CN-121982816-A - Intelligent weak current security monitoring method and system integrating computer vision

CN121982816ACN 121982816 ACN121982816 ACN 121982816ACN-121982816-A

Abstract

The application relates to the technical field of security monitoring, in particular to an intelligent weak current security monitoring method and system integrating computer vision, wherein the method comprises the steps of extracting a core key point set from a monitoring video stream; the method comprises the steps of carrying out dynamic sequence analysis on a core key point set to obtain a direction consistency measure and an action synergy measure, carrying out abnormal behavior identification based on the direction consistency measure and the action synergy measure and according to a preset quantitative combination judgment rule, wherein the quantitative combination judgment rule is configured to meet a direction consistency threshold value condition, an action synergy threshold value condition and a displacement threshold value condition at the same time, and generating and preferentially transmitting alarm information in response to an abnormal behavior identification result. The application can improve the accuracy of identifying the abnormal behavior in the complex weak current scene through light dynamic sequence and collaborative analysis.

Inventors

LI JIANHUI
XIA JUN
CHEN RONGAN
LIU YANG

Assignees

南京图南智慧大数据集团有限公司

Dates

Publication Date: 20260505
Application Date: 20260203

Claims (10)

1. The intelligent weak current security monitoring method integrating computer vision is characterized by comprising the following steps of: extracting a core key point set associated with the dynamic characteristics of the preset abnormal behaviors from the monitoring video stream; In an analysis window with a preset time length, carrying out dynamic sequence analysis on the core key point set in continuous frames to obtain a direction consistency measure reflecting the motion trend persistence of the key points associated with the abnormal behaviors and an action synergy measure reflecting the time sequence relevance of different key points; Carrying out abnormal behavior identification based on the direction consistency measurement and the action consistency measurement and according to a preset quantitative combination judgment rule, wherein the quantitative combination judgment rule is configured to simultaneously meet a direction consistency threshold condition based on the direction consistency measurement, an action consistency threshold condition based on the action consistency measurement and a displacement threshold condition based on the displacement change of the core key point set; In response to the abnormal behavior recognition result, generating and preferentially transmitting alarm information, wherein the alarm information at least comprises the recognized abnormal behavior type and the direction consistency metric and the action synergy metric for judging the behavior.
2. The method of claim 1, wherein the dynamic sequence analysis is performed in a sliding window manner, comprising: performing displacement vector tracking on the core key point set based on the continuous frame sequence in the analysis window; According to the tracking result, calculating a direction consistency rate representing the continuation of the movement direction trend of the key point associated with the abnormal behavior as the direction consistency measure; And identifying a preset time sequence association mode between actions represented by different core key points in the analysis window, and calculating the statistical frequency of occurrence of the time sequence association mode as the action cooperativity measurement.
3. The method of claim 2, wherein the quantitative combination decision rule further incorporates torso pose features as decision conditions; Aiming at a preset climbing behavior type, the quantitative combination judgment rule is configured to meet a first trunk gesture threshold condition based on a trunk inclination angle; For a preset intrusion behavior type, the quantitative combination judgment rule is configured to meet a second trunk posture threshold condition based on the trunk displacement direction.
4. A method according to claim 3, wherein the first torso pose threshold condition and the second torso pose threshold condition are configured by performing a collaborative decision by performing a associative match between the torso pose features and the direction consistency and motion synergy metrics obtained by the dynamic sequence analysis, and based on a constrained relationship between torso motion and limb motion defined by a normalized sequence of actions of the abnormal behavior type.
5. The method of claim 4, wherein the constrained relationship between the torso motion and limb motion upon which the torso pose features are based in the associative matching process is further configured to dynamically modify the decision threshold of the first torso pose threshold condition or the second torso pose threshold condition based on a trend of variation of the directional uniformity metric and the motion cooperativity metric with respect to respective corresponding threshold conditions over a plurality of consecutive analysis windows of the preset length of time.
6. The method of claim 3, wherein the quantitative combination decision rule comprises a multi-level logic decision structure; Wherein, for the climbing behavior type, the multi-stage logic determination structure is configured to preferentially determine the direction consistency threshold condition and the action cooperativity threshold condition, and then determine the first torso gesture threshold condition; for the break-in behavior type, the multi-stage logic determination structure is configured to preferentially determine the displacement threshold condition and the action cooperativity threshold condition, and then determine the second torso pose threshold condition.
7. The method according to claim 2, wherein the preset time sequence association pattern is configured to define a concomitant relation between a hand rising motion and a foot pedaling motion for a climbing behavior type, define a sequential relation between a body center of gravity displacement and a foot moving direction for an intrusion behavior type, and the action synergetic measure is calculated by counting a proportion of consecutive frames within the analysis window while satisfying the concomitant relation or the sequential relation defined for a corresponding abnormal behavior type.
8. A method according to claim 3, further comprising: After the abnormal behavior is primarily identified based on the quantitative combination judgment rule, continuously checking the abnormal behavior identification result based on the direction consistency measurement, the action cooperativity measurement and the change trend of the trunk gesture characteristic in the subsequent analysis window adjacent in time; the persistence check is configured to correct or cancel a preliminary recognition result when a matching degree of a variation trend of at least one of the direction consistency metric, the action cooperativity metric and the trunk gesture feature in the subsequent analysis window and an expected dynamic feature variation trend corresponding to an abnormal behavior type is lower than a matching degree threshold.
9. The method of claim 8, wherein the persistence check is further configured to include a collaborative logic check: based on a plurality of continuous subsequent analysis windows, respectively calculating window matching degrees among the direction consistency metric, the action cooperativity metric and the change trend of the trunk gesture characteristics and the expected dynamic characteristic change trend of the corresponding abnormal behavior types; and when the window matching degree of the plurality of continuous subsequent analysis windows is continuously lower than the matching degree threshold value, or the window proportion of the window matching degree lower than the matching degree threshold value in a preset check period exceeds a preset proportion threshold value, executing correction or cancellation of the preliminary identification result.
10. An intelligent weak current security monitoring system integrating computer vision, which is characterized by being used for executing the method of any one of claims 1-9, and comprising: the feature extraction module is used for extracting a core key point set associated with the dynamic feature of the preset abnormal behavior from the monitoring video stream; The sequence analysis module is connected with the feature extraction module and is used for carrying out dynamic sequence analysis on the core key point set in the continuous frames in an analysis window with a preset time length so as to obtain a direction consistency measurement reflecting the continuity of the key point movement trend associated with the abnormal behavior and an action synergetic measurement reflecting the action time sequence association of different key points; the fusion judging module is connected with the sequence analyzing module and is used for identifying abnormal behaviors based on the direction consistency measurement and the action consistency measurement and according to a preset quantitative combination judging rule, and the quantitative combination judging rule is configured to simultaneously meet a direction consistency threshold condition based on the direction consistency measurement, an action consistency threshold condition based on the action consistency measurement and a displacement threshold condition based on the displacement change of the core key point set; And the priority alarm module is connected with the sequence analysis module and the fusion judgment module and is used for responding to the abnormal behavior identification result and generating and preferentially transmitting alarm information, wherein the alarm information at least comprises the identified abnormal behavior type and the direction consistency measurement and the action synergy measurement for judging the behavior.

Description

Intelligent weak current security monitoring method and system integrating computer vision Technical Field The application relates to the technical field of security monitoring, in particular to an intelligent weak current security monitoring method and system integrating computer vision. Background Along with the deep fusion of artificial intelligence and the internet of things technology, a security monitoring system gradually evolves towards an intelligent and automatic direction. Perimeter precautions are used as a first barrier of a security system of places such as schools, industrial parks, residential communities and the like, and the intelligent monitoring demands are increasingly highlighted. By carrying out real-time and automatic identification and early warning on abnormal behaviors such as climbing and intrusion of personnel, the security response speed can be greatly improved, and the dependence of manpower is reduced, so that the intelligent perimeter intrusion detection technology based on computer vision has become a key direction of industry development. Prior art has been provided with identification schemes for rail climbing activities. For example, chinese patent No. CN115294610B, entitled "method, apparatus, and electronic device for identifying fence climbing", provides a technical path based on head detection and tracking. The method comprises the steps of firstly detecting a human head target in a video by using a model, generating a detection frame, continuously tracking the human head target, setting a virtual warning line attached to the top of a fence in a video picture, calculating a central point of the human head detection frame and a climbing judgment point which is calculated according to a fixed human body proportion and represents the approximate position of a trunk, and finally comprehensively judging climbing behaviors by analyzing the position relation (such as whether crossing occurs) between the two points and the warning line in a tracking sequence. However, when the technical scheme is deployed in the actual weak current security monitoring scenes such as communities and parks, the scenes are generally characterized by long monitoring distance, uneven ambient illuminance and possibly complex situations of multiple parallel people, and the scheme highlights the core problem of insufficient recognition accuracy. Specifically, the scheme takes the human head detection frame and the tracking track thereof as the whole analysis basis, and under common scenes such as that people wear a cap shirt, backlight is generated due to insufficient illumination at night, or mutual shielding is caused by close contact of multiple people with a fence, the detection frame is extremely easy to lose, jump or merge, tracking is interrupted, the whole judging process is invalid due to loss of continuous and reliable data input, and missing report is directly caused. In addition, the core criteria of the method depend on a geometric relationship between a static climbing judgment point and a fixed warning line estimated according to a fixed proportion, in practical application, individual body type differences (such as children and adults) of personnel and body posture diversity (such as sideways and curling) in the climbing process lead to obvious and irregular deviation between the estimation point and the actual position of a key part (such as the center of gravity of a trunk) of an actual human body, so that the analysis of the position relationship based on the estimation point loses physical significance, and misjudgment is caused. Moreover, the analysis logic of the scheme essentially infers based on the spatial positional relationship of points and lines over several discrete moments, a geometrical interpretation of behavior transient states that is static. It cannot model and understand the core elements that constitute climbing, intrusion, etc., i.e., continuous, coordinated movement trends and timing of specific body parts (e.g., hands, feet) over a period of time. For example, it is difficult to distinguish between the climbing action of the arms which continue to extend upwards with foot pedaling and the normal action of putting down after a brief lifting of the hand to finish the clothing, and the normal action of leaving after a brief stay beside the fence from the intrusion process of the body centre of gravity which continues to move into the fence. The cognition lack of the dynamic time sequence characteristics and the cooperative semantics of the behavior makes the prior art have weak generalization capability and coexistence of misjudgment and missed judgment when facing complex and changeable behavior patterns in real scenes, and finally can not realize high-reliability abnormal behavior identification in weak current security monitoring scenes with limited resources and complex environments. Disclosure of Invention In order to improve the accuracy of abnormal behavi