CN-122024321-A - Important person abnormal behavior identification method based on AI visual analysis

CN122024321ACN 122024321 ACN122024321 ACN 122024321ACN-122024321-A

Abstract

The invention discloses a key person abnormal behavior identification method based on AI visual analysis, which relates to the field of key person abnormal behavior identification, and comprises the steps of extracting a behavior characteristic sequence of a designated key person, establishing a personal behavior baseline model and outputting an abnormal behavior likelihood value; setting an abnormality detection mechanism and model updating according to a personal behavior baseline model, extracting a high-level semantic feature vector of fusion space-time association of the current behavior by utilizing a pre-trained space-time diagram neural network, constructing a behavior consistency reasoning map, outputting consistency check scores, mapping the scores into alarm grades based on the consistency check scores output by the behavior consistency reasoning map, and generating grading early warning information and an interpretability analysis report according to the alarm grades. The invention has the advantages of realizing the intelligent identification of the abnormal behaviors of the heavy personnel, which not only precisely quantizes but also accords with the judgment of common sense, and effectively solving the pain points which are easily deceived on one side of thousands of people in the traditional detection model.

Inventors

ZENG HAOFENG
LAI TANGSHENG
CHENG HUI

Assignees

广州视海拓界科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260202

Claims (8)

1. The key person abnormal behavior identification method based on AI visual analysis is characterized by comprising the following steps: Acquiring multi-mode monitoring video data of appointed key personnel, extracting a behavior characteristic sequence of the multi-mode monitoring video data, establishing a personal behavior baseline model, and outputting an abnormal behavior likelihood value; according to the personal behavior baseline model, setting an abnormality detection mechanism and model updating in a real-time monitoring stage; Based on the real-time video stream, extracting a high-level semantic feature vector of fusion space-time association of the current behavior by utilizing a pre-trained space-time diagram neural network; The expert knowledge rules are fused, a behavior consistency reasoning map is constructed, multidimensional logic consistency verification is carried out on the current behavior, and consistency verification scores are output; mapping the score into three alarm levels of high risk, medium risk and low risk based on consistency check scores output by the behavior consistency reasoning atlas; and generating hierarchical early warning information and an interpretability analysis report according to the warning level.
2. The method for identifying abnormal behaviors of key personnel based on AI visual analysis according to claim 1, wherein the steps of obtaining the multi-mode monitoring video data of the appointed key personnel, extracting the behavior characteristic sequence thereof, establishing a personal behavior baseline model, and outputting the likelihood value of the abnormal behavior comprise the following steps: intercepting continuous video clips of key personnel in a non-abnormal period from a historical monitoring video as baseline learning data; Multi-dimensional feature extraction is carried out on the baseline learning data, wherein the multi-dimensional feature extraction comprises a skeleton joint point coordinate sequence, an appearance feature vector and motion features; Performing time alignment and normalization processing on the extracted multidimensional features to form a standardized personal historical behavior feature sequence; modeling the personal historical behavior feature sequence by adopting a Gaussian mixture model, constructing a personal behavior baseline model, and outputting an abnormal behavior likelihood value.
3. The method for identifying abnormal behaviors of key personnel based on AI visual analysis according to claim 2, wherein the setting of the abnormality detection mechanism and the model update in the real-time monitoring stage according to the personal behavior baseline model specifically comprises: In a real-time monitoring stage, a standardized personal historical behavior characteristic sequence extracted at the current moment is obtained; Calculating the log likelihood value of the characteristic sequence of the personal historical behavior by using the trained Gaussian mixture model; After model training, calculating the distribution of the log likelihood values by using all training samples or a set of verification sets, and setting a% low score points of the distribution as an abnormal threshold; judging whether the log likelihood value is smaller than an abnormal threshold value, if so, judging that the current behavior deviates from a normal baseline of the individual, and if not, judging that the current behavior does not belong to the abnormality; And setting a sample buffer area with a fixed size, merging the sample buffer area with a historical sample based on a fixed period or event triggering, retraining the Gaussian mixture model, and updating model parameters.
4. The method for identifying abnormal behaviors of key personnel based on AI visual analysis as claimed in claim 3, wherein the extracting the high-level semantic feature vector of fusion space-time association of the current behaviors by using a pre-trained space-time diagram neural network based on real-time video stream specifically comprises: performing target detection and tracking on the real-time video frame, and positioning and continuously tracking key personnel to be analyzed; Extracting real-time skeleton joint point coordinates of a person region tracked in each frame, and constructing a dynamic time-space diagram by taking the same joint point between the previous frame and the following frame as a time edge and physically connecting human bodies in the same frame as a space edge; Splicing the appearance characteristic vector of each frame with the coordinates of the joint point of the frame skeleton to serve as the enhancement characteristic of the joint point; And inputting the dynamic space-time diagram into a pre-trained space-time diagram neural network, and fusing space-time associated high-level semantic feature vectors of the current behaviors.
5. The method for identifying abnormal behaviors of key personnel based on AI visual analysis according to claim 4, wherein the steps of fusing expert knowledge rules, constructing a behavior consistency reasoning map, performing multidimensional logic consistency check on the current behavior, and outputting consistency check scores comprise: presetting a behavior logic rule base defined by a domain expert, wherein the rule base comprises a gesture-behavior consistency rule, a behavior-scene suitability rule, a time sequence behavior mode rule and a false behavior identification rule; According to the real-time analysis result, instantiating a behavior consistency reasoning map, wherein the map nodes comprise nodes based on high-level semantic feature and nodes in a dynamic space-time diagram; the sides of the map consist of logical relation sides defined by experts and preset distribution deviation degree calculation relation sides; the expert-defined logical relationship edge at least comprises a AND, OR, contradiction, implication and time sequence; Calculating a relation edge by the preset distribution deviation, wherein the relation edge is connected with the real-time characteristic node and the personal behavior baseline model; and running a rule reasoning engine on the reasoning map, checking whether logic contradiction or low probability combination exists among all elements detected in real time, and outputting consistency check scores.
6. The method for identifying abnormal behaviors of key personnel based on AI visual analysis according to claim 5, wherein the mapping the scores into three alert levels of high risk, medium risk and low risk based on the consistency check scores output by the behavior consistency reasoning atlas specifically comprises: The consistency check score output by the behavior consistency reasoning map is directly used as an abnormal probability index; And setting a multi-level risk threshold according to the abnormal probability value by combining service experience and statistical distribution characteristics, and mapping the probability value into three alarm levels of high risk, medium risk and low risk.
7. An electronic device comprising at least one processor, and a memory communicatively coupled to the at least one processor, wherein, The memory stores instructions executable by the at least one processor to enable the at least one processor to perform a key person abnormal behavior identification method based on AI visual analysis as set forth in any one of claims 1-6.
8. A computer-readable storage medium storing a computer program, wherein the computer program when executed by a processor implements a key person abnormal behavior recognition method based on AI visual analysis as set forth in any one of claims 1 to 6.

Description

Important person abnormal behavior identification method based on AI visual analysis Technical Field The invention relates to the field of recognition of abnormal behaviors of key personnel, in particular to a recognition method of abnormal behaviors of key personnel based on AI visual analysis. Background Along with the progress of security monitoring to intellectualization and refinement, the recognition of abnormal behaviors of key personnel becomes a core requirement in the fields of public security, financial security, park management and the like, and the AI visual analysis technology becomes a mainstream technical means in the field by virtue of non-contact, real-time and automation advantages. The current video monitoring enters a multi-mode acquisition stage, a depth camera and a high-definition camera can synchronously acquire multidimensional data such as human bones, appearances, motion tracks and the like, a data basis is provided for behavior feature extraction, and meanwhile, the rapid development of AI technologies such as graphic neural networks, target tracking, probability modeling and the like enables quantitative analysis of complex human behaviors to be possible. Under the background, the technical requirement of identifying abnormal behaviors of key personnel is changed from 'general behavior detection' to 'personalized accurate monitoring', the model is required to adapt to behavior habits of different personnel, hidden abnormalities deviating from the normal state of the personnel are identified, the model is required to have self-adaption to cope with the natural evolution of the behaviors of the personnel, meanwhile, the interpretability of the detection result is considered, and a clear decision basis is provided for security and protection treatment. The existing technology for identifying abnormal behaviors of key personnel in the AI visual analysis class mostly adopts a general behavior detection model, does not build a personalized behavior baseline aiming at the key personnel, has the problem of thousands of people, and cannot accurately capture the behavior habit and mode of an individual based on group behavior feature training, so that the behavior conforming to the normal state of the individual is easily misjudged as abnormal, and the implicit abnormal behavior deviating from the personal baseline is also difficult to identify; in addition, most of the prior art is a pure data driving architecture, lacks an inference mechanism for fusing expert knowledge and common sense logic in the field, cannot effectively identify camouflage and fraud abnormal behaviors possibly existing in the heavy personnel, and is difficult to deeply mine and verify logic rationality behind the behaviors. Disclosure of Invention In order to solve the technical problems, the technical scheme provides an important person abnormal behavior identification method based on AI visual analysis, which solves the problems that behavior habits and modes of individuals cannot be accurately captured, behaviors conforming to the normal state of the individuals are easily misjudged to be abnormal, hidden abnormal behaviors deviating from the baseline of the individuals are difficult to identify, camouflage and fraud abnormal behaviors possibly existing in important persons cannot be effectively identified, and logic rationality behind the behaviors is difficult to deeply excavate and verify. In order to achieve the above purpose, the invention adopts the following technical scheme: an important person abnormal behavior identification method based on AI visual analysis comprises the following steps: Acquiring multi-mode monitoring video data of appointed key personnel, extracting a behavior characteristic sequence of the multi-mode monitoring video data, establishing a personal behavior baseline model, and outputting an abnormal behavior likelihood value; according to the personal behavior baseline model, setting an abnormality detection mechanism and model updating in a real-time monitoring stage; Based on the real-time video stream, extracting a high-level semantic feature vector of fusion space-time association of the current behavior by utilizing a pre-trained space-time diagram neural network; The expert knowledge rules are fused, a behavior consistency reasoning map is constructed, multidimensional logic consistency verification is carried out on the current behavior, and consistency verification scores are output; mapping the score into three alarm levels of high risk, medium risk and low risk based on consistency check scores output by the behavior consistency reasoning atlas; and generating hierarchical early warning information and an interpretability analysis report according to the warning level. Preferably, the acquiring the multi-mode monitoring video data of the appointed key personnel, extracting the behavior characteristic sequence thereof, establishing a personal behavior baseline model, and o