Campus abnormal behavior detection method based on complex theory

The campus abnormal behavior detection method uses advanced data collection and deep neural networks to analyze complex student behaviors, improving accuracy and timeliness of abnormal behavior identification and response.

CN119068549APending Publication Date: 2024-12-03GUANGXI PUBLIC INFORMATION IND CO LTD
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411182327.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-27
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

Existing campus monitoring systems struggle to accurately identify and predict student behavior patterns, particularly subtle abnormalities, due to the lack of comprehensive analysis of individual and environmental factors, and fail to provide timely and personalized responses, leading to increased false positives and delayed interventions.

Method used

A campus abnormal behavior detection method utilizing advanced data collection through video surveillance and sensors, combined with complex system theory and deep neural networks, to analyze three-dimensional motion trajectories, crowd density, emotional responses, and social interactions, enabling real-time identification and prediction of abnormal behaviors.

Benefits of technology

The method achieves accurate and timely detection of abnormal behaviors by considering individual and environmental factors, reducing false positives and enabling targeted interventions, thus enhancing campus safety and response efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119068549A_ABST
    Figure CN119068549A_ABST
Patent Text Reader

Abstract

The invention discloses a campus abnormal behavior detection method based on a complex theory. The method comprises the steps of collecting student activity data, converting the student activity data into an analyzable behavior mode data set, carrying out data normalization processing, reflecting an evolution process of student group behaviors by adopting graph theory analysis and a dynamic system theory, identifying common behavior modes and abnormal behavior modes in the behavior mode data set, judging abnormal behaviors, sending an alarm and the like. Dynamic changes of student behaviors and complex interaction behind the dynamic changes can be comprehensively analyzed and understood, such as social interaction between individuals and interaction between the individuals and the environment. Meanwhile, by using the deep neural network model, the accuracy of behavior pattern recognition is improved, and real-time monitoring and analysis of student behavior data are realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of campus safety and network technology, and in particular relates to a campus abnormal behavior detection method based on complex theory. Background Art

[0002] In the modern educational environment, ensuring campus safety and promoting the healthy growth of students are the top priorities of educators and administrators. In recent years, with the development of technology, especially the advancement of video surveillance, sensor technology, data analysis and machine learning, campus security monitoring and student behavior analysis have become more feasible and effective. However, despite the progress of existing technologies in some aspects, there are still some major challenges in student behavior pattern recognition and abnormal behavior detection.

[0003] Existing technologies mainly rely on simple motion detection, rule-based behavior recognition, or traditional machine learning methods to analyze video surveillance data. Although these methods can identify some obvious abnormal behaviors, such as running and fighting, they are generally unable to effectively understand and analyze the complexity and dynamics of student behavior. Students' behavior in a campus environment is influenced by many factors, including individual psychological state, social interactions, and environmental conditions. Existing methods often ignore the interaction of these factors, resulting in the inability to accurately identify and predict student behavior patterns, especially those that are subtle but may indicate potential problems.

[0004] In addition, existing technologies are also limited in terms of personalized identification and real-time response. Due to the lack of effective data analysis frameworks and algorithms, it is difficult for existing systems to monitor and analyze each student's behavior in a personalized manner, and it is impossible to take into account each student's unique behavioral characteristics and background. This not only increases the risk of false positives, but also makes it impossible for the system to identify and respond to possible safety issues in a timely and accurate manner. In addition, due to the lack of effective real-time data processing and analysis capabilities, existing systems often cannot respond immediately, missing the best time for prevention and intervention. Summary of the invention

[0005] In view of the problems that the existing campus monitoring system cannot accurately identify and predict students' behavior patterns, cannot timely and accurately identify and respond to possible safety issues, and lacks effective real-time data processing and analysis capabilities, the present invention provides a campus abnormal behavior detection method based on complex theory. The method can comprehensively analyze and understand the dynamic changes of children's behavior and the complex interactions behind them. By introducing advanced data analysis and machine learning algorithms, the system can deeply understand the complexity and dynamics of students' behavior, consider the interaction of multiple factors such as individual psychological state, social interaction and environmental conditions, so as to more accurately identify and predict students' behavior patterns; adopt personalized identification and real-time response technology, use effective data analysis framework and algorithm to realize personalized monitoring and analysis of each student's behavior, timely and accurately identify and respond to possible safety issues, reduce false alarms, and ensure that appropriate actions are taken at critical times of prevention and intervention.

[0006] The technical solution of the present invention is as follows:

[0007] A campus abnormal behavior detection method based on complex theory includes the following steps:

[0008] S1. Collect student activity data on campus through video surveillance systems and sensors integrated with advanced algorithms, including three-dimensional motion trajectories, changes in group density, dynamic changes in distance between individuals, and emotional responses captured through micro-expression recognition technology; and transmit the collected activity data to the campus data processing center in real time; select key areas on campus with large traffic and frequent student activities, including library entrances, classrooms, and cafeterias, as target areas for data collection, and install video surveillance equipment with high resolution and wide viewing angle in each target area, so that the equipment can cover the entire area and capture student activities without blind spots;

[0009] S2, transform the collected student activity data into analyzable behavioral pattern data sets;

[0010] S3, performing noise filtering and data normalization on the behavior pattern dataset;

[0011] S4. Establish a complex system model to simulate the interaction between individual students and their interaction with the environment, and use graph analysis and dynamic system theory to reflect the evolution of student group behavior;

[0012] S5. Use the deep neural network model to train the behavior pattern data set, and guide the deep neural network model to identify common behavior patterns and abnormal behavior patterns in the behavior pattern data set through the theoretical basis and behavior dynamic analysis provided by the complex system model;

[0013] S6. Set an abnormal behavior threshold, and monitor student behavior data in real time based on the deep neural network model trained in step S5. Once the detected behavior data exceeds the abnormal behavior threshold, the system determines it as abnormal behavior;

[0014] S7. When the system detects abnormal behavior, it automatically sends alerts to campus administrators and provides relevant behavioral data and recommended intervention measures.

[0015] To further illustrate the present invention, step S1 includes the following steps:

[0016] S11, the video monitoring system extracts the student's three-dimensional motion trajectory from the video stream in real time through an image recognition algorithm; the image recognition algorithm uses an inter-frame difference method to extract moving objects, applies an optical flow method to track the object's motion, and combines stereo vision technology to reconstruct the student's three-dimensional motion trajectory. xyz ;

[0017] Let the continuous video frames be F t and F t+1 , where t represents time, and the 3D position of the student in frame F t and F t+1 P t (x t ,y t , z t ) and Pt +1 (x t+1 ,y t+1 , z t+1 );

[0018] Motion vector V t→t+1 Calculated by inter-frame difference and optical flow estimation:

[0019] V t→t+1 =P t+1 -P t =(x t+1 -x t ,y t+1 -y t , z t+1 -z t );

[0020] Combined with stereo vision, three-dimensional motion trajectory T xyz The calculation is:

[0021]

[0022] Where n is the total number of video frames;

[0023] High-precision reconstruction of three-dimensional motion trajectories is achieved through inter-frame difference method, optical flow method and stereo vision technology. Specifically, the moving objects are extracted through inter-frame difference method, the object movement is tracked by optical flow method, and the three-dimensional motion trajectory of children is reconstructed by combining stereo vision technology. In this way, high-precision and real-time monitoring of children's activities is achieved, the accuracy of data collection is improved, the robustness of the system in complex environments is enhanced, and the problems of high misrecognition rate, incomplete data collection and insufficient dimensions of traditional monitoring systems are solved.

[0024] S12. Integrate a crowd density analysis algorithm in the video surveillance system, segment the crowd area in the picture through image segmentation technology, calculate the number of pixels in each segmented area, and estimate the number of people in each area by the ratio of the number of pixels to the preset average pixel ratio of the human body. Combined with the number of pixels in each video frame F t Crowd segmentation area A m , m represents the mth area, and the number of people in each area is estimated by the number of pixels P in the area Am Compared with the preset average number of pixels per person P avg The change in population density is expressed as the sum of the number of people in all regions at each time point t:

[0025]

[0026] Population density change D t It is expressed as the sum of the number of people in all regions at each time point t:

[0027]

[0028] S13. Ultrasonic or infrared distance sensors are arranged in each target area of ​​the campus to collect the dynamic changes of the distance between individuals. The distance d between two students i and j at time t measured by ultrasonic or infrared sensors is ij (t) can be expressed as the round trip time T of the signal ij And the signal propagation speed c is calculated as:

[0029]

[0030] S14. Use high-definition cameras to capture students’ facial expressions and use micro-expression recognition technology to analyze students’ emotional reactions, including eyebrows rising and mouth corners drooping, to classify students’ basic emotional states. f .

[0031] To further illustrate the present invention, step S2 comprises the following steps:

[0032] S21, extracting the three-dimensional motion trajectory T from the data collected in step S1 xyz 、 Population density change Dt , dynamic changes in the distance between individuals ij (t) and emotional state E f As raw data input;

[0033] S22. Create a data record for each student containing all activity data during the observation period. For each student i’s data record, consider its time series data X i Multidimensionality:

[0034] X i ={T xyz,i (t), D t,i , d ij,i (t), E f,i (t)};

[0035] Among them, t is the time variable, which represents the observation value at different time points t;

[0036] S23, for each data record X i The data is structured to reflect the behavior patterns of students in a specific time period. The continuous time point data is clustered into discrete behavior events. Each behavior event is defined by four dimensions: the student's position, the density of the group, the relative position to other students, and the emotional state. Dynamic time warping is used to identify and classify behavior patterns. For any two time series X i and X j It is expressed as:

[0037]

[0038] Where T is the length of the time series, x it and x jt They are the time series X i and X j The value at time t;

[0039] S24. The behavior pattern data sets of all students are combined into a comprehensive behavior pattern data set B for subsequent behavior pattern recognition and abnormal behavior detection:

[0040]

[0041] Where N is the number of participating students, ensuring that the dataset B covers all observed student behaviors on campus, and ∪ represents the merge operation, so that the dataset covers all observed student behaviors on campus;

[0042] S25. Define metadata description for Dataset B, including the time range, location, number of participating students, and timestamp and duration of each behavioral event. The purpose of metadata is to provide background information for the dataset, so that the data content can be quickly located and understood during subsequent analysis:

[0043] Metadata(B)={"Collection Period": P, "Locations": L, "Participants": N, "Event Timestamps": T e ,"Duration":D e};

[0044] Where P represents the time frame of data collection, L represents the location of data collection, N is the number of participating students, and T e and D e Represents the timestamp and duration of each behavioral event, respectively.

[0045] By utilizing a variety of sensors and data processing technologies, a comprehensive behavioral pattern dataset was constructed. This method overcomes the limitations of traditional single modality data, achieves accurate identification of complex behavioral patterns, and improves the adaptability and accuracy of the system.

[0046] To further illustrate the present invention, step S4 comprises the following steps:

[0047] S41. Define the model parameters of the student individual and the environment. The state of the student individual i at time t is determined by the position P i (t) = (x i (t), y i (t), z i (t)), velocity vector V i (t)=(v x,1 (t), v y,i (t), v z,i (t)), emotional state E f,i (t) and the social interaction parameter S with student j ij (t), the environmental parameters are composed of the population density D t and location-specific characteristics L k Represents, where k specifies the type of place;

[0048] S42. Construct a graph model G(V, E) to represent the interactions between individual students, where nodes V represent individual students, edges E represent social interactions between students, and edge weights w ij (t) is determined by the social interaction parameter \(S_{ij}(t)\), which reflects the intensity and frequency of social interactions between individuals:

[0049] w ij (t) = α·exp(-β·d ij (t))·E f,i (t)·E f,j (t);

[0050] Among them, α and β are adjustment parameters, d i j(t) is the distance between students, reflecting the interaction between the intensity of social interaction and emotional state;

[0051] S43. Using dynamic system theory, a dynamic behavior equation is defined for each student in the graphical model to simulate the time evolution of student behavior. The dynamic equation is based on the individual's position parameters, velocity vector, emotional state, and interaction with other individuals, and is formally expressed as:

[0052]

[0053]

[0054] Among them, U(P i (t) is a potential energy field that reflects the impact of the environment on the individual's movement, f inter is an influence function of interactions between individuals, which depends on social interaction parameters, emotional states, and location characteristics; represents a vector operator in vector calculus, called the gradient operator; d represents the differential operator, which is used to represent the small change in a function with respect to a variable;

[0055] S44. Using graph theory analysis methods, we analyze the connectivity, clustering coefficient and path length of the graph model G(V, E) to reveal the network characteristics and social structure of student group behavior, including the average path length L(G) and clustering coefficient C(G):

[0056]

[0057]

[0058] Among them, δ ij is the shortest path distance between nodes i and j, e ijk is the number of triangles connecting nodes i, j, and k, k i is the degree of node i;

[0059] S45. Integrate the results of dynamic system theory and graph theory analysis to predict the evolution trend of student group behavior and possible abnormal behavior patterns. By simulating different social interactions and environmental change scenarios, evaluate the impact of factors on the evolution of student group behavior:

[0060]

[0061] Among them, Ψ is a comprehensive analysis function that takes into account the evolution of the graph model in the time series, the behavior pattern dataset B, the time series of students' emotional states, and the changes in group density to predict future behavior patterns and potential abnormal behaviors.

[0062] Through in-depth analysis of the network characteristics and evolutionary trends of children's group behavior, and proposing a graph sequence equation to analyze the connectivity, clustering coefficient and path length of the graph model, the social structure and evolutionary laws of children's group behavior are revealed; by effectively solving the problem that traditional static models cannot accurately describe dynamic behavior, accurate prediction of group behavior evolutionary trends and abnormal behaviors is achieved.

[0063] To further illustrate the present invention, step S5 comprises the following steps:

[0064] S51, construct a deep neural network model to process and analyze the behavior pattern data set B obtained from steps S1 to S4, the deep neural network model is designed as a multi-layer architecture, including an input layer, multiple hidden layers and an output layer, to identify the complex patterns of student behavior, the deep neural network has L layers, each layer has n l neurons, for the multidimensional data of student behavior patterns, the input layer of the network receives the feature vector X extracted from the behavior pattern dataset B = [T xyz , D t , d ij , E f ], the output of layer 1 is expressed as:

[0065] O (l) =σ(W (l) O (l-1) +b (l) );

[0066] Among them, O (0) =X and W (l) is the weight matrix of the \(l\)th layer, \b (l) is the bias vector, σ is the activation function;

[0067] S52, dividing the behavior pattern data set B into a training set and a test set, the training set is used for training the model, and the test set is used for evaluating the performance and accuracy of the model;

[0068] S53, train the deep neural network, input the position parameters P of the student individuals i (t), velocity vector V i (t), emotional state E f,i (t) and social interaction parameter S ij (t) and environmental parameters such as population density D t and location characteristics Lk During the training process, the model minimizes the difference between the predicted results and the actual behavior pattern by adjusting the network weights. The cross entropy loss function is used to optimize the model during the training process. Combined with the classification problem of student behavior, the regularization term of the graph model characteristics is added to promote the model to learn social structure characteristics. The loss function is defined as:

[0069]

[0070] Where C(G) and L(G) represent the clustering coefficient and average path length of the graph model calculated from the student behavior pattern dataset predicted by the model, respectively. λ1 and λ2 are regularization coefficients used to balance different parts of the loss function. o,c Indicates the true situation that the observed value belongs to category c. It is an indicator function. When the observed value does belong to category c, it is 1, otherwise it is 0. log(p o,c ) is the probability (p o,c ) takes the logarithm, where (p o,c ) refers to the probability that the model predicts that the observation belongs to category c;

[0071] S54. Apply the theoretical basis and behavioral dynamics analysis provided by complex system theory to guide the structural design and training process of deep neural network models, and introduce the environmental impact function E(P i , L k ) and the interaction influence function I(S ij , E f,i , E f,j ), the dynamic behavior equation is expressed as:

[0072]

[0073] in, The influence of the environment on individual behavior. Represents the combined effect of interactions between individuals.

[0074] To further illustrate the present invention, the S6 comprises the following steps:

[0075] S61. After the deep neural network model training is completed, an abnormal behavior threshold T is set based on the model performance index obtained from step S5. ab The threshold is used to distinguish normal behavior patterns from abnormal behavior patterns. The optimal threshold is determined by calculating the F1 score of the model for abnormal behavior recognition and the area under the receiver operating characteristic curve:

[0076]

[0077] Among them, Precision(τ) and Recall(τ) are the precision and recall of the model at the threshold τ, respectively, and the threshold τ varies from 0 to 1;

[0078] S62. Real-time monitoring of student behavior data. Through sensors and video surveillance systems deployed on campus, students’ location, speed, emotional state, and social interaction data are continuously collected. Real-time data constitutes a real-time behavior feature vector:

[0079] X realtime =concat([T xyz (t), D t , d ij (t), E f (t)]);

[0080] Among them, concat represents the operation of concatenating data of each dimension into a high-dimensional feature vector, and each dimension is calculated from the real-time monitoring data through a specific function;

[0081] S63, the real-time behavior feature vector X realtime Input into the deep neural network model trained in step S5, the model outputs the classification probability of the behavior pattern of each real-time observation \(P_{ab}(X_{realtime})\), that is, the probability that the observation is judged as abnormal behavior;

[0082] S64. For each real-time observation, compare its behavior pattern classification probability P ab (X realtime ) and abnormal behavior threshold T ab , combining model uncertainty and classification probability:

[0083]

[0084] Among them, Uncertainty(X realtime ) represents the prediction uncertainty of the model for a given real-time feature vector, which can be estimated by the entropy of the model’s output probability distribution. λ is a tuning parameter used to balance the impact of prediction probability and uncertainty.

[0085] By deploying the model and training the behavior pattern data set, the deep neural network model is guided by the theory and behavior dynamics analysis provided by the complex system model of children's group behavior to identify common behaviors and abnormal behaviors in the behavior pattern data. A child behavior feature vector X is proposed. realtime =concat([T xyz (t), D t , d ij (t), E f(t)]) to construct data and propose an uncertain row discrimination formula IsAbnormal(X realtime ) is used to calculate entropy and balance the effects of probability and uncertainty on the prediction of children's abnormal behavior.

[0086] To further illustrate the present invention, the S7 comprises the following steps:

[0087] S71. For any abnormal behavior detected, the system automatically triggers an alarm mechanism and sends an alarm message to relevant campus administrators through the campus communication network, including the type, time, location, and anonymous identifier of the abnormal behavior. At the same time, the system provides preliminary response suggestions, which are classified according to the nature and urgency of the abnormal behavior.

[0088] S72. Establish an abnormal behavior log database to record all detected abnormal behavior events, including a detailed description of the event, time, location, and system response.

[0089] Beneficial effects of the present invention:

[0090] By introducing complex system theory, the present invention can comprehensively analyze and understand the dynamic changes of children's behavior and the complex interactions behind them, such as social interactions between individuals and interactions between individuals and the environment. At the same time, by using a deep neural network model, it not only improves the accuracy of behavioral pattern recognition, but also realizes real-time monitoring and analysis of children's behavior data.

[0091] The present invention can more accurately identify and predict individual behavior patterns by performing personalized analysis of each child's behavior data, taking into account each child's unique behavioral characteristics and background. This personalized analysis not only reduces the risk of false positives, but also provides the possibility of providing targeted intervention measures. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0093] The present invention will be further described below in conjunction with specific embodiments.

[0094] Example:

[0095] A campus abnormal behavior detection method based on complex theory includes the following steps:

[0096] S1. Collect student activity data on campus through video surveillance systems and sensors integrated with advanced algorithms, including three-dimensional motion trajectories, changes in group density, dynamic changes in distance between individuals, and emotional responses captured through micro-expression recognition technology; and transmit the collected activity data to the campus data processing center in real time; select key areas on campus with large traffic and frequent student activities, including library entrances, classrooms, and cafeterias, as target areas for data collection, and install video surveillance equipment with high resolution and wide viewing angle in each target area, so that the equipment can cover the entire area and capture student activities without blind spots;

[0097] S2, transform the collected student activity data into analyzable behavioral pattern data sets;

[0098] S3, performing noise filtering and data normalization on the behavior pattern dataset;

[0099] S4. Establish a complex system model to simulate the interaction between individual students and their interaction with the environment, and use graph analysis and dynamic system theory to reflect the evolution of student group behavior;

[0100] S5. Use the deep neural network model to train the behavior pattern data set, and guide the deep neural network model to identify common behavior patterns and abnormal behavior patterns in the behavior pattern data set through the theoretical basis and behavior dynamic analysis provided by the complex system model;

[0101] S6. Set an abnormal behavior threshold, and monitor student behavior data in real time based on the deep neural network model trained in step S5. Once the detected behavior data exceeds the abnormal behavior threshold, the system determines it as abnormal behavior;

[0102] S7. When the system detects abnormal behavior, it automatically sends alerts to campus administrators and provides relevant behavioral data and recommended intervention measures.

[0103] The step S1 comprises the following steps:

[0104] S11, the video monitoring system extracts the student's three-dimensional motion trajectory from the video stream in real time through an image recognition algorithm; the image recognition algorithm uses an inter-frame difference method to extract moving objects, applies an optical flow method to track the object's motion, and combines stereo vision technology to reconstruct the student's three-dimensional motion trajectory. xyz ;

[0105] Let the continuous video frames be F t and F t+1 , where t represents time, and the 3D position of the student in frame F t and F t+1 P t (x t ,yt , z t ) and P t+1 (x t+1 ,y t+1 , z t+1 );

[0106] Motion vector V t→t+1 Calculated by inter-frame difference and optical flow estimation:

[0107] V t→t+1 =P t+1 -P t =(x t+1 -x t ,y t+1 -y t , z t+1 -z t );

[0108] Combined with stereo vision, three-dimensional motion trajectory T xyz The calculation is:

[0109]

[0110] Where n is the total number of video frames;

[0111] S12. Integrate a crowd density analysis algorithm in the video surveillance system, segment the crowd area in the picture through image segmentation technology, calculate the number of pixels in each segmented area, and estimate the number of people in each area by the ratio of the number of pixels to the preset average pixel ratio of the human body. Combined with the number of pixels in each video frame F t Crowd segmentation area A m , m represents the mth area, and the number of people in each area is estimated by the number of pixels P in the area Am Compared with the preset average number of pixels per person P avg The change in population density is expressed as the sum of the number of people in all regions at each time point t:

[0112]

[0113] Population density change D t It is expressed as the sum of the number of people in all regions at each time point t:

[0114]

[0115] S13. Ultrasonic or infrared distance sensors are arranged in each target area of ​​the campus to collect the dynamic changes of the distance between individuals. The distance d between two students i and j at time t measured by ultrasonic or infrared sensors is ij (t) can be expressed as the round trip time T of the signal ijAnd the signal propagation speed c is calculated as:

[0116]

[0117] S14. Use high-definition cameras to capture students’ facial expressions and use micro-expression recognition technology to analyze students’ emotional reactions, including eyebrows rising and mouth corners drooping, to classify students’ basic emotional states. f .

[0118] The step S2 comprises the following steps:

[0119] S21, extracting the three-dimensional motion trajectory T from the data collected in step S1 xyz 、 Population density change D t , dynamic changes in the distance between individuals ij (t) and emotional state E f As raw data input;

[0120] S22. Create a data record for each student containing all activity data during the observation period. For each student i’s data record, consider its time series data X i Multidimensionality:

[0121] X i ={T xyz,i (t), D t,i , d ij,i (t), E f,i (t)};

[0122] Among them, t is the time variable, which represents the observation value at different time points t;

[0123] S23, for each data record X i The data is structured to reflect the behavior patterns of students in a specific time period. The continuous time point data is clustered into discrete behavior events. Each behavior event is defined by four dimensions: the student's position, the density of the group, the relative position to other students, and the emotional state. Dynamic time warping is used to identify and classify behavior patterns. For any two time series X i and X j It is expressed as:

[0124]

[0125] Where T is the length of the time series, x it and x jt They are the time series X i and X j The value at time t;

[0126] S24. The behavior pattern data sets of all students are combined into a comprehensive behavior pattern data set B for subsequent behavior pattern recognition and abnormal behavior detection:

[0127]

[0128] Where N is the number of participating students, ensuring that the dataset B covers all observed student behaviors on campus, and ∪ represents the merge operation, so that the dataset covers all observed student behaviors on campus;

[0129] S25. Define metadata description for Dataset B, including the time range, location, number of participating students, and timestamp and duration of each behavioral event. The purpose of metadata is to provide background information for the dataset, so that the data content can be quickly located and understood during subsequent analysis:

[0130] Metadata(B)={"Collection Period": P, "Locations": L, "Participants": N, "Event Timestamps": T e ,"Duration":D e};

[0131] Where P represents the time frame of data collection, L represents the location of data collection, N is the number of participating students, and T e and D e Represents the timestamp and duration of each behavioral event, respectively.

[0132] The step S4 comprises the following steps:

[0133] S41. Define the model parameters of the student individual and the environment. The state of the student individual i at time t is determined by the position P i (t) = (x i (t), y i (t), z i (t)), velocity vector V i (t)=(v x,i (t), v y,i (t), v z,i (t)), emotional state E f,i (t) and the social interaction parameter S with student j ij (t), the environmental parameters are composed of the population density D t and location-specific characteristics L k Represents, where k specifies the type of place;

[0134] S42. Construct a graph model G(V, E) to represent the interactions between individual students, where nodes V represent individual students, edges E represent social interactions between students, and edge weights w ij (t) is determined by the social interaction parameter \(S_{ij}(t)\), which reflects the intensity and frequency of social interactions between individuals:

[0135] w ij (t) = α·exp(-β·d ij (t))·E f,i (t)·E f,j (t);

[0136] Among them, α and β are adjustment parameters, d ij (t) is the distance between students, reflecting the interaction between the intensity of social interaction and emotional state;

[0137] S43. Using dynamic system theory, a dynamic behavior equation is defined for each student in the graphical model to simulate the time evolution of student behavior. The dynamic equation is based on the individual's position parameters, velocity vector, emotional state, and interaction with other individuals, and is formally expressed as:

[0138]

[0139]

[0140] Among them, U(P i (t) is a potential energy field that reflects the impact of the environment on the individual's movement, f inter is an influence function of interactions between individuals, which depends on social interaction parameters, emotional states, and location characteristics; represents a vector operator in vector calculus, called the gradient operator; d represents the differential operator, which is used to represent the small change in a function with respect to a variable;

[0141] S44. Using graph theory analysis methods, we analyze the connectivity, clustering coefficient and path length of the graph model G(V, E) to reveal the network characteristics and social structure of student group behavior, including the average path length L(G) and clustering coefficient C(G):

[0142]

[0143]

[0144] Among them, δ ij is the shortest path distance between nodes i and j, e ijk is the number of triangles connecting nodes i, j, and k, k i is the degree of node i;

[0145] S45. Integrate the results of dynamic system theory and graph theory analysis to predict the evolution trend of student group behavior and possible abnormal behavior patterns. By simulating different social interactions and environmental change scenarios, evaluate the impact of factors on the evolution of student group behavior:

[0146]

[0147] Among them, Ψ is a comprehensive analysis function that takes into account the evolution of the graph model in the time series, the behavior pattern dataset B, the time series of students' emotional states, and the changes in group density to predict future behavior patterns and potential abnormal behaviors.

[0148] The step S5 comprises the following steps:

[0149] S51, construct a deep neural network model to process and analyze the behavior pattern data set B obtained from steps S1 to S4, the deep neural network model is designed as a multi-layer architecture, including an input layer, multiple hidden layers and an output layer, to identify the complex patterns of student behavior, the deep neural network has L layers, each layer has n l neurons, for the multidimensional data of student behavior patterns, the input layer of the network receives the feature vector X extracted from the behavior pattern dataset B = [T xyz , D t , d ij , E f ], the output of layer 1 is expressed as:

[0150] O (l) =σ(W (l) O (l-1) +b (l) );

[0151] Among them, O (0) =X and W (l) is the weight matrix of the \(l\)th layer, \b (l) is the bias vector, σ is the activation function;

[0152] S52, dividing the behavior pattern data set B into a training set and a test set, the training set is used for training the model, and the test set is used for evaluating the performance and accuracy of the model;

[0153] S53, train the deep neural network, input the position parameters P of the student individuals i (t), velocity vector V i (t), emotional state E f,i (t) and social interaction parameter S ij (t) and environmental parameters such as population density D t and location characteristics L kDuring the training process, the model minimizes the difference between the predicted results and the actual behavior pattern by adjusting the network weights. The cross entropy loss function is used to optimize the model during the training process. Combined with the classification problem of student behavior, the regularization term of the graph model characteristics is added to promote the model to learn social structure characteristics. The loss function is defined as:

[0154]

[0155] Where C(G) and L(G) represent the clustering coefficient and average path length of the graph model calculated from the student behavior pattern dataset predicted by the model, respectively. λ1 and λ2 are regularization coefficients used to balance different parts of the loss function. o,c Indicates the true situation that the observed value belongs to category c. It is an indicator function. When the observed value does belong to category c, it is 1, otherwise it is 0. log(p o,c ) is the probability (p o,c ) takes the logarithm, where (p o,c ) refers to the probability that the model predicts that the observation belongs to category c;

[0156] S54. Apply the theoretical basis and behavioral dynamics analysis provided by complex system theory to guide the structural design and training process of deep neural network models, and introduce the environmental impact function E(P i , L k ) and the interaction influence function I(S ij , E f,i , E f,j ), the dynamic behavior equation is expressed as:

[0157]

[0158] in, The influence of the environment on individual behavior. Represents the combined effect of interactions between individuals.

[0159] The S6 comprises the following steps:

[0160] S61. After the deep neural network model training is completed, an abnormal behavior threshold T is set based on the model performance index obtained from step S5. ab The threshold is used to distinguish normal behavior patterns from abnormal behavior patterns. The optimal threshold is determined by calculating the F1 score of the model for abnormal behavior recognition and the area under the receiver operating characteristic curve:

[0161]

[0162] Among them, Precision(τ) and Recall(τ) are the precision and recall of the model at the threshold τ, respectively, and the threshold τ varies from 0 to 1;

[0163] S62. Real-time monitoring of student behavior data. Through sensors and video surveillance systems deployed on campus, students’ location, speed, emotional state, and social interaction data are continuously collected. Real-time data constitutes a real-time behavior feature vector:

[0164] X realtime =concat([T xyz (t), D t , d ij (t), E f (t)]);

[0165] Among them, concat represents the operation of concatenating data of each dimension into a high-dimensional feature vector, and each dimension is calculated from the real-time monitoring data through a specific function;

[0166] S63, the real-time behavior feature vector X realtime Input into the deep neural network model trained in step S5, the model outputs the classification probability of the behavior pattern of each real-time observation \(P_{ab}(X_{realtime})\), that is, the probability that the observation is judged as abnormal behavior;

[0167] S64. For each real-time observation, compare its behavior pattern classification probability P ab (X realtime ) and abnormal behavior threshold T ab , combining model uncertainty and classification probability:

[0168]

[0169] Among them, U ncertainty (X realtme ) represents the prediction uncertainty of the model for a given real-time feature vector, which can be estimated by the entropy of the model’s output probability distribution. λ is a tuning parameter used to balance the impact of prediction probability and uncertainty.

[0170] The S7 comprises the following steps:

[0171] S71. For any abnormal behavior detected, the system automatically triggers an alarm mechanism and sends an alarm message to relevant campus administrators through the campus communication network, including the type, time, location, and anonymous identifier of the abnormal behavior. At the same time, the system provides preliminary response suggestions, which are classified according to the nature and urgency of the abnormal behavior.

[0172] S72. Establish an abnormal behavior log database to record all detected abnormal behavior events, including a detailed description of the event, time, location, and system response.

Claims

1. A campus abnormal behavior detection method based on complex theory, characterized by: The following steps are involved: S1. Collect student activity data on campus through video surveillance systems and sensors integrated with advanced algorithms, including three-dimensional movement trajectories, changes in group density, dynamic changes in distance between individuals, and emotional reactions captured through micro-expression recognition technology; and transmit the collected activity data to the campus data processing center in real time; S2, transform the collected student activity data into analyzable behavioral pattern data sets; S3, performing noise filtering and data normalization on the behavior pattern dataset; S4. Establish a complex system model to simulate the interaction between individual students and their interaction with the environment, and use graph analysis and dynamic system theory to reflect the evolution of student group behavior; S5. Use the deep neural network model to train the behavior pattern data set, and guide the deep neural network model to identify common behavior patterns and abnormal behavior patterns in the behavior pattern data set through the theoretical basis and behavior dynamic analysis provided by the complex system model; S6. Set an abnormal behavior threshold, and monitor student behavior data in real time based on the deep neural network model trained in step S5. Once the detected behavior data exceeds the abnormal behavior threshold, the system determines it as abnormal behavior; S7. When the system detects abnormal behavior, it automatically sends alerts to campus administrators and provides relevant behavioral data and recommended intervention measures.

2. The campus abnormal behavior detection method based on complex theory according to claim 1 is characterized by: The step S1 comprises the following steps: S11, the video monitoring system extracts the student's three-dimensional motion trajectory from the video stream in real time through an image recognition algorithm; the image recognition algorithm uses an inter-frame difference method to extract moving objects, applies an optical flow method to track the object's motion, and combines stereo vision technology to reconstruct the student's three-dimensional motion trajectory. xyz ; Let the continuous video frames be F t and F t+1 , where t represents time, and the 3D position of the student in frame F t and F t+1 P t (x t ,y t , z t ) and P t+1 (x t+1 ,y t+1 , z t+1 ); Motion vector V t→t+1 Calculated by inter-frame difference and optical flow estimation: V t→t+1 =P t+1 -P t =(x t +1-x t ,y t+1 -y t ,z t+1 -z t ); Combined with stereo vision, three-dimensional motion trajectory T xyz The calculation is: Where n is the total number of video frames; S12. Integrate a crowd density analysis algorithm in the video surveillance system, segment the crowd area in the picture through image segmentation technology, calculate the number of pixels in each segmented area, and estimate the number of people in each area by the ratio of the number of pixels to the preset average pixel ratio of the human body. Combined with the number of pixels in each video frame F t Crowd segmentation area A m , m represents the mth area, and the number of people in each area is estimated by the number of pixels P in the area Am Compared with the preset average number of pixels per person P avg The change in population density is expressed as the sum of the number of people in all regions at each time point t: Population density change D t It is expressed as the sum of the number of people in all regions at each time point t: S13. Ultrasonic or infrared distance sensors are arranged in each target area of ​​the campus to collect the dynamic changes of the distance between individuals. The distance d between two students i and j at time t measured by ultrasonic or infrared sensors is ij (t) can be expressed as the round trip time T of the signal ij And the signal propagation speed c is calculated as: S14. Use high-definition cameras to capture students’ facial expressions and use micro-expression recognition technology to analyze students’ emotional reactions, including eyebrows rising and mouth corners drooping, to classify students’ basic emotional states. f .

3. The campus abnormal behavior detection method based on complex theory according to claim 1 is characterized by: The step S2 comprises the following steps: S21, extracting the three-dimensional motion trajectory T from the data collected in step S1 xyz , Population density change D t , dynamic changes in the distance between individuals ij (t) and emotional state E f As raw data input; S22. Create a data record for each student containing all activity data during the observation period. For each student i’s data record, consider its time series data X i Multidimensionality: X i ={T xyz,i (t),D t,i ,d ij,i (t),E f,i (t)}; Among them, t is the time variable, which represents the observation value at different time points t; S23, for each data record X i The data is structured to reflect the behavior patterns of students in a specific time period. The continuous time point data is clustered into discrete behavior events. Each behavior event is defined by four dimensions: the student's position, the density of the group, the relative position to other students, and the emotional state. Dynamic time warping is used to identify and classify behavior patterns. For any two time series X i and X j It is expressed as: Where T is the length of the time series, x it and x jt They are the time series X i and X j The value at time t; S24. The behavior pattern data sets of all students are combined into a comprehensive behavior pattern data set B for subsequent behavior pattern recognition and abnormal behavior detection: Where N is the number of participating students, ensuring that the dataset B covers all observed student behaviors on campus, and ∪ represents the merge operation, so that the dataset covers all observed student behaviors on campus; S25. Define metadata description for Dataset B, including the time range, location, number of participating students, and timestamp and duration of each behavioral event. The purpose of metadata is to provide background information for the dataset, so that the data content can be quickly located and understood during subsequent analysis: Metadata(B)={″Collection Period″:P,″Locations″:L,″Participants″:N,″EventTimestamps″:T e ,″Duration″:D e }; Where P represents the time frame of data collection, L represents the location of data collection, N is the number of participating students, and T e and D e Represents the timestamp and duration of each behavioral event, respectively.

4. The campus abnormal behavior detection method based on complex theory according to claim 1 is characterized by: The step S4 comprises the following steps: S41. Define the model parameters of the student individual and the environment. The state of the student individual i at time t is determined by the position P i (t) = (x i (t), y i (t), z i (t)), velocity vector V i (t)=(v x,i (t), v y,i (t), v z,i (t)), emotional state E f,i (t) and the social interaction parameter S with student j ij (t), the environmental parameters are composed of the population density D t and location-specific characteristics L k Represents, where k specifies the type of place; S42. Construct a graph model G(V, E) to represent the interactions between individual students, where nodes V represent individual students, edges E represent social interactions between students, and edge weights w ij (t) is determined by the social interaction parameter \(S_{ij}(t)\), which reflects the intensity and frequency of social interactions between individuals: w ij (t)=α·exp(-β·d ij (t))·E f,i (t)·E f,j (t); Among them, α and β are adjustment parameters, d ij (t) is the distance between students, reflecting the interaction between the intensity of social interaction and emotional state; S43. Using dynamic system theory, a dynamic behavior equation is defined for each student in the graphical model to simulate the time evolution of student behavior. The dynamic equation is based on the individual's position parameters, velocity vector, emotional state, and interaction with other individuals, and is formally expressed as: Among them, U(P i (t) is a potential energy field that reflects the impact of the environment on the individual's movement, f inter is an influence function of interactions between individuals, which depends on social interaction parameters, emotional states, and location characteristics; represents a vector operator in vector calculus, called the gradient operator; d represents the differential operator, which is used to represent the small change in a function with respect to a variable; S44. Using graph theory analysis methods, we analyze the connectivity, clustering coefficient and path length of the graph model G(V, E) to reveal the network characteristics and social structure of student group behavior, including the average path length L(G) and clustering coefficient C(G): Among them, δ ij is the shortest path distance between nodes i and j, e ijk is the number of triangles connecting nodes i, j, and k, k i is the degree of node i; S45. Integrate the results of dynamic system theory and graph theory analysis to predict the evolution trend of student group behavior and possible abnormal behavior patterns. By simulating different social interactions and environmental change scenarios, evaluate the impact of factors on the evolution of student group behavior: Among them, Ψ is a comprehensive analysis function that takes into account the evolution of the graph model in the time series, the behavior pattern dataset B, the time series of students' emotional states, and the changes in group density to predict future behavior patterns and potential abnormal behaviors.

5. The campus abnormal behavior detection method based on complex theory according to claim 1 is characterized by: The step S5 comprises the following steps: S51, construct a deep neural network model to process and analyze the behavior pattern data set B obtained from steps S1 to S4, the deep neural network model is designed as a multi-layer architecture, including an input layer, multiple hidden layers and an output layer, to identify the complex patterns of student behavior, the deep neural network has L layers, each layer has n l neurons, for the multidimensional data of student behavior patterns, the input layer of the network receives the feature vector X extracted from the behavior pattern dataset B = [T xyz , D t , d ij , E f ], the output of layer 1 is expressed as: The (l) =σ(W (l) The (l-1) +b (l) ); Among them, O (0) =X and W (l) is the weight matrix of the \(l\)th layer, \b (l) is the bias vector, σ is the activation function; S52, dividing the behavior pattern data set B into a training set and a test set, the training set is used for training the model, and the test set is used for evaluating the performance and accuracy of the model; S53, train the deep neural network, input the position parameters P of the student individuals i (t), velocity vector V i (t), emotional state E f,i (t) and social interaction parameter S ij (t) and environmental parameters such as population density D t and location characteristics L k During the training process, the model minimizes the difference between the predicted results and the actual behavior pattern by adjusting the network weights. The cross entropy loss function is used to optimize the model during the training process. Combined with the classification problem of student behavior, the regularization term of the graph model characteristics is added to promote the model to learn social structure characteristics. The loss function is defined as: Where C(G) and L(G) represent the clustering coefficient and average path length of the graph model calculated from the student behavior pattern dataset predicted by the model, respectively. λ1 and λ2 are regularization coefficients used to balance different parts of the loss function. o,c Indicates the true situation that the observed value belongs to category c. It is an indicator function. When the observed value does belong to category c, it is 1, otherwise it is 0. log(p o,c ) is the probability (p o,c ) takes the logarithm, where (p o,c ) refers to the probability that the model predicts that the observation belongs to category c; S54. Apply the theoretical basis and behavioral dynamics analysis provided by complex system theory to guide the structural design and training process of deep neural network models, and introduce the environmental impact function E(P i , L k ) and the interaction influence function I(S ij , E f,i , E f,j ), the dynamic behavior equation is expressed as: in, The influence of the environment on individual behavior. Represents the combined effect of interactions between individuals.

6. The campus abnormal behavior detection method based on complex theory according to claim 1 is characterized by: The S6 comprises the following steps: S61. After the deep neural network model training is completed, an abnormal behavior threshold T is set based on the model performance index obtained from step S5. ab The threshold is used to distinguish normal behavior patterns from abnormal behavior patterns. The optimal threshold is determined by calculating the F1 score of the model for abnormal behavior recognition and the area under the receiver operating characteristic curve: Among them, Precision(τ) and Recall(τ) are the precision and recall of the model at the threshold τ, respectively, and the threshold τ varies from 0 to 1; S62. Real-time monitoring of student behavior data. Through sensors and video surveillance systems deployed on campus, students’ location, speed, emotional state, and social interaction data are continuously collected. Real-time data constitutes a real-time behavior feature vector: X realtime =concat([T xyz (t),D t ,d ij (t),E f (t)]); Among them, concat represents the operation of concatenating data of each dimension into a high-dimensional feature vector, and each dimension is calculated from the real-time monitoring data through a specific function; S63, the real-time behavior feature vector X realtime Input into the deep neural network model trained in step S5, the model outputs the classification probability of the behavior pattern of each real-time observation \(P_{ab}(X_{realtime})\), that is, the probability that the observation is judged as abnormal behavior; S64. For each real-time observation, compare its behavior pattern classification probability P ab (X realtime ) and abnormal behavior threshold T ab , combining model uncertainty and classification probability: Among them, Uncertainty(X realtime ) represents the prediction uncertainty of the model for a given real-time feature vector, which can be estimated by the entropy of the model’s output probability distribution. λ is a tuning parameter used to balance the impact of prediction probability and uncertainty.

7. The campus abnormal behavior detection method based on complex theory according to claim 1 is characterized by: The S7 comprises the following steps: S71. For any abnormal behavior detected, the system automatically triggers an alarm mechanism and sends an alarm message to relevant campus administrators through the campus communication network, including the type, time, location, and anonymous identifier of the abnormal behavior. At the same time, the system provides preliminary response suggestions, which are classified according to the nature and urgency of the abnormal behavior. S72. Establish an abnormal behavior log database to record all detected abnormal behavior events, including a detailed description of the event, time, location, and system response.

Citation Information

Cited By

  • Video monitoring method and system for visual operation

    CN119600547A

  • Social engineering attack detection method based on context awareness

    CN119652639A