A campus safety risk prevention and control cloud service platform

Abstract

The application provides a kind of campus security risk prevention and control cloud service platform, it is related to cloud platform technical field, including: data processing and transmission module, for based on cloud platform architecture data storage, processing and secure transmission, and support the automatic reconnection and concurrent control of data transmission;Safety assurance module, connect data processing and transmission module, for the digital management of the civil air defense, physical protection and technical support information of campus;Security risk prevention and control module, connect data processing and transmission module and safety assurance module, for the information processing and linkage of campus safety management, safety education and safety emergency process;Joint defense and control module, connect security risk prevention and control module, for based on the processing result of security risk prevention and control module, realize the data linkage and emergency response command between multi-level education management department and school. By constructing the campus security risk prevention and control cloud service platform, the intelligentization and collaboration level of campus safety management is significantly improved.

Description

A campus safety risk prevention and control cloud service platform Technical Field

[0001] The present invention relates to the technical field of cloud platforms, and particularly to a campus safety risk prevention and control cloud service platform. Background Art

[0002] With the increasing prominence of campus safety issues, traditional campus safety management methods face many challenges, such as information silos, lagging responses, and decentralized resource management.

[0003] Currently, many schools rely on manual inspections and paper records to manage safety affairs, resulting in low efficiency in detecting potential safety hazards, untimely emergency responses, and a lack of systematic data support and cross-departmental collaboration mechanisms. In addition, the dissemination of safety education content is single, making it difficult to effectively cover all teachers and students. The formulation and rehearsal of emergency plans also mostly remain on paper, lacking digital support and real-time linkage functions.

[0004] At present, the rapid development of cloud computing, big data, and Internet of Things technologies has provided new solutions for campus safety management. By constructing a campus safety risk prevention and control system based on a cloud platform, centralized storage and processing of data, digital management of safety information, and efficient linkage among multiple departments can be achieved. However, existing technical solutions still have deficiencies in data transmission stability, concurrent control, and multi-module collaboration, making it difficult to meet the safety management requirements in complex campus scenarios.

[0005] Therefore, there is an urgent need for a cloud service platform that integrates data processing, security guarantee, risk prevention and control, and joint prevention and control to improve the intelligent level of campus safety management and enhance the effectiveness of safety education and emergency response. Summary of the Invention

[0006] The purpose of the present invention is to provide a campus safety risk prevention and control cloud service platform to solve the problems pointed out in the background art.

[0007] A campus safety risk prevention and control cloud service platform provided by an embodiment of the present invention includes:

[0008] A data processing and transmission module, which is used for data storage, processing, and secure transmission based on the cloud platform architecture, and supports automatic reconnection and concurrent control of data transmission;

[0009] A security guarantee module, connected to the data processing and transmission module, which is used for digital management of the information on campus human defense, physical defense, and technical guarantee;

[0010] A security risk prevention and control module, connected to the data processing and transmission module and the security guarantee module, which is used for information processing and linkage of campus safety management, safety education, and safety emergency processes;

[0011] The joint prevention and control module is connected to the security risk prevention and control module and is used to realize data linkage and emergency response command between multi-level education management departments and schools based on the processing results of the security risk prevention and control module.

[0012] Optionally, the data processing and transmission module is further configured to: use a three-tier software architecture and a relational database for data storage and processing, be deployed on a server operating system, and use the HTTP protocol and Socket communication mechanism to select data transmission paths and automatically reconnect.

[0013] Optionally, the security module includes:

[0014] The civil defense support unit is used to provide an interface for organizational responsibility management and to receive and store configuration information of full-time security personnel;

[0015] The physical security unit is used to establish a materials database and manage the configuration information of campus defense equipment and fire-fighting facilities;

[0016] The technical support unit is used to receive technical security information from associated perimeter alarm, video surveillance, and one-button alarm devices.

[0017] Optionally, the civil defense protection unit is further configured to: generate a personnel information reporting interface, receive and structurally store the reported security work organizational structure and basic information of full-time security guards.

[0018] Optionally, the physical protection unit is further configured to: generate a material information reporting interface, receive and classify the reported configuration status information of anti-collision facilities, fire-fighting facilities and rescue facilities.

[0019] Optionally, the technical support unit is further configured to: generate a technical security equipment information reporting interface, receive and associate the reported location and status information of video surveillance devices and perimeter alarm devices.

[0020] Optionally, the security risk prevention and control module includes:

[0021] The safety management unit is used to provide a streamlined process for reporting safety hazards, distributing tasks, and tracking them.

[0022] The safety education unit is used to build a safety knowledge resource base and push safety education information and test content to end users.

[0023] The safety emergency unit is used to provide electronic emergency plan development, drill reporting, and emergency response support functions.

[0024] Optionally, the safety emergency unit is further configured to: provide a plan editing interface, receive input emergency plan data, and store it in association with emergency response stages, so as to generate corresponding phased command and dispatch plans when an event is triggered.

[0025] Optionally, the campus security risk prevention and control cloud service platform also includes:

[0026] The intelligent risk control module for large-scale events is used to achieve precise dynamic blind spot mapping and early warning of bullying behavior through a cloud-driven drone swarm during large-scale events on campus. Specifically, it performs the following operations:

[0027] Based on the event's electronic map, real-time personnel distribution heat map, and historical safety management database, the event's future timeline is dynamically divided into multiple non-equal length adaptive time periods using a spatiotemporal segmentation optimization engine.

[0028] Whenever any adaptive time period is entered, the activity type label of the current adaptive time period is obtained through the activity arrangement information parsing unit. Combined with the real-time location data of the management personnel and the individual physiological response feature library stored in the historical safety management database, the real-time fatigue coefficient of the management personnel is quantified by the multi-source fusion fatigue dynamic modeling algorithm.

[0029] When the real-time fatigue coefficient exceeds the preset dynamic threshold, the field of view reduction algorithm for three-dimensional spatial occlusion correction is triggered. The field of view reduction algorithm identifies fixed blind spots caused by physical obstacles based on the three-dimensional point cloud model constructed from the site electronic map, and calculates the dynamic effective field of view radius based on the geographical location of the management personnel and the real-time fatigue coefficient. The dynamic monitoring blind spot geographic coordinate set is generated by Boolean operation between the dynamic effective field of view radius and the three-dimensional point cloud model.

[0030] The drone swarm collaborative control unit drives the drones equipped with multispectral sensors to adjust the hovering height and gimbal angle in real time according to the dynamic monitoring blind zone geographic coordinate set, and capture the multimodal behavior data stream of students in the blind zone;

[0031] The multimodal behavioral data stream is input into the bullying risk causal reasoning engine, and the possibility of school bullying is determined through a four-stage cascaded verification model. The four-stage cascaded verification model performs group dynamics anomaly detection to identify non-random clustered behaviors, initiates causal chain construction to detect power imbalance behavior chains, implements micro-expression action cross-verification to strengthen the confidence of causal chains, and performs spatiotemporal constraint verification to match spatiotemporal patterns of historical events. When the four-stage cascaded verification results form a closed-loop evidence chain in the spatiotemporal dimension, a high-risk bullying signal is output.

[0032] Optionally, the intelligent risk control module for large-scale events may also perform the following operations:

[0033] In response to the bullying high-risk signal, the high-risk behavior implementation area is generated through a spatiotemporal coupled trajectory extrapolation model;

[0034] A dual-modal tracking system is implemented through a drone swarm collaborative control unit. The primary drone performs wide-area trajectory tracking, while the backup drone focuses on key behavioral details. The system compares the actual movement trajectories of relevant students with the spatial relationship of high-risk behavior areas in real time. When a student is detected entering a high-risk behavior area and remaining there for more than a dynamic threshold, a tiered early warning generation mechanism is triggered. Other features and advantages of this invention will be set forth in the following description and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained by means of the structures particularly pointed out in the written description and drawings.

[0035] The invention has achieved the following beneficial effects:

[0036] This technical solution, by constructing a campus security risk prevention and control cloud service platform, achieves efficient data storage and secure transmission, significantly improving the intelligence and collaboration level of campus security management. The security assurance module digitally manages information on human, physical, and technological security measures, optimizing resource allocation and response efficiency. The security risk prevention and control module, through process-oriented hazard handling, the delivery of security education resources, and electronic emergency plan support, greatly improves the pertinence of security management and emergency response capabilities. The joint prevention and control module enables real-time data sharing and command linkage among multiple departments, effectively shortening emergency response time and comprehensively enhancing campus security prevention and control capabilities.

[0037] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0038] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0039] Figure 1 is a schematic diagram of a campus security risk prevention and control cloud service platform according to an embodiment of the present invention. Detailed Implementation

[0040] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0041] To address the issues of low informatization and poor collaboration in traditional campus security management, this technical solution utilizes cloud computing as its core, developing a cloud service platform that integrates data processing, security assurance, risk prevention and control, and joint prevention and control functions. During development, the data processing and transmission module was prioritized to ensure efficient data storage and stable transmission, and network instability was addressed through automatic reconnection and concurrency control technologies. The security assurance module employs a modular design, constructing digital management units for human, physical, and technical security measures to improve information management efficiency. The security risk prevention and control module focuses on the streamlined integration of hazard investigation, educational outreach, and emergency support processes to enhance management effectiveness. The joint prevention and control module, through data interfaces and multi-level linkage mechanisms, enables cross-departmental collaboration and rapid response, comprehensively meeting the needs of campus security management.

[0042] Figure 1 is a schematic diagram of a campus security risk prevention and control cloud service platform provided in an embodiment of this application. As shown in Figure 1, the platform includes:

[0043] Data processing and transmission module 1 is used for data storage, processing and secure transmission based on cloud platform architecture, and supports automatic reconnection and concurrency control of data transmission.

[0044] The data processing and transmission module is based on a cloud platform architecture, utilizing a distributed storage and computing framework to store, process, and securely transmit data, aiming to ensure the real-time performance, integrity, and security of campus security data. Distributed storage uses a key-value database to shard data across cloud nodes, ensuring high availability and fault tolerance. Data processing employs a streaming computing framework to clean and aggregate real-time data, generating structured data for subsequent modules. Secure transmission utilizes the TLS encryption protocol to protect data transmission between the cloud and the terminal, preventing data leakage or tampering. An automatic reconnection mechanism monitors network connection status through heartbeat detection, automatically initiating a reconnection request when a disconnection is detected, with a maximum of 5 retries and a 3-second retry interval. Concurrency control limits the number of concurrent requests using a token bucket algorithm, setting the maximum number of requests per second to 1000 to ensure system stability. The data shard size is calculated based on the storage node's disk capacity and data access frequency, with a default of 64MB. The heartbeat detection period is set to 5 seconds based on network latency statistics. The token bucket capacity is set to 1000 tokens based on server processing capacity.

[0045] Security module 2, connected to the data processing and transmission module, is used for the digital management of campus security, physical security, and technical support information. Security module 2 includes:

[0046] The civil defense support unit is used to provide an interface for organizational responsibility management and to receive and store configuration information of full-time security personnel.

[0047] The civil defense support unit receives and stores the configuration information of full-time security personnel through an organizational responsibility management interface, realizing the digital management of campus civil defense resources. The organizational responsibility management interface is designed based on a RESTful API, supporting the uploading of security personnel information (such as name, qualification number, and duty time) via POST methods. Data is stored in a relational database, including fields such as personnel ID, job assignment, and training records. Stored procedures verify data integrity (such as qualification number format validation) through triggers and record operation logs to trace modification history. Interface response time is monitored through an API gateway, with a target of less than 200ms. Data storage capacity is estimated based on school size, averaging 100 records per school, with each record approximately 1KB. Log retention is set at 180 days according to regulations. These parameters are collected via API gateway (such as Kong) to assess request latency, combined with database capacity monitoring tools (such as MySQL Workbench) to statistically analyze storage requirements and optimize interface performance periodically.

[0048] The physical security unit is used to establish a materials database and manage the configuration information of campus defense equipment and fire-fighting facilities.

[0049] The physical security unit manages the configuration information of campus defense equipment and fire-fighting facilities by establishing a materials database, ensuring the traceability and maintainability of physical security resources. The materials database uses a relational database to store equipment information (such as equipment ID, type, installation location, and maintenance cycle), and optimizes query performance through indexing. The management process includes equipment registration, status updates, and maintenance reminders. Maintenance reminders are sent via a scheduled task that scans the database to check if the equipment maintenance cycle (default 1 year) is approaching and sends notifications through a message queue. Technical parameters include: database query response time (collected through database monitoring tools, with a target of less than 100ms), message queue throughput (estimated based on the number of devices, set at 100 notifications per second), and maintenance cycle (set according to the equipment type manual, e.g., 1 year for fire hydrants). These parameters are monitored for query latency using database performance analysis tools (such as pgAdmin), and throughput is statistically analyzed using message queue monitoring tools (such as the RabbitMQ management interface). The maintenance cycle is determined in conjunction with the equipment vendor manual.

[0050] The technical support unit is used to receive technical security information from associated perimeter alarm, video surveillance, and one-button alarm devices.

[0051] The technical support unit receives and correlates security information from perimeter alarms, video surveillance, and one-button alarm devices. It integrates multi-source heterogeneous data through a data association algorithm to achieve unified management of the campus security system. The data association algorithm constructs a device relationship graph based on a graph database. Nodes represent devices (e.g., camera ID001), and edges represent relationships (e.g., the camera and alarm share area A). The status of associated devices is queried using a graph traversal algorithm (based on depth-first search). Received data is pushed from devices to the cloud via the MQTT protocol and stored in a time-series database to support high-frequency writes. Key technical parameters include: data write frequency (estimated based on the number of devices, set at 1000 records per second), graph query response time (monitored via the graph database, with a target of less than 50ms), and data retention period (set at 30 days according to regulations). These parameters are monitored for data throughput via an MQTT broker (e.g., Mosquitto), query latency is collected by a graph database (e.g., Neo4j), and the retention period is set in accordance with regulatory requirements.

[0052] The security risk prevention and control module 3 connects the data processing and transmission module and the security assurance module, and is used for information processing and linkage of campus safety management, safety education, and safety emergency procedures. The security risk prevention and control module includes:

[0053] The safety management unit is used to provide a streamlined process for reporting safety hazards, distributing tasks, and tracking them.

[0054] The safety management unit automates and ensures traceability of hazard management by providing a streamlined process for hazard reporting, task distribution, and tracking, based on a workflow engine. The workflow engine uses a Directed Acyclic Graph (DAG) to define the process, where nodes represent processing status (e.g., "report" or "distribute"), edges represent state transition conditions (e.g., "hazard confirmed"), and a state machine manages task flow. Hazard reporting data is submitted via a web form and stored in a relational database. Task distribution is based on hazard priority (high, medium, low) and assigned to responsible personnel. Tracking is achieved through scheduled task checks. Technical parameters include: workflow execution time (collected from engine logs, target within 500ms), task allocation accuracy (verified manually, target 95%), and database write frequency (estimated based on reporting frequency, set at 50 records per second). These parameters are analyzed using workflow engine (e.g., Activiti) logs to determine execution time, manual sampling to verify allocation accuracy, and database monitoring tools (e.g., MySQLWorkbench) to statistically analyze write frequency.

[0055] The safety education unit is used to build a safety knowledge resource base and push safety education information and test content to end users.

[0056] The safety education unit builds a safety knowledge resource library to push safety education information and test content to end users, improving the relevance of the educational content based on recommendation algorithms. The resource library uses a document database to store safety knowledge (such as fire escape guides) and optimizes retrieval speed through inverted indexes. The recommendation algorithm is built on collaborative filtering, collecting user interaction data (such as browsing history) to generate user profiles, calculating the matching degree (cosine similarity) between content and user interests, and pushing relevant educational content and test questions. Technical parameters include: retrieval response time (monitored through the database, with a target of less than 100ms), push frequency (set to once a day based on user activity), and recommendation accuracy (based on user feedback, with a target of 90%). These parameters are collected using database monitoring tools (such as MongoDB Atlas) to monitor retrieval latency, log analysis to assess user activity, and manual verification of the matching degree of recommended content.

[0057] The safety emergency unit is used to provide electronic emergency plan development, drill reporting, and emergency response support functions.

[0058] The safety emergency response unit provides electronic emergency plan development, drill reporting, and emergency response support functions, and automates emergency process management based on a rule engine. The rule engine uses the Rete algorithm to define triggering conditions (e.g., "fire alarm triggered") and response actions (e.g., "initiate evacuation"), and caches rules in an in-memory database to improve execution efficiency. Plans and drill data are stored in a relational database, and response support sends real-time instructions via message queues. Technical parameters involved in this process include: rule execution time (target within 100ms via engine logs), message queue latency (target within 50ms via monitoring tools), and data storage capacity (set to 50MB per school based on the number of plans). These parameters are collected by logging execution time using a rule engine (e.g., Drools), calculating latency using message queue monitoring (e.g., RabbitMQ), and estimating storage requirements using database tools (e.g., MySQL Workbench). The joint prevention and control module 4, connected to the safety risk prevention and control module, is used to achieve data linkage and emergency response command between multi-level education management departments and schools based on the processing results of the safety risk prevention and control module.

[0059] The joint prevention and control module 4, based on the processing results of the security risk prevention and control module, enables data linkage and emergency response command between multi-level education management departments and schools through a data sharing interface and an event-driven architecture. The data sharing interface, based on the gRPC protocol, supports bidirectional streaming of security event data (such as alarm logs), ensuring low latency and high throughput. The event-driven architecture subscribes to security events via message queues, triggering multi-level responses (such as the district education bureau directing school evacuations). Technical parameters include: interface response time (monitored via gRPC, with a target of less than 200ms), message queue throughput (set to 500 messages per second based on event frequency), and data synchronization frequency (set to once per minute based on network bandwidth). These parameters are collected using gRPC monitoring tools (such as Prometheus) to collect response time, message queue tools (such as Kafka) to calculate throughput, and network monitoring tools (such as Zabbix) to analyze bandwidth usage.

[0060] This technical solution, by constructing a campus security risk prevention and control cloud service platform, achieves efficient data storage and secure transmission, significantly improving the intelligence and collaboration level of campus security management. The security assurance module digitally manages information on human, physical, and technological security measures, optimizing resource allocation and response efficiency. The security risk prevention and control module, through process-oriented hazard handling, the delivery of security education resources, and electronic emergency plan support, greatly improves the pertinence of security management and emergency response capabilities. The joint prevention and control module enables real-time data sharing and command linkage among multiple departments, effectively shortening emergency response time and comprehensively enhancing campus security prevention and control capabilities.

[0061] In some embodiments, the data processing and transmission module 1 is further configured to: use a three-tier software architecture and a relational database for data storage and processing, be deployed on a server operating system, and use the HTTP protocol and Socket communication mechanism to select data transmission paths and automatically reconnect.

[0062] The data processing and transmission module 1 utilizes a three-tier software architecture (presentation layer, business logic layer, and data access layer) combined with a relational database to store, process, and securely transmit campus security data, ensuring system efficiency and reliability. The presentation layer receives user input via a web interface, the business logic layer handles data cleaning and aggregation, and the data access layer interacts with the relational database via SQL queries to store structured data (such as security event logs). The database uses a relational database (such as MySQL), storing data in table structures containing fields such as event ID, time, and type, with indexes optimizing query performance. Data transmission paths are selected based on the HTTP protocol and Socket communication mechanism. The HTTP protocol is used for short-connection RESTful API requests (such as data queries), while Socket communication is used for long-connection real-time data pushes (such as alarm events). The automatic reconnection mechanism monitors connection status via heartbeat detection at 3-second intervals. If a disconnection is detected, a reconnection is initiated using an exponential backoff algorithm (initial retry interval of 3 seconds, maximum of 5 retries).

[0063] In some embodiments, the civil defense protection unit 21 is further configured to: generate a personnel information reporting interface, receive and structurally store the reported security work organizational structure and basic information of full-time security guards.

[0064] The Civil Defense Support Unit 21 receives and structures the campus security organizational structure and full-time security guard information through a personnel information entry interface, achieving digitalization and traceability of personnel management. The entry interface is based on a RESTful API, using the POST method to receive JSON-formatted input data, including fields such as security guard name, qualification number, and duty time. Data storage employs a relational database (e.g., [database name missing]).

[0065] PostgreSQL stores information through table structures. Table fields include personnel ID, job title, and training records. Primary key and foreign key constraints ensure data consistency. Stored procedures use triggers to verify the integrity of input data (e.g., qualification numbers must conform to a 16-digit number format) and generate operation logs to record modification history.

[0066] In some embodiments, the physical protection unit 22 is further configured to: generate a material information reporting interface, receive and classify the reported configuration status information of anti-collision facilities, fire-fighting facilities and rescue facilities.

[0067] The physical security unit 22 receives and categorizes the configuration status information of anti-collision facilities, fire protection facilities, and rescue facilities through a material information reporting interface, ensuring the manageability and maintainability of campus physical security resources. The reporting interface is based on a RESTful API, supporting the POST method to receive JSON format data, including fields such as device ID, type (e.g., fire hydrant), installation location, and status (normal / faulty). Data storage uses a relational database (e.g., MySQL), categorizing and storing device information through table structures. Table fields include device ID, type, and maintenance time, with indexes optimizing query performance. Stored procedures use triggers to check data format (e.g., device ID must be an 8-digit alphanumeric combination), and the categorization logic automatically assigns data to the corresponding sub-table (e.g., the fire protection facilities table) based on the device type field.

[0068] In some embodiments, the technical support unit 23 is further configured to: generate a technical security equipment information reporting interface, receive and associate the reported location and status information of video surveillance devices and perimeter alarm devices.

[0069] Technical support unit 23 generates an interface for filling in security equipment information, receiving and associating the location and status information of video surveillance devices and perimeter alarm devices to achieve unified management of security equipment. The filling interface is based on a RESTful API, supporting the POST method to receive JSON format data, including fields such as device ID, type (e.g., camera), location, and status (online / offline). Data storage uses a relational database (e.g., PostgreSQL), storing device information through table structures. Table fields include device ID, type, and location, and association tables store the relationships between devices (e.g., the regional association between cameras and alarms). Stored procedures use triggers to validate the data format (e.g., device IDs must be 10 digits), and the association logic uses SQL foreign key constraints to bind device relationships.

[0070] In some embodiments, the safety emergency unit 33 is further configured to: provide a plan editing interface, receive input emergency plan data, and store it in association with emergency response stages, so as to generate a corresponding phased command and dispatch plan when an event is triggered.

[0071] The safety emergency unit 33 provides a plan editing interface, receives input emergency plan data, and stores it in association with emergency response stages. A rule engine generates phased command and dispatch plans when events are triggered. The editing interface is based on a web form, supporting user input of plan data (such as fire alarm plans and evacuation procedures). Data is submitted in JSON format via a RESTful API, containing fields such as plan ID, stage (warning / response), and action (notification / evacuation). Data storage uses a relational database (such as MySQL), storing plan and stage information in a table structure. Table fields include plan ID, stage, and action, and related tables record the mapping between stages and actions. The rule engine is built based on the Rete algorithm, defining triggering conditions (such as "fire alarm signal") and response actions (such as "initiate evacuation"), and caching rules using an in-memory database (such as Redis) to improve execution efficiency.

[0072] During large-scale campus events, such as sports meets and cultural performances, students gather in large numbers, and the diverse activities and high mobility significantly increase the difficulty of management. Traditional campus security management mainly relies on manual patrols and fixed camera surveillance, which has blind spots and delayed responses. Especially during group gatherings or free time, students may take advantage of the crowd density or dispersion to engage in retaliatory or bullying behaviors. Meanwhile, administrators, burdened by heavy workloads and accumulated fatigue, find it difficult to comprehensively cover all areas, leading to regulatory loopholes. For example, latent conflicts often occur between students during activity breaks or rest periods. Historical data shows that such incidents are more frequent during periods of high population density or greater freedom of movement, and due to a lack of real-time data support, administrators struggle to accurately identify risk points. Furthermore, existing technologies cannot dynamically adapt to changes in activity intensity and lack sufficient analysis of the combined impact of administrators' physiological state and environmental constraints, making it difficult to achieve accurate risk warnings and rapid responses.

[0073] Therefore, in some embodiments, the campus security risk prevention and control cloud service platform also includes:

[0074] The intelligent risk control module for large-scale events is used to achieve precise dynamic blind spot mapping and early warning of bullying behavior through a cloud-driven drone swarm during large-scale events on campus. Specifically, it performs the following operations:

[0075] A. Based on the event's electronic map, real-time personnel distribution heat map, and historical safety management database, the event's future timeline is dynamically divided into multiple non-equal length adaptive time periods using a spatiotemporal segmentation optimization engine. The division of these adaptive time periods is adjusted adaptively based on the event's intensity variation characteristics, which include the personnel density during the collective gathering phase, the degree of freedom of movement during the dispersed activity phase, and the frequency of events during the rest transition period.

[0076] The spatiotemporal segmentation optimization engine dynamically divides the activity timeline into multiple non-equal-length adaptive time periods by analyzing data from the activity's electronic map, real-time personnel distribution heatmap, and historical safety management database to adapt to changes in activity intensity. These changes include personnel density during group gatherings, freedom of movement during dispersed activities, and event frequency during rest transition periods. Personnel density is calculated through heatmap analysis, specifically by dividing the activity area into 1m x 1m grids and counting the number of people in each grid to obtain a density value (unit: people / square meter). Freedom of movement is calculated using personnel trajectory data, extracting personnel movement speed and direction vectors based on Simultaneous Localization and Mapping (SLAM) technology, and calculating the displacement entropy per unit time (unit: meters / second). Event frequency is extracted from the historical safety management database, counting the number of alarms or abnormal events per unit time (e.g., per minute) to obtain a frequency value (unit: times / minute). The spatiotemporal segmentation optimization engine employs a dynamic programming-based segmentation algorithm to construct a time-axis segmentation model. Taking activity intensity variation characteristics as input, the optimization objective is to minimize the variance of intensity variation within a time period, generating non-uniform adaptive time periods. The segmentation algorithm uses a weighting of 0.4 for personnel density, 0.3 for movement freedom, and 0.3 for event frequency to comprehensively calculate intensity variation characteristic values. Segmentation points are determined using a greedy strategy that minimizes variance. Technical parameters involved in this process include: grid resolution (determined by the electronic map resolution, set to 1 meter), displacement entropy calculation cycle (based on the SLAM data sampling rate, set to 1 second), event frequency statistics cycle (based on the database record granularity, set to 1 minute), and minimum segmentation time period length (based on the activity duration, set to 5 minutes). These parameters are obtained through a Geographic Information System (GIS) tool on the activity's electronic map, the SLAM system (such as ORB-SLAM3) collects trajectory data to calculate displacement entropy, and a database query tool (such as MySQLWorkbench) counts event frequencies.

[0077] Here is an implementation example:

[0078] During a school sports meet at a middle school, a spatiotemporal segmentation optimization engine dynamically divided the event timeline based on an electronic map (1-meter resolution), a real-time personnel distribution heatmap, and a historical safety management database. The sports meet lasted 4 hours. The electronic map covered the playground and spectator area (5000 square meters). The heatmap was generated in real-time using a camera mounted on a drone, with a grid resolution of 1 meter. Statistics showed that the density of people in the center of the playground during the opening ceremony was 5 people / square meter, and during dispersed activities, it was 1 person / square meter. The SLAM system captured personnel trajectories through cameras, calculating the displacement entropy as 0.2 meters / second during the opening ceremony and 1.5 meters / second during the competition phase. The historical database showed an event frequency of 0.5 times / minute during the opening ceremony and 0.1 times / minute during breaks. The engine operates using a dynamic programming algorithm, taking density, displacement entropy, and event frequency as inputs, with weights of 0.4, 0.3, and 0.3 respectively. It calculates intensity change characteristic values ​​to determine segmentation points, dividing the 4-hour event into an opening ceremony (30 minutes), a competition phase (120 minutes), a break phase (60 minutes), and a closing ceremony (30 minutes). The segmentation results are stored on a cloud platform for subsequent modules to access. The minimum segmentation time is 5 minutes to ensure appropriate granularity. Parameter acquisition involves parsing electronic maps using GIS tools (such as ArcGIS), sampling trajectory data every second using a SLAM system (ORB-SLAM3), and using MySQL Workbench to statistically analyze historical event frequencies, verifying that the goal of minimizing segmentation variance has been achieved and that the segmentation results are consistent with the actual event schedule.

[0079] B. Whenever any adaptive time period is entered, the activity type label of the current adaptive time period is obtained through the activity arrangement information parsing unit. Combined with the real-time location data of the management personnel and the individual physiological response feature library stored in the historical safety management database, the real-time fatigue coefficient of the management personnel is quantified using a multi-source fusion fatigue dynamic modeling algorithm. The multi-source fusion fatigue dynamic modeling algorithm generates the real-time fatigue coefficient by weighted fusion of physiological state perception factor, task load evolution factor and individual adaptive compensation factor. The physiological state perception factor is based on the facial micro-expression features of the management personnel extracted from the high-definition video stream of the UAV. The task load evolution factor is dynamically calculated based on the working time attenuation weight and event processing frequency. The individual adaptive compensation factor is personalized according to the parameters of the management personnel's historical fatigue recovery curve.

[0080] The activity schedule information parsing unit extracts activity type labels (such as "opening ceremony" and "competition") from the activity schedule, combining this with real-time location data of management personnel and an individual physiological response feature library from the historical safety management database. A multi-source fusion fatigue dynamic modeling algorithm is then used to generate real-time fatigue coefficients. Activity type labels are extracted from the schedule text using a Natural Language Processing (NLP) model based on the BERT architecture, pre-trained and fine-tuned for the campus activity schedule corpus, outputting labels such as "high density" and "low degree of freedom." Management personnel location data is acquired using an RTK-GPS module mounted on a drone, achieving centimeter-level accuracy and a sampling frequency of 1Hz. The individual physiological response feature library stores historical fatigue data (such as heart rate and cadence) and is managed through a relational database. The fatigue dynamic modeling algorithm generates fatigue coefficients through a weighted fusion of physiological state perception factors, task load evolution factors, and individual adaptive compensation factors. The physiological state perception factor is based on high-definition video streams from drones (1080p resolution, 30fps), extracting fatigue features (such as eyelid ptosis, range 0-1) through a facial micro-expression recognition model (based on ResNet-50, trained on the FER2013 dataset). The task load evolution factor is calculated based on working duration (statistics from positioning data, unit: minutes) and event processing frequency (queried from the database, unit: times / hour), with a decay weight of 0.8. The individual adaptive compensation factor is based on historical fatigue recovery curves (fitted from fatigue data within 30 days in the database, using an exponential decay model). Technical parameters involved include: label extraction accuracy (validated by an NLP model, target 95%), positioning accuracy (calibrated via RTK-GPS, set to 0.05 meters), video frame rate (configured via camera, set to 30fps), and fatigue coefficient range (normalized via algorithm, set to 0-1). Parameter acquisition involves validating label accuracy using the BERT model test set, calibrating positioning accuracy using the RTK-GPS device, setting the frame rate using the camera firmware, and recording the fatigue coefficient using algorithm logs.

[0081] Continuing with the above implementation example, in the sports meet scenario, the activity arrangement information parsing unit extracts the activity type label "high density" for the opening ceremony time slot (30 minutes) from the schedule. The parsing unit uses a BERT model (pre-trained on a general Chinese corpus, fine-tuned on 1000 campus activity schedule data points), calls the cloud platform API, inputs the schedule text, and outputs a label with an accuracy of 96%. Management personnel location data is provided by a drone RTK-GPS module (DJI M300 RTK), with a positioning accuracy of 0.05 meters and a sampling frequency of 1Hz, recording the real-time location of 10 security guards in the playground area. A historical security management database (MySQL) stores 30 days of security guard fatigue data (e.g., average heart rate of 80 beats / minute). A fatigue dynamic modeling algorithm extracts facial micro-expressions of security guards from a 1080p video stream (30fps) from the drone. A ResNet-50 model (trained on FER2013, fine-tuned on 5000 campus security guard images) calculates eyelid ptosis as 0.7, generating a physiological state perception factor. The task load evolution factor is calculated as 0.6 based on the record of security guards performing 30 minutes of duty and handling 2 incidents, with a decay weight of 0.8. The individual adaptive compensation factor is based on the fatigue recovery curves of security guards in the database (exponential decay, half-life of 2 hours), with a value of 0.5. The algorithm fuses the three factors with weights of 0.4, 0.3, and 0.3, outputting a fatigue coefficient of 0.65, which is stored in the cloud platform for subsequent steps. The parameters are verified for label accuracy using the BERT test set, the positioning accuracy is confirmed by RTK-GPS device logs, the frame rate is set in the camera firmware, and the fatigue coefficient calculation process is recorded in the algorithm log.

[0082] C. When the real-time fatigue coefficient exceeds a preset dynamic threshold, a three-dimensional spatial occlusion correction field-of-view reduction algorithm is triggered. The field-of-view reduction algorithm identifies fixed blind spots caused by physical obstacles based on a three-dimensional point cloud model constructed from the site's electronic map, and calculates the dynamic effective field-of-view radius based on the manager's geographical location and real-time fatigue coefficient. A dynamic monitoring blind spot geographic coordinate set is generated through Boolean operations between the dynamic effective field-of-view radius and the three-dimensional point cloud model. The dynamic monitoring blind spot geographic coordinate set accurately reflects the combined influence of the manager's real-time physiological state and physical environmental constraints.

[0083] When the real-time fatigue coefficient exceeds a dynamic threshold (dynamically adjusted based on the activity type label, such as 0.6 for "high density"), a 3D spatial occlusion correction field-of-view reduction algorithm is triggered to identify fixed and dynamic blind spots. Fixed blind spots are calculated using a 3D point cloud model constructed from a site electronic map. This point cloud model is generated by LiDAR sensor scanning (0.1-meter resolution), and occlusion areas caused by obstacles (such as walls and trees) are identified using a stereoscopic geometry algorithm. The dynamic effective field-of-view radius is calculated based on the manager's geographical location (provided by RTK-GPS, accuracy 0.05 meters) and fatigue coefficient using a linear decay model. The formula is: Radius = Base Field-of-View Radius × (1 - Fatigue Coefficient). The base field-of-view radius is estimated based on the average visibility of the activity area (set to 20 meters using GIS tools). The field-of-view reduction algorithm generates a set of geographic coordinates for the dynamic blind spots through Boolean operations (intersection of the point cloud model and the field-of-view radius). The coordinates are expressed in latitude and longitude with an accuracy of 0.01 meters. Relevant technical parameters include: point cloud resolution (calibrated using a LiDAR device, set to 0.1 meters), field of view radius (estimated using GIS, set to 5-20 meters), coordinate accuracy (calibrated using GPS, set to 0.01 meters), and algorithm execution time (collected from logs, target within 100ms). Parameter acquisition involves calibrating the point cloud resolution using a LiDAR device (such as Velodyne Puck), estimating visibility using a GIS tool (ArcGIS), verifying coordinate accuracy using RTK-GPS logs, and recording execution time using algorithm logs (Python log module).

[0084] For example, during the opening ceremony of the sports meet, a security guard's fatigue coefficient was 0.65, exceeding the threshold of 0.6 for the "high density" label, triggering a field-of-view reduction algorithm. The electronic map of the venue was scanned using LiDAR (Velodyne Puck, 0.1-meter resolution) to generate a point cloud model. A stereoscopic geometry algorithm identified a fixed blind spot (50 square meters) formed by trees behind the stands. The security guard's geographical location was determined by RTK-GPS (0.05-meter accuracy) to be at the center of the field (latitude and longitude: 121.12345, 31.67890). Using a basic field-of-view radius of 20 meters (field visibility estimated via ArcGIS) and a fatigue coefficient of 0.65, the algorithm calculated a dynamic effective field-of-view radius of 7 meters (20 × (1-0.65)). Boolean operations were used to intersect the point cloud model with a sphere of 7 meters radius, generating a dynamic blind spot coordinate set containing 10 coordinate points, including a corner of the field (121.12350, 31.67895), with an accuracy of 0.01 meters. The algorithm execution time was recorded as 80ms via Python logging, meeting the 100ms target. The coordinate set was stored on a cloud platform for use by the drone swarm. Parameters were calibrated for point cloud resolution using a LiDAR device, visibility was analyzed using ArcGIS, coordinate accuracy was verified using RTK-GPS logs, and the logs confirmed the algorithm's execution efficiency.

[0085] D. Driving a drone equipped with a multispectral sensor through a drone swarm collaborative control unit, the drone adjusts its hovering height and gimbal angle in real time based on the geographic coordinate set of the dynamic monitoring blind zone, capturing multimodal behavioral data streams of students within the blind zone; wherein, the multimodal behavioral data streams include high-precision spatial trajectory sequences, behavioral causal graph feature matrices, and micro-expression action coupling feature vectors. The high-precision spatial trajectory sequences generate centimeter-level positioning data through visual synchronous positioning and mapping technology. The behavioral causal graph feature matrix extracts the causal dependencies of student interactions based on a temporal graph convolutional network. The micro-expression action coupling feature vectors synchronously analyze the temporal correlation between facial emotion features and body movements.

[0086] The drone swarm collaborative control unit coordinates multiple drones equipped with multispectral sensors (including visible light and infrared cameras) to adjust hovering height (range 5-20 meters) and gimbal angle (range 0-360°) based on the geographic coordinates of the dynamic monitoring blind spot, capturing multimodal behavioral data streams of students within the blind spot. The multimodal data stream includes high-precision spatial trajectory sequences, behavioral causal graph feature matrices, and micro-expression action coupling feature vectors. The high-precision spatial trajectory sequences are generated using visual SLAM technology (based on ORB-SLAM3) and combined with multispectral video streams (1080p resolution, 30fps frame rate) to calculate centimeter-level positioning data of students (accuracy 0.05 meters). The behavioral causal graph feature matrix extracts the causal dependencies of student interactions using a temporal graph convolutional network (T-GCN), with inputs of trajectory sequences and video frames, and outputs an adjacency matrix (dimension: number of students × number of students). Micro-expression motion coupling feature vectors are jointly analyzed using a ResNet-50 model (trained on FER2013) and an OpenPose model to detect facial emotions (smile intensity, range 0-1) and body movements (joint angles, range 0-180°). Technical parameters include: hovering height range (set via drone firmware, 5-20 meters), gimbal angle accuracy (calibrated via servo motors, 0.1°), trajectory accuracy (calibrated via SLAM, 0.05 meters), and data stream acquisition frequency (set via sensors, 30Hz). Parameter acquisition involves setting altitude and angle via drone firmware (e.g., DJI SDK), verifying trajectory accuracy using ORB-SLAM3 logs, and recording the acquisition frequency using camera and model logs.

[0087] For example, in a blind spot in the corner of the sports field (121.12350, 31.67895), a drone swarm collaborative control unit drives three DJIM300 RTK drones (equipped with 1080p visible light and infrared cameras) to adjust their hovering height to 10 meters and gimbal angle to 45° (servo motor accuracy 0.1°) based on the coordinate set. A SLAM system (ORB-SLAM3) processes the video stream, generating a spatial trajectory sequence of five students with a positioning accuracy of 0.05 meters, recording coordinate changes (e.g., 121.12351, 31.67896). A T-GCN model (trained on 1000 interactive videos of the campus) analyzes the trajectories and videos, generating a 5×5 adjacency matrix to represent the proximity behavior between students. A ResNet-50 model (fine-tuned for the campus videos) and OpenPose jointly extract the students' smile intensity (0.3) and arm angle (60°) to generate coupled feature vectors. Data is collected at a frequency of 30Hz, transmitted to the cloud platform via the MQTT protocol, and stored in a time-series database (InfluxDB). Parameters such as height and angle are set using the DJI SDK, trajectory accuracy is verified using ORB-SLAM3 logs, the acquisition frequency is confirmed using camera logs, and feature extraction efficiency is recorded using model logs.

[0088] E. Input the multimodal behavior data stream into the bullying risk causal reasoning engine, and determine the possibility of school bullying through a four-stage cascaded verification model; wherein, the four-stage cascaded verification model performs group dynamics anomaly detection to identify non-random clustering behavior, initiates causal chain construction to detect power imbalance behavior chain, implements micro-expression action cross-verification to strengthen the confidence of causal chain, and performs spatiotemporal constraint verification to match historical event spatiotemporal patterns. When the four-stage cascaded verification results form a closed-loop evidence chain in the spatiotemporal dimension, a high-risk bullying signal is output.

[0089] The bullying risk causal reasoning engine analyzes multimodal behavioral data streams using a four-stage cascaded validation model to determine the likelihood of school bullying. The first stage, group dynamics anomaly detection, uses the DBSCAN algorithm to identify non-random clustering behaviors. Input trajectory sequences are set with a radius parameter of 0.5 meters and a minimum number of points of 3, outputting anomalous clusters. The second stage, causal chain construction, uses a Bayesian network (trained on 1000 school event data points) to detect power imbalance behavioral chains. Input a behavioral causal graph feature matrix, outputting causal probability (range 0-1). The third stage, micro-expression and action cross-validation, uses a logistic regression model (trained on FER2013 and OpenPose data) to fuse smile intensity and body angle, strengthening the causal chain confidence (range 0-1). The fourth stage, spatiotemporal constraint validation, matches current data with historical event spatiotemporal patterns (extracted from the database, including location and time), outputting a matching degree (range 0-1). When the overall confidence score (weighted average, weight 0.25×4) of the four stages exceeds 0.8, a high-risk bullying signal is output. Technical parameters include: DBSCAN radius (0.5 meters through trajectory density analysis), causal probability threshold (0.7 through model validation), confidence threshold (0.8 through experiments), and validation execution time (within 100ms through logs). The parameters are determined by the DBSCAN experiment (radius), validation probability using a Bayesian network test set, confidence level determined by logistic regression experiments, and execution time recorded in the logs.

[0090] In the blind spot of the sports meet, multimodal data streams from 5 students were input into a bullying risk causal inference engine. DBSCAN (radius 0.5 meters, minimum number of points 3) analyzed the trajectory sequence, identifying a cluster of 3 students and marking it as an anomaly. A Bayesian network (trained on 1000 campus events) analyzed the behavioral causal graph, detecting a power imbalance dominated by one person, outputting a causal probability of 0.75. A logistic regression model fused a smile intensity of 0.3 and an arm angle of 60°, outputting a confidence score of 0.8. Spatiotemporal constraint verification matched historical bullying events in a corner of the playground (occurrence rate 0.1 times / hour) in the database, outputting a matching degree of 0.85. The overall confidence score (0.25×0.75+0.25×0.8+0.25×0.85+0.25×1) was 0.85, exceeding the threshold of 0.8, outputting a high-risk signal. The execution time, logged by Python, was 90ms. The parameters were determined by the radius through DBSCAN experiment (100 sets of trajectories), the probability was verified by Bayesian network test set, the confidence level was determined by logistic regression experiment (500 sets of data), and the historical event matching degree was queried by MySQL Workbench.

[0091] F. In response to the bullying high-risk signal, a high-risk behavior implementation area is generated through a spatiotemporal coupled trajectory extrapolation model; wherein, the spatiotemporal coupled trajectory extrapolation model integrates kinematic differential equations and a dynamic risk map of the campus geographic information system. The kinematic differential equations predict the student's future trajectory, and the dynamic risk map of the campus geographic information system marks and fixes high-risk areas and integrates historical event heat map data for spatial correction. When the intersection area of ​​the predicted trajectory and the dynamic risk map meets the preset spatial conditions, the high-risk behavior implementation area is determined.

[0092] A spatiotemporally coupled trajectory extrapolation model integrates kinematic differential equations and a dynamic risk map from a campus geographic information system (GIS) to predict areas where high-risk behaviors may occur. The kinematic differential equations, based on trajectory sequences (velocity, acceleration, accuracy 0.05 m / s), predict students' trajectories over the next 5 minutes using a second-order Runge-Kutta method, outputting a predicted coordinate set (accuracy 0.01 m). The dynamic risk map, generated by GIS, marks fixed high-risk areas (e.g., a corner of the playground, area 50 square meters) and incorporates historical event heatmaps (extracted from a database, event density unit: times / square meter). The model identifies high-risk areas by performing intersection operations between the trajectory and the map (based on spatial Boolean operations), determining areas where the intersection area exceeds 10 square meters. Technical parameters include: prediction time window (based on activity duration, 5 minutes), trajectory accuracy (via SLAM, 0.05 m / s), map resolution (via GIS, 0.1 meters), and intersection area threshold (experimentally, 10 square meters). The parameters were verified for trajectory accuracy using SLAM logs, map resolution was analyzed using ArcGIS, event density was queried from the database, and area thresholds were determined through Boolean operations.

[0093] For example, in a sports meet scenario, after a high-risk bullying signal is triggered, a spatiotemporal coupled trajectory extrapolation model analyzes the trajectories of five students (speed 0.5 m / s, SLAM accuracy 0.05 m / s). The Runge-Kutta method predicts the trajectories within 5 minutes, outputting a coordinate set (e.g., 121.12352, 31.67897). A GIS dynamic risk map (resolution 0.1 m) marks a corner of the playground as a high-risk area (area 50 square meters), and integrates 10 historical bullying events from the database (density 0.2 events / square meter). Boolean operations calculate the intersection of the trajectory and the map; an area of ​​15 square meters exceeds the threshold of 10 square meters, thus identifying the high-risk area as a corner of the playground (121.12350-121.12355, 31.67895-31.67900). The results are stored on a cloud platform for subsequent tracking. The parameters were verified for trajectory accuracy using SLAM logs, the map resolution was analyzed using ArcGIS, the event density was queried using MySQL Workbench, and the area threshold was determined through experiments (50 sets of data).

[0094] G. A dual-modal tracking system is implemented through a drone swarm collaborative control unit. The primary drone performs wide-area trajectory tracking, while the backup drone focuses on key behavioral details. The system compares the actual movement trajectory of relevant students with the spatial relationship between the high-risk behavior implementation area in real time. When a relevant student is detected to have entered the high-risk behavior implementation area and stayed for more than the dynamic judgment threshold, a graded early warning generation mechanism is triggered. The graded early warning generation mechanism generates an early warning level based on the weighted result of the bullying risk probability value and the stay time. When the weighted result is lower than the first threshold, an early warning message is generated and pushed to the mobile terminal of the on-site management personnel. When the weighted result is between the first threshold and the second threshold, the campus sound and light alarm system is activated. When the weighted result exceeds the second threshold, an emergency response instruction is sent to multi-level education management departments through the joint prevention and control unit. The emergency response instruction includes the student's real-time location coordinates and behavioral evidence information.

[0095] The drone swarm collaborative control unit implements dual-modal tracking using a primary drone (wide-area tracking, covering 100m x 100m) and a backup drone (detailed focusing, covering 10m x 10m), comparing student trajectories (SLAM-generated, accuracy 0.05m) with high-risk areas (latitude and longitude range, accuracy 0.01m) in real time. The primary drone uses a wide-angle lens (120° field of view), and the backup drone uses a zoom lens (30° field of view). When a student enters a high-risk area and stays there for more than a dynamic judgment threshold (5-30 seconds depending on the activity type), a tiered early warning generation mechanism is triggered. The early warning mechanism is based on the bullying risk probability (from sub-step E, range 0-1) and the stay time (unit: seconds), weighted (weights 0.6 and 0.4). If the result is below 0.5, it is pushed to the management terminal (via MQTT, latency within 50ms); if it is between 0.5 and 0.8, an audible and visual alarm is activated (latency within 100ms); if it exceeds 0.8, an emergency command (including coordinates and video evidence) is sent through the joint prevention and control unit. Technical parameters include: tracking accuracy (0.05 meters via SLAM), judgment threshold (5-30 seconds via experiment), push latency (within 50ms via MQTT), and alarm latency (within 100ms via device). Accuracy was verified using SLAM logs, thresholds were determined experimentally (100 sets of data), and latency was recorded using MQTT and device logs.

[0096] For example, in a high-risk area (121.12350-121.12355, 31.67895-31.67900) in a corner of the sports field, the main drone (DJI M300 RTK, 120° wide-angle lens) tracks the trajectories of 5 students (accuracy 0.05 meters), while a backup drone (30° zoom lens) focuses on details. The SLAM system detects one student entering the area and staying for 15 seconds, exceeding the "high-density" activity threshold of 10 seconds. The risk probability is 0.81 (from sub-step E), the dwell time is 15 seconds, and the weighted result (0.6×0.81+0.4×15 / 30) is 0.686, between 0.5 and 0.8, triggering an audible and visual alarm (80ms delay). The alarm is pushed to the security terminal via MQTT (40ms delay), containing coordinates (121.12352, 31.67897) and video footage. The accuracy of the parameters was verified using SLAM logs, the threshold was determined by experiments (100 sets of dwell data), and the latency was recorded by MQTT logs (Mosquitto) and device logs.

[0097] This technical solution achieves precise prevention and control of safety risks for large-scale campus events through drone swarms, multi-source data fusion, and intelligent algorithms. The system can dynamically divide event time periods, quantify the fatigue coefficient of management personnel in real time, accurately identify blind spots in supervision, and accurately warn of bullying risks through multimodal behavior analysis and four-stage cascaded verification. It generates a tiered response mechanism, effectively reducing the incidence of student retaliation and bullying incidents, and improving safety management efficiency and response speed.

[0098] In particular, step A utilizes a spatiotemporal segmentation optimization engine to dynamically divide the activity timeline into non-uniform adaptive time periods based on the activity's electronic map, real-time personnel distribution heatmap, and historical safety management database. This division fully considers the activity's intensity variation characteristics (such as personnel density, freedom of movement, and event frequency), and optimizes the time period segmentation through dynamic programming algorithms to ensure that the division results highly match the actual activity rhythm. It can accurately capture the safety risk characteristics of different stages of the activity, providing a precise time frame for subsequent blind spot mapping and behavioral warnings, significantly improving the targeting and efficiency of safety management. For example, in a sports event scenario, the risk characteristics of the opening ceremony, competition, and rest phases are effectively distinguished, avoiding the blindness of traditional fixed-time divisions and reducing regulatory loopholes caused by improper time allocation.

[0099] Step B utilizes an activity scheduling information analysis unit and a multi-source fusion fatigue dynamic modeling algorithm to quantify the fatigue coefficient of management personnel in real time, comprehensively considering multiple dimensions such as physiological state, task load, and individual adaptability. By employing technologies such as high-definition video streams from drones and RTK-GPS, the real-time status of management personnel is accurately perceived, overcoming the deficiency of traditional safety management methods that neglect manager fatigue. This quantification method not only improves the scientific rigor of management personnel status monitoring but also provides crucial input for subsequent calculations of dynamic monitoring blind spots.

[0100] Step C uses a 3D spatial occlusion correction-based field-of-view reduction algorithm, combined with a 3D point cloud model of the site and real-time fatigue coefficients, to generate a dynamic set of geographic coordinates for blind spots in surveillance. The algorithm comprehensively considers the dual constraints of physical obstacles and the physiological state of management personnel, accurately identifying both fixed and dynamic surveillance blind spots and overcoming the limitations of traditional fixed cameras and manual patrols. For example, in a sports event scenario, the algorithm identifies blind spots in the corners of the playground caused by tree obstruction and adjusts the field-of-view radius based on the fatigue state of security personnel, generating a precise set of blind spot coordinates. This method significantly improves the accuracy and real-time performance of blind spot identification, providing a reliable spatial basis for subsequent precise monitoring by drones.

[0101] Step D utilizes a drone swarm collaborative control unit to capture multimodal behavioral data streams of students within blind spots using multispectral sensors. This includes spatial trajectories, behavioral causal graphs, and micro-expression features. Through advanced technologies such as high-precision SLAM and temporal graph convolutional networks, comprehensive and dynamic monitoring of student behavior is achieved, overcoming the shortcomings of traditional monitoring methods in densely populated or dispersed environments. For example, in a blind spot in a corner of a sports field during a sports meet, the system successfully captured student trajectories and interactive behaviors, generating a high-precision data stream, providing a rich data foundation for the accurate identification of bullying risks. This multimodal data acquisition method significantly improves the comprehensiveness and reliability of behavioral analysis.

[0102] More importantly, step E analyzes multimodal behavioral data streams using a four-stage cascaded verification model of the bullying risk causal reasoning engine to accurately determine the likelihood of school bullying. Through four stages—group dynamics anomaly detection, causal chain construction, micro-expression action cross-validation, and spatiotemporal constraint verification—a closed-loop evidence chain is formed, significantly improving the accuracy and reliability of bullying risk identification. This multi-stage verification mechanism effectively reduces the false positive rate and provides a scientific basis for timely intervention.

[0103] Step F uses a spatiotemporally coupled trajectory extrapolation model, combined with kinematic differential equations and GIS dynamic risk maps, to predict areas where high-risk behaviors occur. By integrating student trajectory predictions and historical event heatmaps, the system can accurately locate potential bullying areas, overcoming the problem of insufficient future risk prediction in traditional safety management. For example, in a sports meet scenario, the model predicts a corner of the playground as a high-risk area, and verifies the intersection area through Boolean operations to generate precise coordinates of the danger zone. This predictive capability significantly improves the foresight of safety management, providing crucial support for subsequent accurate tracking and early warning.

[0104] In some embodiments, the spatiotemporal segmentation optimization engine synchronously runs an online learning optimizer. After completing the monitoring operation for each adaptive time period, it updates the weight parameters of the fatigue dynamic modeling algorithm using a deep reinforcement learning algorithm based on blind zone coverage verification data and successful intervention cases in the historical safety management database. When the actual coverage efficiency of the blind zone is lower than the preset standard, it triggers online optimization of the model parameters.

[0105] The spatiotemporal segmentation optimization engine, through synchronously running an online learning optimizer, dynamically updates the weight parameters of the fatigue dynamic modeling algorithm after the monitoring operation is completed in each adaptive time period, thereby improving blind zone coverage efficiency and safety management effectiveness. Blind zone coverage verification data refers to the actual blind zone monitoring coverage rate collected by the UAV swarm, defined as the ratio (unit: percentage, range 0-100%) of the area of ​​the dynamic monitoring blind zone geographic coordinate set covered by the UAV multispectral sensor to the expected blind zone area. This data is extracted from the UAV sensor data stream through the cloud platform log analysis module, specifically by comparing the intersection area of ​​the actual coverage coordinate set (latitude and longitude accuracy 0.01 meters) and the dynamic monitoring blind zone coordinate set generated in sub-step C. The historical safety management database stores successful intervention cases, including intervention time, location, event type, and result (e.g., bullying event prevention rate, unit: percentage), extracted using database query tools (e.g., MySQL Workbench) indexed by event ID. The deep reinforcement learning algorithm employs a Q-learning-based framework, constructing a state space (blind zone coverage, activity type label, and personnel density), an action space (adjusting the weights of fatigue dynamic modeling algorithms, such as physiological state perception factor weights, task load evolution factor weights, and individual adaptive compensation factor weights, ranging from 0 to 1), and a reward function (blind zone coverage improvement, ranging from -1 to 1). The algorithm runs on high-performance computing nodes (such as NVIDIA A100 GPUs) on a cloud platform, iteratively updating weights to maximize the reward function value. The preset standard for triggering online optimization is a blind zone coverage efficiency threshold (determined experimentally, set to 85%). When the actual coverage efficiency is lower than this threshold, the algorithm updates the weight parameters using blind zone coverage verification data for the current time period and historical successful intervention cases as input. Technical parameters include: blind zone coverage (calculated from drone logs, with 1% accuracy), weight adjustment step size (experimentally set to 0.01), reward function convergence threshold (experimentally set to 0.1), and optimization execution time (based on logs, targeting less than 200ms). These parameters are obtained by extracting coverage through the drone log module (based on ROS2), determining the step size and threshold through Q-learning algorithm experiments, and recording execution time through cloud platform logs (Prometheus).

[0106] Here is an implementation example:

[0107] In a scenario of a middle school sports meet, after monitoring the opening ceremony (30 minutes), the spatiotemporal segmentation optimization engine ran an online learning optimizer to optimize the fatigue dynamic modeling algorithm. A drone swarm (3 DJI M300RTKs) collected blind spot coverage verification data, covering a dynamic monitoring blind spot in a corner of the playground (area 50 square meters, latitude and longitude 121.12350-121.12355, 31.67895-31.67900, accuracy 0.01 meters). The actual coverage area was calculated to be 40 square meters using the cloud platform log analysis module (based on ROS2), representing a coverage rate of 80% (40 / 50×100%), lower than the preset threshold of 85%. The historical safety management database (MySQL) queried 10 successful intervention cases (prevention rate 90%) in the playground corner over the past 30 days, extracting the intervention time (average 15 minutes) and event type (3 bullying incidents). The Q-learning algorithm (based on the PyTorch framework, running on an NVIDIA A100 GPU, with a state space including 80% coverage, "high-density" activity type, and a density of 5 people / square meter) uses a weight adjustment step size of 0.01. After 10 iterations, the weights of the physiological state perception factor are increased from 0.4 to 0.42, the task load evolution factor weight is decreased from 0.3 to 0.28, the individual adaptive compensation factor weight remains at 0.3, and the reward function value increases from 0.05 to 0.12, exceeding the convergence threshold of 0.1. The optimization process, recorded in cloud platform logs (Prometheus), has an execution time of 180ms, meeting the 200ms target. The new weights are stored in the cloud platform for use in the next time period (competition phase), and verification shows that the coverage increased to 88% in the next time period. Parameters are extracted using ROS2 logs to determine coverage, PyTorch experiments (100 sets of data) to determine the step size and threshold, and Prometheus logs to verify the execution time.

[0108] The bullying risk causal reasoning engine performs multi-source data conflict resolution during the judgment process. When the spatial trajectory shows abnormal interaction but micro-expression analysis does not detect the expected emotion, the activity arrangement information parsing unit is called to verify the activity type label in order to correct the risk probability value.

[0109] The bullying risk causal inference engine addresses inconsistencies between spatial trajectories and micro-expression analysis results through a multi-source data conflict resolution mechanism, improving the accuracy of bullying risk assessment. When the four-stage cascaded verification model in sub-step E detects abnormal interactions (identifying non-random clusters using the DBSCAN algorithm, radius 0.5 meters, minimum number of points 3) but micro-expression analysis (extracting smile intensity using the ResNet-50 model, range 0-1) fails to detect the expected emotion (e.g., negative emotion, smile intensity below 0.3), conflict resolution is triggered. Conflict resolution extracts activity type labels (e.g., "high density," "low degree of freedom") for the current adaptive time period by calling the activity scheduling information parsing unit (based on the BERT model, pre-trained on Chinese corpus, fine-tuned on campus activity schedule corpus). Label accuracy is validated using a test set (target 95%). Activity type labels are used to correct risk probability values. The correction method is as follows: The causal chain confidence weights are adjusted based on the labels (0.6 for high-density labels and 0.4 for low-degree-of-freedom labels). Trajectory and micro-expression features are re-integrated using a logistic regression model (trained on 5000 sets of campus behavior data) to output the corrected risk probability (range 0-1). Technical parameters include: abnormal interaction detection radius (0.5 meters through DBSCAN experiments), smile threshold (0.3 through ResNet-50 experiments), label accuracy (95% through BERT testing), and probability correction execution time (within 100ms through logs). Parameter acquisition involves determining the radius through DBSCAN experiments (100 sets of trajectory data), validating the smile threshold on the ResNet-50 test set (FER2013), validating the label accuracy on the BERT test set (1000 schedule data points), and recording the execution time in cloud platform logs (Prometheus). The corrected probability values ​​are stored in the cloud platform for subsequent alerts.

[0110] For example, in the competition phase (120 minutes) of the aforementioned sports meet scenario, the bullying risk causal inference engine processes the multimodal data stream from the blind spot in the corner of the playground (121.12350-121.12355, 31.67895-31.67900). The DBSCAN algorithm (radius 0.5 meters, minimum number of points 3) analyzes the trajectory sequences of 5 students, detecting 3 students not randomly clustered and marking them as abnormal interactions. The ResNet-50 model (trained on FER2013, fine-tuned on 5000 campus images) analyzes the video stream, extracting a smile intensity of 0.5, which is higher than the expected negative emotion threshold of 0.3, triggering conflict resolution. The engine calls the activity schedule information parsing unit, and the BERT model (fine-tuned on 1000 campus schedule data) parses the competition phase schedule, outputting the activity type label "high degree of freedom" with an accuracy of 96%. A logistic regression model (trained on 5000 sets of behavioral data) with high-degree-of-freedom label weights of 0.4 was used to re-integrate trajectory (weight 0.6) and smile score (weight 0.4), correcting the initial causal probability from 0.75 to 0.65. The correction process, recorded in cloud platform logs (Prometheus), took 90ms, meeting the 100ms target. The corrected probability was stored in the cloud platform for use in the four-stage validation of sub-step E. Validation showed a 5% reduction in misclassification rate after correction. Parameters were determined using a DBSCAN experiment (100 sets of trajectories), the smile score threshold was validated using a ResNet-50 test set, label accuracy was confirmed using a BERT test set, and execution time was verified using Prometheus logs.

[0111] The hierarchical early warning generation mechanism ensures real-time transmission of early warning instructions through a priority scheduling protocol for data processing and transmission units. When network fluctuations cause transmission delays, it automatically activates the emergency communication channel and executes a timestamp verification and retransmission mechanism.

[0112] The tiered early warning generation mechanism ensures that early warning commands (including high-risk bullying signals, student location coordinates, and video evidence) are transmitted in real time to management personnel terminals, audible and visual alarm systems, or multi-level education management departments through a priority scheduling protocol for data processing and transmission units. The priority scheduling protocol is based on the MQTT protocol and defines three priority levels: high priority (bullying risk probability > 0.8, transmission delay target within 50ms), medium priority (probability 0.5-0.8, delay within 100ms), and low priority (probability < 0.5, delay within 200ms). The protocol allocates bandwidth through a message queue on the cloud platform (based on Mosquitto), prioritizing the transmission of high-priority commands. Network fluctuation detection monitors transmission delay in real time (via Ping test, unit: milliseconds). When the delay exceeds the target value (e.g., high priority exceeds 50ms), it automatically switches to the emergency communication channel (based on 4G / 5G cellular network, bandwidth 10Mbps). The timestamp verification and retransmission mechanism verifies the data packet timestamp (1ms accuracy, synchronized via NTP) after a switchover. If packet loss is detected (checked via sequence number), the missing data packet is retransmitted from the cloud platform cache (Redis database). Technical parameters include: transmission latency (50-200ms via Ping test), timestamp accuracy (1ms via NTP), retransmission success rate (95% through experiment), and switchover time (within 100ms via logs). Parameters are obtained by testing latency using the Ping tool, calibrating timestamps using the NTP server, calculating the retransmission success rate using Redis logs, and recording the switchover time in the cloud platform logs (Prometheus).

[0113] For example, in a high-risk area in the corner of the sports field (121.12350-121.12355, 31.67895-31.67900), sub-step G generates a high-risk bullying signal (probability 0.81), triggering a tiered early warning generation mechanism. The data processing and transmission unit transmits the early warning command (containing coordinates 121.12352, 31.67897 and a video clip) with high priority via the MQTT protocol (Mosquitto server), with a target latency of 50ms. Ping testing detects network fluctuations, increasing the latency to 60ms, exceeding the high-priority threshold, and automatically switches to the 5G emergency channel (10Mbps bandwidth). The switching process is recorded in the cloud platform logs (Prometheus) at 80ms, meeting the 100ms target. The timestamp verification mechanism (NTP synchronization, 1ms accuracy) detects a lost data packet (missing sequence number), retransmits it from the Redis cache, and achieves a 96% success rate in retransmission. The warning command was successfully pushed to the security terminal (45ms delay), simultaneously activating the audible and visual alarms (80ms delay) and sending an emergency command (including coordinates and video) to the education management department through the joint prevention and control unit. Parameters were tested for latency using Ping, the NTP server was used to calibrate the timestamp, Redis logs were used to verify the retransmission success rate, and Prometheus logs were used to confirm the switchover time.

[0114] The aforementioned technical solutions significantly enhance the intelligence and real-time performance of safety management for large-scale campus events through online learning optimizers, multi-source data conflict resolution, and priority scheduling protocols. By dynamically updating the weights of the fatigue dynamic modeling algorithm using deep reinforcement learning, management efficiency is optimized based on blind spot coverage verification data and historical intervention cases. For example, in sports events, this improves blind spot coverage and reduces monitoring loopholes caused by manager fatigue. The multi-source data conflict resolution mechanism resolves conflicts between trajectory and micro-expression analysis, corrects risk probabilities, reduces misjudgment rates, and ensures the accuracy of bullying risk assessment. Priority scheduling and emergency communication channels ensure real-time transmission of warning commands, effectively addressing network fluctuations and improving response speed and reliability. These technologies collectively achieve accurate risk warnings and rapid responses, significantly reducing the incidence of campus bullying incidents and enhancing the safety management effectiveness of large-scale events.

[0115] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A campus security risk prevention and control cloud service platform, characterized in that, include: The data processing and transmission module is used for data storage, processing, and secure transmission based on the cloud platform architecture, and supports automatic reconnection and concurrency control of data transmission. The system includes: a security module connected to the data processing and transmission module for digital management of campus security, including human, physical, and technical support information; a security risk prevention and control module connected to both the data processing and transmission module for information processing and linkage of campus security management, security education, and emergency response procedures; a joint prevention and control module connected to the security risk prevention and control module for data linkage and emergency response command between multi-level education management departments and schools based on the processing results of the security risk prevention and control module; and a large-scale event intelligent risk control module for using cloud-driven drone clusters during large-scale events on campus. The system achieves precise mapping of dynamic blind spots and early warning of bullying behavior by performing the following operations: Based on the activity's electronic map, real-time personnel distribution heatmap, and historical safety management database, the system dynamically divides the future timeline of the activity into multiple unequal-length adaptive time periods using a spatiotemporal segmentation optimization engine. Whenever any adaptive time period is entered, the activity type label for the current adaptive time period is obtained through the activity arrangement information parsing unit. Combined with real-time location data of management personnel and the individual physiological response feature library stored in the historical safety management database, a multi-source fusion fatigue dynamic modeling algorithm is used to quantify the real-time fatigue coefficient of management personnel. This multi-source fusion fatigue dynamic modeling algorithm generates the real-time fatigue coefficient through a weighted fusion of physiological state perception factors, task load evolution factors, and individual adaptive compensation factors. The physiological state perception factor is based on facial micro-expression features extracted from high-definition video streams from drones. The task load evolution factor is dynamically calculated based on the attenuation weight of working hours and the frequency of event processing. The individual adaptive compensation factor is personalized based on the historical fatigue recovery curve parameters of management personnel. When the real-time fatigue coefficient exceeds a preset dynamic threshold, a three-dimensional spatial occlusion correction field-of-view reduction algorithm is triggered. This field-of-view reduction algorithm identifies fixed blind spots caused by physical obstacles based on a three-dimensional point cloud model constructed from the site's electronic map. The dynamic effective field of view radius is calculated based on the geographical location of the management personnel and the real-time fatigue coefficient. A dynamic monitoring blind spot geographic coordinate set is generated through Boolean operations between the dynamic effective field of view radius and the 3D point cloud model. The dynamic effective field of view radius is calculated using a linear decay model based on the geographical location and fatigue coefficient of the management personnel. The formula is: radius = basic field of view radius × (1 - fatigue coefficient). The basic field of view radius is estimated based on the average visibility of the activity area. The drones equipped with multispectral sensors are driven by the drone swarm collaborative control unit to adjust the hovering height and gimbal angle in real time according to the dynamic monitoring blind spot geographic coordinate set, thereby capturing the multimodal behavioral data stream of students in the blind spot.The multimodal behavioral data stream includes a high-precision spatial trajectory sequence, a behavioral causal graph feature matrix, and micro-expression action coupling feature vectors. The high-precision spatial trajectory sequence generates centimeter-level positioning data using visual synchronous localization and mapping technology. The behavioral causal graph feature matrix extracts the causal dependencies between student interactions based on a temporal graph convolutional network. The micro-expression action coupling feature vectors synchronously analyze the temporal correlation between facial emotion features and body movements. This multimodal behavioral data stream is input into a bullying risk causal inference engine, and a four-stage cascaded verification model is used to determine the likelihood of school bullying. The first stage, group dynamics anomaly detection, identifies non-random clustered behaviors using the DBSCAN algorithm, inputting the trajectory sequence and outputting abnormal clusters. The second stage, causal chain construction, detects power imbalance behavioral chains using a Bayesian network, inputting the behavioral causal graph feature matrix and outputting causal probability. The third stage, micro-expression action cross-validation, integrates smile intensity and body angle using a logistic regression model to strengthen the confidence of the causal chain. The fourth stage, spatiotemporal constraint verification, matches current data with historical event spatiotemporal patterns and outputs a matching degree. When the four-stage cascaded verification results form a closed-loop evidence chain in the spatiotemporal dimension, a high-risk bullying signal is output.

2. The campus security risk prevention and control cloud service platform as described in claim 1, characterized in that, The data processing and transmission module is further configured to use a three-tier software architecture and a relational database for data storage and processing, be deployed on a server operating system, and use the HTTP protocol and Socket communication mechanism to select data transmission paths and automatically reconnect.

3. The campus security risk prevention and control cloud service platform as described in claim 1, characterized in that, The security module includes: a human security unit, which provides an interface for organizational responsibility management and receives and stores configuration information of full-time security personnel; a physical security unit, which establishes a material database and manages the configuration information of campus defense equipment and fire-fighting facilities; and a technical security unit, which receives and associates technical security information from perimeter alarms, video surveillance, and one-button alarm devices.

4. The campus security risk prevention and control cloud service platform as described in claim 3, characterized in that, The civil defense support unit is further configured to: generate a personnel information reporting interface, receive and structurally store the reported security work organizational structure and basic information of full-time security guards.

5. The campus security risk prevention and control cloud service platform as described in claim 3, characterized in that, The physical protection unit is further configured to: generate a material information reporting interface, receive and classify the reported configuration status information of anti-collision facilities, fire-fighting facilities and rescue facilities.

6. The campus security risk prevention and control cloud service platform as described in claim 3, characterized in that, The technical support unit is further configured to: generate a technical security equipment information reporting interface, receive and associate the reported location and status information of video surveillance devices and perimeter alarm devices.

7. The campus security risk prevention and control cloud service platform as described in claim 1, characterized in that, The security risk prevention and control module includes: a security management unit, which provides a streamlined process for reporting security risks, distributing tasks, and tracking them; a security education unit, which builds a security knowledge resource base and pushes security education information and test content to end users; and a security emergency unit, which provides electronic emergency plan development, drill reporting, and emergency response support functions.

8. The campus security risk prevention and control cloud service platform as described in claim 7, characterized in that, The safety emergency unit is further configured to: provide a plan editing interface, receive input emergency plan data, and store it in association with emergency response stages, so as to generate corresponding phased command and dispatch plans when an event is triggered.

9. The campus security risk prevention and control cloud service platform as described in claim 8, characterized in that, The intelligent risk control module for large-scale events also performs the following operations: in response to the bullying high-risk signal, it generates the high-risk behavior implementation area through a spatiotemporal coupled trajectory extrapolation model; it implements dual-modal tracking through a drone cluster collaborative control unit, with the main drone performing wide-area trajectory tracking and the backup drone focusing on key behavioral details, and compares the spatial relationship between the actual movement trajectory of the relevant students and the high-risk behavior implementation area in real time. When it detects that the relevant students have entered the high-risk behavior implementation area and their stay time exceeds the dynamic judgment threshold, it triggers a graded early warning generation mechanism.

Images