An online self-identification system for campus bullying

Abstract

The application provides a campus bullying online self-identification system, which comprises a scene marking module, a scene correlation analysis module, a data acquisition module, and a suspiciousness analysis module. The scene marking module marks the bullying possibility of the scene corresponding to the position of each camera, and obtains a scene table. The scene correlation analysis module takes the scene with high bullying possibility in the external input monitoring blind area as a bullying high-incidence scene, analyzes the contact degree of the bullying high-incidence scene and each scene in the scene table, and updates the scene table to obtain an optimized table according to the contact degree. The data acquisition module acquires the current scene image photographed by each camera as an image to be analyzed. The suspiciousness analysis module analyzes the image to be analyzed through a dynamic analysis strategy, a semantic analysis strategy, and a shape analysis strategy to obtain a bullying suspiciousness value. The decision output module compares the bullying suspiciousness value with a threshold value, and outputs voice warning of the camera in the scene and / or pushes an instruction to a guardian. The application has the advantage of analyzing whether bullying occurs in the current scene according to the bullying possibility of the scene.

Description

Technical Field

[0001] This invention relates to the technical field of academic affairs systems, and more specifically to an online self-identification system for school bullying. Background Technology

[0002] School bullying takes many forms, including but not limited to physical violence, verbal bullying, damage to personal property, and cyberbullying. Among these, physical violence is the most harmful to teenagers, and it can cause long-term psychological damage to victims. Bullying often occurs on or around the school, and is characterized by its hidden nature and difficulty in detection.

[0003] Current identification methods rely on traditional surveillance, which has limitations. Typically, cameras are used to monitor various areas of the school, requiring personnel to monitor multiple cameras simultaneously, making comprehensive and detailed monitoring difficult. With the continuous development of artificial intelligence (AI) technology, AI-based image recognition and audio monitoring technologies offer new solutions for monitoring school bullying. By deploying high-definition cameras and intelligent audio monitoring equipment in key areas such as teaching buildings, playgrounds, and canteens, real-time monitoring and early warning of abnormal situations on campus can be achieved. For example, the existing invention patent with publication number CN117237155A discloses an AI-based smart campus student behavior analysis system. This system primarily acquires audio from suspicious areas, performs semantic analysis to determine if the statements are bullying, and analyzes voiceprints to identify the number of people involved. However, this method is prone to misjudgments due to unclear speech or difficulty in identifying hidden locations. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide an online self-identification system for school bullying. This system can further analyze whether bullying is occurring in the current scenario based on the likelihood of bullying in the given situation, enabling timely and accurate judgment and warning.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] An online self-identification system for school bullying, including

[0007] The scene labeling module labels the probability of bullying in the corresponding scene based on the location of each camera, historical data, and / or expert experience, to obtain a scene table.

[0008] The scene association analysis module identifies scenes with a high probability of bullying in the blind spots of external input monitoring as high-incidence bullying scenes, analyzes the degree of connection between high-incidence bullying scenes and each scene in the scene table, and updates the scene table based on the degree of connection to obtain an optimized table.

[0009] The data acquisition module acquires the current scene images captured by each camera on campus as images to be analyzed, according to the scene sorting in the optimization table.

[0010] The suspicion analysis module analyzes the image to be analyzed using dynamic analysis strategy, semantic analysis strategy and morphological analysis strategy to obtain bullying suspicion value, and attaches the bullying suspicion value to the optimization table;

[0011] The decision output module outputs a voice warning from the camera and / or pushes instructions to the guardian based on a comparison between the suspected bullying value and the threshold.

[0012] Furthermore, the dynamic analysis strategy includes target detection and tracking steps and target behavior analysis steps.

[0013] The target detection and tracking step involves obtaining the target object through target detection based on the image to be analyzed, and then continuously tracking the target by extracting target features to obtain the behavioral trajectory of the target object.

[0014] The target behavior analysis step involves obtaining the target object's behavior type through deep learning or database matching based on the behavior trajectory, determining whether there is a bullying action in the behavior type, and outputting a first scoring instruction or an ignore instruction based on the judgment result.

[0015] Furthermore, the semantic analysis strategy includes semantic analysis steps and word / sentence matching steps.

[0016] The semantic analysis step involves automatically translating the audio data corresponding to the image to be analyzed into text segments based on a semantic model.

[0017] The word matching step involves filtering the text segments by bullying text matching or bullying text synonym matching to determine whether bullying information exists in the current scenario, and outputting a second scoring instruction or an ignore instruction based on the judgment result.

[0018] Furthermore, the morphological analysis strategy includes a face analysis step and a part analysis step.

[0019] The face analysis step involves capturing facial images of each target object from the image to be analyzed using face recognition, obtaining the current object's expression type through deep learning or database matching based on the facial images, and outputting a third scoring instruction based on the expression type.

[0020] The part analysis step involves obtaining the head information and / or clothing information of the target object through feature extraction based on the image to be analyzed, and outputting a fourth scoring instruction based on the head information and / or clothing information through deep learning or database matching.

[0021] Furthermore, the decision output module, when the bullying suspicion value is greater than or equal to the threshold, filters the bullied object and the bullying object based on the target object's behavior type and / or the current object's facial expression type and / or head information and / or clothing information, connects to an external academic affairs system, indexes the personal file information of the bullied object and the bullying object in the academic affairs system, and obtains the bullying object's violent tendency value and the bullied object's mental health value through weight analysis based on the personal file information.

[0022] Furthermore, the personal file information includes a learning record section, a family record section, and a past evaluation section. The learning record section reflects the student's academic performance over the years, the family record section reflects the student's family members and home visit records, and the past evaluation section reflects the past evaluation records of the student by teachers, parents, and classmates.

[0023] Furthermore, the suspicion analysis module filters the time each student spends entering and exiting high-incidence bullying scenes based on the images to be analyzed, determines the bullying suspicion value based on the time spent, and attaches the bullying suspicion value corresponding to this analysis to the optimization table.

[0024] Furthermore, in the word matching step, the text segment is filtered by matching the distress message or the synonym of the distress message to determine whether there is a distress message in the current scenario. If there is, the distress message is sent to the decision output module.

[0025] Furthermore, the scenario association analysis module obtains a 3D model of the campus from the external academic affairs system, marks the high-incidence bullying scenarios in the 3D model to obtain first marker points, and marks each scenario in the scenario table in the 3D model to obtain several second marker points. The module analyzes the lateral distance between the first marker point and each second marker point, where the lateral distance reflects the degree of connection.

[0026] Furthermore, the high-incidence scenarios for bullying include toilets or storage rooms.

[0027] The beneficial effects of this invention are as follows: By using scene association analysis, high-incidence bullying scenes in blind spots are associated with monitored scenes. This allows for analysis of videos from monitored scenes to identify suspicious individuals entering high-incidence bullying scenes. Specifically, by comprehensively applying multiple strategies such as dynamic analysis, semantic analysis, and morphological analysis, the system can more accurately identify potential bullying behaviors from images captured by cameras. Furthermore, the system can process monitoring data in real time and respond immediately upon detecting suspicious behavior, significantly shortening the time interval from the occurrence of an event to its identification. Additionally, for bullying incidents involving high vigilance, the system can analyze the morphology of the bullied student to prevent them from failing to report to their parents or teachers in a timely manner due to panic, thus preventing a decline in the student's mental health.

[0028] In addition, data acquisition can be achieved through multi-channel data collection, such as facial, behavioral, and trajectories obtained from camera video analysis; voice and semantic analysis; student health records, etc., providing comprehensive analysis and decision-making capabilities, including multi-dimensional information integration: scoring based on scenarios and correlations, optimizing table mechanisms; proactive identification and early warning for early prevention, personalized and precise intervention, and other advantages. Attached Figure Description

[0029] Figure 1 This is the overall flowchart of the present invention;

[0030] Figure 2 This is the optimized table display diagram in this invention;

[0031] Figure 3 This is a screenshot of the system in this invention. Detailed Implementation

[0032] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0033] Current identification methods rely on traditional surveillance, which has limitations. These methods typically use cameras to monitor various areas of the school, requiring personnel to monitor multiple cameras simultaneously, making comprehensive and detailed monitoring difficult. Therefore, this invention designs an online self-identification system for school bullying. Figure 1-3 As shown, including

[0034] The scene labeling module labels the corresponding scenes based on the location of each camera, according to historical data and / or expert experience, to determine the probability of bullying. This results in a scene table. By accurately locating high-risk areas, the system can allocate resources more effectively and improve monitoring efficiency. For example, if the scene includes corridors, the top floor, and the corner of the top floor staircase, historical data and expert experience indicate that the top floor is more likely to be a bullying location. Therefore, the top floor is listed as the first scene in the scene table, and the scenes are sorted according to their probability of bullying.

[0035] The scene association analysis module considers that places like restrooms and changing rooms are private spaces where surveillance cameras are not readily installed. Similarly, scenes with fewer people, such as storage rooms, are also unlikely to be monitored by surveillance cameras. Therefore, scenes with a high probability of bullying within the blind spots of external surveillance are designated as high-incidence bullying scenes (including restrooms, changing rooms, and storage rooms). The module analyzes the correlation between these high-incidence bullying scenes and the scenes in the scene table. Specifically, the correlation analysis involves: first, obtaining a 3D model of the campus from the external academic affairs system; marking the high-incidence bullying scenes in the 3D model to obtain first marker points; and marking each scene in the scene table in the 3D model to obtain several second marker points. The lateral distance between the first marker points and each second marker point is then analyzed. The horizontal distance refers to the horizontal distance. Since the first and second marker points may have a minimum vertical distance, and the vertical distance may reflect the distance between the upper and lower floors, this invention mainly uses the horizontal distance as a reference factor. The horizontal distance reflects the degree of connection. Assuming the toilet is located in the middle of the corridor, the toilet is the first marker point. There is a visual camera A in the corridor that can capture the boundary between the corridor and the toilet. Then the camera scene A1 in the corridor is the second marker point. The horizontal distance between the first and second marker points is the smallest, which reflects the highest degree of connection. There is also a visual camera B in another corner of the corridor. Then the camera scene B1 is the second marker point. Compared with the horizontal distance, it is larger, which means the degree of connection is smaller, corresponding to the fact that A1 is closer to the toilet.

[0036] The data acquisition module retrieves images of the current scene captured by each camera on campus according to the scene sorting in the optimization table. These images are then used as the images to be analyzed. Each image consists of several consecutive images and includes matching audio. The images are also attached to the corresponding columns of each scene in the optimization table. (In the data acquisition module, the scenes on campus are sorted according to the optimized scene table, and the real-time image data of the scenes with the highest ranking is obtained first. The system then captures the currently captured scene images from each camera according to this ranking, which serve as the image source for subsequent analysis. This module ensures key monitoring of high-risk areas for bullying and provides more accurate and timely data support for the analysis module.)

[0037] The suspiciousness analysis module analyzes the image to be analyzed using dynamic analysis, semantic analysis, and morphological analysis strategies to obtain bullying suspiciousness values. Dynamic analysis is used to capture abnormal movement trajectories, semantic analysis identifies aggressiveness in speech or facial expressions, and morphological analysis focuses on physical contact or abnormal crowd behavior. Finally, the calculated bullying suspiciousness values ​​are appended to an optimization table to facilitate subsequent risk assessment and early warning response. Specifically, the bullying suspiciousness values ​​obtained by each of the three analysis strategies can be appended to the optimization table separately, or the bullying suspiciousness values ​​obtained by the three analysis strategies can be weighted to obtain the total bullying suspiciousness value.

[0038] Specifically, the dynamic analysis strategy includes target detection and tracking steps and target behavior analysis steps.

[0039] The target detection and tracking step preprocesses the input image to be analyzed, including scaling, cropping, and denoising. Then, the target object is obtained through target detection. The target object is mainly students (people). All other parts in the image are ignored. The student's position is then initialized, that is, the student's position is determined in the first frame. In subsequent image frames, the student's position is continuously tracked using methods such as feature matching and optical flow to finally obtain the student's behavioral trajectory.

[0040] The target behavior analysis steps assume that the bullying behavior occurs under surveillance. Key features, such as speed, acceleration, and trajectory shape, are extracted based on the student's behavior. Deep learning models (such as convolutional neural networks and recurrent neural networks) or database matching methods are used to identify the behavioral trajectory. The trajectory is then matched with predefined behavior types to determine the target's behavior type. For example, if student A's behavior features include raising a fist or kicking, the matching indicates bullying behavior, and a first scoring instruction is output. If at least one student is simultaneously identified as exhibiting bullying behavior within one minute, the first score is higher. If no bullying behavior is identified in this scenario, an ignore instruction is output, excluding this scenario from the dynamic analysis strategy.

[0041] Specifically, the semantic analysis strategy includes semantic analysis steps and word / sentence matching steps.

[0042] The semantic analysis step involves automatically translating the audio data corresponding to the image to be analyzed into text segments using a semantic model. Specifically, audio processing involves matching the audio data captured by the camera with the image to be analyzed, using audio processing techniques to remove interference sounds from the audio, and then converting the audio data into text. The semantic model further processes the converted text using semantic models in natural language processing techniques (such as deep learning models, statistical language models, etc.) to ensure the accuracy and readability of the text. The semantic model can handle problems such as grammatical errors and spelling errors, and attempts to understand the semantic meaning in the text. Finally, the output result, the text (i.e., text segments) processed by the semantic model, will be used as input for the subsequent word and sentence matching step.

[0043] The word and phrase matching step filters text segments through bullying word matching or bullying word synonym matching to determine whether bullying information exists in the current scene. Based on the judgment result, a second scoring instruction or an ignore instruction is output. Specifically, bullying word matching: the speech corresponding to the image to be analyzed captured by the camera is automatically converted into text segments by audio translation methods, and a database containing known bullying words and phrases is established. The converted text segments are matched with this database to check whether there are any direct bullying expressions. Bullying word synonym matching: since there may be different words or phrases in the Chinese word semantic database to express similar meanings, it is also necessary to consider the synonyms or related expressions of bullying words, and expand the matching range through the thesaurus or more complex semantic analysis techniques. Judgment and output: if bullying information is found in the word and phrase matching step, the second scoring instruction is output; if no bullying information is found, the ignore instruction is output.

[0044] Bullying word matching rules include: whole word matching or keyword matching. Whole word matching is more accurate. For example, if the bullying word "xxx" appears in a word segment, it is considered that there is a possibility of bullying. This method is whole word matching. In addition, if the keyword in the bullying word "x1 x2x3" is "x2", according to the meaning of the bullying word, the keyword "x2" can only be located after the word "x1" or before the word "x3". That is, the order of the keywords is the judgment condition. When this order appears in the word segment, the condition is triggered to determine the possibility of bullying.

[0045] In addition, the same semantic analysis described above is used to analyze the preceding and following paragraphs identified as containing bullying words. If the bullying word "xxx" or other bullying words are identified, the preceding paragraphs "xxxxx1" and the following paragraphs "xxxxx2" are analyzed to verify whether the bullying words are consecutive bullying paragraphs or whether the bullying words constitute bullying behavior in the consecutive paragraphs. The purpose is to reduce the error rate in the analysis of bullying words.

[0046] The purpose of the semantic analysis strategy is primarily to analyze the presence of bullying words in the current scene through audio analysis, and to assign a second score based on the frequency and number of occurrences of bullying words, without considering who utters the bullying words. Compared to traditional single-speech analysis that uses voiceprints to accurately identify the speaker of bullying words, the semantic analysis in this invention is only to ensure whether there is any suspicion of bullying in the current scene. It can be combined with dynamic analysis strategies to conduct specific bullying probability analysis. Of course, this invention can combine voiceprint analysis to accurately locate the speaker of bullying words in order to verify the bullying targets obtained from dynamic analysis.

[0047] In addition, the word matching step can also filter by matching the text segment with the distress text or the synonyms of the distress text to determine whether there is a distress message in the current scene. If there is, the distress message is sent to the decision output module. The distress text may include "help", "stop hitting", "sorry", etc.

[0048] Specifically, the morphological analysis strategy includes face analysis steps and part analysis steps.

[0049] In the face analysis step, facial images of each target object are extracted from the image to be analyzed using face recognition. The expression type of the current object is then determined through deep learning or database matching based on these facial images. Finally, a third scoring instruction is output based on the expression type. Specifically, firstly, face recognition and extraction utilize existing face recognition technology to identify the facial regions of all target objects (people) in the image to be analyzed, and then these facial images are extracted as expression analysis images. Secondly, expression recognition uses deep learning models or database matching technology to perform expression recognition on the extracted facial images, analyzing subtle changes in facial muscles to determine the expression type of the current object, such as happiness, sadness, or anger. In typical cases of school bullying, the expression of the bullied is usually sadness, anger, or blankness, while the bully's expression is usually happiness or anger. Finally, a third scoring instruction is output. Based on the identified expression type, the system can output a corresponding scoring instruction. When a bullied expression appears, a score is assigned to that scene.

[0050] In the part analysis step, the system extracts features from the image to be analyzed, focusing on capturing the head features and / or clothing features of the target object. It identifies postures (such as abnormal head position, face obscuring) or clothing features (such as gang logos, specific ways of wearing uniforms, etc.) related to bullying behavior. When features matching the bullying context are detected, a fourth scoring instruction is output, adding points to the bullying suspicion value of that scene; if no relevant features are found, the analysis results are ignored. This step, by refining feature analysis, further enhances the system's ability to identify risks in a scene, contributing to comprehensive early warning.

[0051] Specifically, based on the image to be analyzed, the head information and / or clothing information of the target object are obtained through feature extraction. Based on this head information and / or clothing information, a fourth scoring instruction is output through deep learning or database matching. Specifically, firstly, feature extraction is performed in the part analysis step, where feature extraction technology is used to obtain the head information and / or clothing information of the target object from the image to be analyzed. Secondly, feature analysis is performed. Assuming that student A and student B first appear in scene 1, the system captures the image of their first appearance, recording their head and clothing information. When student A and student B reappear in scene 1 after entering a high-incidence bullying scene, their monitoring image is captured. At this time, the system analyzes the features of student A and student B at this time, such as hairstyle and clothing position, and compares them with the hairstyle and clothing position of student A and student B when they first entered scene 1. This allows for a preliminary determination of whether bullying has occurred and who the bully and the bullied are. If bullying is determined, a fourth scoring instruction is output, and a score is assigned.

[0052] The aforementioned morphological analysis strategy primarily addresses the monitoring of both bullies and victims in surveillance scenarios, whether the scene of bullying occurs or when the victim moves from a high-incidence bullying area into a surveillance environment. This strategy aims to monitor the faces and various parts of the body of both bullies and victims to prevent victims from failing to report bullying to their guardians due to threats or intimidation, thus preventing them from living under the shadow of school bullying for an extended period. Furthermore, the morphological analysis strategy is combined with dynamic analysis and semantic analysis strategies to conduct specific bullying probability analysis.

[0053] All of the above scores represent the presence of suspected bullying.

[0054] The decision output module compares the bullying suspicion value for each scenario with a preset threshold. When the suspicion value exceeds the threshold, a corresponding output mechanism is triggered: either playing an audio warning near the scene's camera to intervene in potential bullying behavior in real time, or sending an early warning notification and processing instructions to the guardian for rapid response. This module achieves automated early warning and tiered response, improving the efficiency of monitoring and intervening in bullying risks.

[0055] In addition, when the bullying suspicion value is greater than or equal to the threshold, the bullied object and the bullying object are filtered according to the target object's behavior type and / or the current object's facial expression type and / or head information and / or clothing information. The external academic affairs system is connected, and the personal file information of the bullied object and the bullying object is indexed in the academic affairs system. The personal file information includes a learning record column, a family record column, and a past evaluation column. The learning record column reflects the student's academic performance over the years, the family record column reflects the student's family members and home visit records, and the past evaluation column reflects the past evaluation records of the student by teachers, parents, and classmates. Based on the personal file information, the violent tendency value of the bullying object and the mental health value of the bullied object are obtained through weight analysis.

[0056] Typically, school bullying occurs among students with poor academic performance. Therefore, academic records can be used for initial screening. However, bullying can also occur among students with average or above-average academic performance. Therefore, family records and past evaluations can also be used for screening. For example, if there is parental violence or family conflict in the family, this factor can be used as a weight in the evaluation of violent tendency or mental health. In addition, the evaluations of teachers and classmates are also relatively important in the past evaluations. The above evaluations were all recorded in a non-public manner, so the data is relatively reliable.

[0057] In addition to dynamic analysis, semantic analysis, and morphological analysis strategies, the suspiciousness analysis module also needs to analyze the time spent entering high-incidence bullying scenes. If several students enter the toilet or storage room at the same time and the time exceeds the preset threshold, abnormal situations may occur, such as smoking or bullying. The module then filters the time spent by each student entering and exiting high-incidence bullying scenes based on the images to be analyzed, judges the bullying suspiciousness value based on the time spent, and attaches the bullying suspiciousness value corresponding to this analysis to the optimization table.

[0058] Assignment weight logic: such as Figure 2 As shown, assuming that student A and student B first appear in scene 1 (corridor A) at 11:24, and that potential bullying behavior can be indexed in the model based on morphological analysis in the surveillance images of scene 1 (corridor A), and assigned a score of F1;

[0059] In the surveillance footage captured at 11:24 in Scene 1 (Corridor A), the verbal interaction between student A and student B is recorded. Based on semantic analysis, if offensive, threatening, or insulting words can be indexed in the database, the corresponding score in the database is obtained as F2.

[0060] Record the facial images of students A and B in the surveillance footage taken in Scene 1 (Corridor A) at 11:24, and record the facial images of students A and B when they reappear in Scene 1 (Corridor A) after entering the high-incidence bullying scene at 11:39. Prioritize analyzing whether the students' faces have an expression type (such as anger, fear, pain, etc.). Assuming that student A's expression type at 11:24 is fear, the index score in the database is f1. If student A's expression type at 11:39 is pain, the index score in the database is f2. At this time, only student A needs to be scored, and the score is F3 = f1 + f2. At this time, student A can be identified as a suspected victim of bullying.

[0061] In the surveillance footage captured at 11:24 in Scene 1 (Corridor A), the head and clothing information of students A and B are recorded as normal. When students A and B reappear in Scene 1 (Corridor A) after entering the high-incidence bullying scene, the time is 11:39. During this period, the two students spent 15 minutes in the blind spot. Based on the abnormal occlusion or face covering of student A's head, a score of f3 is obtained by matching with the existing similarity in the matching database. Based on whether student A's clothing is damaged or neat, a score of f4 is obtained by matching with the existing similarity in the matching database. Therefore, student A can be identified as a suspected victim of bullying. The total score for morphological analysis is F4 = f3 + f4.

[0062] The weighting formula is configured as follows:

[0063]

[0064] Where S is the bullying suspicion value, λ is the normalization coefficient used to standardize the scores of the four different analyses to the same order of magnitude. e is the base of the natural logarithm, λ is the decay parameter that controls the shape and sensitivity of the exponential function, F1(t) is the morphological analysis scoring function, which gives a score based on behavioral patterns, F2(s) is the semantic analysis scoring function, which gives a score based on speech content, depending on the frequency and intensity of aggressive, threatening, or insulting words in the speech, F3(f,t) is the facial expression analysis scoring function, which gives a score based on expression type and temporal changes, where f represents the expression type (e.g., anger, fear, pain, etc.), and this function considers the changes and persistence of expressions, and F4(p,c) is the partial analysis scoring function, which gives a score based on abnormal head position and clothing condition, where p represents the degree of occlusion or covering of abnormal head position and c represents the degree of damage or neatness of clothing.

[0065] The range of S is (0,1], where values ​​close to 1 indicate highly suspicious bullying behavior, while values ​​close to 0 indicate lower suspicion of bullying.

[0066] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An online self-identification system for school bullying, characterized in that: include The scene labeling module labels the probability of bullying in the corresponding scene based on the location of each camera, historical data, and / or expert experience, to obtain a scene table. The scene association analysis module identifies scenes with a high probability of bullying in the blind spots of external input monitoring as high-incidence bullying scenes, analyzes the correlation between high-incidence bullying scenes and each scene in the scene table, and updates the scene table based on the correlation to obtain an optimized table. The scenario association analysis module obtains a 3D model of the campus from the external academic affairs system, marks the high-incidence bullying scenarios in the 3D model to obtain first marker points, and marks each scenario in the scenario table in the 3D model to obtain several second marker points. The module analyzes the lateral distance between the first marker point and each second marker point, and the lateral distance reflects the degree of connection. The data acquisition module acquires the current scene images captured by each camera on campus as images to be analyzed, according to the scene sorting in the optimization table. The suspiciousness analysis module analyzes the image to be analyzed using dynamic analysis strategy, semantic analysis strategy and morphological analysis strategy to obtain bullying suspiciousness value, and attaches the bullying suspiciousness value to the optimization table. The dynamic analysis strategy includes target detection and tracking steps and target behavior analysis steps. The semantic analysis strategy includes semantic analysis steps and word matching steps. The morphological analysis strategy includes face analysis steps and part analysis steps. The decision output module outputs a voice warning from the camera and / or pushes instructions to the guardian based on a comparison between the suspected bullying value and the threshold.

2. The online self-identification system for school bullying according to claim 1, characterized in that: The target detection and tracking step involves obtaining the target object through target detection based on the image to be analyzed, and then continuously tracking the target by extracting target features to obtain the behavioral trajectory of the target object. The target behavior analysis step involves obtaining the target object's behavior type through deep learning or database matching based on the behavior trajectory, determining whether there is a bullying action in the behavior type, and outputting a first scoring instruction or an ignore instruction based on the judgment result.

3. The online self-identification system for school bullying according to claim 2, characterized in that: The semantic analysis step involves automatically translating the audio data corresponding to the image to be analyzed into text segments based on a semantic model. The word matching step involves filtering the text segments by bullying text matching or bullying text synonym matching to determine whether bullying information exists in the current scenario, and outputting a second scoring instruction or an ignore instruction based on the judgment result.

4. The online self-identification system for school bullying according to claim 3, characterized in that: The face analysis step involves capturing facial images of each target object from the image to be analyzed using face recognition, obtaining the current object's expression type through deep learning or database matching based on the facial images, and outputting a third scoring instruction based on the expression type. The part analysis step involves obtaining the head information and / or clothing information of the target object through feature extraction based on the image to be analyzed, and outputting a fourth scoring instruction based on the head information and / or clothing information through deep learning or database matching.

5. The online self-identification system for school bullying according to claim 1, characterized in that: The suspicion analysis module filters the time taken for each student to enter and exit high-incidence bullying scenes based on the images to be analyzed, determines the bullying suspicion value based on the time taken, and attaches the bullying suspicion value corresponding to the analysis to the optimization table.

6. The online self-identification system for school bullying according to claim 3, characterized in that: The word matching step involves filtering the text segments by matching the distress message or the synonyms of the distress message to determine whether there is a distress message in the current scenario. If there is, the distress message is sent to the decision output module.

7. The online self-identification system for school bullying according to claim 1, characterized in that: The locations where bullying is most frequent include toilets or storage rooms.

Images