A method for intelligently identifying bullying behavior in a campus monitoring video

Abstract

The present application relates to the technical field of bullying behavior intelligent identification, and particularly relates to a campus monitoring video bullying behavior intelligent identification method.The method comprises the following steps: extracting bullying history videos of a campus monitoring control center, marking the behaviors of bullies, and generating behavior marking data; based on the marking data, pre-compression behaviors are inductively learned, so that pre-compression behaviors of bullying are identified, and their escalation trends are analyzed; subsequently, the trends are used to evaluate the confidence interval of bullying compression escalation behaviors, and the possibility interval thereof is output; finally, the evaluation results are used to perform information gain memory learning on the pre-compression behaviors of bullying, an automatic alarm architecture is designed, and alarm information is fed back to a control terminal.The present application optimizes the intelligent identification technology of bullying behaviors, so that the intelligent identification technology of bullying behaviors is more perfect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bullying behavior intelligent identification, and particularly relates to a bullying behavior intelligent identification method for campus monitoring video. BACKGROUND

[0002] With the development of big data and machine learning technology, the bullying behavior intelligent identification method based on monitoring video gradually has the ability of self-learning and self-optimization. The system can constantly update and optimize the identification algorithm in the process of processing more video data, and enhance its identification ability for different types of bullying behavior. Combined with information gain memory learning and intelligent alarm mechanism, the system can automatically issue a warning when it detects the precursor of bullying behavior in real time, and timely feedback to the campus security or management department to ensure the safety and harmony of the campus environment. However, the traditional bullying behavior intelligent identification method for campus monitoring video has the problem of inaccurate analysis of bullying oppression precursor behavior, resulting in large bullying behavior intelligent identification error. SUMMARY

[0003] Therefore, it is necessary to provide a bullying behavior intelligent identification method for campus monitoring video to solve at least one of the above technical problems.

[0004] To achieve the above purpose, a bullying behavior intelligent identification method for campus monitoring video, the method comprising the following steps:

[0005] Step S1: extracting bullying behavior history video from the campus monitoring control center; marking the behavior of the bully according to the bullying behavior history video, and outputting the bully behavior marking data;

[0006] Step S2: inducing learning of the precursor oppression behavior according to the bully behavior marking data to obtain the bullying precursor oppression behavior; analyzing the oppression behavior escalation trend according to the bullying precursor oppression behavior;

[0007] Step S3: performing confidence interval evaluation of the bullying oppression escalation behavior based on the oppression behavior escalation trend to output the escalation trend behavior confidence interval;

[0008] Step S4: performing information gain memory learning on the bullying precursor oppression behavior based on the escalation trend behavior confidence interval, then designing an automatic alarm architecture, and feeding back to the control terminal to perform bullying behavior intelligent identification.

[0009] The present application has the beneficial effect that by extracting historical videos of bullying behavior from the campus monitoring control center and marking the behavior of bullies, the system can automatically identify and record the occurrence of bullying behavior. Historical video data can be systematically organized to provide detailed marker data for subsequent analysis. Traditional manual inspection methods often fail to comprehensively cover every potential bullying event in the campus, while automated behavior marking makes monitoring more efficient and reduces human annotation errors and biases. These data not only provide a basis for identifying existing bullying behavior, but also help establish various types of bullying behavior patterns, providing valuable references for further analysis. By accurately marking the behavior of bullies, the system can effectively analyze the evolution trend of oppressive behavior in subsequent steps and conduct targeted prevention and intervention. Based on the bullying behavior marker data, the system can identify the oppressive behavior actions that may occur before the bullying behavior actions through inductive learning of pre-oppressive behavior. The core of this step is to identify the premonitory signs of oppressive behavior actions through machine learning models, providing a theoretical basis for predicting the occurrence of bullying incidents. Traditional bullying identification methods usually focus on direct violent behavior, ignoring the pre-oppressive behavior. By learning pre-oppressive behavior, the system can identify potential violence risks in advance and avoid passive response to bullying behavior. The process of inductive learning can analyze historical data to identify common characteristics of oppressive behavior, thereby effectively improving prediction accuracy in practical applications. Through in-depth mining of oppressive behavior, schools can take intervention measures earlier to prevent school violence incidents. Based on the escalation trend of oppressive behavior, the system can quantify and predict the escalation trend of oppressive behavior through confidence interval evaluation of bullying behavior. Through statistical methods, a data-based evaluation standard is provided for campus administrators to help judge the possibility of behavior escalation. This evaluation not only makes the bullying behavior warning more accurate, but also predicts future bullying incidents based on historical behavior patterns, allowing for early intervention and reducing potential risks and harm. At the same time, the confidence interval evaluation can provide different possibilities of escalating behavior, allowing for corresponding preventive measures at different risk levels. This method breaks through the limitations of traditional methods by digitally and quantitatively warning of bullying in schools, contributing to the precise management and intervention of school violence. Based on the escalation trend behavior confidence interval, information gain memory learning is performed on the pre-oppressive behavior of bullying, thereby optimizing the learning ability of the model and designing an automatic alarm architecture. It can continuously update and optimize itself according to new data, improving the identification accuracy and response speed of bullying behavior. Information gain memory learning enables the system to automatically update and strengthen the memory of risky behavior when dealing with new oppressive behavior, allowing the system to maintain high efficiency in identifying changing bullying patterns. At the same time, the design of the automatic alarm architecture ensures that information can be fed back to the campus monitoring terminal in real time when the premonitory signs of bullying behavior are discovered.Through the intelligent early warning mechanism, the school manager can take immediate measures to prevent further escalation of bullying incidents. The introduction of the automatic alarm system greatly reduces the time and cost of manual intervention, improves the processing efficiency, and ultimately provides an efficient and sustainable technical solution for campus safety management. Therefore, the present application is an optimized processing of the traditional school monitoring video bullying behavior intelligent identification method, which solves the problem of inaccurate analysis of bullying oppression pre-behavior in the traditional school monitoring video bullying behavior intelligent identification method, thereby causing large intelligent identification error of bullying behavior, improves the accuracy of bullying oppression pre-behavior analysis, and reduces the intelligent identification error of bullying behavior. BRIEF DESCRIPTION OF DRAWINGS

[0010] Figure 1 It is a step flowchart of a school monitoring video bullying behavior intelligent identification method.

[0011] Figure 2 It is Figure 1 The detailed implementation step flowchart of step S2 in the embodiment. DETAILED DESCRIPTION

[0012] Please refer to Figure 1 A school monitoring video bullying behavior intelligent identification method, the method comprises the following steps:

[0013] Step S1: extracting bullying behavior history video from the campus monitoring control center; marking the bullying behavior of the bully according to the bullying behavior history video, and outputting the bullying behavior marking data of the bully;

[0014] In the embodiment of the present application, the video index management subsystem is called in the data storage module of the campus monitoring control center, and the video data containing the determined historical bullying event is extracted in time sequence and place number order. After extraction, all video frames are frame-level split according to a sampling interval of 3 seconds, and the dynamic particle noise in each frame picture is removed by using a median filtering algorithm based on time domain noise energy constraint. The algorithm obtains the denoising video with smooth visual signal by calculating the pixel gray change rate between consecutive frames and constraining the non-structural random fluctuation. After denoising, an interpolation algorithm based on optical flow field motion compensation is used to generate interpolation frames between each adjacent two frames to make up for the insufficient frame rate or fuzzy motion area, forming a bullying behavior compensation video with time sequence continuity. The compensation video is input into a convolution layer to extract key point features, a skeleton recognition algorithm based on differential pose estimation is used to label key joint points of individuals appearing in each picture, potential aggressors and victims are identified through body pose scale proportion and motion trajectory change, and bullying behavior marker data is automatically generated according to action continuity and spatial proximity. The data is output in the form of time stamp, coordinate point sequence and behavior classification label, providing a structured data basis for subsequent learning steps.

[0015] Step S2: Preceding oppression behavior induction learning is performed according to the bullying behavior marker data to obtain a bullying preceding oppression behavior; and an oppression behavior escalation trend is analyzed according to the bullying preceding oppression behavior;

[0016] In the embodiment of the present application, an action aggregation algorithm based on pose approximation threshold calculation is used to measure the approximation space of body pose changes in each time period, and the body pose approximation behavior is identified by calculating the Euclidean distance convergence trend between the body center of gravity and the target individual in three dimensions. Then, the audio signal is separated from the denoising video, the short-time Fourier transform is used to perform spectral decomposition on continuous speech segments, and the sound pressure peak value and speech speed amplitude change features are extracted. The pose approximation behavior features and audio aggressiveness features are input into a preceding behavior induction network based on hidden layer state clustering, and the bullying preceding oppression behavior is obtained by behavior sequence clustering analysis. In order to further analyze the stage change trend between behaviors, a time sliding window clustering analysis method is used to construct a preceding oppression behavior transition sequence, the behavior form change proportion in adjacent time periods is measured by gradient, the continuous difference trend of time sequence features is calculated, and the escalation trend of oppression behavior is obtained.

[0017] In another embodiment, based on the bully behavior label data output in step S1, by time period division of the behavior label sequence, the segment with action change intensity exceeding the preset threshold in each segment label is analyzed as a pre-pressing candidate segment. In order to extract the pre-pressing behavior features, first, the posture approximation behavior analysis is performed on the label sequence of the candidate segment, wherein the posture approximation degree of each frame is quantified according to the relative position difference between the key points, and then the quantification result is continuously tracked in the time dimension, thereby obtaining the bully posture approximation behavior sequence. On this basis, the bully audio corresponding to the time period is synchronously extracted from the original historical video, and after the audio signal is time-aligned, it is input into the behavior induction learning process. In the behavior induction learning, the peak change segment of the posture approximation behavior sequence is identified as the body approaching action, and the audio signal with the attack tendency is identified based on the instantaneous intensity change. Then, the body action sequence is clustered, and the segments with similar posture change patterns are clustered into multiple pre-pressing behavior clusters. By expanding the analysis on the time advancing mode of the action in each cluster, the escalation tendency features of the behavior pattern are obtained. Then, according to the action change span, action extension direction and audio attack amplitude between the clusters, the pressing behavior escalation trend sequence is constructed. The trend sequence is arranged in time sequence, which is used to reflect the gradual change process from mild pressing, approaching action to obvious attack tendency action, thereby forming the pressing behavior escalation trend.

[0018] Step S3: performing confidence interval evaluation of the bully pressing escalation behavior based on the pressing behavior escalation trend, to output an escalation trend behavior confidence interval;

[0019] In the embodiment of the application, the obtained pressing behavior escalation trend sequence is evaluated for confidence. First, a feature matrix of the escalation trend is constructed by a time sequence pattern recognition algorithm, and the matrix elements are time correlation values between different behavior stages. The matrix is convoluted by using a multi-scale sliding window convolution calculation method, the correlation convolution response of the trend sequence is calculated in a fixed step window, and the continuous change pulse signals between different behavior patterns are extracted. The convolution data is statistically analyzed for confidence, an interval evaluation algorithm based on mean shift estimation is used to calculate the concentration and dispersion of the overall trend sequence, and the fluctuation range of the trend distribution under multiple iterations is obtained by a pseudo-random disturbance repeated sampling method. According to the obtained sample distribution boundary, the upper and lower limits of the confidence interval of the bully pressing escalation behavior are determined, and the confidence interval evaluation result of the escalation trend is formed. The evaluation result is used to quantify the statistical boundary of the development of the pressing behavior to the high-risk stage, and provides a stable weight interval basis for the adaptive learning stage.

[0020] In another embodiment, the escalation trend of the compression behavior obtained in step S2 is taken as input, and a time sequence pattern structure of the escalation trend is constructed by performing time sequence pattern recognition on the trend sequence. The construction of the structure adopts a continuous segment division method based on frame-level change amplitude, splits the trend sequence into multiple time sub-segments with different action promotion strengths, and classifies the trend increase amplitude in each sub-segment to form a trend time sequence pattern composed of multiple time sequence segments. Then, convolution processing is performed on the trend time sequence pattern to strengthen the trend increase amplitude feature. The convolution processing starts from the first segment of the time sequence pattern with a fixed window length, and slides segment by segment. The trend change value in each window is point-by-point corresponding to the preset trend direction template, and the trend pattern convolution data is formed by numerical convolution, so that the trend change strength of each time segment presents more prominent differences after convolution. After obtaining the trend pattern convolution data, the convolution data is evaluated in a confidence interval. In the evaluation process, the amplitude gradient of the trend change, the length of the continuous rising segment and the fluctuation density are taken as the input of the confidence evaluation, so as to comprehensively calculate the stability degree of the trend segment, the upward promotion strength of the trend and the influence degree of the fluctuation on the overall trend. The numerical results of multiple continuous trends form the confidence interval data. The confidence interval represents the stability degree of the escalation behavior at different trend stages and the certainty degree of the trend change. Finally, the escalation trend behavior confidence interval is output, which provides a basis for weighted processing for subsequent information gain memory learning of the pre-pressing behavior.

[0021] Step S4: performing information gain memory learning on the bullying pre-pressing behavior based on the escalation trend behavior confidence interval, then designing an automatic alarm architecture, and feeding back to a control terminal to perform intelligent identification of bullying behavior;

[0022] In the embodiment of the application, the pre-bullying oppression behavior generated in step S2 is information gain memory learning with the escalation trend behavior confidence interval obtained in step S3 as the weighted basis. First, the pre-bullying oppression behavior is normalized. The normalization uses interval scaling method to directly convert the amplitude change and time sequence position of each behavior data to a unified scale, so that different categories of action change sequences participate in subsequent learning under a unified scale. Then, the normalized sequence and the escalation trend behavior confidence interval are aligned according to the time period, and the normalized sequence is processed by dynamic weighted sampling. In the weighted sampling, the time period with a higher trend confidence is given a higher weight during sampling to enhance the influence of key oppression action fragments. The weighted sequence is divided into multiple time subsegments of fixed sliding window length, and the segmented information gain is calculated in each sliding window. Through joint analysis of the behavior change amplitude, action direction change and action continuity in the sliding window, a local information gain vector corresponding to the sliding window is obtained. The information gain vector is multiplied point by point with the activation values of the forget gate, input gate and output gate in the long short-term memory structure, and then a scaling coefficient for controlling the injection degree is added, so that the sliding window gain information is injected into the hidden state sequence. The hidden state sequence is iteratively trained in the time dimension to gradually accumulate high-gain behavior characteristics and suppress low-gain redundancy, thereby forming escalation trend learning data between pre-bullying oppression behaviors. Finally, an automatic alarm architecture is constructed according to the learning data, the alarm threshold and the trend learning result are combined to form an alarm condition judgment rule, and the rule is output to the campus monitoring control terminal for execution of intelligent identification and timely alarm of bullying behavior.

[0023] The construction of the automatic alarm architecture is specifically: after the generation of the escalation trend learning data, the increasing sequence of oppression actions, the cumulative sequence of tone aggressiveness and the sequence of posture approaching acceleration contained in the learning data are mapped into an alarm trigger matrix. The trigger matrix is composed of three types of sub-matrices, including an action increasing sub-matrix, an audio attack sub-matrix and a posture acceleration sub-matrix, each sub-matrix generates an independent threshold interval according to the feature change direction, change amplitude and change duration in the time window. Then, the three sub-matrices are cascaded by an automatic rule generation algorithm to encode the cross-modal trend association relationship in the form of a directed trigger chain, forming a joint trigger graph for alarm judgment. In the joint trigger graph, a threshold crossing detection algorithm is used to scan the node sequence. When it is detected that the action progression node, the tone attack node and the posture acceleration node complete cross triggering in the same sliding window, the trigger state is written into the alarm state register, and an alarm instruction data packet is generated. Finally, the automatic alarm architecture transmits the alarm instruction data packet to the control terminal via the control bus to complete the bullying behavior identification feedback.

[0024] It needs to be explained: by extracting historical bullying video from the campus monitoring control center and marking the behavior, the structured processing of data is realized, and the basic data source after visual feature extraction and time labeling is provided for subsequent analysis. Then, step S2 uses the labeled behavior data to recognize the oppression signs before the violent behavior through multi-modal fusion learning of posture, action and voice features, and extracts the escalation path of the oppression behavior according to the behavior clustering and time trend analysis, realizes the transition from single action recognition to trend understanding. After entering step S3, the statistical confidence interval evaluation is carried out on the obtained escalation trend of the oppression behavior, the reliable range of the trend is determined through the time sequence mode convolution and interval estimation method, and the risk boundary with probability confidence is formed. Finally, step S4 carries out information gain memory learning on the pre-oppression behavior under the guidance of the confidence interval, deepens the long-term memory of the high-risk behavior characteristics of the model through dynamic weighting, time sequence sliding window and gate memory mechanism, and designs an automatic alarm architecture based on the learning result, feeds back the recognition signal to the monitoring terminal, and realizes the optimization of the early warning ability.

[0025] Step S1 includes the following steps:

[0026] Step S11: extracting bullying behavior history video from the campus monitoring control center;

[0027] Step S12: video denoising is performed on the bullying behavior history video to obtain a bullying behavior denoising video;

[0028] Step S13: compensating and inserting frames on the bullying behavior denoising video to obtain a bullying behavior compensation video;

[0029] Step S14: bullying behavior labeling is performed according to the bullying behavior compensation video, and bullying behavior labeling data is output.

[0030] In the embodiment of the application, first, the video storage index table is called through the centralized management system of the campus monitoring control center, and the historical video segment is extracted according to the three-dimensional index mode of event occurrence date, monitoring camera number and area number. The extraction process adopts a multi-channel video data retrieval algorithm based on time synchronization identification, integrates the frame sequence corresponding to the synchronization time stamp of each camera, and ensures that the video data of different angles is strictly aligned on the time axis. After all the video data is extracted into the cache queue, a fixed frame rate of 30 frames per second is used for frame-by-frame analysis, and the gray histogram of each frame is executed. The frames with a brightness saturation lower than the threshold are removed to prevent extreme light from causing the feature extraction to deviate. The analyzed data is stored in the video frame database in order, and each frame data is attached with a time label, a camera identification number and an area coordinate index, which is used for subsequent video quality preprocessing and target behavior analysis.

[0031] The bullying behavior history video of step S11 is processed by using a double-layer filtering algorithm based on spatio-temporal joint constraint. The first layer uses median filtering in the time domain, selects two frames before and after in the frame sequence, and performs median calculation on the pixel gray level change rate in the current frame to weaken the instantaneous noise interference; the second layer uses spatial domain guided filtering to perform weight distribution on the structure boundary preservation factor in the local neighborhood of the pixel to retain the human outline and action details. After completing the double-layer denoising, the time sequence consistency detection is performed on the full frame data, the smoothness is judged by calculating the mean difference of the optical flow of adjacent frames, and the abnormal frame is replaced and corrected. The video sequence output after processing by the algorithm is the bullying behavior denoising video, which has balanced image signal and reduced background interference, and provides a basis for subsequent frame insertion compensation operation.

[0032] The frame insertion algorithm based on bidirectional optical flow field estimation compensates and inserts the bullying behavior denoising video. First, the motion vector field between pixel points of every two frames of video is calculated, and forward and reverse motion vectors are generated by the direction and speed of the optical flow. Then, the intermediate frame is interpolated between adjacent frames according to the motion vector field to ensure the continuity of the motion trajectory. In order to avoid optical flow distortion, boundary pruning operation is performed on the extreme points in the motion vector field to make the interpolation position keep smooth and incremental distribution. The generated intermediate frame and the original video frame are re-sequenced in time sequence to obtain the bullying behavior compensation video with complete frame rate and natural action transition, and then the linear interpolation correction algorithm is used to smooth the color difference of the picture brightness to prevent flicker at the fusion position of the front and rear frames. After this process, the video frame sequence is uniform in time sequence and stable in frame rate, and can be directly used for subsequent behavior recognition labeling.

[0033] The human pose estimation algorithm is used to detect the key points of each individual in the video, including the coordinates of the head, shoulder, elbow, wrist, knee and ankle. Then, the behavior recognition algorithm based on the change amplitude of the action and the time sequence continuity is used to analyze the speed, angle, acceleration and other dynamic parameters of the key points of each individual in consecutive frames. The algorithm distinguishes the aggressor from the victim by time sequence similarity clustering, locks the aggressor individual by the stable contraction mode of the body approaching speed and spatial distance, and extracts the behavior label by combining the direction and action aggressiveness. The finally generated bullying behavior label data is stored in a structured form, each record contains a timestamp, an individual identification number, a posture motion sequence and a behavior category code, which provides accurate input for the subsequent pre-compression behavior induction learning step.

[0034] Reference Figure 2 As described above, step S2 includes the following steps:

[0035] Step S21: performing posture approaching behavior analysis on the bullying behavior label data to obtain bullying posture approaching behavior;

[0036] In the embodiment of the present application, the bully behavior marker data output in step S14 is input into the posture approximation analysis process. First, according to the skeleton key point sequence of each individual in the marker, the joint three-dimensional coordinate trajectory is constructed, and the spatial contraction ratio of the body posture is calculated through the Euclidean distance between the key points and the angular velocity change. Then, taking the body center of gravity projection point at each time as a reference, the center of gravity offset of the previous and subsequent frames is calculated by difference, and the motion trend of the body approaching the target individual is obtained. The convergence speed change is calculated by the posture contraction curve in multiple time periods, and the approximation degree of the individual and the target object is judged. The process uses a time sliding window to traverse each frame of data, calculates the average approximation speed and distance change rate in the window. If the approximation rate remains increasing in multiple consecutive windows, it is determined as a posture approximation behavior. All calculation results are output in the form of a time sequence to form a posture approximation behavior, wherein each record contains the body joint motion direction, the center of gravity offset value and the spatial distance relative to the target, so as to be called in subsequent oppression behavior induction learning.

[0037] Step S22: extracting bully audio in the bully behavior history video;

[0038] In the embodiment of the present application, the audio signal is separated from the extracted history bully behavior video. The extraction process uses a synchronous separation method to split the video data stream into image frame signals and audio frame signals, and records all audio samples through a fixed sampling rate. The audio stream is processed by short-time framing, each frame is 25 milliseconds long, and the frame shift is 10 milliseconds. The windowing operation is performed on each frame to suppress the truncation effect. Then, each frame is converted to the frequency domain using short-time Fourier transform, and the energy distribution and sound pressure peak value are extracted to form an audio feature matrix. In order to enhance the accuracy of the tone feature recognition, the high-energy sub-band in the spectrum is filtered to remove environmental noise and smoothed, and then the zero-crossing rate and fundamental frequency energy ratio are calculated to judge whether the speech segment contains aggressive tone. Finally, the continuous acoustic parameters are integrated into a time-aligned sequence, the time stamp is kept synchronous with the posture behavior time sequence, and the basis for subsequent induction of audio features and posture features is provided.

[0039] Step S23: pre-oppression behavior induction learning according to the bully posture approximation behavior and the bully audio to obtain a bully pre-oppression behavior;

[0040] In the embodiment of the present application, in the implementation process of step S23, the posture approximation behavior data obtained in step S21 and the audio feature sequence obtained in step S22 are jointly calculated. First, the two types of features are aligned according to the time stamp, and the smoothing and completion of different synchronization sections are completed through linear interpolation. Then, the change rate of each joint node trajectory in the posture feature is standardized, and four types of core indexes such as the approximation space reduction rate, the upper limb opening angle, the trunk tilt angle and the leg support ratio are extracted; at the same time, the high sound pressure band energy increment and the occurrence rate of repetitive derogatory words in the audio sequence are counted as the speech attack strength index. The relative distance difference of the motion trajectory from the shoulder to the chest and from the arm to the head is calculated, and the speech attack strength time curve is combined to form a joint feature vector. In the joint feature space, the time similarity aggregation algorithm is used to identify the behavior combination features, and the segments with strong spatial approximation and synchronous features of the speech attack in the continuous time period are extracted, which are summarized as the pre-bullying oppression behavior. The summarized data is output in the form of time start and end frame number, corresponding action type and speech energy value, which provides ordered input for subsequent clustering analysis.

[0041] Step S24: performing behavior clustering processing on the pre-bullying oppression behavior to obtain a pre-oppression clustering behavior;

[0042] In the embodiment of the present application, the pre-oppression behavior data obtained in the previous step is subjected to clustering analysis. First, all behavior samples are combined into a feature matrix according to the time period, and the matrix row represents different behavior instances, and the list represents the extracted behavior parameters in each dimension, including the posture trajectory angle change, the center of gravity offset rate, the speech intensity increment and the interaction time interval. After eliminating the order of magnitude difference through feature standardization processing, the correlation degree between behaviors is calculated by using the hierarchical clustering algorithm based on time similarity measure, and a distance matrix between behaviors is constructed. Then, the distance matrix threshold is used for step-by-step aggregation, and several pre-oppression behavior clusters are formed. In order to ensure the time sequence consistency of the clustering results, the time continuity constraint is added when merging adjacent behavior clusters, and the pre-oppression clustering behavior data containing cluster number and sample time period are output after clustering.

[0043] Step S25: analyzing the oppression behavior escalation trend according to the pre-oppression clustering behavior.

[0044] In the embodiment of the present application, the pre-pressing cluster behavior data output in step S24 is taken as input, and a behavior transition sequence is constructed in time sequence. By calculating the difference vector between the adjacent cluster center features, the behavior change gradient is obtained. The gradient is used to analyze the trend direction of the change over time, smooth the short-term fluctuations by moving average curve, and extract the overall rising and falling segments. The rising segment is subjected to segmented slope difference analysis to determine whether the pressing behavior performance has a continuously increasing trend. Then, according to the growth rate and duration of the gradient value in the continuous time, the trend interval is divided into low, medium and high three levels. The trend number, time span and behavior intensity change ratio are recorded in each interval to form a pressing behavior escalation trend data sequence. The data is taken as the input of subsequent confidence interval evaluation to provide a structured trend basis for bullying behavior risk dynamic monitoring.

[0045] It should be noted that: step S21 first quantifies the spatial contraction degree, proximity speed change and local posture offset between key points based on the bully behavior marker data, to obtain a posture approach behavior sequence that can reflect the body gradually approaching the victim of the attacker; step S22 extracts the corresponding bullying audio from the same time axis to make the motion approach information have a synchronous tone of attack reference; step S23 establishes a joint expression between the motion approach behavior and the audio signal, and by cross-corresponding the approach amplitude sudden increase segment, body motion mutation node and audio tone of attack enhancement point, the motion features and voice features are induced into pre-pressing behaviors with attack precursors; step S24 clusters the pre-pressing behaviors, and divides different types of pressing modes into multiple stable behavior clusters according to the action structure, action propulsion rhythm and voice attack growth mode, so that the pressing behaviors converge from scattered time sequence segments into structured analysis of pressing behavior groups; step S25 analyzes the time advancement relationship, action intensity incremental chain and cross-cluster linkage transition mode between different pressing clusters according to the clustering results, to form a pressing behavior escalation trend.

[0046] Step S23 includes the following steps:

[0047] Step S231: identifying the approach space reduction amplitude in the time dimension of the bullying posture approach behavior, to obtain the approach space reduction amplitude;

[0048] Step S232: analyzing the limb motion mutation of the bullying posture approach behavior, to obtain bullying limb motion mutation data;

[0049] Step S233: inducing the trajectory pre-behavior between the bullying postures according to the approach space reduction amplitude and the bullying limb motion mutation data, to obtain the body posture trajectory pre-behavior;

[0050] Step S234: quantifying the tone of attack of the bullying audio, and outputting the bullying tone of attack features;

[0051] Step S235: Preceding oppressive behavior induction learning is performed according to the body posture trajectory preceding behavior and the bullying tone aggressive feature, to obtain a bullying preceding oppressive behavior.

[0052] In the embodiment of the present application, the bullying posture approximation behavior obtained in step S21 is input into analysis according to the video sampling time sequence. First, the spatial coordinates of the body skeleton key points of the violent individual in each frame are extracted, and a joint coordinate set with the head, chest center, pelvic center and foot bottom as nodes is constructed. Then, with the chest center point and the chest center point of the oppressed individual as references, the Euclidean distance of the two in the consecutive frames is frame-by-frame differentiated to obtain a distance change sequence in the time dimension. The distance change sequence is subjected to time sliding window smoothing processing to convert the distance reduction trend into an approximation space reduction curve. The local extreme point distribution interval of the curve is calculated, the distance minimum change rate of the three consecutive time periods is extracted, and the reduction amplitude is obtained. The calculation of the reduction amplitude is based on the projection rate difference of the body center of gravity movement vector in the time dimension. If the reduction ratio in the consecutive window is increasing, it is defined as an approximation space reduction event. All identified reduction amplitude results are recorded according to the three indexes of time stamp, distance increment and shrinkage rate to form an approximation space reduction amplitude.

[0053] In the implementation process of step S232, the skeleton sequence of the posture approximation behavior is used to calculate the angle change of the main parts of the body in each key frame. First, the shoulder, elbow, wrist, hip, knee and ankle are selected as the analysis targets, the change rate of the angles of the joints between adjacent key frames is calculated, and an angle change time curve is constructed. Then, the turning point position in the angle change curve is analyzed by using the time sequence difference method, the local mean deviation of the transition region is analyzed, and the mutation occurrence point is determined. The basis for determining the mutation is that the angular velocity increment exceeds the set standard deviation threshold, and the continuous acceleration directions between adjacent frames are opposite, which is determined as a mutation action. After each mutation detection, the number of frames in which the upper and lower limbs and the trunk mutate is counted and the corresponding angle value interval is recorded to describe the strength of the action mutation. In order to avoid the influence of noise, the angle change rate curve is subjected to secondary smoothing correction processing to ensure that the detected mutation points correspond to real motion changes. Finally, the bullying limb action mutation data including the body part name, the mutation frame number, the angular velocity jump amplitude and the mutation duration are output, providing high dynamic feature input for subsequent trajectory behavior induction.

[0054] The approximation space reduction amplitude in step S231 is time-aligned and fused with the body movement mutation data in step S232. First, according to the time stamp identification in the two groups of data, the approximation reduction rate at the same time is paired with the mutation angle increment. By calculating the time sequence correlation coefficient of the two, the coupling degree of the approximation trend and the sudden action is determined, and when the correlation coefficient is higher than the preset threshold, the time point is determined as the key behavior node of the body posture trajectory pre-stage. Subsequently, taking the body center of gravity position as the reference, the moving direction of the shoulder and hand trajectory in the continuous frame is tracked, and the trajectory vector field is constructed combined with the leg support offset. The main direction extraction operation is performed on the trajectory vector field, the main motion trajectory in the approximation process is identified, and the key behavior node is taken as the boundary to linearly fit the front and rear motion paths to form the pre-contact body posture trajectory curve. The curve contains three data of approximation rate, direction change and stable duration time, which are used to represent the spatial evolution characteristics of bullying action. All the trajectory curve data are stored in time sequence to form the body posture trajectory pre-behavior result set, which is used as the input of subsequent audio tone feature fusion analysis.

[0055] The audio data extracted in step S22 is quantified for tone aggressiveness. First, short-time energy calculation and zero-crossing rate statistics are performed on the audio signal to distinguish speech activity regions and noise intervals. Then, the speech activity frame is spectrally decomposed to extract the energy density and spectral slope parameters of the high sound pressure frequency band. When identifying repeatedly occurring derogatory words in the speech segment, the duration, stress ratio and drawn-out reading features of adjacent syllables are analyzed, and the energy sudden increase amplitude of each frame of speech is recorded as the tone aggressiveness intensity factor. According to the corresponding relationship between the high sound pressure spectral slope and the short-time energy increment, a tone aggressiveness time sequence feature vector is formed. For the speech segment with continuous high sound pressure and high frequency peak value, the peak duration and average energy difference are calculated to quantify the continuity of aggressive expression in tone. Finally, the time stamp, peak intensity, speech duration and aggressiveness score are structured into data output to form the bullying tone aggressiveness feature, providing an acoustic feature basis for pre-compression behavior induction learning.

[0056] The body posture trajectory pre-behavior data obtained in step S233 is fused and analyzed with the tone aggressiveness features generated in step S234. First, the posture trajectory features and the speech features are aligned according to the time stamps, so that the corresponding time segments in the two sequences are strictly synchronized. Then, three types of features, including the approaching space reduction amplitude, the limb movement mutation rate, and the tone aggressiveness, are taken as inputs to establish a time-continuous sample set. The sample set is subjected to multivariate statistical aggregation according to the frame time sequence, and the joint change gradient of the three types of features in the calculation window is calculated to identify the stage in which the force application tendency is enhanced in the behavior, and the time segment in which the gradient peak value is located is defined as the bullying pre-compression interval. Then, the continuous compression segments are determined according to the time sequence similarity between the samples, and are integrated into a complete compression behavior sequence. In the sequence output stage, the average value of the approaching rate, the number of action mutations, the tone intensity coefficient, and the behavior duration frame length are recorded for each time interval to accurately represent the formation process of the compression chain. Finally, the comprehensive data including the time sequence, the dynamic amplitude, and the semantic aggressiveness are output as the bullying pre-compression behavior, which is used as the input for the subsequent clustering and trend analysis stage.

[0057] The trajectory pre-behavior induction between the bullying postures includes:

[0058] The reduction rate variance of the approaching space is calculated based on the approaching space reduction amplitude, and the dynamic identification of the double-arm outward radius is performed on the bullying limb movement mutation data to obtain the dynamic radius of the double-arm outward radius.

[0059] The dynamic forward inclination angle of the thorax is evaluated based on the dynamic radius of the double-arm outward radius.

[0060] The dynamic forward inclination angle of the thorax is subjected to short-time angle surge splitting fitting to obtain the short-time surge data of the forward inclination angle.

[0061] The pre-contact posture trajectory is quantified according to the reduction rate variance and the short-time surge data of the forward inclination angle.

[0062] The trajectory pre-behavior induction between the bullying postures is performed according to the pre-contact posture trajectory to obtain the body posture trajectory pre-behavior.

[0063] In the embodiment of the present application, firstly, the approximation space reduction amplitude obtained in step S231 is called to perform variance calculation on the distance change sequence in the time period, which is used to evaluate the stability degree of the approximation space reduction rate. The process takes the center of gravity distance difference of each time as the input, performs sliding window statistics on the contraction rate between continuous frames, calculates the mean value and deviation of the rate in the window, and then obtains the reduction rate variance value through the deviation. When the variance value is high, it indicates that the approximation action has strong volatility, reflecting the unstable situation of the space motion of the violent individual before contact. Then, using the limb action mutation data obtained in step S232, the double-arm extension radius of the extension motion state of the left and right upper arms of the violent individual in each frame is calculated. The calculation takes the midpoint of the two shoulder connecting lines as the rotation center and the wrist joint coordinates as the trajectory endpoints, extracts the distance of the wrist relative to the center point in each frame, and forms a double-arm extension radius sequence. The radius change rate is calculated by time derivation, the extension action is judged whether it is continuously expanding or contracting, and the peak point in the high-speed extension stage is dynamically identified, so as to generate the dynamic radius of double-arm extension. The above dynamic radius and the reduction rate variance jointly constitute the space expansion constraint condition of the action before contact, which is used to guide the analysis of the chest forward inclination angle.

[0064] According to the output dynamic radius data of double-arm extension, the chest dynamic forward inclination angle of the upper body posture of the violent individual is estimated. In the analysis, the chest center line is first determined, that is, the space connecting line from the bottom point of the neck to the center point of the pelvis. The chest inclination angle is calculated by the angle change of the line and the ground plane in the continuous frame. Since the upper limb extension and the chest forward movement occur at the same time in the dynamic process, in order to prevent angle error accumulation, a dynamic radius adjustment method is used to correct the angle data of each frame. The method uses the change trend of the double-arm extension radius to judge the center of gravity moving direction of the body posture, when the double-arm extension radius shows a shrinking trend, the chest inclination angle increment is set to be positive, to reflect the body forward trend; otherwise, the corresponding offset is subtracted. After the adjustment, the forward inclination angle of continuous frames is executed in time sequence smoothing, and the time curve of the chest dynamic forward inclination angle is obtained by calculating the angle change mean value. The rising interval in the curve represents the body active forward behavior stage, and the falling interval represents the body back or posture recovery stage. The finally output chest dynamic forward inclination angle data contains the angle mean value, change rate and time sequence identifier, which is the input basis for subsequent short-time surge interval identification.

[0065] The obtained thoracic dynamic forward inclination angle time curve is analyzed for short-time surge and segmented fitting. First, the angle increment between adjacent frames is calculated by a difference algorithm and the angle rising mutation point is detected to locate the angle mutation peak. Then, a short-time analysis window is set, the angle rising interval is extracted in each window, the angle change in the interval is linearly fitted, the rising rate of each stage is obtained, and the continuous amplitude segment in the rate change curve is summed to extract the surge motion characteristics. If the amplitude proportion of the angle rising rate in continuous multiple time windows exceeds the set threshold, it is defined as a forward inclination angle short-time surge event. The time periods of all surge events are merged and arranged to generate the forward inclination angle short-time surge data. The data mainly records the start and end frames, peak angle and rising rate, and is accompanied by the duration of the time interval and the total surge amplitude. The short-time surge data identified through this process reflects the body tilt mutation behavior of the individual in the preparation action before contact, which is a key feature index for posture trajectory quantization.

[0066] According to the reduced rate variance and forward inclination short-time surge data input trajectory quantization calculation process, first, the frame-level synchronization of the two types of data is performed according to the time stamp to ensure the overlap accuracy of the variance peak value and the forward inclination surge period. Then, taking the dynamic change trajectory of the center of gravity coordinates of the attacker as a reference, the geometric characteristics of the body forward movement path are calculated. By sampling the spatial trajectory coordinates of the center of gravity, shoulder point and knee point, a three-dimensional trajectory line is constructed and the path distance between consecutive frames is calculated to generate the pre-contact posture trajectory. Subsequently, the weighted calculation method is used to integrate the reduced rate variance weight and the forward inclination surge amplitude weight to quantify the body movement trend of each time period. If the comprehensive weight is in a continuous increasing state, it indicates that the body is continuously leaning forward and approaching, otherwise it is adjusted to posture stagnation or retreat. Finally, the trajectory time sequence is structured and arranged according to the motion direction, angle change and distance increment to form the pre-contact posture trajectory, providing specific motion data basis for the induction of trajectory pre-position behavior.

[0067] The quantized trajectory sequence is segmented according to the time interval, and the average inclination angle, motion direction offset and center of gravity acceleration of each segment of the trajectory are used as dimension features to form a feature set. Then, all feature sets are subjected to time sequence clustering induction, and the trajectory direction difference and posture stability of adjacent segments are calculated using the trajectory shape similarity measurement method. The clustering process maintains the time continuity constraint to ensure the integrity of the same body posture transformation process in the clustering result. The induced trajectory mode uses the approaching direction feature and the forward inclination rate distribution as the basis for judgment, and defines the change process of the body from the stable state to the forward leaning state before contact as the body posture trajectory pre-position behavior. The output data records the start and end time, trajectory direction vector and average inclination rate of each body posture change segment, realizing detailed dynamic representation of the attacker from the pre-contact preparation to the pre-action sequence stage, and obtaining the complete body posture trajectory pre-position behavior result.

[0068] The tone aggression quantification comprises:

[0069] Repeated derogatory words in the bullying audio are extracted, and high sound pressure spectrum slope calculation is performed on the bullying audio according to the repeated derogatory words to obtain high sound pressure spectrum slope;

[0070] Logarithmic nonlinear fitting is performed based on the high sound pressure spectrum slope to obtain slope logarithmic fitting data;

[0071] Derogatory word lengthening and repetition feature analysis is performed on the bullying audio according to the repeated derogatory words to obtain derogatory word lengthening and repetition feature;

[0072] Short-time spectral amplitude increment evaluation is performed on the derogatory word lengthening and repetition feature to obtain short-time amplitude increment data;

[0073] The tone aggression quantification comprises:

[0074] In the embodiment of the application, first, the extracted bullying audio signal is framed according to a fixed sampling rate, and the length of each frame is set to 25 milliseconds and overlaps by 10 milliseconds to ensure spectral continuity. Windowing processing is performed on each frame to eliminate boundary effects, and then short-time Fourier transform is performed on the full frame signal to convert the time domain signal into a frequency spectrum amplitude matrix. On the basis of spectral analysis, syllable-level speech segmentation is performed, the energy peak distribution is used to identify the start and end points of the word, and the acoustic features of each continuous speech segment are extracted. By matching with the established derogatory vocabulary library, the repeated derogatory word information in the audio is detected. The repeated frequency and continuous frame number of the derogatory words on the time axis are counted, and the corresponding energy peak and sound pressure level data when appearing are recorded. Then, for the audio segment containing derogatory words, the spectral slope is calculated according to the energy distribution of the spectral amplitude feature in the high sound pressure frequency region. The calculation method is to extract the frequency peak point sequence in the amplitude spectrum of the corresponding frame, measure the spectral line slope rate by linear interpolation method, and finally obtain the high sound pressure spectrum slope data. This data reflects the slope degree of the position of the speech energy concentration changing with frequency, which is an acoustic basic index for quantifying tone aggression, and provides input for subsequent logarithmic nonlinear fitting. Derogatory words mainly refer to words with clear insulting, derogatory, mocking, discriminatory or personal attack nature in language expression, which are often used in school bullying scenes to oppress others mentally or humiliate others personally.

[0075] The obtained high sound pressure spectrum slope data is subjected to logarithmic nonlinear fitting. First, the sequence of high sound pressure spectrum slope change over time is input into the fitting algorithm, and a natural logarithm mapping transformation is performed on it in the time domain, so that the slope values in different amplitude intervals are uniformly distributed in the logarithmic domain, so as to be subsequently smoothed. Then, the nonlinear curve fitting is performed on the logarithm converted data by the least square estimation algorithm, and the fitting function adopts a polynomial form to capture the periodic variation characteristics in the spectrum slope. After the fitting is completed, the monotonicity correction processing is performed on the fitting residual, and the abnormal oscillation section in the data is eliminated, so that the output fitting curve is smoothly transitioned in time. The local extreme points of the fitting curve represent the attack burst time of the sound pressure change over time, and the amplitude change between these extreme points can reflect the continuous increase of the amplitude of the speech attack. The final obtained slope logarithmic fitting data includes the average fitting strength, the fitting curve variance and the extreme point distribution time index, which are used to represent the nonlinear distribution characteristics of the energy attack in the derogatory speech, and provide the curve feature input for the overall quantification of the tone attack.

[0076] The lengthening and repetition characteristics of the identified repetitive derogatory word segments are analyzed. First, the start and end frames of the derogatory words are located in the time domain waveform of the audio, and the duration of the single word is calculated to extract the lengthening ratio. The specific calculation method is to compare the time length of the current word with the average time length of the standard pronunciation by taking the average speed of the basic pronunciation of the same speaker as the reference, and obtain the time length ratio. Then, the repetition degree of each syllable of the word is calculated by using the frame-level energy distribution, and the stress intensity is determined by finding the energy peak points and evaluating the difference between adjacent peak values. The lengthening and repetition characteristics of each derogatory word are jointly constituted by the lengthening and repetition characteristics. In order to prevent the interference of pronunciation fluctuation, the error value of abnormal speech segment is corrected by short-time smoothing processing. After all the derogatory words are calculated one by one, the lengthening and repetition characteristics of the derogatory words are formed, which include the text identification of the word, the speech time length gain, the repetition intensity value and the occurrence time. These characteristic values reflect the syllable prominence and tone manipulation, and provide specific speech region positioning input for subsequent short-time frequency spectrum amplitude increment evaluation.

[0077] The output derogatory word lengthening and re-reading feature time sequence is mapped to the corresponding audio frame interval, and the spectral amplitude data in the corresponding interval is extracted. Then, short-time spectral analysis is performed on these spectral segments, and the amplitude change rate of consecutive frames is calculated to obtain a short-time spectral amplitude increment sequence. In the analysis process, a sliding window method is applied to scan the amplitude sequence frame by frame with a fixed window length, and the amplitude dynamic increment is obtained by calculating the maximum and minimum amplitude difference in the window. For derogatory speech frames with re-reading features, the amplitude increment usually shows a sustained upward trend. To improve the calculation accuracy, the amplitude curve inside each re-reading segment is corrected for linear trend, and the background noise effect is removed. Finally, the short-time amplitude increment values of all re-reading corresponding time domain segments are summarized to generate short-time amplitude increment data, which is defined by four parameters: timestamp, increment amplitude, frame length, and frequency index. This data is used to evaluate the degree of energy burst and aggressive time domain features in subsequent tone aggression quantification.

[0078] The generated slope logarithmic fitting data and the obtained short-time amplitude increment data are quantified for tone aggression. First, the two types of data are synchronized at the frame level using the timestamp as the index, ensuring that the spectral slope change and the amplitude increment in the same time period correspond to each other. Then, the covariance of the two features in each time window is calculated to determine the synchronization of energy slope increase and sound pressure peak enhancement. According to the size of the covariance, the tone aggression intensity level is defined. The larger the covariance, the higher the coupling degree of energy change and spectral slope rate, indicating stronger aggressive expression. Next, the interval weighting summary is performed on the entire time sequence data to form the overall tone aggression score. Finally, the output bullying tone aggression features include average aggression intensity, high aggression segment proportion, speech frequency band concentration, and time duration parameters. This result comprehensively reflects the tone burst and oppressive degree in the speaking process, providing basic input data for pre-oppression behavior induction learning.

[0079] Step S25 includes the following steps:

[0080] Step S251: Constructing a time sequence behavior transition graph for each cluster in the pre-oppression clustering behavior;

[0081] Step S252: Performing multi-scale linkage analysis between adjacent clusters based on the time sequence behavior transition graph to obtain adjacent oppression behavior linkage data;

[0082] Step S253: Performing multi-segment trend decomposition on the adjacent oppression behavior linkage data, and then assigning a trend escalation risk level to obtain the trend escalation risk level;

[0083] Step S254: Analyzing the oppression behavior escalation trend based on the trend escalation risk level.

[0084] In the embodiment of the present application, the pre-pressing clustering behavior output in step S24 is classified and processed according to the cluster number. Each cluster is composed of the same pre-pressing behavior, and contains information such as time sequence identifier, action type number, posture parameter mean value, and speech attack feature mean value. First, the samples in the same cluster are arranged based on time sequence, and the behavior node set is established according to the inter-frame time interval. Then, the time span and feature change amplitude between any adjacent nodes in the same cluster are calculated, the time span is defined as the transition weight, and the feature change amplitude is defined as the edge weight. All nodes are connected in time increasing order to form a behavior transition path, and each path is represented as a directed edge, and a complete time sequence behavior transition graph is constructed. The node in the transition graph represents the average behavior state in the time slice, and the edge represents the dynamic change direction and correlation strength of the behavior in the continuous period. In order to depict the transition characteristics between different clusters, the inter-cluster connection relationship needs to be extracted, and the cross-cluster edge link is established between the starting node of the adjacent number cluster and the end node of the last cluster, so as to ensure the time continuity and cross-cluster traceability of the transition graph structure, and finally the time sequence behavior transition graph corresponding to each cluster is obtained.

[0085] The time sequence behavior transition graph obtained in step S251 is subjected to multi-scale linkage analysis to obtain adjacent pre-pressing behavior linkage data. First, the time scale division is determined, and the transition graph time axis is divided into three short, medium and long windows according to the number of frames. The average value of the edge weight, the time weight ratio and the direction change amount between adjacent nodes are calculated in each time window, and the multi-scale behavior change sequence is obtained in a window sliding manner. Then, the transition graph of adjacent clusters is matched across the graph, and the coupling degree of the transition path between different clusters is calculated according to the node feature similarity. The linkage strength between two behavior clusters is quantified by the path coupling rate, and the greater the linkage strength value, the more continuous the evolution relationship between the two behavior clusters in the time track. Then, the linkage strength in each scale window is superimposed and averaged by using the time sequence coherence analysis method, and a multi-dimensional linkage matrix containing time scale dimension and behavior feature dimension is generated. Each numerical point in the matrix reflects the synchronization level of adjacent behavior clusters at different time scales, and the overall distribution result of the matrix is defined as adjacent pre-pressing behavior linkage data.

[0086] The multi-dimensional linkage matrix is sorted by time axis, and the average linkage intensity curve of each time period is extracted. Then the curve is processed by multi-section trend decomposition algorithm, and the overall trend is separated into three categories: slow increasing section, sudden rising section and decreasing section. The decomposition process uses local linear regression method to identify the slope change point progressively, so as to delimit the trend boundary. The average slope and change rate are calculated in each section interval to determine the continuous direction of the trend. By comparing all section trends, the continuity and escalation rate of the oppression behavior evolution are determined. According to the trend slope and the mean square deviation value of the slope change, a risk assessment index is defined, and the risk level is divided into three categories: rising risk, stable risk and weakening risk. According to the scoring results, the trend escalation risk level is assigned. The trend escalation risk level table generated by the process includes time interval, risk category number, trend slope average and behavior duration time parameters, which are used to guide the subsequent oppression behavior escalation trend judgment.

[0087] The trend escalation risk level data obtained according to step S253 is analyzed to analyze the escalation trend of the oppression behavior. First, the risk level points of all time periods are summarized in time sequence to form a risk level time curve. Then the grade fluctuation amplitude of the curve in the continuous interval is calculated, and the ascending section from low to high risk level is identified. In the ascending section, the number of continuous frames and the behavior intensity increment are counted to determine the escalation continuity of the oppression behavior. The difference accumulation method is used to calculate the change gradient value of the risk level on the time axis. When the gradient value is continuously positive and exceeds the set threshold, the continuous escalation trend is confirmed. To prevent short-term fluctuations from interfering, the curve is processed by moving average to smooth the mutation, and then the wave peak position is extracted as the trend vertex. The escalation start and end times of all trend intervals are marked, and the risk transition time and behavior intensity growth rate between the upper and lower nodes are calculated. The final output of the oppression behavior escalation trend includes four parameters: time sequence, risk level change, escalation amplitude and trend duration, which are used as input and verification in the subsequent confidence interval evaluation stage.

[0088] Step S3 comprises:

[0089] Step S31: performing time sequence pattern recognition processing on the escalation trend of the oppression behavior to obtain an escalation trend time sequence pattern;

[0090] Step S32: performing convolution processing on the escalation trend time sequence pattern to obtain escalation trend pattern convolution data;

[0091] Step S33: performing confidence interval evaluation of the bullying oppression escalation behavior based on the escalation trend pattern convolution data to output an escalation trend behavior confidence interval.

[0092] In the embodiment of the present application, the whole escalation trend of the oppression behavior is time-standardized, and the missing segments are filled by equal-interval resampling to make the sequence continuous in the time scale. Then, the trend feature vector at each time point is calculated, and the feature content includes risk level, slope amplitude, oppression intensity difference value, and behavior duration, etc. In order to extract the key stage change in the trend, the dynamic time warping algorithm is used to align and calculate the trend waveform in different periods, and the risk level change rate is used as the matching reference in the alignment process, and the feature point position and amplitude response are recorded. Subsequently, the time series data is divided into short-term fluctuations and long-term trends by using hierarchical decomposition method, the short-term fluctuations contain small amplitude behavior intensity random changes, and the long-term trend reflects the overall escalation direction of the oppression behavior. Then, the self-similarity ratio and delay response of the trend sequence in each period are calculated by using the sliding window similarity analysis method, and the periodic escalation mode is extracted. The self-similarity curve is described in structure, and the feature mode corresponding to each high-risk interval is labeled. Finally, the risk level change, trend periodicity feature and time series stability parameter are integrated and output to form the escalation trend time series mode, which provides preparation data for subsequent convolution feature extraction.

[0093] The escalation trend time series mode obtained in step S31 is subjected to convolution processing to extract feature response information. First, a fixed-width sliding convolution kernel is set on the time axis, and point-by-point interval convolution operation is performed on the time series mode data. Each convolution operation is based on the weighted summation of the trend feature vectors in the time window, and the weight value depends on the comprehensive value of the trend amplitude change rate and the risk level deviation. The output result after convolution reflects the local cumulative effect of the trend change. In order to analyze the behavior change characteristics of different frequencies, a multi-scale convolution kernel group is constructed, and short-scale is used to capture sudden oppression behavior jumps, and long-scale is used to capture slow incremental continuous oppression trend. By sliding convolution on the whole time series, the superimposed integral of the results of each scale is performed to obtain the convolution data of the escalation trend mode. The response peak value of each time period in the convolution data corresponds to the aggregation degree of the oppression behavior intensity at a certain time point, which is used to reflect the potential escalation outbreak segment. In order to avoid noise interference, the convolution result is filtered and normalized for several times, so that the energy of the continuous behavior characteristics on the time axis remains stable distribution. Finally, the escalation trend mode convolution data containing convolution peak value, trend response mean value, behavior expansion width and time coordinate index is formed as the input of confidence interval evaluation.

[0094] The confidence interval evaluation is performed on the upgrade trend pattern convolution data generated in step S32 to output the upgrade trend behavior confidence interval. First, the statistical distribution analysis is performed on the peak sequence in the convolution result with time series as the index to extract the mean value and standard deviation of the convolution response. The statistical distribution curve of the stress behavior trend intensity is obtained by calculating the probability density estimation of the convolution intensity in each time period. Then, the quantile analysis is performed on the curve to divide the distribution interval into three sections of low risk, medium risk and high risk, and determine the probability boundary according to the position of the quantile point. Then, the repeated sampling method is used to perform multiple disturbance sampling on the convolution peak value data, and the mean value offset and interval upper and lower limit difference are calculated under each sampling, and the stable range of the confidence interval boundary is obtained by averaging the results. The correction is provided for the random fluctuation of the stress behavior trend waveform, and the confidence interval curve is smoothed after the interval boundary is obtained, so that the confidence interval curve has continuity. The final output of the upgrade trend behavior confidence interval includes the lower limit boundary, the upper limit boundary, the interval width and the corresponding time index, and is bound with the stress behavior time sequence characteristics for subsequent information gain learning stage to dynamically weight the pre-stress behavior characteristics and set the safety threshold.

[0095] It needs to be explained: first, according to the time series data in the escalation trend of oppressive behavior, the characteristics of continuous growth segment, stage sudden increase segment, and shortening of oppressive action interval are identified, the coarse-grained information originally representing the trend direction is converted into the escalation trend time sequence mode with clear paragraph structure, node distribution and change rhythm; Step S32 uses the time sequence mode as input, and convolves the local change segment in the mode by using the sliding convolution kernel, so that the details such as the implicit local acceleration segment, the continuous approaching slope enhancement segment, and the short-time attack intention sudden rise segment in the trend are amplified and restructured, thereby forming the escalation trend mode convolution data containing the comprehensive characteristics of stable trend segment and violent fluctuation segment; Step S33 further quantifies the stability, growth intensity, fluctuation amplitude and trend persistence of the trend based on the structured data after convolution, and outputs the final result as the escalation trend behavior confidence interval by comprehensively integrating the confidence of multiple local trend segments in time dimension. The role of confidence interval evaluation of bullying oppression escalation behavior is to give the escalation trend of oppressive behavior after trend identification and convolution enhancement a clear intervalized confidence degree, so that the trend is converted from the fuzzy judgment of "whether to escalate" to the structured quantitative result of "how much escalation possibility in which intensity range with how much confidence". This step comprehensively evaluates the continuous growth segment intensity, local acceleration segment stability, trend fluctuation amplitude and time persistence in the trend mode convolution data, represents the rising degree of the trend in an interval manner, so that different intensity intervals correspond to different escalation confidence levels, thereby ensuring that the subsequent information gain memory learning can allocate different weights to various pre-oppression behaviors according to the confidence interval. Through the introduction of confidence interval, the trend judgment is no longer a single point value participating in the subsequent gain calculation, but a structured data containing uncertainty range as the weight basis, so that the system can distinguish the influence of high confidence trend and low confidence trend in the learning process, thereby significantly reducing the deviation accumulation caused by trend misjudgment. The role of this step is to establish a reliable confidence range basis for subsequent memory learning, alarm decision and quantitative processing of oppression behavior escalation intensity, to realize controllable transmission and enhancement of trend information to alarm information, and to make the whole system have continuous, stable and interpretable trend basis.

[0096] Step S4 comprises the following steps:

[0097] Step S41: normalizing the bullying pre-oppression behavior to obtain pre-oppression behavior normalized data;

[0098] Step S42: performing information gain memory learning on the pre-oppression behavior normalized data based on the escalation trend behavior confidence interval to obtain escalation trend learning data between pre-oppression behaviors;

[0099] Step S43: Design an automatic alarm architecture based on the escalation trend learning data and feedback to the control terminal to perform intelligent identification of bullying behavior.

[0100] In the embodiment of the application, the pre-bullying compression behavior obtained in step S23 is taken as input, and unified normalization processing is performed according to the value range of each behavior characteristic. First, the main characteristic items in the data are extracted, including the approaching space reduction amplitude, the limb motion mutation rate, the tone attack intensity and the forward angle change rate. For each characteristic, the minimum value and the maximum value are calculated, and then a linear proportional transformation is used to map all characteristic values to the interval of 0 to 1, so that they have comparability under the condition of unified scale. In order to maintain time continuity, the sliding average smoothing processing is performed on the normalized sequence to prevent gradient interruption caused by mutation points. Then, all characteristics are renumbered according to the time stamp, so that the characteristics of different proportions are strictly aligned on the time axis, ensuring that the changes of each behavior characteristic have a corresponding relationship at the same time. Finally, a structured feature matrix is generated, the rows of the matrix represent time series, the columns represent standardized characteristic values, and frame numbers and behavior category labels are attached to each time step to form pre-bullying behavior normalized data, which provides a standard input condition for the subsequent information gain memory learning session.

[0101] In the implementation process of step S42, the pre-bullying compression behavior normalized data normalized in step S41 is subjected to information gain memory learning with reference to the escalation trend behavior confidence interval output in step S33. First, a dynamic weighting function is established according to the upper and lower limits of the confidence interval, and the data points outside the confidence interval are assigned a lower weight, while the data points inside the interval are assigned an increasing weight according to their proximity to the upper limit. A weighted pre-bullying behavior sequence is formed through this weighting strategy. Then, the weighted data is segmented by a fixed length time sliding window, and the information entropy difference is calculated in each sliding window to obtain a local information gain vector, which is used to quantify the contribution value of the behavior characteristics in the current time period to the overall trend change. In the memory learning process, the information gain vector is input into the gated memory structure, and through element-by-element multiplication with the activation signals of the forget gate, the input gate and the output gate, and multiplication with an adjustable scaling factor, the long-time preservation of high information gain sequences and the suppression of low information gain sequences are realized. The network updates the hidden state in multiple rounds of cyclic iteration training to retain key behavior patterns while suppressing noise features. The accumulated time-dependent features and trend intensity difference between the pre-bullying compression behaviors in the output hidden state sequence form the escalation trend learning data between the pre-bullying compression behaviors, providing judgment parameters for the alarm architecture design.

[0102] In the implementation process of step S43, the escalation trend learning data output by step S42 is input into the alarm logic design process. First, the alarm trigger threshold is established according to the trend score of the learning data and the time weight. When the comprehensive score in the trend learning data continuously exceeds the upper limit of the confidence interval and the duration exceeds the set threshold frame number, an alarm event is triggered. The alarm logic structure is composed of a signal analysis layer, a threshold judgment layer and an execution feedback layer. The signal analysis layer reads the real-time value of the trend score sequence and calculates the average growth amplitude of the continuous time period; the threshold judgment layer compares the growth amplitude and the duration according to the fixed threshold, and outputs an alarm flag signal when the condition is met; the execution feedback layer generates an alarm instruction signal immediately after receiving the alarm flag, and the instruction signal is sent to the monitoring terminal device through the unified control protocol and the corresponding warning information is displayed on the terminal interface. The terminal automatically records the alarm time, risk position and corresponding video segment number after receiving, realizing real-time identification of bullying behavior. Thus, an automatic alarm architecture is formed, which can maintain a dynamic alert state by continuously receiving updated learning data, thereby completing intelligent detection and real-time feedback of bullying behavior.

[0103] It needs to be explained that: all feature sequences of the pre-pressing behavior are normalized to make features from different sources (such as action approach amplitude, tone aggressiveness value, and body mutation intensity) comparable as standardized behavior vectors in the same dimension, so as to ensure that subsequent learning is not disturbed by the difference in feature scale; step S42 further uses the escalation trend behavior confidence interval output by step S33 as a weight basis to perform information gain memory learning with the normalized behavior vector in the time dimension, and by giving stronger memory gain to high-confidence-interval behaviors and suppressing low-confidence-interval behaviors, the learning result can highlight those pre-pressing behavior sequences that are truly at risk of escalation, and finally form escalation trend learning data with trend sensitivity; step S43 constructs an automatic alarm architecture based on the learning data, maps the learned structured trend pattern to a specific alarm condition, establishes the trigger relationship between the action escalation path, the tone aggressiveness accumulation and the posture approach acceleration chain, and enables the alarm output to be automatically fed back to the control terminal.

[0104] Step S42 includes:

[0105] Step S421: based on the escalation trend behavior confidence interval, dynamically weighting and sampling the pre-pressing behavior normalized data to obtain weighted pre-pressing behavior sequence data;

[0106] Step S422: performing time sliding window-based segmented information gain calculation on the weighted pre-pressing behavior sequence data to obtain a local information gain vector in each sliding window;

[0107] Step S423: selectively injecting information gain into the gating units of the long short-term memory network according to the local information gain vectors in each sliding window, to obtain a gain-enhanced hidden state sequence, wherein the selective information gain injection is achieved by Hadamard multiplication of the local information gain vectors and the original activation values of the forget gate, the input gate and the output gate, and appending a learnable scaling coefficient, so as to realize high-gain memory retention and low-gain memory inhibition of the historical oppression behavior pattern;

[0108] Step S424: performing residual memory iterative training based on the hidden state sequence, and then performing information gain memory learning, to obtain upgrade trend learning data between the pre-oppression behaviors.

[0109] In the embodiment of the present application, in the implementation process of step S421, the pre-oppression behavior normalization data obtained in step S41 is taken as input, and a dynamic sampling weight is calculated according to the upgrade trend behavior confidence interval of step S33. First, the normalized data is arranged in chronological order, and the sample features corresponding to each time point include posture approximation rate, limb action mutation frequency, tone attack intensity value and body tilt angle change rate. Then, the lower limit of the confidence interval is taken as the minimum risk benchmark, and the upper limit is taken as the high-risk boundary, to construct an interval mapping function. According to the relative position of the sample features in the interval, the weighting coefficient is calculated, and the closer to the upper limit, the higher the weight, and the closer to the lower limit, the lower the weight. In order to ensure the smoothness of the weight change in the continuous time period, the weight sequence is modified by moving average. The weighted sampling algorithm is used to extract a representative sample set from the overall sequence by taking the weight value as the sampling probability. After extraction, the sampling results are rearranged into a time-continuous structure, and the corresponding time stamp is retained, so that the data maintains progressive consistency in the time dimension. Finally, the weighted pre-oppression behavior sequence data is formed, which includes time index, feature value combination and weight distribution, providing basic input for segmented information gain analysis.

[0110] The weighted pre-oppression behavior sequence data generated in step S421 is taken as input, and the information gain of each time interval is calculated using the time sliding window mechanism. First, the sliding window length and step value are set, so that the analysis window slides along the time axis frame by frame, and a fixed number of continuous samples are analyzed each time. In each sliding window, the entropy value of the distribution of the normalized features is calculated, and the information gain value is obtained by comparing the difference between the feature distribution in the current window and the overall distribution. In order to prevent abnormal values from causing gain deviation, low-pass filtering smoothing operation is performed on the feature entropy change curve, so that the gain change between windows is continuous. By traversing the entire time sequence, independent local information gain vectors are generated window by window. Each vector contains the numerical value of the feature gain in the window and the corresponding time position identifier, which is used to characterize the intensity of the contribution of the oppression behavior features in different stages to the overall trend. All local information gain vectors are arranged in chronological order to form a gain matrix, which provides time sequence feature basis for the information injection stage of the gating unit.

[0111] The local information gain vector obtained by step S422 is used to perform selective information gain injection on the gating units of the long short-term memory network. First, the original activation values of the forget gate, input gate and output gate in the network at the current time step are read. Then the local information gain vector is multiplied element by element with the three types of gating activation values to form gain modulation signals. To control the intensity of different gains, a learnable scaling coefficient is introduced based on the product result, and the modulation signal of each gating channel is amplified or suppressed according to the scaling ratio. This process is completed by Hadamard multiplication to maintain the dimensional consistency of the output tensor. The injected gating activation signal will adjust the memory retention and forgetting ratio of the hidden state in the state update phase. The signal in the high information gain interval will obtain a larger retention coefficient in the forget gate, while the low gain interval will be enhanced with a forgetting weight to reduce the accumulation of invalid information. The updated output hidden state sequence records the priority mode of the behavior characteristics in the high gain stage, forming a gain-enhanced hidden state sequence, which lays a structural foundation for subsequent residual memory training.

[0112] The gain-enhanced hidden state sequence obtained by step S423 is input to perform residual memory iterative training to complete information gain memory learning. First, for each time step, the difference between the current hidden state and the hidden state at the previous time step is calculated, and the difference is used as a residual signal to retain the time-dependent correlation. By accumulating the residual signal back to the current hidden layer, the historical features are continuously retained in multiple iterations. Subsequently, multiple forward and backward propagation training is performed on the entire sequence, and the gating parameters and scaling coefficients are updated in each training to make the output state gradually approach a stable feature distribution. To prevent gradient vanishing and overfitting, the state output after each iteration is normalized to constrain the output range within a set interval. After completing all iterations, the main feature correlation information of each period is extracted from the final training stable hidden state, recording the time index and response strength. The output data formed by integration has the effects of continuous time feature retention and information gain weight layering, and serves as learning data for the escalation trend of the pre-compression behavior, providing accurate learning result input for subsequent alarm judgment structure design.

[0113] The above described is only a specific implementation of the present application, enabling those skilled in the art to understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for intelligent recognition of bullying behavior in campus surveillance videos, characterized in that, Includes the following steps: Step S1: Extract historical videos of bullying behavior from the campus monitoring and control center; mark the bullies' behavior based on the historical videos of bullying behavior, and output the bully behavior marking data; Step S2: Based on the bully behavior labeling data, perform pre-bullying behavior inductive learning to obtain pre-bullying behavior; Analysis of the escalation trend of oppressive behavior based on the preceding oppressive behaviors; Step S3: Based on the escalation trend of the oppressive behavior, evaluate the confidence interval of the bullying and oppressive escalation behavior, and output the confidence interval of the escalation trend behavior; Step S4: Based on the confidence interval of the escalation trend behavior, information gain memory learning is performed on the pre-bullying oppression behavior, and then an automated alarm architecture is designed and fed back to the control terminal to perform intelligent recognition of bullying behavior; Step S4 includes: Step S41: Normalize the pre-bullying oppression behavior to obtain normalized data on the pre-bullying oppression behavior; Step S42: Based on the confidence interval of the escalation trend behavior, perform information gain memory learning on the normalized data of the preceding oppression behavior to obtain the escalation trend learning data among the preceding oppression behaviors; Step S43: Design an automated alarm architecture based on upgrade trend learning data and feed it back to the control terminal to perform intelligent identification of bullying behavior; Step S42 includes: Step S421: Based on the confidence interval of the escalation trend behavior, perform dynamic weighted sampling on the normalized data of the pre-oppression behavior to obtain weighted pre-oppression behavior sequence data; Step S422: Perform segmented information gain calculation based on time sliding window on the weighted pre-compression behavior sequence data to obtain the local information gain vector within each sliding window; Step S423: Selectively inject information gain into the gating units of the Long Short-Term Memory Network according to the local information gain vector in each sliding window to obtain a gain-enhanced hidden state sequence. The selective information gain injection achieves high-gain memory retention and low-gain memory suppression of historical oppression behavior patterns by performing Hadamard product of the local information gain vector with the original activation values ​​of the forget gate, input gate and output gate and adding a learnable scaling coefficient. Step S424: Perform residual memory iterative training based on the hidden state sequence, and then perform information gain memory learning to obtain escalation trend learning data between the preceding oppressive behaviors.

2. The intelligent recognition method for bullying behavior in campus surveillance videos according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Extract historical videos of bullying behavior from the campus monitoring and control center; Step S12: Denoise the historical bullying videos to obtain denoised bullying videos; Step S13: Perform frame interpolation compensation on the bullying behavior denoised video to obtain the bullying behavior compensated video; Step S14: Mark the bully's behavior based on the bullying behavior compensation video and output the bully behavior marking data.

3. The intelligent recognition method for bullying behavior in campus surveillance videos according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Perform pose approximation behavior analysis on the bully behavior labeling data to obtain bullying pose approximation behavior; Step S22: Extract bullying audio from historical bullying videos; Step S23: Based on the bullying posture approach behavior and bullying audio, summarize and learn the pre-bullying oppression behavior to obtain the pre-bullying oppression behavior; Step S24: Perform behavioral clustering on the pre-bullying oppression behaviors to obtain pre-bullying oppression clustered behaviors; Step S25: Analyze the escalation trend of oppression behavior based on the clustering behavior of prior oppression.

4. The intelligent recognition method for bullying behavior in campus surveillance videos according to claim 3, characterized in that, Step S23 includes the following steps: Step S231: Identify the reduction magnitude of the approximation space in the time dimension of the bullying posture approximation behavior to obtain the reduction magnitude of the approximation space; Step S232: Perform abrupt change analysis of the bullying posture approach behavior to obtain bullying body movement change data; Step S233: Based on the shrinkage of the approximation space and the data on abrupt changes in bullying limb movements, summarize the trajectory pre-before behavior between bullying postures to obtain the pre-before behavior of body trajectory. Step S234: Quantify the aggressive tone of the bullying audio and output the aggressive tone features of the bullying. Step S235: Based on the pre-bullying behavior of body trajectory and the aggressive characteristics of bullying tone, perform pre-bullying oppression behavior induction learning to obtain pre-bullying oppression behavior.

5. The intelligent recognition method for bullying behavior in campus surveillance videos according to claim 4, characterized in that, The preceding behaviors that precede bullying postures can be summarized as follows: The variance of the reduction rate of the approximation space is calculated based on the reduction magnitude of the approximation space; the dynamic identification of the outward radius of the arms is performed on the abrupt change data of bullying limb movements to obtain the dynamic radius of the outward extension of the arms; The dynamic forward tilt angle of the thorax is evaluated based on the dynamic radius of the outward extension of the arms. Short-term angle surge decomposition and fitting were performed on the dynamic thoracic anteversion angle to obtain short-term anteversion angle surge data; Quantify the pre-contact attitude trajectory based on the reduction rate variance and short-term surge data of the forward tilt angle; Based on the pre-contact posture trajectory, the pre-trajectory behavior between bullying postures is summarized to obtain the pre-trajectory behavior of body posture.

6. The intelligent recognition method for bullying behavior in campus surveillance videos according to claim 4, characterized in that, Quantifying the aggressiveness of a tone includes: Extract repetitive derogatory words from bullying audio; calculate the high sound pressure level (HPL) spectral slope of the bullying audio based on the repetitive derogatory words to obtain the HPL spectral slope; Logarithmic nonlinear fitting is performed based on the slope of the high sound pressure spectrum to obtain logarithmic fitting data of the slope. Based on the repetitive derogatory words, the bullying audio was analyzed for the characteristics of prolonged and stressed derogatory words, and the characteristics of prolonged and stressed derogatory words were obtained. The short-time spectral amplitude increment is evaluated based on the prolongation and repetition features of the derogatory words to obtain short-time amplitude increment data; Based on the slope logarithmic fitting data and short-time amplitude increment data, the tone of aggression is quantified, and the bullying tone of aggression features are output.

7. The intelligent recognition method for bullying behavior in campus surveillance videos according to claim 6, characterized in that, Step S25 includes the following steps: Step S251: Construct a temporal behavior transition graph for each cluster in the pre-compression clustering behavior; Step S252: Perform multi-scale linkage analysis between adjacent clusters based on the time-series behavior transition diagram to obtain linkage data of adjacent compression behaviors; Step S253: Perform multi-segment trend decomposition on the linkage data of adjacent compression behaviors, and then assign a trend upgrade risk level to obtain the trend upgrade risk level. Step S254: Analyze the escalation trend of oppressive behavior based on the aforementioned trend escalation risk level analysis.

8. The intelligent recognition method for bullying behavior in campus surveillance videos according to claim 1, characterized in that, Step S3 includes: Step S31: Perform time-series pattern recognition processing on the escalation trend of oppressive behavior to obtain the escalation trend time-series pattern; Step S32: Perform convolution processing on the upgrade trend time series pattern to obtain the upgrade trend pattern convolution data; Step S33: Evaluate the confidence interval of bullying and oppression escalation behavior based on the escalation trend pattern convolutional data, and output the confidence interval of escalation trend behavior.

Images