidate area, carrying out time sequence weighting on the abnormal events in different time periods through an exponential decay function, calculating the level of the monitoring signal, and dynamically distributing the monitoring signal according to a preset risk level interval.
  5. 5. The method according to claim 2, wherein the method for mapping the spatial positions of the objects by the data twinning cockpit is that the abnormal objects in the park security image, the identified abnormal types and the spatial positions of the abnormal objects are mapped into a digital park scene in real time, twinning mirror image display of a real park and the digital park is constructed, and the positions of the abnormal objects in the twinning scene are positioned through regional mapping and coordinate calibration.
  6. 6. The method of claim 5, wherein the constructing method of the abnormal intelligent judgment model is as follows: Performing feature extraction on the dynamic monitoring signals through a time sequence decomposition and feature vectorization method, converting unstructured image data into a computable input feature set, and dynamically associating the distribution of anomaly types and spatial positions by combining behavior data in a historical anomaly record to generate a multidimensional anomaly data driving feature matrix; The multidimensional abnormal data driving feature matrix receives and processes image fluctuation features according to signal driving, converges the judgment result through a linear programming algorithm, and generates a judgment scheme; And executing abnormal execution data and feedback results acquired in real time by the evaluation scheme, dynamically correcting the parameter weight of the abnormal intelligent evaluation model through a rolling update strategy, and constructing the abnormal intelligent evaluation model by taking the change of the park policy and the constraint condition of the environmental change as external mapping parameters.
  7. 7. The method of claim 4, wherein the outputting of the anomaly monitoring alert is: Inputting the dynamic monitoring signal into the abnormal intelligent judgment model, analyzing logic association between the park security image and the abnormal record based on a specific recognition algorithm, and generating an abnormal monitoring alarm through numerical association between interpolation time sequence image characteristic data and the abnormal history record; Pushing the generated abnormal monitoring alarm to a park security control center for secondary verification and risk scoring, updating a digital scene in the data twin cockpit according to a verification result, and synchronously recording mapping coordinates and behavior types in the digital twin cockpit.
  8. 8. The method of claim 2, wherein the structured event report includes a risk level and a judgment basis as a data basis for abnormal event identification to transform a risk distribution into a visual analysis chart.
  9. 9. The method of claim 4, wherein the generating process of the evaluation enhancement sample training set is to label the behavior characteristics and types of various abnormal events in the park based on the structured event report to form initial abnormal sample data, identify the abnormal type and high-frequency abnormal region with higher occurrence frequency in the park according to the initial abnormal sample data set, construct the evaluation enhancement sample training set through time sequence characteristic expansion and abnormal behavior disturbance simulation, and dynamically update the spatial attention weight of the abnormal intelligent evaluation model.
  10. 10. The method of claim 1, wherein the method for adjusting the spatial attention weight is characterized in that the abnormal behavior contribution degree of different time periods is calculated based on an exponential decay function, risk stratification is carried out on the spatial regions of the park according to the abnormal event occurrence frequency of each spatial region in the evaluation enhancement sample training set, an initial weight coefficient is given to the spatial attention weight through the change rate of the abnormal density in the region, so that the high-frequency abnormal region obtains higher attention degree, and the low-risk region obtains lower weight.
  11. 11. The method of claim 8, wherein the risk scoring criteria is to set a basic risk coefficient for different abnormal events in combination with environmental correction factors of a campus where the event is located, perform time sequence priority calibration on risk scores based on occurrence modes of the abnormal events in different time periods and event burstiness indexes, and perform adaptive amplification on scores of short-time high-frequency or burstiness abnormal events.
  12. 12. A campus intelligent security AI image intelligent analysis system for performing the method of any of claims 1-11, comprising: the abnormal candidate region detection module is used for acquiring a park security image, encoding behavior features of video frames in the park security image based on a preset violation event-image feature relation network, calculating pixel change rates of a current frame and a historical reference frame through an inter-frame difference method to obtain an abnormal candidate region, and distributing dynamic monitoring signals according to the area of the abnormal candidate region and the occurrence frequency of the abnormal behavior; The data twinning mapping module is used for constructing an abnormal intelligent judging model based on the dynamic analysis signals, inputting the park security image into a violation event-image characteristic relation network, judging the behavior type of an abnormal object through a multi-scale area matching algorithm, mapping the identified abnormal type and object space position to a data twinning cockpit, carrying out twinning mirror image display on a real park scene and a digital park scene through an area mapping method, and outputting an abnormal monitoring alarm; the alarm report generation pushing module pushes the abnormal monitoring alarm to a park security control center for secondary verification and risk scoring, synchronously updates a mirror image scene in the data twin cockpit and automatically generates a structured event report containing a time stamp and a time type; And the model training and adjusting module is used for synchronously counting high-frequency abnormal areas and types of various abnormal events in the park through the structured event report to obtain a judging enhancement sample training set, dynamically adjusting the space attention weight of the abnormal intelligent judging model according to the high-frequency abnormal areas and carrying out self-adaptive matching on the actual risk distribution of the park and the space attention weight.

Description

Intelligent analysis method and system for intelligent security and protection AI image of park Technical Field The invention belongs to the technical field of intelligent security monitoring, and particularly relates to a intelligent analysis method and system for intelligent security AI images of a park. Background Along with the rapid development of the park security system from the traditional video monitoring to the intelligent and digital directions, how to realize the accurate identification and real-time early warning of abnormal behaviors in a large-scale, multi-scene and continuous space-time environment becomes an important research direction of the current intelligent security technology. Traditional park security mainly relies on manual inspection or a rule-based video monitoring system to judge abnormal conditions through fixed thresholds, static trigger conditions or simple motion detection algorithms. However, such methods generally have the problems of poor adaptability, high false alarm rate, strong manual dependence, difficulty in coping with complex scene changes, and the like. For example, under environments such as illumination change, occlusion change, crowd density fluctuation, background dynamic interference, etc., the traditional detection mode often cannot effectively distinguish subtle differences between normal behaviors and abnormal behaviors, so that the system is difficult to meet the requirements of a modern park on safety management in terms of response efficiency and accuracy. In recent years, with the rapid development of artificial intelligence, deep learning and behavior recognition algorithms, intelligent security technology based on image and video content understanding is gradually matured. However, the prior art still has the outstanding problems that firstly, most models are easy to generate characteristic distortion in a park-level multi-scene and large-scale space, especially when cameras are distributed in a complex mode and a target behavior span is large, the capturing capability of the models to context relations is still insufficient, secondly, the existing abnormal behavior recognition models usually treat different event types in an isolated mode, modeling on relevance among abnormal events is lacking, collaborative judgment cannot be carried out according to semantic dependency, co-occurrence mode or risk relevance among different events, so that recognition precision and risk assessment capability are insufficient, thirdly, the existing systems are difficult to form a complete closed loop structure of 'detection-assessment-feedback', especially after abnormal events occur, an automatic risk scoring mechanism and a structured event generation flow are lacking, and subsequent data accumulation, model enhancement and system iterative optimization are difficult to support. In addition, with the wide application of the data twinning technology in urban management and park operation and maintenance scenes, the construction of synchronous mapping between a real park and a digital park has become an important trend of intelligent security development. However, most security systems still stay in the video image layer, digital mirror image display of the real environment, the spatial position and the event type cannot be realized, and the capability of dynamically superposing the abnormal recognition result to the digital twin cockpit for visual expression is also lacking. Due to the lack of an effective spatial mapping and online judging mechanism, the existing system is difficult to realize intelligent decision based on spatial risk dynamic distribution in a complex park scene. In summary, the existing intelligent security system still has significant shortcomings in terms of abnormal behavior recognition accuracy, event correlation reasoning, dynamic risk assessment, data twinning visual presentation, model self-adaptation capability and the like. Therefore, there is a need for a smart analysis method for intelligent security and protection AI images of a campus, which can integrate anomaly detection, associated modeling of illegal events, multi-scale behavior recognition, risk scoring, data twin mapping and dynamic model optimization, so as to improve the level of intelligence and real-time response capability of security and protection of the campus. Disclosure of Invention In order to solve the problems in the prior art, the invention provides an intelligent analysis method for intelligent security AI images of a park, The aim of the invention can be achieved by the following technical scheme: S1, acquiring a park security image, encoding behavior characteristics of video frames in the park security image based on a preset violation event-image characteristic relation network, calculating pixel change rates of a current frame and a historical reference frame through an inter-frame difference method to obtain an abnormal candidate region, and distributing dy