Campus spoofing monitoring system based on millimeter wave radar

Abstract

The invention belongs to the technical field of campus spoofing monitoring, and particularly relates to a campus spoofing monitoring system based on a millimeter wave radar. The invention discloses a campus spoofing monitoring system based on millimeter wave radar, which is characterized in that the millimeter wave radar collects radar signals of a private space and transmits the radar signals to a data return module, so that a data collector receives the radar signals transmitted by the data return module and returns the radar signals to an abnormal point cloud cluster detection module; the abnormal point cloud cluster detection module performs DBSCAN algorithm processing on the received radar signal to generate a cluster; the deception detection module analyzes the abnormal clusters for three times through the personnel counting detection module, the action abnormity detection module and the personnel toppling detection module. The campus spoofing monitoring system based on the millimeter wave radar disclosed by the invention can accurately detect whether a spoofing behavior occurs in a privacy area in real time, and is high in detection efficiency and high in detection precision.

Description

Technical Field

[0001] The present invention belongs to the technical field of campus bullying monitoring, and particularly relates to a campus bullying monitoring system based on millimeter-wave radar. Background Art

[0002] In today's educational environment, the scale of students is showing an increasing trend, and the campus is populated by student groups from different backgrounds and with diverse personalities. While this diverse student composition enriches campus culture, it also brings new challenges. Among them, unfriendly behaviors on campus, such as bullying and conflicts, occur from time to time, having a negative impact on students' physical and mental health and the campus environment. When faced with these new situations, there are inevitably some omissions in campus management. Traditional monitoring means such as cameras, although helpful to a certain extent in maintaining campus security, are limited in their coverage due to privacy protection needs and cannot function in areas not covered by monitoring, such as school restrooms and dormitories, which are precisely the high-incidence areas of unfriendly behaviors, posing great difficulties to school management.

[0003] In this context, millimeter-wave radar technology, as a new type of detection means, has shown great application potential. Millimeter-wave radar uses electromagnetic waves to detect the movement trajectories of objects and can effectively monitor the above-mentioned privacy areas without infringing on students' privacy. It can not only accurately sense the presence, quantity, location, and movement trajectories of people in the area, but also identify abnormal behavior patterns, such as staying for a long time, abnormal gatherings, and large-scale limb movements, through intelligent algorithms, and trigger the early warning mechanism in a timely manner, sending early warning information to relevant personnel for intervention and handling. This intelligent campus bullying monitoring technology not only overcomes the limitations of traditional monitoring means but also provides a more accurate, efficient, and user-friendly solution for campus security management, contributing to creating a safer and more harmonious campus environment. Summary of the Invention

[0004] The main purpose of the present invention is to provide a campus bullying monitoring system based on millimeter-wave radar, which can accurately detect in real time whether bullying behaviors occur in privacy areas, with high detection efficiency and high detection accuracy.

[0005] To achieve the above object, a campus bullying monitoring system based on millimeter-wave radar includes a millimeter-wave radar, a data backhaul module, a data collector, an abnormal point cloud cluster detection module, a bullying detection module, and an early warning level judgment and response module, wherein: The millimeter-wave radar collects radar signals in a private space and transmits the radar signals to the data backhaul module, so that the data collector receives the radar signals transmitted by the data backhaul module and transmits them back to the abnormal point cloud cluster detection module; The abnormal point cloud cluster detection module processes the received radar signals using the DBSCAN algorithm to generate clustering clusters, screens the generated clustering clusters using the GNN algorithm, filters out abnormal clustering clusters by setting thresholds, and transmits the data to the bullying detection module for analysis; The bullying detection module analyzes the abnormal clustering clusters three times through the personnel counting detection module, the action anomaly detection module, and the personnel dumping detection module respectively, so as to obtain corresponding data and transmit it to the warning level judgment and response module after data fusion; The warning level judgment and response module processes the received fusion data, judges the warning level, and takes corresponding response measures.

[0006] As a further preferred technical solution of the above technical solution, the abnormal point cloud cluster detection module is specifically implemented as follows: First, the received radar signals are processed including clutter and noise removal. The processed radar signals are then discretely sampled to generate a frame sequence, and the frame sequence is cut and recombined to obtain a frame matrix. The DBSCAN algorithm is applied to the generated frame matrix, and clustering clusters are generated from the data in the frame matrix. Then, the generated clustering clusters are screened using the GNN algorithm to screen out clustering clusters with the number of people greater than the first preset number of people. In addition, when the number of people in the clustering cluster is greater than the second preset number of people, a warning reminder is directly triggered, a notice of prohibiting gathering is issued, and the teaching staff is notified to check the situation.

[0007] As a further preferred technical solution of the above technical solution, the analysis of the personnel counting detection module is specifically implemented as follows: According to the regional movement amplitude, limb reflection angle, speed and acceleration of personnel, different movement states are distinguished for subsequent identification. The movement states include walking, running, retreating, jumping and standing still, and the personnel in different movement states are respectively counted and statistically processed. The root mean square value is used to calculate the Doppler velocity of multiple targets obtained from the radar signals as the body movement index; The algorithm adopted by the personnel counting detection module is CFAR target detection and Kalman filter tracking and positioning. The relevant content is as follows: First, calculate the initial radar image, and the formula is: , where represents the radar image at the r distance and θ direction at time t, represents the radar image intensity value at time t, distance r and direction θ, which is an imaging result obtained by weighted synthesis of signals from multiple channels, reflecting the target echo energy or characteristics at the position, specifically: N: The total number of channels; : The weighting coefficient of the nth channel, which is used to adjust the contribution degree of each channel to the final image; : The beamforming weight in the direction θ, which is related to the radiation pattern or beam pointing of the radar antenna and determines the response intensity of the nth channel in the direction θ; : The original signal of the nth channel at time t and distance r; t: The time variable, corresponding to the moment of radar pulse emission or data acquisition; r: The distance variable, representing the radial distance from the target to the radar; θ: The azimuth angle or direction variable, representing the pointing angle of the radar beam; Second, suppress static clutter. The formula is: , where represents the radar image after clutter suppression. Static clutter is suppressed by the time averaging method, and the signals of dynamic targets are retained. Among them: T: The time window length, that is, the integration interval for calculating the average value; : The index of the time window; t′: The integration variable, representing the historical moment; Integral term: In the time window , the time average of the signal intensities at the same distance r and direction θ is calculated to obtain the estimated value of static clutter; Third, estimate the target position. The formula is: , where and represent the position and azimuth angle of the target at time t respectively, is a mathematical operator, indicating to find the parameter pair that makes the function reach the maximum value; Fourth, calculate the new radar image. The initial radar image represents the instantaneous radar image, that is, the range-azimuth signal intensity obtained at a single time point t. The formula for the new radar image is: , where represents the radar image in the th time period, representing the radar image after time averaging. It is obtained by performing energy integration and averaging on the instantaneous images in the th time period ; : The average radar image in the th time period, with the unit of energy or power; : Instantaneous radar image, representing the complex signal at time t, distance r, and azimuth angle θ; Energy of the instantaneous signal, used to eliminate the phase influence and retain the intensity information; T: Time window length, i.e., the time interval of integration; : Index of the time period; Integrand : Cumulate the energy of all instantaneous signals within the -th time period; Fifth, calculate the body movement index, with the formula: , used to quantify the intensity of body movement. By analyzing the energy of the derivative of the movement signal d(τ), it reflects the severity of the movement, where: : Body movement index at time t, with the unit of speed. The larger the value, the more severe the movement; : Original movement signal; : Time derivative of the movement signal, representing the instantaneous speed or rate of change; Take the modulus square or directly square, used to calculate the energy; Tb: Time window length, i.e., the smoothing interval for calculating the movement index; Integration interval : A symmetric window centered at t with a width of T b , ensuring the causality or real-time requirements; Sixth, calculate the correlation coefficient, with the formula: ; Calculate the correlation coefficient between the body movement indices measured by two radars within the same time period, used to quantify whether the movement signals are synchronous or similar, where: Value range: ; 1: Perfect positive correlation; : Perfect negative correlation; 0: No correlation; : At the -th time period, the body movement indices of radar m and radar m′ and correlation coefficient; and : Body movement indices measured by two independent radars at time t; : The time window length for calculating the correlation coefficient; : The index of the time period; Numerator part: The covariance of the two signals within the window, reflecting the joint change trend; Denominator part: The geometric mean of the energies of the two signals, used to normalize the covariance and eliminate the influence of amplitude differences; Seventh, calculate the target association accuracy rate, and the formula is: , used to calculate the overall accuracy rate of multi-target tracking or association tasks. By statistically calculating the association correct rates of all targets and all time periods and taking their average value, where: : The overall accuracy rate of target association; M: The total number of targets, indicating the number of independent targets that the system needs to track or associate; L: The total number of time periods, indicating the total number of time windows during which the system runs; : The association correct rate of the m-th target in the -th time period, which is a binary value or a probability value; If it is a binary value: = 1: Indicates that target m is correctly associated in the time period ; = 0: Indicates an incorrect association; If it is a probability value: Indicates the confidence level or partial correctness of the association; Eighth, define the target association accuracy rate, and the formula is: ; Binary decision rule, used to determine whether the m-th target in the -th time period is correctly associated. Its core logic is: If the association strength of target m with itself is not lower than its association strength with any other target m′ , then the association is considered correct; Otherwise, the association is incorrect; : The association accuracy rate of the m-th target in the -th time period, and the result is a binary value; The association strength of target m with itself in the time period ; : The association strength of target m with other target m′ in the time period ; : Traverse and compare all target m' that are different from m; Ninth, normalize the subjective evaluation score, and the formula is: ; Among them, represents the normalized subjective evaluation score of the m-th participant in the j-th experiment; : The original score of the m-th participant's specific evaluation of the k'-th item in the j-th experiment; K': The total number of evaluation items for each participant or experiment; M: The total number of participants; J: The total number of experiments; Participant: Refers to the human subjects who actually participate in the experiment, rather than the targets in the radar or sensor; Tenth, calculate the objective body movement index, and the formula is: ; Calculate the objective body movement index, by quantifying the movement energy of the m-th participant in the j-th experiment, reflecting its overall movement intensity, where: : The objective body movement index of the m-th participant in the j-th experiment; : The original movement signal at time t; L: The total number of cycles of the experiment; : The duration of a single cycle; : The total time length of the experiment; K: The number of categories of movement indicators; Use the CFAR target detection algorithm to count the number of people within the radar detection range, analyze the movement states of people, identify the entry and exit events of people in different movement states, and judge the characteristics of people from the radar echo data for counting.

[0008] As a further preferred technical solution of the above technical solution, the analysis of the action anomaly detection module is specifically implemented as: For the selected clustering clusters, perform feature extraction, including extracting the spatial coordinates of each limb part at each time point, obtaining the instantaneous velocity of the limb part by calculating the displacement between adjacent time points, and obtaining the acceleration information of the limb part by differentiating the velocity; obtain the acceleration information of the target based on the point cloud data of the millimeter-wave radar, and the specific implementation is: Point cloud data acquisition: The millimeter-wave radar emits electromagnetic waves with a frequency of 77 - 81 GHz, and obtains the point cloud data of the target by receiving the reflected signal and performing signal processing. Each point contains Three-dimensional space coordinates and timestamp t; Coordinate sequence construction: For the filtered clustering clusters, extract the spatial coordinate sequences of each limb part in consecutive time frames ; where: ; ; Velocity calculation: For time point , calculate the instantaneous velocity through displacement difference of adjacent time points: ; Acceleration calculation: Perform second-order difference on the obtained velocity sequence to calculate the instantaneous acceleration at time point : ; Assume the time interval is uniform, acceleration noise reduction processing: Use the Kalman filter algorithm to perform noise reduction processing on the original acceleration data to filter out the noise generated by environmental interference and measurement errors; Velocity is an important indicator for judging the speed of an action. By calculating the moving speed of each limb part, the speed of the action can be judged. The speed calculation formula is: ; where v is the velocity, d is the displacement, and t is the time; Displacement calculation: For each limb part, calculate the displacement between adjacent time points; Time interval: Record the timestamp of each acquisition and calculate the time interval; Trajectory analysis: Combine the movement trajectory of the limb to judge the amplitude of the action. By calculating the path length and displacement of the limb movement, the amplitude of the action can be effectively identified; The path length calculation formula is as follows: ; where, is the path length, is the coordinate of the i-th point; Path calculation: Perform path calculation on the point cloud data of each limb part to obtain the total path length of each action; Displacement judgment: Judge the action amplitude by comparing the path length with a preset amplitude threshold; Action amplitude judgment: The judgment of the action amplitude depends on the movement range of the limb part. By calculating the maximum displacement of the upper and lower limbs within a certain time, the amplitude of the action is determined; Maximum displacement calculation: Record the initial position of each limb part and calculate the maximum displacement during subsequent acquisitions. The formula is: ; Among them, D is the maximum displacement, is the current position coordinate, is the initial position coordinate; Amplitude threshold setting: If the maximum displacement exceeds the preset amplitude threshold, then this action is considered a valid action; Action frequency judgment: The judgment of the action frequency is based on the number of movements completed by the limb part within a unit time. By counting the movement changes of the limb part within the set time period, the frequency of the action is calculated. The specific method is as follows: Frequency calculation: Within a certain time window, count the number of state changes of the limb part and calculate the frequency. The formula is: , where, is the frequency, is the number of state changes, is the length of the time window; Frequency threshold setting: If the frequency exceeds the preset frequency threshold, then this action is considered a frequent action; Overlapping contact analysis: By analyzing the overlapping contact situation with other people, further confirm the effectiveness of the action. The overlapping contact analysis uses the spatial distance calculation method. When the limb contacts of two people are frequent and overlapping, it means that a certain interactive action is in progress. Set an overlapping threshold. When the distance between two people is less than this threshold, it is considered that there is an overlapping contact; Distance calculation: Calculate the spatial distance between participants, using the Euclidean distance formula: ; where, d is the distance, and are the coordinates of two points; Overlapping judgment: If the calculated distance is less than the set overlapping threshold, then record it as an overlapping contact; Threshold setting: To effectively distinguish general limb movements from specific actions, set a medium threshold. This threshold is based on historical data analysis and experimental results and is finally determined to be a specific speed and amplitude range after multiple adjustments; Through the above action analysis, obtain the amplitude and frequency of the action, and through the setting of the medium threshold, determine it as a valid action and determine whether it is an abnormal action.

[0009] As a further preferred technical solution of the above technical solution, the analysis of the personnel tilt detection module is specifically implemented as: Use 2D-FFT, 2D-CFAR and the point cloud angle based on FFT for the abnormal point cloud cluster to obtain the moving target point cloud information, including distance, height, angle and energy, and perform corresponding calculations: Average speed: Calculate the average speed of the target within a period of time to measure the speed of the target movement. The calculation formula is as follows: ; Among them, represents the i-th speed data, and N represents the total number of data points; Average acceleration: Calculate the average acceleration of the target over a period of time to measure the degree of acceleration or deceleration of the target. The calculation formula is as follows: ; Among them, represents the i-th acceleration data, and N represents the total number of data points; Speed standard deviation: Calculate the standard deviation of the target speed to measure the degree of fluctuation of the target speed. The calculation formula is as follows: ; Maximum speed: Calculate the maximum speed of the target over a period of time to measure the instantaneous fastest speed of the target's movement. The calculation formula is as follows: ; Fall duration: Calculate the duration of the fall event from start to end to measure the severity of the fall event. The calculation formula is as follows: ; Input the obtained data into the AdaBoost algorithm model. The principle and formula of the AdaBoost algorithm are as follows: Initialize the weight distribution of each sample: ; For each round of iteration Use the weight distribution Dt to train the training data to obtain a weak classifier ; Calculate the classification error rate of the weak classifier on the training data. The formula is: ; Among them indicates whether the i-th sample is misclassified. If it is misclassified, it is 1; otherwise, it is 0; Calculate the weight of the weak classifier , and the formula is: ; Update the weight distribution of the training data: ; Among them is the normalization factor, making become a probability distribution. The calculation formula is: ; After T rounds of iterations, T weak classifiers and their weights are obtained, and the final strong classifier is: ; in represents the sign of x, if 1 if yes, -1 otherwise.

[0010] As a further preferred technical solution of the above technical solution, the warning level judgment and response module is based on the multi-dimensional analysis of the bullying detection module, and the warning level is divided into three levels: Level 1 warning Trigger conditions: The number of people is 4-5, and the movement amplitude and frequency are both below the threshold; Brief physical contact was detected, but did not last more than 10 seconds; The tilt is slight and the center of mass velocity is normal; Countermeasures: Local reminder: trigger the on-site sound and light alarm to remind relevant personnel to pay attention to behavioral norms; Log records: automatically save data fragments of abnormal time periods for subsequent verification; Security inspection: notify the nearest security personnel to check and lift the warning after confirming that there is no risk; Second level warning Trigger conditions: The number of people is 6-10, or the action frequency exceeds the threshold; Sustained physical contact is detected with an overlap distance of <0.5m; The tilt amplitude is significant and the center of mass velocity is abnormal; Countermeasures: Academic affairs intervention: push warning information to the academic affairs management platform; Multi-terminal linkage: Mobile terminal: Send location and brief event description to the teacher on duty; Command center: displays three-dimensional heat maps and personnel trajectories to assist in remote analysis and judgment; Broadcast warning: voice prompts are played through regional broadcasting; Level 3 warning Trigger conditions: The number of people is >15, or violent body movements are detected; Overlapping contacts are frequent; Center of mass velocity > 350 pixels / second or trunk tilt angle > 75°; Countermeasures: Emergency Response: Linking with the campus security center, activating the one-button alarm system, and notifying the police to intervene; Broadcast evacuation instructions to avoid secondary risks; Real-time tracking: Automatically label the trajectories of the perpetrators and victims, and push high-definition point cloud data to the command large screen; Dispatch drones or patrol robots to the scene for evidence collection; Post-incident handling: Generate a detailed incident report; Activate the psychological counseling plan to intervene in the involved personnel.

[0011] As a further preferred technical solution of the above technical solution, when the millimeter-wave radar is used for the first time, the clustering deformation detection algorithm is adopted for preliminary inspection to determine the size of the private space, and at the same time, the size of the private space to be measured is confirmed, and the parameters are adjusted to make the millimeter-wave radar in the best range resolution mode. Brief Description of the Drawings

[0012] Figure 1 is a schematic diagram of the system of the present invention.

[0013] Figure 2 is a schematic diagram of the system of the present invention deployed in a privacy area (dormitory).

[0014] Figure 3 is a schematic diagram of the system of the present invention deployed in a privacy area (toilet). Detailed Description of the Preferred Embodiment

[0015] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments in the following description are only examples, and those skilled in the art can think of other obvious variations. The basic principles defined in the following description of the present invention can be applied to other implementation schemes, variant schemes, improvement schemes, equivalent schemes, and other technical schemes without departing from the spirit and scope of the present invention.

[0016] In the preferred embodiment of the present invention, those skilled in the art should note that the millimeter-wave radar and the like involved in the present invention can be regarded as the prior art.

[0017] Preferred embodiment.

[0018] As Figures 1-3 shown, the present invention discloses a campus bullying monitoring system based on a millimeter-wave radar, including a millimeter-wave radar, a data transmission module, a data collector, an abnormal point cloud cluster detection module, a bullying detection module, and a warning level judgment and response module, wherein: The millimeter-wave radar collects radar signals of the private space and transmits the radar signals to the data transmission module, so that the data collector receives the radar signals transmitted by the data transmission module and transmits them back to the abnormal point cloud cluster detection module; The abnormal point cloud cluster detection module processes the received radar signals using the DBSCAN algorithm to generate clustering clusters, screens the generated clustering clusters using the GNN (Global Nearest Neighbor) algorithm, filters out abnormal clustering clusters by setting a threshold, and transfers the data to the bullying detection module for analysis; The bullying detection module analyzes the abnormal clustering clusters three times through the personnel counting detection module, the action anomaly detection module, and the personnel dumping detection module (these three sub-modules) respectively, so as to obtain corresponding data and transmit it to the warning level judgment and response module after data fusion; The warning level judgment and response module processes the received fusion data, judges the warning level, and takes corresponding response measures.

[0019] Specifically, the abnormal point cloud cluster detection module is specifically implemented as follows: First, the received radar signals (i.e., echo data) are processed including removing clutter and noise. The processed radar signals are then discretely sampled to generate a frame sequence, and the frame sequence is cut and reorganized to obtain a frame matrix; the DBSCAN algorithm is performed on the generated frame matrix, and the data in the frame matrix generates clustering clusters, and then the generated clustering clusters are screened using the GNN algorithm to screen out clustering clusters with the number of people greater than the first preset number of people (preferably 3 people); in addition, when the number of people in the clustering cluster is greater than the second preset number of people (preferably 15 people), a warning reminder is directly triggered, a notice of prohibiting gathering is issued, and the teaching staff is notified to check the situation.

[0020] More specifically, the analysis of the personnel counting detection module is specifically implemented as follows: According to the regional movement amplitude, limb reflection angle, speed and acceleration of the personnel (i.e., targets), they are classified into different motion states for subsequent identification; the motion states include walking, running, retreating, jumping and standing still, and the personnel in different motion states are respectively counted and statistically processed; the root mean square value (RMS) is used to calculate the Doppler velocity of multiple targets obtained from the radar signals as the body motion index. The algorithm adopted by the personnel counting detection module is CFAR target detection and Kalman filter tracking and positioning. The relevant content is as follows: First, calculate the initial radar image, and the formula is: , where represents the radar image in the r distance and θ direction at time t, represents the radar image intensity value at time t, distance r and direction θ, which is an imaging result obtained by weighted synthesis of signals from multiple channels (or sensors), reflecting the target echo energy or characteristics at the position, specifically: N: Total number of channels (or sensors, beams, array elements); : Weighting coefficient of the nth channel, used to adjust the contribution degree of each channel to the final image (such as gain, calibration coefficient or noise suppression weight); : Beamforming weight in the direction θ, related to the radiation pattern or beam pointing of the radar antenna, determining the response intensity of the nth channel in the direction θ; : Original signal of the nth channel at time t and distance r (the result after processing such as time delay compensation and filtering of the echo signal); t: Time variable, corresponding to the moment of radar pulse emission or data acquisition; r: Distance variable, representing the radial distance from the target to the radar; θ: Azimuth angle or direction variable, representing the pointing angle of the radar beam; Second, suppress static clutter, the formula is: , where represents the radar image after clutter suppression, suppressing static clutter (such as echoes of stationary targets like the ground and buildings) through time averaging method, and retaining the signals of dynamic targets (such as vehicles and aircraft), where: T: Time window length, that is, the integration interval for calculating the average value (usually select the clutter coherence time or a fixed observation period); : Index of the time window (the th time period, for example = 0 represents the first window t ∈ [0, T]); t′: Integration variable, representing the historical moment; Integral term: Within the time window , perform time averaging on the signal intensities at the same distance r and direction θ to obtain an estimate of the static clutter; Third, estimate the target position, the formula is: , where and respectively represent the position and azimuth angle of the target at time t, is a mathematical operator, indicating to find the parameter pair that makes the function obtain the maximum value (that is, search the entire image plane (all possible r and θ) to locate the point with the highest intensity); Fourth, calculate the new radar image, the initial radar image represents the instantaneous radar image, that is, the range-azimuth signal intensity obtained at a single time point t (which may contain noise and clutter), and the formula for the new radar image is: , where represents the radar image within the th time period, and represents the radar image after time averaging, which is obtained by performing energy integration and averaging on the instantaneous images within the th time period; is obtained by performing energy integration and averaging on the instantaneous images within the : the average radar image of the th time period, with the unit of energy or power; : the instantaneous radar image, representing the complex signal (usually containing amplitude and phase information) at time t, distance r, and azimuth angle θ; The energy of the instantaneous signal (modulus square operation) is used to eliminate the phase influence and retain the intensity information; T: the length of the time window, that is, the integration time interval (such as 1 second, 10 pulse periods, etc.); : the index of the time period; The integration term : accumulates the energy of all instantaneous signals within the th time period; Fifth, calculate the body movement index, and the formula is: , which is used to quantify the intensity of body movement. By analyzing the energy of the derivative (i.e., velocity) of the movement signal d(τ), it reflects the severity of the movement (the faster the movement speed, the larger the absolute value of the derivative , the higher the integrated value after squaring; time averaging (dividing by Tb after integration) smooths the instantaneous fluctuations to obtain a stable movement index; the square root operation restores the energy dimension to a physical quantity consistent with velocity for intuitive understanding), where: : the body movement index at time t, with the unit of velocity (such as m / s), and the larger the value, the more severe the movement; : the original movement signal (which may be physical quantities related to body movement such as displacement, acceleration, radar echo phase change, etc.); : the time derivative of the movement signal, representing the instantaneous velocity or rate of change; Take the modulus square (if d(τ) is a complex signal) or directly square (if it is a real number) to calculate the energy; Tb: the length of the time window, that is, the smoothing interval for calculating the movement index (such as 2 seconds); The integration interval : a symmetric window centered at t with a width of T b to ensure the requirements of causality or real-time performance; Sixth, calculate the correlation coefficient, with the formula: ; Calculate the correlation coefficient between the body movement indices measured by two radars within the same time period, which is used to quantify whether the movement signals are synchronized or similar, where: Value range: ; 1: Perfect positive correlation (the changes of the two radar signals are exactly the same); : Perfect negative correlation (the signal changes are exactly opposite); 0: No correlation; : During the th time period, the body movement indices of radar m and radar m′ are and correlation coefficient; and : The body movement indices (such as speed, energy, etc.) measured by two independent radars (or different channels of the same radar) at time t; : The time window length for calculating the correlation coefficient (e.g., 5 seconds); : The index of the time period (the th window ); The numerator part: The covariance of the two signals within the window, reflecting the joint change trend; The denominator part: The geometric mean of the energies of the two signals, which is used to normalize the covariance and eliminate the influence of amplitude differences; Seventh, calculate the target association accuracy rate, with the formula: , which is used to calculate the overall accuracy rate of multi-target tracking or association tasks. By statistically counting the association correct rates of all targets and all time periods and taking their average value, where: : The overall accuracy rate of target association (the value range is [0,1], and the higher the value, the better the association performance); M: The total number of targets, indicating the number of independent targets that the system needs to track or associate; L: The total number of time periods, indicating the total number of time windows for the system to run (e.g., L = 10 means 10 time periods); : The association correct rate of the mth target during the th time period, which is a binary value (0 or 1) or a probability value (between 0 and 1); If it is a binary value: = 1: indicates that the target m is correctly associated within the time period ; = 0: indicates an incorrect association; If it is a probability value: it represents the confidence level or partial correctness of the association; Eighth, define the target association accuracy rate, with the formula: ; The binary decision rule is used to determine whether the m-th target in the -th time period is correctly associated. Its core logic is: If the association strength of the target m with itself is not lower than its association strength with any other target m', , then the association is considered correct ( = 1); Otherwise, the association is incorrect ( = 0); : the association accuracy rate of the m-th target in the -th time period, and the result is a binary value (0 or 1); The association strength of the target m with itself (or the ideal reference) within the time period (usually quantified by correlation coefficients, matching scores, etc.); : the association strength of the target m with other target m' (m' ≠ m) within the time period ; : traverse and compare all target m' different from m; Ninth, normalize the subjective evaluation scores, with the formula: ; Among them, represents the normalized subjective evaluation score of the m-th participant in the j-th experiment; : the original score of the m-th participant's specific evaluation of the k'-th item in the j-th experiment (such as the score for a certain indicator); K': the total number of evaluation items for each participant or experiment (such as the number of questionnaire questions); M: the total number of participants (note: Mj may be the number of participants in the j-th experiment, and it is necessary to confirm whether it is consistent with M); J: the total number of experiments; Participant (m): Refers to the human subjects who actually participate in the experiment (such as users, subjects), rather than the targets in the radar or sensor (e.g., in a user experience study, the participant may be the test user; in a psychological experiment, it may be the subject group). (Target: If the context involves radar / sensor, "target" refers to the object being detected (such as a pedestrian, vehicle), which is different from "participant". Since the formula here is specifically for subjective scoring, "participant" should be a person); Tenth, calculate the objective body movement index, with the formula: ; Calculate the objective body movement index, which reflects the overall movement intensity by quantifying the movement energy of the m-th participant in the j-th experiment, where: : The objective body movement index of the m-th participant in the j-th experiment (the unit is the same as that of bm(j)(t), such as mm / s, g, etc.); : The original movement signal at time t (which may come from sensors such as accelerometers, radar micro-Doppler, video pose estimation, etc.); L: The total number of cycles of the experiment (such as the number of repeated movements); : The duration of a single cycle (such as one walking cycle is 1 second); : The total time length of the experiment (such as L = 10, Tp = 1 second → total time 10 seconds); K: The number of categories of movement indicators (such as K = 3 can correspond to the movement signals of the head, hands, and legs); Using the CFAR target detection algorithm, count the number of people within the radar detection range, analyze the movement states of people, and identify the entry and exit events of people in different movement states. (Using machine algorithms such as deep learning) judge the characteristics of people from the radar echo data for counting (during the counting process, exclude interference by analyzing the limbs and setting medium thresholds: Limb analysis: Analyze the moving speed, trajectory of the point cloud of the upper and lower limbs, and the overlapping contact situation with other people to determine whether the limb movement is a valid action; Setting of medium thresholds: According to the movement speed range of the human limbs, limb movements with a speed lower than the threshold of 0.5 m / s are regarded as general limb movements and can be excluded from the scope of personnel counting; according to the movement trajectory range of the human limbs, limb movements with a trajectory length lower than the threshold of 0.5 m are regarded as general limb movements; similarly, when the overlapping contact threshold is set above 0.1 square meters, it can be regarded as excessive human contact. By setting these three thresholds, general limb movements can be excluded; Through the above methods, the counting of people can be achieved).

[0021] Furthermore, the analysis of the abnormal action detection module is specifically implemented as follows: For the selected clustering clusters, feature extraction is performed, including extracting the spatial coordinates of each limb part at each time point, obtaining the instantaneous velocity of the limb part by calculating the displacement between adjacent time points, and obtaining the acceleration information of the limb part by differentiating the velocity; obtaining the acceleration information of the target based on the point cloud data of the millimeter-wave radar, which is specifically implemented as follows: Point cloud data acquisition: The millimeter-wave radar emits electromagnetic waves with a frequency of 77 - 81 GHz, and the point cloud data of the target is obtained by receiving the reflected signal and performing signal processing. Each point contains three-dimensional spatial coordinates and time stamp t; Coordinate sequence construction: For the selected clustering clusters, extract the spatial coordinate sequence of each limb part in consecutive time frames ; where: ; Velocity calculation: For time point , calculate the instantaneous velocity by differentiating the displacement between adjacent time points: ; Acceleration calculation: Perform a second-order differentiation on the obtained velocity sequence to calculate the instantaneous acceleration at time point : ; Assuming that the time interval is uniform, acceleration noise reduction processing: Use the Kalman filter algorithm to perform noise reduction processing on the original acceleration data to filter out the noise generated by environmental interference and measurement errors; Parameters: Acceleration amplitude: , used to measure the intensity of the action, acceleration change rate: , used to judge the degree of action mutation, acceleration duration: The time length during which the continuous acceleration exceeds the threshold, acceleration direction change: Calculate the angle between the acceleration vectors of adjacent time points; Acceleration threshold and behavior recognition: Detecting brief limb contact (such as pushing and shoving): The characteristic is that the acceleration peak value > 2.0 m / s² and the duration < 0.5 seconds; Detecting continuous limb contact (such as pulling and blocking); The characteristic is that the acceleration fluctuates within the range of 0.8 - 2.0 m / s² and the duration > 1 second; Detecting violent limb actions (such as beating and kicking): The characteristic is that the acceleration peak value > 2.5 m / s² and shows regular multiple peaks.

[0022] Velocity is an important indicator for judging the speed of an action. By calculating the moving speed of each limb part, the speed of the action is judged. The speed calculation formula is: ; Among them, v is the velocity, d is the displacement, and t is the time; Displacement calculation: For each limb part, calculate the displacement between adjacent time points; Time interval: Record the time stamps of each acquisition and calculate the time interval; Trajectory analysis: Combine the movement trajectories of the limbs to judge the amplitude of the action. By calculating the path length and displacement of the limb movement, the (larger) amplitude of the action can be effectively identified; The formula for calculating the path length is as follows: ; Among them, is the path length, is the coordinate of the i-th point; Path calculation: Perform path calculation on the point cloud data of each limb part to obtain the total path length of each action; Displacement judgment: Judge the action amplitude by comparing the path length with a preset amplitude threshold; Action amplitude judgment: The judgment of the action amplitude depends on the movement range of the limb part. By calculating the maximum displacement of the upper and lower limbs within a certain time, the amplitude of the action is determined; Maximum displacement calculation: Record the initial position of each limb part and calculate the maximum displacement during subsequent acquisitions. The formula is: ; Among them, D is the maximum displacement, is the current position coordinate, is the initial position coordinate; Amplitude threshold setting: If the maximum displacement exceeds the preset amplitude threshold, the action is considered a valid action; Action frequency judgment: The judgment of the action frequency is based on the number of movements completed by the limb part within a unit time. By counting the movement changes of the limb part within a set time period, the frequency of the action is calculated. The specific method is as follows: Frequency calculation: Within a certain time window, count the number of state changes of the limb part and calculate the frequency. The formula is: , where, is the frequency, is the number of state changes, is the length of the time window; Frequency threshold setting: If the frequency exceeds the preset frequency threshold, the action is considered a frequent action; Overlap Contact Analysis: By analyzing the overlap contact with other people, the effectiveness of the action is further confirmed. The overlap contact analysis uses the spatial distance calculation method. When the limb contacts between two people are frequent and overlapping (more), it indicates that a certain interactive action is in progress. A overlap threshold is set. When the distance between two people is less than this threshold, it is considered that there is overlap contact; Distance Calculation: Calculate the spatial distance between participants, using the Euclidean distance formula: ; where d is the distance, and are the coordinates of two points; Overlap Judgment: If the calculated distance is less than the set overlap threshold, it is recorded as overlap contact; Specific Action Recognition in Overlap State: When the system detects the overlap contact state (that is, the overlap threshold with a distance less than 0.5m), the system further conducts specific recognition and analysis on the specific action types in the overlap state (including pushing, pulling, blocking, beating, and kicking). The specific implementation is as follows: Spatiotemporal Feature Extraction and Representation: According to the multi-scale spatiotemporal feature analysis method, construct a comprehensive spatiotemporal feature vector: ; where V represents the speed feature group, a represents the acceleration feature group, represents the time series feature group, represents the direction feature group, and I represents the interaction feature group; These features have different feature manifestations according to different types of actions.

[0023] Construction of Mutual Dynamics Model: According to the human kinematics model and the mutual dynamics model, model and analyze the interaction relationship between two (or more) people in the overlap state, and identify the interaction type through the relationship between force and reaction force.

[0024] 1. Pushing Action Recognition: Pushing Action Feature Extraction: The pushing action has the following significant features: Action Direction Feature: The millimeter-wave radar can measure the target motion direction and speed through the Doppler effect, and can identify the correlation between the action direction of the force-applying party and the motion direction of the pushed party, represented by the direction similarity index: ; where, is the angle between the action direction of the force-applying party and the motion direction of the pushed party; Kinematics Feature: The pushing action is manifested as the rapid transfer of momentum from the force-applying direction to the pushed party, which can be observed through the micro-Doppler feature of the millimeter-wave radar, expressed as: = Observed change in the speed of the pushed party v; Among them, is the motion impulse feature observed by the radar; Temporal feature: The interaction time of the pushing action is usually short and can be identified by analyzing consecutive frames of the millimeter-wave radar: <0.5s; Among them, is the contact duration; Pushing action discrimination criterion: When the following conditions are met, it is determined as a pushing action: 1. The moving speed of the force-applying party in the target direction >0.8m / s; 2. The direction consistency coefficient >0.85; 3. The contact duration <0.5s; 4. The velocity change of the pushed party after the interaction v>0.5m / s.

[0025] 2. Pulling action recognition: Pulling action feature extraction: The pulling action has the following significant features: Continuous contact feature: The pulling action usually has a long contact duration.

[0026] Pull force transmission feature: Capture the pull force transmission feature through the analysis of the motion trajectory correlation. The trajectory correlation degree calculation formula: ; Among them, is the velocity vector of the force-applying party, is the velocity vector of the pulled party, is the time delay parameter; Relative position constraint: During the pulling process, the two people maintain a specific relative position relationship. The relative position constraint metric: ; Among them, represents the standard deviation operation, and respectively represent the position coordinates of the two people; Pulling action discrimination criterion: When the following conditions are met, it is determined as a pulling action: The withdrawal speed vpull of the force-applying party after the initial contact >0.6m / s; The trajectory correlation degree Cpull>0.7; The contact duration tcontact>0.8s; The relative position constraint degree Drel<0.2m; The accuracy rate of such a feature combination in the recognition of pulling behavior can reach 90.5%.

[0027] 3. Recognition of surrounding actions: Feature extraction of surrounding actions: Surrounding actions have the following significant features: Spatial enclosure feature: The enclosers form a closed or semi-closed structure in spatial distribution. The calculation formula for spatial enclosure: ; Where, represents the angular range from the position of the person being surrounded to the i-th encloser; Restriction of movement freedom: The movement range of the person being surrounded is significantly restricted. The calculation formula for movement freedom: ; Where, is the actual movement area, is the movement area that can be moved under normal conditions; Abnormal spatial density: The personnel density in the surrounding area is significantly higher than that in the surrounding areas. The density anomaly coefficient: ; Where, is the personnel density in the surrounding area, is the personnel density in the normal area; Criterion for judging surrounding actions: When the following conditions are met, it is judged as a surrounding action: Spatial enclosure Cenclosure > 0.5 (at least 180° angle is surrounded); Movement freedom Mfreedom < 0.4; Duration of surrounding state tsurrounding > 10s; Density anomaly coefficient Ddensity > 2.0; The accuracy rate of using this combined feature to judge surrounding behavior can reach 95.2%.

[0028] 4. Recognition of beating actions: Feature extraction of beating actions: Beating actions have the following significant features: High-speed impact feature: The high-speed impact feature formed by the rapid back-and-forth movement of the arm, calculated by the peak speed: ; Where, is the observation time window; Acceleration waveform feature: The acceleration curve generated by beating actions has a characteristic peak-valley waveform. The waveform feature coefficient: ; Among them, and are the maximum and minimum values of acceleration respectively, is the time interval between the two; Impact energy: Impact energy characterizes the intensity of beating, and the calculation formula is: ; Among them, and represent the acceleration and velocity at time t respectively; Discrimination criteria for beating actions: When the following conditions are met, it is determined as a beating action: The peak velocity of the arm vpeak > 1.2 m / s; The peak acceleration apeak > 3.5 m / s²; The waveform characteristic coefficient Wacc > 15 m / s³; The impact energy Eimpact > 2.8 J; The accuracy rate of identifying beating actions with these characteristic combinations can reach 93.8%.

[0029] 5. Identification of kicking actions: Feature extraction of kicking actions: Kicking actions have the following significant features: Lower limb trajectory feature: The lower limb of the kicking action shows a characteristic arc trajectory, and the trajectory curvature calculation formula is: ; Among them, v is the velocity vector and a is the acceleration vector; Height feature: The height of the lower limb lifted during kicking is significantly higher than that in the normal walking state.

[0030] Velocity segmentation feature: The kicking action usually includes three stages: preparation - extension - recovery, and the velocity features of each stage are significantly different. The segmentation feature calculation is: ; Among them, , , represent the average velocities of the three stages respectively, is the maximum velocity; Discrimination criteria for kicking actions: When the following conditions are met, it is determined as a kicking action: 1. The lower limb extension velocity vkick > 1.5 m / s; 2. The trajectory curvature K > 0.85; 3. The target acceleration after contact aimpact > 2.0 m / s²; The accuracy rate of identifying kicking actions with this feature combination can reach 89.5%.

[0031] Multi - feature fusion decision mechanism: To improve the robustness and accuracy of recognition, this system adopts a multi - feature fusion method based on Dempster - Shafer evidence theory to construct a comprehensive decision formula: ; Among them, represents the comprehensive evidence support degree of the action type , represents the weight of the j - th feature, represents the support degree of the j - th feature for the action type . When the support degree of a certain action type exceeds the preset threshold θdecision (set to 0.75), the system determines it as the action of this type.

[0032] Threshold setting: To effectively distinguish general limb movements from specific actions, a medium threshold is set. This threshold is based on historical data analysis and experimental results, and after multiple adjustments, it is finally determined as a specific speed and amplitude range (specifically: speed threshold: set to 0.5m / s. Amplitude threshold: set to 0.3m. Frequency threshold: set to 3 times per minute. Overlap threshold: set to 0.5m); Through the above action analysis, the amplitude and frequency of the action are obtained, and through the setting of the medium threshold, it is determined as a valid action and whether it is an abnormal action is determined.

[0033] Furthermore, the analysis of the personnel tilt detection module is specifically implemented as follows: For the abnormal point cloud cluster, 2D - FFT, 2D - CFAR and the point cloud angle based on FFT are used to obtain the moving target point cloud information, including distance, height, angle and energy, and corresponding calculations are carried out: Mean Velocity (MV): Calculate the average velocity of the target within a period of time to measure the speed of the target movement. The calculation formula is as follows: ; Among them, represents the i - th velocity data, and N represents the total number of data points; Mean Acceleration (MA): Calculate the average acceleration of the target within a period of time to measure the degree of acceleration or deceleration of the target. The calculation formula is as follows: ; Among them, represents the i - th acceleration data, and N represents the total number of data points; Standard Deviation of Velocity (SDV): Calculate the standard deviation of the target velocity to measure the degree of fluctuation of the target velocity. The calculation formula is as follows: ; Maximum Velocity (MaxV): Calculate the maximum velocity of the target within a certain period of time to measure the instantaneous fastest velocity of the target. The calculation formula is as follows: ; Fall Duration (FD): Calculate the duration from the start to the end of the fall event to measure the severity of the fall event. The calculation formula is as follows: ; Input the obtained data into the AdaBoost algorithm model (AdaBoost (Adaptive Boosting) is an ensemble learning method that improves the classification performance by combining multiple weak classifiers into a strong classifier. The basic idea of AdaBoost is to weight the training samples according to the classification error rate in the previous round in each iteration, so that the misclassified samples receive more attention in the next round. At the same time, each weak classifier has a weight indicating its importance in the final classifier. In this way, AdaBoost can adaptively adjust the weights of each weak classifier to improve the performance of the overall classifier). The principle and formula of the AdaBoost algorithm are as follows: Initialize the weight distribution of each sample: ; For each iteration (T is the number of iterations) Use the weight distribution Dt to train the training data to obtain a weak classifier ; Calculate the classification error rate of the weak classifier on the training data. The formula is: ; where indicates whether the i-th sample is misclassified. If it is misclassified, it is 1; otherwise, it is 0; Calculate the weight of the weak classifier , and the formula is: ; Update the weight distribution of the training data: ; where is the normalization factor, such that Becomes a probability distribution, and the calculation formula is: ; After T rounds of iteration, T weak classifiers and their weights are obtained, and the final strong classifier is: ; Where Represents the sign of x. If Then it is 1, otherwise it is -1 (judged by the final strong separator for falling, and the correct rate of falling reaches 99.7%, with high accuracy).

[0034] Preferably, the warning level judgment and response module is based on the multi-dimensional analysis of the bullying detection module (number of people, inclination amplitude, movement amplitude and frequency, overlapping contact, etc.), and the warning level is divided into three levels: Level 1 warning (low risk) Trigger conditions: The number of people is 4 - 5, and both the movement amplitude and frequency are lower than the threshold; Brief limb contact (such as pushing and shoving) is detected, but it does not last for more than 10 seconds; The inclination amplitude is slight (trunk inclination angle < 30°), and the centroid velocity is normal; Response measures: Local prompt: Trigger the on-site sound and light alarm to remind relevant personnel to pay attention to behavioral norms; Log record: Automatically save the data segment during the abnormal time period for subsequent verification; Security patrol: Notify the nearby security personnel to go and check, and lift the warning after confirming no risk; Level 2 warning (medium risk) Trigger conditions: The number of people is 6 - 10, or the movement frequency exceeds the threshold (> 5 times / minute); Continuous limb contact (such as pulling and blocking) is detected, and the overlapping distance < 0.5 meters; The inclination amplitude is significant (trunk inclination angle 30° - 60°), and the centroid velocity is abnormal (> 200 pixels / second); Response measures: Educational affairs intervention: Push the warning information to the educational affairs management platform; Multi-terminal linkage: Mobile terminal: Send the location and a brief event description (such as "suspected conflict gathering") to the on-duty teacher; Command center: Display the 3D heat map and personnel trajectory to assist remote judgment; Broadcast warning: Play voice prompts (such as "Please keep order") through the area broadcast; Level 3 warning (high risk) Trigger conditions: The number of people is >15, or violent physical movements (such as beating, kicking, etc.) are detected; Frequent overlapping contacts (distance <0.3 m and duration >30 seconds); The center of mass velocity > 350 pixels / second (determined as falling) or the trunk tilt angle > 75°; Countermeasures: Emergency Response: Linking with the campus security center, activating the one-button alarm system, and notifying the police to intervene; Broadcast evacuation instructions (e.g., “Please leave the area immediately”) to avoid secondary risks; Real-time tracking: Automatically mark the tracks of the perpetrator and the victim, and push high-definition point cloud data to the command screen; Dispatching drones or patrol robots to the scene to collect evidence; Post-event treatment: Generate detailed incident reports (including time, location, type of behavior, video evidence); Initiate a psychological counseling plan and intervene with the people involved.

[0035] It also includes a dynamic downgrade mechanism: if the abnormal behavior terminates on its own within 30 seconds, the system will automatically downgrade or lift the warning.

[0036] False alarm optimization: Through the reinforcement learning model, the threshold sensitivity is optimized in combination with historical disposal data (such as reducing the false positive rate of static aggregation).

[0037] Privacy protection: All warning data is desensitized and only the key bone point information is retained.

[0038] This grading system quantifies abnormal behavior characteristics and combines multi-level response strategies, taking into account both early warning efficiency and privacy security. It is suitable for refined security management of private places on campus.

[0039] Preferably, when the millimeter-wave radar is used for the first time, a clustered deformation detection algorithm is used for preliminary inspection to determine the size of the private space. At the same time, the size of the private space to be tested is confirmed, and parameters are adjusted so that the millimeter-wave radar is in the optimal distance resolution mode.

[0040] It is worth mentioning that the technical features such as millimeter-wave radar involved in the patent application of this invention should be regarded as prior art. The specific structure, working principle and possible control method and spatial layout method of these technical features can be selected by conventional methods in the field, and should not be regarded as the inventive point of this patent. This patent will not be further elaborated.

[0041] Those skilled in the art can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A campus bullying monitoring system based on millimeter-wave radar, characterized in that, It includes a millimeter-wave radar, a data transmission module, a data collector, an abnormal point cloud cluster detection module, a bullying detection module, and a warning level judgment and response module, where: The millimeter-wave radar collects radar signals in a private space and transmits the radar signals to the data transmission module, so that the data collector receives the radar signals transmitted by the data transmission module and transmits them back to the abnormal point cloud cluster detection module; The abnormal point cloud cluster detection module processes the received radar signals using the DBSCAN algorithm to generate clustering clusters, screens the generated clustering clusters using the GNN algorithm, and filters out abnormal clustering clusters by setting thresholds, and transfers the data to the bullying detection module for analysis; The bullying detection module analyzes the abnormal clustering clusters three times through a personnel counting detection module, an action abnormality detection module, and a personnel dumping detection module respectively, so as to obtain corresponding data and transmit it to the warning level judgment and response module after data fusion; The warning level judgment and response module processes the received fusion data, judges the warning level, and takes corresponding response measures.

2. The campus bullying monitoring system based on millimeter-wave radar according to claim 1, wherein For the specific implementation of the abnormal point cloud cluster detection module: First, the received radar signals are processed including removing clutter and noise. The processed radar signals are then discretely sampled to generate a frame sequence, and the frame sequence is cut and reorganized to obtain a frame matrix; the DBSCAN algorithm is applied to the generated frame matrix, and the data in the frame matrix generates clustering clusters, and then the generated clustering clusters are screened using the GNN algorithm to screen out clustering clusters with the number of people greater than the first preset number of people; In addition, when the number of people in a clustering cluster is greater than the second preset number of people, a warning reminder is directly triggered, a notice of prohibiting gathering is issued, and the teaching staff is notified to check the situation.

3. The campus bullying monitoring system based on millimeter wave radar according to claim 2, wherein, For the analysis of the personnel counting detection module, the specific implementation is as follows: According to the regional movement amplitude, limb reflection angle, speed and acceleration of personnel, different movement states are distinguished for subsequent identification; the movement states include walking, running, retreating, jumping and standing still, and the personnel in different movement states are respectively counted and statistically processed; the root mean square value is used to calculate the Doppler velocity of multiple targets obtained from the radar signals as the body movement index; The algorithm adopted by the personnel counting detection module is CFAR target detection and Kalman filter tracking and positioning. The relevant content is as follows: First, calculate the initial radar image using the formula: , where represents the radar image at distance r and direction θ at time t, represents the radar image intensity value at time t, distance r, and direction θ, which is an imaging result obtained by weighted synthesis of signals from multiple channels, reflecting the target echo energy or characteristics at the position, specifically: N: The total number of channels; : The weighting coefficient of the nth channel, which is used to adjust the contribution degree of each channel to the final image; : The beamforming weight in the direction θ, which is related to the radiation pattern or beam pointing of the radar antenna, determines the response intensity of the nth channel in the direction θ; : The original signal of the n-th channel at time t and distance r; t: The time variable, corresponding to the moment of radar pulse emission or data collection; r: The distance variable, indicating the radial distance from the target to the radar; θ: The azimuth angle or direction variable, indicating the pointing angle of the radar beam; Second, suppress static clutter, and the formula is: , where represents the radar image after clutter suppression. Static clutter is suppressed by the time averaging method, and the signals of dynamic targets are retained, where: T: The time window length, that is, the integration interval for calculating the average value; : Index of the time window; t′: The integration variable, indicating the historical moment; Integral term: Within the time window the time average of the signal intensity at the same distance r and direction θ is calculated to obtain an estimate of the static clutter; Third, estimate the target position, with the formula: , where and represent the position and azimuth of the target at time t respectively, is a mathematical operator, indicating finding the parameter pair that maximizes the function ; Fourth, calculate the new radar image, the initial radar image represents the instantaneous radar image, that is, the range-azimuth signal intensity obtained at a single time point t. The formula for the new radar image is: , where represents the radar image within the th time period, represents the radar image after time averaging, and is obtained by performing energy integration and averaging on the instantaneous images within the th time period; : The average radar image of the th time period, in units of energy or power; : An instantaneous radar image representing a complex signal at time t, range r, and azimuth angle θ; The energy of the instantaneous signal is used to eliminate the phase influence and retain the intensity information; T: The time window length, that is, the integration time interval; : Index of the time period; Integral term : Cumulate the energy of all instantaneous signals within the th time period; Fifth, calculate the body movement index, with the formula: , which is used to quantify the intensity of body movement. By analyzing the energy of the derivative of the movement signal d(τ), it reflects the intensity of the movement, where: : Body movement index at time t, with the unit of speed. The larger the value, the more intense the movement; : original motion signal; : The time derivative of a motion signal, representing the instantaneous velocity or rate of change; Modulo squaring or direct squaring is used to calculate energy; Tb: The time window length, that is, the smoothing interval for calculating the movement index; Integration interval : A symmetric window centered at t with a width of T b to ensure causality or real-time requirements; Sixth, calculate the correlation coefficient, and the formula is: ; Calculate the correlation coefficient between the body movement indices measured by two radars within the same time period to quantify whether the movement signals are synchronized or similar, where: Range of values: ; 1: Completely positively correlated; : Perfect negative correlation; 0: No correlation; : During the th time period, the correlation coefficient of the body movement indices of radar m and radar m' and ; and : Body movement indices measured by two independent radars at time t; : The time window length for calculating the correlation coefficient; : Index of the time period; Molecular part: The covariance of the two signals within the window, reflecting the joint change trend; Denominator part: The geometric mean of the energies of the two signals, used to normalize the covariance and eliminate the influence of amplitude differences; Seventh, calculate the target association accuracy rate, and the formula is: , which is used to calculate the overall accuracy of multi-target tracking or association tasks. By counting the correct association rates for all targets and all time periods and taking their average, where: : Overall accuracy associated with the target; M: The total number of targets, indicating the number of independent targets that the system needs to track or associate; L: The total number of time periods, indicating the total number of time windows during which the system runs; : The correlation accuracy of the m-th target in the -th time period is a binary value or a probability value; If it is a binary value: =1: indicates that the target m is correctly associated within the time period and = 0: Indicates an association error; If it is a probability value: Indicating the confidence level or partial correctness of the association; Eighth, define the target association accuracy rate, and the formula is: ； The binary decision rule is used to determine whether the mth target is correctly associated in the th time period. Its core logic is as follows: If the association strength of the target m with itself is not lower than its association strength with any other target m′ , it is considered that the association is correct; Otherwise, the association is incorrect; : The association accuracy of the m-th target in the -th time period, and the result is a binary value; The correlation strength of the target m with itself within the time period ; : The association strength between the target m and other targets m' within the time period ; : Traverse and compare all target m' that are different from m; Ninth, normalize the subjective evaluation score, and the formula is: ； Among them, represents the normalized subjective evaluation score of the m-th participant in the j-th experiment; : The original score of the k'-th specific evaluation of the m-th participant in the j-th experiment; K′: The total number of evaluation items for each participant or experiment; M: The total number of participants; J: The total number of experiments; Participant: Refers to the actual human subjects participating in the experiment, rather than the targets in the radar or sensor; Tenth, calculate the objective body movement index, and the formula is: ； Calculate the objective body movement index by quantifying the movement energy of the m-th participant in the j-th experiment, reflecting their overall movement intensity, where: : The objective body movement index of the m-th participant in the j-th experiment; : The original motion signal at time t; L: The total number of cycles of the experiment; : The duration of a single cycle; : The total time length of the experiment; K: The number of categories of movement indicators; Use the CFAR target detection algorithm to count the number of people within the radar detection range, analyze the movement states of people, and identify the entry and exit events of people in different movement states, and count by judging the human characteristics from the radar echo data.

4. The campus bullying monitoring system based on millimeter wave radar according to claim 3, characterized in that, The specific implementation of the analysis for the action anomaly detection module is as follows: For the selected clustering clusters, feature extraction is performed, including extracting the spatial coordinates of each limb part at each time point, obtaining the instantaneous velocity of the limb part by calculating the displacement between adjacent time points, and obtaining the acceleration information of the limb part by differentiating the velocity; obtaining the acceleration information of the target based on the point cloud data of the millimeter-wave radar. The specific implementation is as follows: Point cloud data acquisition: The millimeter-wave radar emits electromagnetic waves with a frequency of 77-81 GHz, and obtains the point cloud data of the target by receiving the reflected signal and performing signal processing. Each point contains three-dimensional spatial coordinates and time stamp t; Coordinate sequence construction: For the filtered clustering clusters, extract the spatial coordinate sequences of each limb part in consecutive time frames , where:​ ； Speed calculation: For time points , the instantaneous speed is calculated by the displacement difference between adjacent time points: ； Acceleration calculation: Perform a second-order difference on the obtained velocity sequence to calculate the instantaneous acceleration at the time point : ； Assume that the time interval is uniform, and perform acceleration noise reduction processing: Use the Kalman filter algorithm to perform noise reduction processing on the original acceleration data to filter out the noise generated by environmental interference and measurement errors; Speed is an important indicator for judging the speed of an action. By calculating the moving speed of each limb part, the formula for judging the speed of the action is: ； Among them, v is the speed, d is the displacement, and t is the time; Displacement calculation: For each limb part, calculate the displacement between adjacent time points; Time interval: Record the time stamp of each acquisition and calculate the time interval; Trajectory analysis: Combine the movement trajectories of the limbs to judge the amplitude of the action. By calculating the path length and displacement of the limb movement, the amplitude of the action can be effectively identified; The formula for the path length is as follows: ； wherein, is the path length, is the coordinate of the i-th point; Path calculation: Perform path calculation on the point cloud data of each limb part to obtain the total path length of each action; Displacement judgment: Judge the action amplitude by comparing the path length with the preset amplitude threshold; Action amplitude judgment: The judgment of the action amplitude depends on the movement range of the limb part. By calculating the maximum displacement of the upper and lower limbs within a certain period of time, the amplitude of the action is determined; Maximum displacement calculation: Record the initial position of each limb part and calculate the maximum displacement during subsequent acquisitions. The formula is: ； Where D is the maximum displacement, is the current position coordinate, is the initial position coordinate; Amplitude threshold setting: If the maximum displacement exceeds the preset amplitude threshold, then the action is considered a valid action; Action frequency judgment: The judgment of the action frequency is based on the number of movements completed by the limb part within a unit time. By counting the movement changes of the limb part within the set time period, the frequency of the action is calculated. The specific method is as follows: Frequency calculation: Within a certain time window, count the number of state changes of the limb part and calculate the frequency. The formula is: , where is the frequency, is the number of state changes, is the length of the time window; Frequency threshold setting: If the frequency exceeds the preset frequency threshold, then the action is considered a frequent action; Overlapping contact analysis: By analyzing the overlapping contact with other people, the validity of the action is further confirmed. Overlapping contact analysis uses the spatial distance calculation method. When the physical contact between two people is frequent and overlapping, it means that some kind of interactive action is taking place. An overlapping threshold is set. When the distance between the two people is less than the threshold, it is considered that overlapping contact exists. Distance calculation: Calculate the spatial distance between participants using the Euclidean distance formula: ； where d is the distance, and are the coordinates of two points; Overlap judgment: If the calculated distance is less than the set overlap threshold, it is recorded as overlapping contact; Threshold setting: In order to effectively distinguish general body movements from specific movements, a medium threshold is set. This threshold is based on historical data analysis and experimental results. After multiple adjustments, it is finally determined to be a specific speed and amplitude range; Through the above motion analysis, the amplitude and frequency of the motion are learned, and through the setting of the medium threshold, it is determined to be a valid motion, and whether it is an abnormal motion is determined.

5. The campus bullying monitoring system based on millimeter-wave radar according to claim 4, characterized in that, The specific implementation of the analysis of the personnel tilt detection module is as follows: Use 2D-FFT, 2D-CFAR and FFT-based point cloud angle to obtain the moving target point cloud information, including distance, height, angle and energy, and perform corresponding calculations: Average speed: Calculate the average speed of the target over a period of time to measure the speed of the target's movement. The calculation formula is as follows: ； Among them, represents the i-th speed data, and N represents the total number of data points; Average acceleration: Calculates the average acceleration of a target over a period of time. It is used to measure the degree of acceleration or deceleration of the target. The calculation formula is as follows: ； Among them, represents the i-th acceleration data, and N represents the total number of data points; Speed standard deviation: Calculate the standard deviation of the target speed to measure the degree of fluctuation of the target speed. The calculation formula is as follows: ； Maximum speed: Calculates the maximum speed of a target within a period of time. It is used to measure the instantaneous maximum speed of the target's movement. The calculation formula is as follows: ； Fall duration: Calculate the duration from the beginning to the end of a fall event to measure the severity of the fall event. The calculation formula is as follows: ； The obtained data is passed into the AdaBoost algorithm model. The principle and formula of the AdaBoost algorithm are as follows: Initialize the weight distribution of each sample: ; For each iteration Train the training data using the weight distribution Dt to obtain a weak classifier ; Calculate the weak classifier The classification error rate on the training data, with the formula: ； where indicates whether the i-th sample is misclassified, being 1 if misclassified and 0 otherwise; Calculate the weight of the weak classifier as follows , and the formula is: ； Update the weight distribution of training data: ； where is a normalization factor such that becomes a probability distribution, and the calculation formula is as follows: ； After T rounds of iterations, T weak classifiers and their weights are obtained, and the final strong classifier is: ； where represents the sign of x, if then it is 1, otherwise it is -1.

6. The campus bullying monitoring system based on millimeter-wave radar according to claim 5, characterized in that The warning level judgment and response module is based on the multi-dimensional analysis of the bullying detection module, and the warning level is divided into three levels: Level 1 warning Trigger conditions: The number of people is 4-5, and the movement amplitude and frequency are both below the threshold; Brief physical contact was detected, but did not last more than 10 seconds; The tilt is slight and the center of mass velocity is normal; Countermeasures: Local reminder: trigger the on-site sound and light alarm to remind relevant personnel to pay attention to behavioral norms; Log records: automatically save data fragments of abnormal time periods for subsequent verification; Security inspection: notify the nearest security personnel to check and lift the warning after confirming that there is no risk; Second level warning Trigger conditions: The number of people is 6-10, or the action frequency exceeds the threshold; Sustained physical contact is detected with an overlap distance of <0.5m; The tilt amplitude is significant and the center of mass velocity is abnormal; Countermeasures: Academic affairs intervention: push warning information to the academic affairs management platform; Multi-terminal linkage: Mobile terminal: Send location and brief event description to the teacher on duty; Command Center: Display 3D heat maps and personnel trajectories to assist remote judgment. Broadcast Warning: Play voice prompts through area broadcasts. Level 3 Early Warning Trigger Conditions: Number of people > 15, or violent body movements detected. Frequent overlapping contacts. Centroid velocity > 350 pixels / second or torso tilt angle > 75°. Countermeasures: Emergency Response: Link with the campus security center, activate the one-key alarm system, and notify the police to intervene. Broadcast evacuation instructions to avoid secondary risks. Real-time Tracking: Automatically mark the trajectories of the assailant and the victim, and push high-definition point cloud data to the command large screen. Dispatch drones or patrol robots to the scene for evidence collection. Post-incident Handling: Generate a detailed incident report. Activate the psychological counseling plan to intervene in the involved personnel.

7. The campus bullying monitoring system based on millimeter wave radar according to claim 1, characterized in that When the millimeter-wave radar is used for the first time, the clustering deformation detection algorithm is used for preliminary inspection to determine the size of the private space. At the same time, the size of the private space to be measured is confirmed, and the parameters are adjusted so that the millimeter-wave radar is in the best range resolution mode.

Images