Multi-angle campus panoramic intelligent monitoring system

Abstract

The invention discloses a multi-angle campus panoramic intelligent monitoring system, and relates to the technical field of security monitoring systems, and the system comprises a sensing module which is used for obtaining sensing data based on sensor networks deployed in all regions of a campus, and dynamically distributing sensing tasks according to region functions; wherein the sensor network comprises a camera, a millimeter wave radar, a pressure sensor, an environmental noise sensor and a Bluetooth device; the processing module is used for processing the sensing data in real time through the reconfigurable hardware nodes and generating compliance features and abnormal identifiers; the risk deduction module is used for constructing a semantic grid, associating regional function attributes, time authority and a historical risk thermodynamic diagram and executing global risk causal chain deduction based on the compliance features and the abnormal identifiers output by the processing module; and the response execution module is used for triggering a dynamic response according to the deduction result, and the dynamic response comprises hierarchical privacy protection and cross-campus coordination action.

Description

Technical Field

[0001] The present application relates to the technical field of security monitoring systems, and in particular to a multi-angle campus panoramic intelligent monitoring system. Background Art

[0002] As campus security incidents become more complex and diverse, building an intelligent security system covering all scenarios has become a core need of educational institutions.

[0003] Existing campus security systems typically deploy multimodal sensing devices such as cameras, sensors, and access control systems, attempting to achieve risk early warning by integrating video surveillance, environmental monitoring, and personnel trajectory data. However, current technical architectures face significant bottlenecks in the deep integration and scenario adaptation of multimodal data. Traditional multimodal fusion solutions often involve simple data splicing, such as time-synchronizing video footage with sensor alarm signals, but lack mechanisms for deep correlation and collaborative analysis.

[0004] To address the above-mentioned problems, no effective solutions have been proposed so far. Summary of the Invention

[0005] The embodiment of the present application provides a multi-angle campus panoramic intelligent monitoring system to solve the above technical problems.

[0006] This application provides a multi-angle campus panoramic intelligent monitoring system, including: A perception module, which is used to acquire perception data based on a sensor network deployed in various areas of the campus and dynamically assign perception tasks based on regional functions; wherein the sensor network includes cameras, millimeter-wave radars, pressure sensors, environmental noise sensors, and Bluetooth devices; A processing module is configured to process the perception data in real time through a reconfigurable hardware node to generate compliance features and anomaly identifiers, wherein the processing method includes: encoding campus rules into neural network differentiable rule synapses using a temporal rule embedded causal inference network; fusing millimeter wave gait entropy features and Bluetooth connection entropy features using an asymmetric modal contrast learning network in a privacy area; A risk deduction module is used to construct a semantic grid based on the compliance features and anomaly identification output by the processing module, associate regional functional attributes, time permissions and historical risk heat maps, and perform global risk causal chain deduction; The response execution module is used to trigger dynamic responses based on the deduction results, including hierarchical privacy protection and cross-campus collaborative actions.

[0007] Furthermore, in the temporal rule embedded causal reasoning network, the generation of neural network differentiable rule synapses includes: The two-person authorization rule in the laboratory is decomposed into spatial conditions and temporal conditions, which are encoded into the weight matrix and bias vector of the neural network respectively. When the millimeter-wave radar detects a single person in the laboratory area and the timestamp is late at night, the convolution kernel operation of the regular synapse is activated; The differentiable knowledge graph is used to verify whether the signal characteristics of the electrical equipment match the hazardous chemical operation mode. If they match, the abnormal identification is generated and the access control locking instruction is triggered.

[0008] Furthermore, after triggering the access lock instruction, the method further includes: The camera in the laboratory area is used to capture facial images of people, and non-sensitive contour features are extracted through edge nodes; Dynamically time-warping the non-sensitive profile features with the authorized personnel model in the compliance feature library; If the matching similarity is lower than the safety threshold, a two-person verification request and a heat map of the person's movement trajectory will be sent to the security terminal; Synchronously cut off the power supply of high-risk devices and record the operation logs to the risk heat map of the semantic grid.

[0009] Furthermore, the operation of the asymmetric modal contrast learning network in the private area of ​​the dormitory building includes: The millimeter-wave radar is used to collect gait waveforms, and the sliding window variance of the step length variation coefficient is extracted as the millimeter-wave gait entropy feature. Synchronously obtain the Bluetooth device connection sequence and calculate the spectrum entropy value of the signal strength fluctuation as the Bluetooth connection entropy feature; A dual-channel attention mechanism is built at the edge nodes: the gait entropy channel focuses on joint swing frequency, and the device entropy channel strengthens the connection interval characteristics; The dual-channel outputs are fused into an anonymized compliance feature vector, which is only verified if the vector's Euclidean distance is lower than the adaptive threshold of the dormitory member template library; If the verification fails continuously, the corridor camera will be activated to generate a low-resolution thermal image and push it to the dormitory management terminal.

[0010] Furthermore, the template update mechanism of the anonymized compliance feature vector includes: Gait data of thick-shoe gait in winter were collected, and the seasonal offset of joint swing frequency entropy was calculated using a Bayesian probability model. When the seasonal offset exceeds a preset critical value, the gait entropy weight coefficient is reduced and the decision-making proportion of Bluetooth connection entropy is increased; The updated template library is encrypted and distributed to the edge nodes of each campus, and the permission attributes in the semantic grid are updated synchronously.

[0011] Furthermore, the construction of the semantic grid includes: The three-dimensional geographic model of the campus is discretized into cubic grid cells with a side length of 30 cm; Bind the mobile phone signal blocking rule to the teaching building grid during the exam week. If a mobile device radio frequency signal is detected during the effective period of the rule, an abnormal flag will be generated. The thermal weight values ​​of historical fall events are loaded into the stair grid, and the thermal weight values ​​increase in real time with rain and snow weather; A four-dimensional index structure is formed with grid coordinates as the core and time authority, function label, and risk coefficient as attributes.

[0012] Furthermore, the implementation steps of the global risk causal chain deduction in the dormitory fire scenario include: When the current sensor detects that the load on the power strip continues to exceed the limit and the infrared sensor identifies that the density of people exceeds the limit, a composite abnormality mark is marked; Retrieve the escape route topology and flammable material distribution matrix associated with the dormitory grid; The campus event knowledge base was used to simulate the fire evolution chain: a circuit overload caused sparks to ignite the bed curtains, which in turn caused smoke to spread along the ventilation ducts; A dynamic risk vector field is generated based on smoke diffusion particle swarm simulation, and the evacuation path is planned using the tangential component of the dynamic risk vector field as a constraint. Push vector maps with real-time route updates to teachers and students' mobile phones, and overlay a semi-transparent warning layer of smoke-covered areas on corridor displays; Automatically remove privacy blur from escape route cameras based on risk levels and generate a heat map of occupant density to assist in rescue dispatch.

[0013] Furthermore, the dynamic task allocation of the sensor network includes: The node value coefficient is calculated through a bionic optimization algorithm, which is determined based on the target distance, remaining power and network delay; When the density of people gathering on the playground exceeds the set density threshold, the neighboring node group starts the face tracking mode and compresses non-critical background pixels; The remote node group dispatches drones to capture a bird's-eye view of crowd distribution, and the edge node generates a crowd density heat map that is projected onto the ring screen in the security center. High-density areas in the crowd density heat map are automatically marked as stampede risk areas and trigger evacuation broadcasts.

[0014] Furthermore, the implementation of the hierarchical privacy protection includes: Deploy a piezoelectric sensor array at the entrance of the psychological counseling room to analyze the periodic characteristics of pedaling duration; When the environmental noise sensor detects an abnormal fundamental frequency of the sobbing sound pattern, it generates an encrypted warning message and sends it to the school doctor terminal; A dynamic pixel reconstruction strategy is used in real time for indoor video streams to transform human forms into abstract geometric blocks while preserving motion vectors.

[0015] Furthermore, the implementation of the cross-campus collaborative action includes: Compress the millimeter-wave radar waveform features of climbing walls into a 128-bit fingerprint code; After receiving the 128-bit fingerprint code, the mountain campus matches similar waveform patterns in the local feature library through Hamming distance; If the matching difference value is lower than the dynamic threshold, the lightweight detection model is downloaded to the edge device and the risk factor of the semantic grid is updated synchronously.

[0016] Based on the embodiments provided by this application, the perception module acquires perception data based on the sensor network deployed in various areas of the campus and dynamically allocates perception tasks according to regional functions. It can flexibly configure perception resources according to the functional requirements of different areas, improve the pertinence and effectiveness of perception, avoid resource waste, and better adapt to the diverse security needs of various areas of the campus. The processing module processes perception data in real time through reconfigurable hardware nodes, which can achieve rapid processing of perception data and timely generate compliance features and anomaly identification; uses a temporal rule embedded causal reasoning network to encode campus rules into neural network differentiable rule synapses, so that campus rules are deeply integrated into the neural network processing process, and achieve in-depth analysis and reasoning of the compliance of perception data, rather than simple data-level splicing; uses an asymmetric modal contrast learning network to fuse millimeter-wave gait entropy features and Bluetooth connection entropy features in the privacy area, which not only achieves deep fusion analysis of multimodal data in the privacy area, but also avoids the direct use of video and other methods that may leak privacy, thereby improving the ability of privacy area anomaly detection while ensuring privacy. The risk deduction module constructs a semantic grid based on the output of the processing module, associating regional functional attributes, temporal permissions, and historical risk heat maps to perform global risk causal chain deduction. This allows for global correlation analysis of risks from multiple dimensions, including regional functions, temporal permissions, and historical risks, exploring causal relationships between risks and achieving a more comprehensive and in-depth deduction of campus risks. This addresses the existing lack of deep correlation and collaborative analysis. The response execution module triggers dynamic responses based on the deduction results, including hierarchical privacy protection and cross-campus collaborative actions. This enables hierarchical privacy protection based on risk conditions, better balancing security and privacy when responding to risks. Cross-campus collaborative actions enable campuses to collaborate on security incidents, improving the coordination and effectiveness of the campus' overall security response. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings described herein are used to provide a further understanding of the embodiments of the present invention and constitute a part of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation of this application. In the drawings: Figure 1 1 is a structural diagram of an optional multi-angle campus panoramic intelligent monitoring system according to an embodiment of the present application; Figure 2 This is a flow chart of an optional generation of differentiable regular synapses in a neural network according to an embodiment of the present application; Figure 3 This is a flowchart of the operation of an optional asymmetric modal contrast learning network in a private area of ​​a dormitory building according to an embodiment of the present application.

[0018] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0020] Alternatively, as Figure 1 As shown, this application provides a multi-angle campus panoramic intelligent monitoring system, including: The perception module 101 is used to acquire perception data based on the sensor network deployed in various areas of the campus and dynamically assign perception tasks according to regional functions. The sensor network includes cameras, millimeter-wave radars, pressure sensors, environmental noise sensors, and Bluetooth devices. In this embodiment, the campus buildings are divided into functionally differentiated areas based on their use, and sensors dynamically adjust their sensing tasks according to the area's function. For example, in the laboratory area (3rd floor of the Chemistry Building), millimeter-wave radars and current sensors are deployed to focus on monitoring the compliance of hazardous chemical operations; in the private area of ​​the dormitory building (corridor on the 4th floor of Building 1), a combination of millimeter-wave radars and Bluetooth devices is deployed to balance identity verification and privacy protection; in the open area of ​​the playground, cameras and environmental noise sensors are deployed for crowd density monitoring and sudden noise warnings. The processing module 102 is configured to process the sensory data in real time through a reconfigurable hardware node to generate compliance features and anomaly indicators. The processing method includes: encoding campus rules into neural network differentiable rule synapses using a temporal rule embedded causal inference network; fusing millimeter wave gait entropy features and Bluetooth connection entropy features using an asymmetric modal contrast learning network in the privacy area; In this embodiment, the reconfigurable hardware nodes are hardware modules based on FPGAs / edge computing units. For example, during exam week, the edge nodes automatically switch computing resources from facial recognition to mobile phone signal spectrum analysis, dynamically loading the RF signal detection module by reconfiguring the hardware logic.

[0021] The temporal rule-embedded causal inference network breaks down campus security rules (such as two-person authorization and time-of-day permissions) into temporal and spatial conditions computable by neural networks. This network then implements rule triggering and causal inference through differentiable synapses. For example, the library closing hours (22:00-6:00) are encoded as follows: the temporal condition corresponds to the neural network bias vector (bias value + 1 after 22:00), and the spatial condition corresponds to the entrance area weight matrix. When the infrared sensor detects a single person entering, a synaptic convolution operation outputs an anomaly flag.

[0022] The asymmetric modal contrastive learning network uses an asymmetric strategy to fuse different sensor features in the privacy zone: millimeter-wave radar acquires physical features (gait), and Bluetooth devices acquire behavioral features (device connection patterns), achieving anonymous identity verification through contrastive learning. For example, at the entrance of the psychological counseling room, the pressure waveform (gait) collected by the piezoelectric sensor array and the Bluetooth MAC address change sequence of the visitor's mobile phone are used by the contrastive learning network to generate a 128-dimensional anonymous feature vector, retaining only the temporal correlation features and removing the biometric information. The risk deduction module 103 is used to construct a semantic grid based on the compliance features and anomaly identification output by the processing module, associate regional functional attributes, time permissions and historical risk heat maps, and perform global risk causal chain deduction; Among them, compliance features include feature vectors extracted from sensor data that comply with safety rules. For example, in a two-person authorization scenario in a laboratory, the millimeter-wave radar simultaneously detects the spatial distribution characteristics of two physical features.

[0023] The response execution module 104 is used to trigger dynamic responses based on the deduction results, including hierarchical privacy protection and cross-campus collaborative actions.

[0024] Based on the embodiments provided by this application, the perception module acquires perception data based on the sensor network deployed in various areas of the campus and dynamically allocates perception tasks according to regional functions. It can flexibly configure perception resources according to the functional requirements of different areas, improve the pertinence and effectiveness of perception, avoid resource waste, and better adapt to the diverse security needs of various areas of the campus. The processing module processes perception data in real time through reconfigurable hardware nodes, which can achieve rapid processing of perception data and timely generate compliance features and anomaly identification; uses a temporal rule embedded causal reasoning network to encode campus rules into neural network differentiable rule synapses, so that campus rules are deeply integrated into the neural network processing process, and achieve in-depth analysis and reasoning of the compliance of perception data, rather than simple data-level splicing; uses an asymmetric modal contrast learning network to fuse millimeter-wave gait entropy features and Bluetooth connection entropy features in the privacy area, which not only achieves deep fusion analysis of multimodal data in the privacy area, but also avoids the direct use of video and other methods that may leak privacy, thereby improving the ability of privacy area anomaly detection while ensuring privacy. The risk deduction module constructs a semantic grid based on the output of the processing module, associating regional functional attributes, temporal permissions, and historical risk heat maps to perform global risk causal chain deduction. This allows for global correlation analysis of risks from multiple dimensions, including regional functions, temporal permissions, and historical risks, exploring causal relationships between risks and achieving a more comprehensive and in-depth deduction of campus risks. This addresses the existing lack of deep correlation and collaborative analysis. The response execution module triggers dynamic responses based on the deduction results, including hierarchical privacy protection and cross-campus collaborative actions. This enables hierarchical privacy protection based on risk conditions, better balancing security and privacy when responding to risks. Cross-campus collaborative actions enable campuses to collaborate on security incidents, improving the coordination and effectiveness of the campus' overall security response.

[0025] Furthermore, if Figure 2 As shown in the figure, the generation of differentiable rule synapses in the temporal rule embedded causal reasoning network includes: S201, decompose the laboratory two-person authorization rule into spatial conditions and temporal conditions, which are encoded into the weight matrix and bias vector of the neural network respectively; In this embodiment, the rule decomposition logic includes the following: spatial conditions: the spatial feature extraction kernel of the neural network weight matrix corresponding to the laboratory area coordinates (e.g., X=123, Y=456, Z=0); and temporal conditions: the timestamp weighting coefficients of the late-night period (23:00-5:00) converted into bias vectors. For example, when the millimeter-wave radar detects a single person at 01:30 in the chemical laboratory (X=123, Y=456), the weight matrix activates the spatial feature convolution operation, and the bias vector triggers the time condition check. When both conditions are met, the electrical device signal analysis is initiated.

[0026] S202, when the millimeter-wave radar detects a single person in the laboratory area and the timestamp is late at night, the convolution kernel operation of the regular synapse is activated; S203, using a differentiable knowledge graph to verify whether the signal characteristics of the electrical equipment match the hazardous chemicals operation mode. If they match, an abnormality mark is generated and an access control lock instruction is triggered.

[0027] In this embodiment, the differentiable knowledge graph verification involves constructing hazardous chemical operation modes (e.g., ether distillation requires continuous heating at 300W) as knowledge graph nodes and verifying current characteristics through a differentiable graph neural network. Examples of hazardous chemicals include: flammable and explosive ether (operation requires anti-static grounding and heating power ≤ 300W); highly toxic potassium cyanide (two-person, two-lock management, access requires simultaneous activation of two RFID tags).

[0028] In a specific embodiment, the timing rule synaptic activation function is:

[0029] in, is the spatial condition factor (the result of single-person body sign detection in the laboratory area, 1 if a single person is detected, otherwise 0); is the time condition factor (late night time coding, 23:00-05:00 time period is 1, and other time periods are 0); is the spatial condition weight matrix (based on the two-person authorization rule, the risk weight of a single person's physical sign is 2.0); is the time condition bias vector (the risk weight of the late night period is 1.5); Embedding vector of the knowledge graph of electrical equipment signal features (dimension 128, generated by graph convolutional network); The baseline atlas vector for compliant hazardous chemicals operation modes (dimension 128); is the knowledge graph verification coefficient (dynamically determined according to the level of hazardous chemicals, 1.2 for highly toxic storage areas and 0.8 for ordinary laboratories); is the Sigmoid activation function (mapping the input to the activation probability in the [0,1] interval); is the 2-norm of the vector.

[0030] Based on the above formula, through the linear combination of spatial conditions (single person's physical signs) and temporal conditions (late night time), combined with the knowledge graph verification item (the normalized Euclidean distance between the device signal and the compliance mode), the traditional Boolean rules are converted into differentiable neural network operations, solving the problem that static rules cannot adapt to complex scenarios.

[0031] Based on the embodiment provided by this application, by decomposing the laboratory's two-person authorization rules into spatial conditions and temporal conditions and encoding them into the weight matrix and bias vector of the neural network, the system can accurately identify abnormal situations in which a single person's physical signs appear in the laboratory area late at night. When the millimeter-wave radar detects a single person's physical sign and the timestamp is in the late night period, the convolution kernel operation of the rule synapse is activated, and the matching verification of the signal characteristics of the electrical equipment and the operation mode of hazardous chemicals is combined with the differentiable knowledge graph. In a short period of time, an abnormal identification can be generated and an access control lock instruction can be triggered, achieving accurate detection and real-time control of illegal laboratory operations, effectively improving the safety protection capabilities of high-risk scenarios in the laboratory, and avoiding safety accidents caused by a single person operating hazardous chemicals.

[0032] Furthermore, after triggering the access lock instruction, the method further includes: The camera in the laboratory area is used to capture facial images of people, and non-sensitive contour features are extracted through edge nodes; In this embodiment, the feature extraction mechanism for non-sensitive contour feature recognition includes: edge nodes use CNN to extract the geometric parameters of the facial contour (16 non-biological feature points such as the curvature of the jawline and the slope of the nose bridge), removing sensitive information such as texture and skin color. For example, after the laboratory access is locked, the camera captures a person wearing a mask. The system only extracts the contour curve from the eyebrows to the jaw (ignoring the area covered by the mask), generating an 8-dimensional contour vector (such as [1.2, 0.8, 1.5, ...]). Dynamic time-warping matching of non-sensitive profile features with authorized personnel models in the compliance feature library; Among them, the compliance feature library stores non-sensitive contour templates of authorized personnel, such as the contour vector set {V1, V2, V3} of laboratory administrator A corresponding to different angles).

[0033] If the matching similarity is lower than the safety threshold, a two-person verification request and a heat map of the person's movement trajectory will be sent to the security terminal; Examples of security thresholds include: normal working hours threshold: 0.7 (Euclidean distance < 0.7 is considered an authorized person); late night abnormal period threshold: 0.9 (to improve verification accuracy and reduce the risk of misplacement).

[0034] Synchronously cut off the power supply of high-risk equipment and record the operation log to the risk heat map of the semantic grid. Based on the embodiment provided by the present application, after the access control lock command is triggered, the camera in the laboratory area is called to capture the facial image of the person and extract the non-sensitive contour features through the edge node, which not only ensures the privacy protection of the person's facial information, but also obtains the key features for identity authentication. The non-sensitive contour features are dynamically time-regularized matched with the authorized personnel model in the compliance feature library. If the matching similarity is lower than the safety threshold, a two-person verification request and a personnel movement trajectory heat map can be quickly sent to the security terminal. At the same time, the power supply of high-risk equipment is synchronously cut off and the operation log is recorded to the risk heat map of the semantic grid. This series of operations not only ensures the safe power-off of laboratory equipment, but also provides security personnel with accurate verification information and personnel trajectories, further strengthens the subsequent processing capabilities of laboratory abnormalities, and forms a complete safety closed loop from detection, response to verification.

[0035] Furthermore, if Figure 3 As shown in Figure 2, the operation of the asymmetric modal contrast learning network in the privacy area of ​​the dormitory building includes: S301, collecting gait waveforms through millimeter-wave radar, and extracting the sliding window variance of the stride length variation coefficient as the millimeter-wave gait entropy feature; In this embodiment, the sliding window variance of the stride length coefficient of variation is calculated by calculating the standard deviation of 10 consecutive strides and then calculating the variance over a 30-second sliding window (e.g., variance is <0.5 for normal gait and >0.8 for thick shoes). For example, in a dormitory building during winter, a student wearing snow boots may experience an increase in the variance of the stride length coefficient of variation from 0.4 to 1.1, triggering a gait entropy feature deviation warning.

[0036] S302, synchronously obtain a Bluetooth device connection sequence, and calculate a spectrum entropy value of signal strength fluctuation as a Bluetooth connection entropy feature; At step S303, a dual-channel attention mechanism is constructed at the edge node: the gait entropy channel focuses on joint swing frequency, while the device entropy channel enhances connection interval features. At step S304, the dual-channel outputs are fused into an anonymized, compliant feature vector. This vector is only validated if its Euclidean distance is below an adaptive threshold in the dormitory member template library. The adaptive thresholds of the dormitory member template library include but are not limited to 1.2, 1.5, etc.

[0037] In this embodiment, the gait entropy channel outputs a 16-dimensional joint swing frequency feature vector, and the device entropy channel outputs an 8-dimensional Bluetooth connection interval feature vector. These are then weighted and fused into a 24-dimensional anonymous vector through an attention mechanism. If the Euclidean distance between the fused vector and the dormitory member template library is less than 1.2, the member is considered a legitimate member. Otherwise, the corridor camera is activated to generate a heat map (for example, if the Bluetooth connection intervals of outsiders are irregular, the calculated distance value may reach 2.5). S305: If the verification fails continuously, the corridor camera is activated to generate a low-resolution thermal image and push it to the dormitory management terminal.

[0038] In a specific embodiment, the asymmetric modality contrast learning fusion function is:

[0039] in, is the millimeter wave gait entropy feature vector (including the sliding window variance of the step length variation coefficient); is the Bluetooth connection entropy feature vector (including the spectrum entropy value of signal strength fluctuation); The attention mechanism for the gait entropy channel (focusing on the 0.8-1.2Hz joint swing frequency characteristics); An attention mechanism for device entropy channels (enhancing the 10-30 second Bluetooth connection interval feature); is the gait entropy weight coefficient (dynamic range [0.2, 0.8], automatically reduced to 0.3 in winter).

[0040] Based on the above formula, a dual-channel attention mechanism extracts key features of gait entropy (joint swing frequency) and device entropy (connection interval stability). A temperature-sensitive weight coefficient α dynamically adjusts the modal fusion ratio: When winter temperatures fall below a baseline, α automatically decreases (e.g., to 0.3), minimizing the impact of gait variations in thick shoes on verification while increasing the weight of Bluetooth connection entropy in decision-making. Compared to fixed-weight fusion (e.g., traditional methods that prioritize gait entropy at 50%), this mechanism reduces verification failure rates in winter scenarios and protects user privacy by anonymizing feature vectors (with a 128-dimensional fusion) (transmitting only abstract features, not raw signals). Furthermore, the attention mechanism focuses on specific frequency / interval features (e.g., gait entropy prioritizes joint swings between 0.8-1.2Hz, while device entropy prioritizes connection intervals exceeding 10 seconds). This effectively distinguishes the behavioral patterns of dormitory residents from those of outsiders, improving security accuracy in private areas.

[0041] Based on the embodiment provided by the present application, in the privacy area of ​​the dormitory building, the millimeter-wave radar is used to collect gait waveforms to extract millimeter-wave gait entropy features, the Bluetooth device connection sequence is simultaneously obtained to calculate the Bluetooth connection entropy features, and a dual-channel attention mechanism is constructed at the edge node, focusing on the joint swing frequency and strengthening the connection interval features respectively, and the dual-channel output is fused into an anonymous compliance feature vector. This multimodal fusion method can accurately identify the identity of people while protecting the privacy of dormitory members. The verification is passed only when the Euclidean distance of the vector is lower than the adaptive threshold of the dormitory member template library. If the verification fails continuously, the corridor camera can be activated to generate a low-resolution thermal imaging map and push it to the dormitory management terminal. This solution effectively balances the needs of privacy protection and security verification, improves the accuracy and security of identity verification of people in the dormitory area, and reduces the risk of outsiders illegally entering the dormitory.

[0042] Furthermore, the template update mechanism for the anonymized compliance feature vector includes: Gait data from thick winter shoes is collected, and the seasonal offset of joint swing frequency entropy is calculated using a Bayesian probability model. When the seasonal offset exceeds a preset critical value, the gait entropy weight coefficient is reduced and the decision-making proportion of Bluetooth connection entropy is increased. In this embodiment, the dynamic weight adjustment mechanism includes: when the winter offset exceeds 15%, the gait entropy weight is reduced from 60% to 40%, and the Bluetooth connection entropy is increased from 30% to 50%. Due to the large gait variation in thick shoes, it depends on the device connection pattern. Example threshold values: Winter mode: 15% (triggered when temperature is less than 5°C); Rainy season mode: 10% (triggered when air humidity is greater than 80%).

[0043] The updated template library is encrypted and distributed to the edge nodes of each campus, and the permission attributes in the semantic grid are updated synchronously.

[0044] Based on the embodiment provided by this application, for the template update mechanism of anonymized compliance feature vectors, by collecting winter thick shoe gait data, the seasonal offset of the joint swing frequency entropy is calculated using the Bayesian probability model. When the offset exceeds the preset critical value, the weight coefficients of the gait entropy and the Bluetooth connection entropy are automatically adjusted, and the updated template library is distributed to the edge nodes of each campus after encryption, and the authority attributes in the semantic grid are updated synchronously. This mechanism can adaptively respond to the impact of seasonal changes on the gait characteristics of people, and ensure that the accuracy of dormitory member identity verification is not significantly affected in different seasonal environments. By dynamically adjusting weights and distributed template updates, the system can maintain the ability to accurately identify the identities of people in the dormitory area for a long time, avoid verification misjudgments due to seasonal factors, and improve the environmental adaptability and stability of the system.

[0045] Furthermore, the construction of the semantic grid includes: The 3D campus geographic model is discretized into cubic grid cells with a side length of 30 cm. A mobile phone signal shielding rule for exam week is bound to the teaching building grid. If a mobile device radio frequency signal is detected during the effective period of the rule, an abnormal flag is generated. In this example, physical spaces such as campus buildings, roads, and green spaces are converted into a 3D campus geographic model, discretized into cubic grid cells with a precision of 30 cm for spatial positioning and rule binding. For example, the main library building is modeled as a 100×80×20 (unit: grid) 3D structure. The bookshelf area grid is bound to the "RFID book positioning rule," while the reading area grid is loaded with the "people density monitoring task."

[0046] The rule's effective period includes the blocking period set according to the class schedule, typically 8:00-12:00 and 14:00-17:00 on exam days. For example, on May 30, 2025 (the third day of exam week), the blocking rule will be activated for all grids on the third floor of the teaching building between 8:30-11:30. If a mobile phone signal strength > -70dBm is detected during this period, an abnormal flag will be generated.

[0047] The thermal weight values ​​of historical fall events are loaded into the stair grid, and the thermal weight values ​​increase in real time with rain and snow weather; Thermal weights are assigned to stair grids based on historical accident data. For example, the east staircase of the administration building has an initial weight of 0.7 due to two falls in the winter of 2024. During moderate rain, the weight is multiplied by 1.8 to 1.26, triggering a "install anti-slip mat" reminder. The dormitory staircase has a weight multiplied by 1.5 after evening study (10:00 PM to 11:00 PM) due to insufficient light. If it rains, the weight is multiplied by another 1.8, resulting in a final weight of 0.5 × 1.5 × 1.8 = 1.35.

[0048] A four-dimensional index structure is formed with grid coordinates as the core and time authority, function label, and risk coefficient as attributes.

[0049] Based on the embodiment provided by this application, when constructing a semantic grid, the three-dimensional geographic model of the campus is discretized into cubic grid units with a side length of 30 cm. The corresponding rules are bound to the grids of different functional areas and the historical risk thermal weight values ​​are loaded to form a four-dimensional index structure with grid coordinates as the core. This refined grid division and attribute binding enables the system to accurately associate campus rules (such as mobile phone signal blocking rules during exam weeks) and historical risk data (such as the thermal weight value of stair fall events that increases in real time with rain and snow) with specific spatial locations and time dimensions. During real-time monitoring, it is possible to determine whether to generate an abnormality mark based on the attribute information of the grid unit, providing accurate and efficient basic data support for the deduction of the global risk causal chain, and realizing refined modeling and dynamic management of risks in various areas of the campus.

[0050] Furthermore, the implementation steps of the global risk causal chain deduction in the dormitory fire scenario include: When the current sensor detects that the load on the power strip continues to exceed the limit and the infrared sensor identifies that the density of people exceeds the limit, a composite abnormality mark is marked; In this embodiment, power strip loads that continuously exceed the limit include: For ordinary dormitory power strips: continuously >10A (for example, using a 1600W hair dryer and a 500W hand warmer simultaneously, the current is ≈ (1600+500) / 220≈9.5A, approaching the limit). For power strips for high-risk laboratory equipment: continuously >5A (for example, using a 1200W electric hot pot illegally, the current is ≈1200 / 220≈5.45A, triggering a limit violation).

[0051] The population density exceeds the standard, including: standard dormitory (20㎡): >8 people (average area per person <2.5㎡, for example, 10 people live in a 6-person room); corridor (2m wide): >3 people / ㎡ (for example, 60 people gather in a 10m corridor, the density is 60 / (10×2)=3 people / ㎡).

[0052] A compound exception requires at least two independent exception conditions to be met simultaneously, while a common exception is triggered by a single condition. For example, a compound exception: a power strip current of 12A (exceeding the limit) + 10 people in the dormitory (exceeding the density limit) generates a flag of 0x1101. A common exception: a single current of 12A generates a flag of 0x1001, and a single person with 10 people generates a flag of 0x0101.

[0053] Retrieve the escape route topology and flammable material distribution matrix associated with the dormitory grid; For example, the escape route topology map associated with a dormitory grid shows that the nearest safe exit is 15 meters to the east, and the alternative route requires a 20-meter detour through the stairwell; the flammable material distribution matrix is ​​marked with: bed curtains (burning speed 0.2m / s), bedding (calorific value 15MJ / kg), and books (combustion smoke toxicity coefficient 0.8).

[0054] The campus event knowledge base was used to simulate the fire evolution chain: a circuit overload caused sparks to ignite the bed curtains, which in turn caused smoke to spread along the ventilation ducts; For example, the fire evolution chain simulation includes: circuit overload (15A for 5 minutes) → insulation layer melts and produces sparks (temperature > 300°C) → ignites bed curtains (0.5 meters away from the power strip) → smoke spreads along the ventilation duct (20 cm in diameter) to the corridor at a speed of 0.8 m / s.

[0055] A dynamic risk vector field is generated based on smoke diffusion particle swarm simulation, and the evacuation path is planned using the tangential component of the dynamic risk vector field as a constraint. For example, the smoke diffusion particle swarm simulation includes: simulating a dormitory fire, releasing 1,000 smoke particles, each particle diffuses according to the law of Brownian motion, and adjusts its direction according to the wind speed of 1.2m / s when encountering a ventilation hole, generating a dynamic risk vector field (the red area indicates a particle concentration of >500 / cubic meter).

[0056] Push vector maps with real-time route updates to teachers and students' mobile phones, and overlay a semi-transparent warning layer of smoke-covered areas on corridor displays; Automatically remove privacy blur from escape route cameras based on risk levels and generate a heat map of occupant density to assist in rescue dispatch.

[0057] In this embodiment, the risk levels are as follows: Level 1 (Low): Smoke concentration <200ppm, temperature <60°C; Level 2 (Medium): 200-500ppm, 60-80°C; Level 3 (High): >500ppm, >80°C. The removal conditions include: When the risk level reaches Level 2 or above, the privacy blurring function on the escape route cameras is automatically removed (for example, changing facial mosaics to outlines).

[0058] Density heatmaps: These maps use color gradients to visually display the density of people in an area. These maps are generated based on infrared sensor or camera data. For example, in a dormitory fire scenario, a heatmap would show corridor corners as red (density > 4 people / m2) and areas near emergency exits as green (< 1.5 people / m2), assisting with rescue dispatch.

[0059] In a specific embodiment, the dynamic risk vector field generation model is:

[0060] Based on the above formula, is the risk vector field at the three-dimensional space coordinate (x, y, z) at time t (unit: person / (s・m³), pointing in the direction of risk reduction); is the smoke diffusion speed (unit: m / s, initial value 0.3m / s, increases by 0.1m / s for every 5℃ increase in temperature); is the personnel density function (unit: person / m³, generated by calculating the drone heat map and edge nodes); The Euclidean distance from the current location to the nearest escape route (unit: m, calculated based on the campus 3D geographic model); is the channel resistance coefficient (dynamically adjusted according to the thermal weight of historical fall events, 0.5 for stairs and 1.2 for corridors); is the length of the escape route (unit: m, pre-stored in the four-dimensional index structure of the semantic grid); is the gradient operator (derivative with respect to spatial coordinates, unit: 1 / m).

[0061] Based on the above formula, the risk flow per unit time and per unit area is quantified by the product term of smoke diffusion speed (positively correlated with temperature, simulating thermal buoyancy effect) and occupant density; the exponential term describes the attenuation effect of escape route distance and resistance on risk (e.g. stairs have a high risk of falls due to historical reasons, Taking 0.5 increases the attractiveness of the escape route. The negative gradient ensures that the vector field points to the location of minimal risk. Combined with the smoke diffusion path generated by particle swarm simulation, intelligent evacuation path planning is achieved. Dynamic avoidance of smoke-covered areas is achieved (for example, the vector field sets a repulsive component in areas with smoke concentrations > 500 ppm). Furthermore, by removing the privacy blur from the escape route cameras, the simultaneously generated occupant density heat map (based on the tangential component of the risk vector field) assists security terminals in real-time adjustment of rescue strategies, improving evacuation efficiency and addressing the issues of delayed risk deduction and extensive path planning inherent in traditional security systems.

[0062] Based on the embodiments provided by this application, in a dormitory fire scenario, when the current sensor detects that the power strip load continues to exceed the limit and the infrared sensor identifies that the human density exceeds the standard, a composite abnormality flag is marked. Subsequently, the escape route topology map and flammable material distribution matrix associated with the dormitory grid are retrieved, and the campus event knowledge base is called to simulate the fire evolution chain. A dynamic risk vector field is generated based on the smoke diffusion particle swarm simulation, and the evacuation path is planned using its tangential component as a constraint. This solution can detect abnormal signals in the early stage of the fire through multimodal fusion, generate a scientific evacuation path, and push a vector map containing real-time path updates to teachers and students' mobile phones. A semi-transparent warning layer of the smoke-covered area is superimposed on the corridor display screen. At the same time, the privacy blur of the escape route camera is automatically removed according to the risk level, and a human density heat map is generated to assist in rescue scheduling. This series of operations realizes early warning of dormitory fires, dynamic risk assessment and efficient evacuation guidance, providing strong support for the safe escape and rescue scheduling of teachers and students in fire scenarios, and minimizing the losses caused by fire.

[0063] Furthermore, the dynamic task allocation of sensor networks includes: The node value coefficient is calculated through a bionic optimization algorithm, which is determined based on the target distance, remaining power and network delay; In this embodiment, the biomimetic optimization algorithm simulates biological behaviors such as ant colonies foraging and bee colonies' division of labor to dynamically optimize sensor task allocation strategies. For example, during a playground gathering, the algorithm prioritizes the node group with the lowest path loss between the sensing node and the crowd center, just as an ant colony searches for the shortest path, while also preventing nodes with low battery power from overworking.

[0064] For example, the node value coefficient is related to the target distance: <50 meters from the target: value +0.4 (such as a camera 30 meters away from the center of the playground); 50-100 meters: value +0.2 (such as a camera 80 meters away from the edge of the playground); >100 meters: value -0.3 (such as a camera 150 meters outside the playground).

[0065] Node value coefficient and remaining power: >70%: value +0.3 (node ​​with 80% power); 30%-70%: value +0.1 (node ​​with 50% power); <30%: value -0.4 (node ​​with 20% power).

[0066] Node value coefficient and network delay: <30ms: value +0.2 (node ​​with 25ms delay); 30-100ms: value 0 (node ​​with 50ms delay); >100ms: value -0.3 (node ​​with 150ms delay).

[0067] When the density of people gathering on the playground exceeds the set density threshold, the neighboring node group starts the face tracking mode and compresses non-critical background pixels; For example, density thresholds are set as follows: Yellow Alert: >2 people / m2 (crowding at the playground entrance); Red Alert: >4 people / m2 (crowding in front of the stage). During a playground gathering, the neighboring node group includes: fixed cameras within 50 meters of the playground (30 meters from the center of the crowd); and millimeter-wave radars installed on the stand railings (20 meters from the crowd).

[0068] Non-critical background pixels: When tracking people, the pixels in the background areas such as trees and fences in the video are compressed into 16×16 pixel blocks (original resolution 1920×1080), and only the high resolution (such as 512×512) of the area where the people are located is retained.

[0069] The remote node group dispatches drones to capture a bird's-eye view of crowd distribution, and the edge node generates a crowd density heat map that is projected onto the ring screen in the security center. For example, during a playground assembly, the remote node group includes: a telephoto camera deployed on the roof of the teaching building (200 meters away from the playground); and a dispatched drone (300 meters above the playground).

[0070] High-density areas in the crowd density heat map are automatically marked as stampede risk areas and trigger evacuation broadcasts.

[0071] Among them, if the density of an area is greater than 2.5 people / ㎡ or greater than 5 people / ㎡, the area is a high-density area.

[0072] Based on the embodiments provided in the present application, for the dynamic task allocation of the sensor network, the node value coefficient is calculated through a bionic optimization algorithm, and the perception tasks are reasonably allocated based on factors such as target distance, remaining power and network delay. When the density of people gathering on the playground exceeds the set threshold, the neighboring node group can start the face tracking mode and compress non-critical background pixels. The remote node group dispatches the drone to take a bird's-eye view of the crowd distribution, and generates a crowd density heat map through the edge node and projects it to the ring screen of the security center. The high-density areas in the heat map are automatically marked as stampede risk areas and trigger evacuation broadcasts. This dynamic task allocation mechanism realizes the efficient use of sensor network resources, can monitor the crowd distribution and density changes in crowded scenes such as playground gatherings in real time and comprehensively, quickly identify stampede risk areas and trigger evacuation measures in a timely manner, effectively improving the safety assurance capabilities of large-scale gatherings on campus and reducing the possibility of stampede accidents.

[0073] Furthermore, the implementation of hierarchical privacy protection includes: Deploy a piezoelectric sensor array at the entrance of the psychological counseling room to analyze the periodic characteristics of pedaling duration; For example, a 3×3 piezoelectric sensor array is embedded in the floor at the entrance of a psychological counseling room. Each sensor is 20 cm apart and arranged in a matrix pattern, which can locate the stepping position and pressure distribution. The analysis of periodic characteristics includes: During normal entry, the pedaling interval is greater than 1.5 seconds (e.g., the "pedal-pedal" interval is 2 seconds), and during emergency situations, the interval is less than 0.8 seconds (e.g., the "pedal-pedal-pedal" interval is 0.5 seconds). The system uses FFT to analyze the frequency periodicity of the pedaling signal.

[0074] When the environmental noise sensor detects an abnormal fundamental frequency of the sobbing sound pattern, it generates an encrypted warning message and sends it to the school doctor terminal; For example, in normal conversation, the fundamental frequency is 180-250Hz for women and 100-150Hz for men; in abnormal sobbing, the fundamental frequency is >280Hz (sharp crying) for women and >180Hz (suppressed crying) for men.

[0075] A dynamic pixel reconstruction strategy is used in real time for indoor video streams to transform human forms into abstract geometric blocks while preserving motion vectors.

[0076] The dynamic pixel reconstruction strategy includes: converting indoor people into abstract geometric color blocks: standing posture → blue cylinder (1.7m high, 0.3m diameter); waving action → adding red fan-shaped dynamic vector (radius 0.5m, direction changes with the action); walking trajectory → green dotted line to retain the motion path (such as a straight line from point A to point B).

[0077] Based on the embodiments provided in this application, in terms of hierarchical privacy protection, the piezoelectric sensor array deployed at the entrance of the psychological counseling room analyzes the periodic characteristics of the trampling duration, and combined with the environmental noise sensor's recognition of abnormal fundamental frequency of the sobbing sound pattern, it can accurately detect possible abnormal conditions in the room, generate encrypted early warning information and send it to the school doctor terminal to ensure the security and timeliness of information transmission. At the same time, the indoor video stream uses a dynamic pixel reconstruction strategy in real time to convert the human body shape into an abstract geometric color block and retain the motion vector. Without leaking sensitive information such as the person's face, it can monitor the movement status of indoor personnel in real time, balancing the needs of privacy protection and security monitoring. This solution provides effective security protection for privacy-sensitive areas such as psychological counseling rooms, which not only protects the privacy of visitors, but also provides timely early warning and intervention when abnormal situations occur, thereby improving the level of security management in private areas on campus.

[0078] Furthermore, the realization of cross-campus collaborative actions includes: Compress the millimeter-wave radar waveform features of climbing walls into a 128-bit fingerprint code; For example, for the millimeter-wave radar waveform characteristics, normal climbing: the waveform shows high-frequency jitter (frequency 10-20Hz), large amplitude variation (peak value > 0.5V), and duration 5-10 seconds; animal climbing (such as cats): the waveform frequency is low (5-8Hz), the amplitude variation is small (peak value < 0.3V), and the duration is irregular.

[0079] After receiving the 128-bit fingerprint code, the mountain campus matches similar waveform patterns in the local feature library using Hamming distance; For example, the climbing code sent by the main campus is 10110101... (128 bits), and the "personnel climbing" code in the local feature library of the mountain campus is 10100101..., and the Hamming distance between the two is 2 (only the 3rd and 6th bits are different), and they are judged to be highly similar.

[0080] If the matching difference value is lower than the dynamic threshold, the lightweight detection model is downloaded to the edge device and the risk factor of the semantic grid is updated synchronously.

[0081] For example, dynamic thresholds include: Campus Difference Threshold: Plain Campus (no vegetation interference): Matching Difference ≤ 10 bits → Climbing is determined; Mountain Campus (tree obstruction): Difference ≤ 15 bits → Climbing is determined. Time Dynamic Threshold: Daytime (good visibility): Threshold 10 bits; Nighttime (noisy): Threshold 15 bits.

[0082] In this example, after receiving a climbing warning, the mountain campus downloaded a 500KB CNN detection model (including a climbing motion recognition layer) from the main campus and deployed it to edge nodes. This model analyzes millimeter-wave radar waveforms in real time, improving recognition rates. The mountain campus increased the risk factor for the grid in the walled area from 0.6 to 0.9, updated time permissions (for example, increasing the millimeter-wave scanning frequency to 5 seconds between 11:00 PM and 5:00 AM), and marked the area with a "High Climbing Risk in the Near Future" label.

[0083] Based on the embodiments provided in this application, in the implementation of cross-campus collaborative actions, the millimeter-wave radar waveform features of wall climbing are compressed into 128-bit fingerprint codes. After the mountain campus receives the code, it matches similar waveform patterns in the local feature library through Hamming distance. If the matching difference value is lower than the dynamic threshold, the lightweight detection model is downloaded to the edge device, and the risk coefficient of the semantic grid is updated synchronously. This cross-campus collaborative mechanism realizes the rapid sharing of security information and the collaborative update of detection models between different campuses, enabling remote areas such as mountain campuses to obtain abnormal event feature information of other campuses in a timely manner, quickly improving their own detection capabilities for abnormal behaviors such as wall climbing, forming a collaborative defense system within the entire campus, and effectively improving the global responsiveness and safety protection level of the campus security system.

[0084] In one embodiment of the present application, a multi-algorithm collaborative prevention and control system for laboratory hazardous chemical operations includes: Sensor Deployment and Dynamic Task Allocation: Millimeter-wave radars (10cm accuracy) are deployed around laboratory workbenches, scanning the thermal profile of the human body every 200ms. Combined with pressure sensors (±5kg accuracy), they determine whether a person is standing in front of a hazardous chemical storage cabinet. A bionic optimization algorithm dynamically schedules sensors: When the 10:00 PM to 6:00 AM timeframe is detected, cameras near the lab bench automatically switch to 120fps high-speed recording mode, while remote corridor cameras reduce the sampling rate to 15fps to conserve bandwidth. (Node value coefficient calculation: distance < 30 meters + 0.4, time period risk level + 0.3). Embedded Causal Reasoning for Temporal Rules: Rule Encoding: The "two-person authorization" rule is broken down into a spatial feature matrix and temporal gating parameters. Spatial conditions: Laboratory coordinates (X=120-130, Y=40-50, Z=0-2) correspond to the high-weight regions of the 3D convolution kernel, generating a specific response to the individual human sign signal from the millimeter-wave radar. Temporal conditions: An LSTM time-gating unit converts timestamps from 22:00-6:00 into bias vectors, reducing the activation threshold of the rule synapse during this period by 30%. Causal Verification: A differentiable knowledge graph is used to construct a hazardous chemical operation state machine. When the current sensor detects that the power curve of a heating device exceeds the rated value by 15% for five consecutive minutes (for example, a 300W device reaches 345W), the graph automatically traverses the state transition path from "power limit exceeded → temperature anomaly → hazardous chemical volatilization." The probability of each state transition is trained using historical accident data (for example, the transition probability from power limit exceeded to temperature anomaly is 0.85). Non-sensitive feature extraction and dynamic matching: Edge nodes use an improved HOG algorithm to extract the gradient direction histogram of the facial contour, retaining only six geometric feature points (non-sensitive features), such as the jawline and brow bone, and compressing the dimensions to 16. The dynamic time warping algorithm incorporates an attention mechanism: historical profile data of authorized personnel in the laboratory is sorted by time, forming a timestamped feature sequence. When the angle deviation between the feature to be matched and the template exceeds 25°, the time axis is automatically stretched by ±10% for elastic matching, keeping matching time within 80ms. Semantic grid-driven risk evolution simulation: A 30cm-precision laboratory semantic grid is constructed, with each grid bound to attributes such as equipment power thresholds and hazardous chemical storage locations. When a single person's physical sign and power overlimit are detected, the "Hazardous Chemical Operation Risk Map" for the target grid is retrieved to simulate possible accident chains: primary risk: single-person operation without triggering dual-person authorization; secondary risk: power overlimit may cause container overheating (verified by equipment temperature sensor data); ultimate risk: hazardous chemical volatilization reaching a critical concentration (linked to gas sensor data). When simulating smoke diffusion, a particle swarm simulation model uses real-time ventilation system data (e.g., exhaust duct wind speed of 1.2 m / s) as a guiding vector for particle motion, ensuring a simulated diffusion path error of less than 15%.

[0085] Hierarchical privacy protection and device linkage: When access control is triggered, the laboratory video stream uses "dynamic pixel masking" technology. While preserving the motion trajectory of human joints, the pixels in the facial area are randomly shuffled and reassembled. The reassembled video stream can recognize body movements (such as reaching for a reagent) but cannot restore facial features. High-risk equipment power cuts are controlled using a dual-circuit system: first, the main power is disconnected via a relay, and then the backup power is disconnected via hardware logic within 50ms, ensuring the cut-off is completed within 100ms. Operation logs are simultaneously written to the blockchain for evidence storage.

[0086] In another embodiment of the present application, multimodal identity verification and risk response in a dormitory building's privacy zone includes: Real-time gait and Bluetooth feature collection: Millimeter-wave radar collects gait waveforms at 50Hz, and calculates the sliding window variance of the coefficient of variation of stride length (window size 10 steps, sliding interval 2 steps). The normal variance for thick winter shoes ranges from 0.6 to 1.0, but can reach over 1.3 for outsiders due to varying shoe sole materials. Bluetooth devices monitor the 2.4G / 5G frequency bands, recording signal strength fluctuations during device connection sequences. Spectral entropy is calculated using a fast Fourier transform. The entropy for dormitory occupants' device connections ranges from 0.5 to 0.7 (regular intervals), while the entropy for outsiders' devices is often greater than 0.8 (irregular connections). Dual-channel attention mechanism implementation: Gait entropy channel: 1D-CNN is used to extract joint swing frequency features, focusing on the 5-15Hz frequency band (the primary frequency range for human walking). A self-attention mechanism assigns weights to each frequency component, automatically reducing the weight of ankle swing features in winter (due to the influence of thick shoes) and increasing the weight of knee swing features. Device entropy channel: An LSTM network learns the temporal pattern of Bluetooth connection intervals. When it detects three consecutive connection intervals less than 2 minutes (high-frequency connection behavior by non-dormitory members), an abnormal weight gain is triggered. Fusion strategy: A dynamic routing algorithm automatically adjusts the weights of the two channels based on time of day (60% weight for Bluetooth during late night and 50% for gait during daytime). The fused anonymous feature vector has a 24-dimensional dimension. Bayesian model-driven seasonal adaptation: A prior distribution model for winter gait is established. Based on historical data, the joint swing frequency is determined to have a mean μ = 1.0 Hz and a standard deviation σ = 0.2 Hz. If the frequency mean of the real-time data deviates from μ by more than 1.5σ (i.e., < 0.7 Hz or > 1.3 Hz), it is determined to be affected by thick winter shoes. An incremental Bayesian update algorithm is used to update the prior distribution μ and σ each time valid winter gait data is detected, allowing the model to adapt to changes in winter shoe styles across years. (For example, if the sole of a snow boot in 2025 is 1 cm thicker than that in 2024, the model automatically adjusts the threshold.) Multi-level verification and risk assessment: Initial verification: If the Euclidean distance between the fused feature vector and the template library is less than 1.2 (the dormitory membership threshold), the person passes. A secondary verification check (using low-resolution thermal imaging from corridor cameras) is initiated if the distance is between 1.2 and 1.5. If verification fails three times in a row (distance > 1.5), the "outsider intrusion" risk chain in the semantic grid is activated. Historical intrusion records for that dormitory grid are retrieved, and millimeter-wave radars on adjacent floors are linked to expand the scanning range (from 3 meters to 5 meters). A thermal map of the person's gait trajectory is simultaneously sent to the dormitory management terminal.

[0087] The above are only preferred embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. Multi-angle campus panoramic intelligent monitoring system, characterized by: include: A perception module, which is used to acquire perception data based on a sensor network deployed in various areas of the campus and dynamically assign perception tasks based on regional functions; wherein the sensor network includes cameras, millimeter-wave radars, pressure sensors, environmental noise sensors, and Bluetooth devices; A processing module is configured to process the perception data in real time through a reconfigurable hardware node to generate compliance features and anomaly identifiers, wherein the processing method includes: encoding campus rules into neural network differentiable rule synapses using a temporal rule embedded causal inference network; fusing millimeter wave gait entropy features and Bluetooth connection entropy features using an asymmetric modal contrast learning network in a privacy area; A risk deduction module is used to construct a semantic grid based on the compliance features and anomaly identification output by the processing module, associate regional functional attributes, time permissions and historical risk heat maps, and perform global risk causal chain deduction; The response execution module is used to trigger dynamic responses based on the deduction results, including hierarchical privacy protection and cross-campus collaborative actions.

2. The multi-angle campus panoramic intelligent monitoring system according to claim 1 is characterized in that: In the temporal rule embedded causal reasoning network, the generation of neural network differentiable rule synapses includes: The two-person authorization rule in the laboratory is decomposed into spatial conditions and temporal conditions, which are encoded into the weight matrix and bias vector of the neural network respectively. When the millimeter-wave radar detects a single person in the laboratory area and the timestamp is late at night, the convolution kernel operation of the regular synapse is activated; The differentiable knowledge graph is used to verify whether the signal characteristics of the electrical equipment match the hazardous chemical operation mode. If they match, the abnormal identification is generated and the access control locking instruction is triggered.

3. The multi-angle campus panoramic intelligent monitoring system according to claim 2 is characterized in that: After triggering the access lock instruction, the method further includes: The camera in the laboratory area is used to capture facial images of people, and non-sensitive contour features are extracted through edge nodes; Dynamically time-warping the non-sensitive profile features with the authorized personnel model in the compliance feature library; If the matching similarity is lower than the safety threshold, a two-person verification request and a heat map of the person's movement trajectory will be sent to the security terminal; Synchronously cut off the power supply of high-risk devices and record the operation logs to the risk heat map of the semantic grid.

4. The multi-angle campus panoramic intelligent monitoring system according to claim 1 is characterized in that: The operation of the asymmetric modal contrastive learning network in the private area of ​​the dormitory building includes: The millimeter-wave radar is used to collect gait waveforms, and the sliding window variance of the step length variation coefficient is extracted as the millimeter-wave gait entropy feature. Synchronously obtain the Bluetooth device connection sequence and calculate the spectrum entropy value of the signal strength fluctuation as the Bluetooth connection entropy feature; A dual-channel attention mechanism is built at the edge nodes: the gait entropy channel focuses on joint swing frequency, and the device entropy channel strengthens the connection interval characteristics; The dual-channel outputs are fused into an anonymized compliance feature vector, which is only verified if the vector's Euclidean distance is lower than the adaptive threshold of the dormitory member template library; If the verification fails continuously, the corridor camera will be activated to generate a low-resolution thermal image and push it to the dormitory management terminal.

5. The multi-angle campus panoramic intelligent monitoring system according to claim 4 is characterized in that: The template update mechanism of the anonymized compliance feature vector includes: Gait data of thick-shoe gait in winter were collected, and the seasonal offset of joint swing frequency entropy was calculated using a Bayesian probability model. When the seasonal offset exceeds a preset critical value, the gait entropy weight coefficient is reduced and the decision-making proportion of Bluetooth connection entropy is increased; The updated template library is encrypted and distributed to the edge nodes of each campus, and the permission attributes in the semantic grid are updated synchronously.

6. The multi-angle campus panoramic intelligent monitoring system according to claim 1 is characterized in that: The construction of the semantic grid includes: The three-dimensional geographic model of the campus is discretized into cubic grid cells with a side length of 30 cm; Bind the mobile phone signal blocking rule to the teaching building grid during the exam week. If a mobile device radio frequency signal is detected during the effective period of the rule, an abnormal flag will be generated. The thermal weight values ​​of historical fall events are loaded into the stair grid, and the thermal weight values ​​increase in real time with rain and snow weather; A four-dimensional index structure is formed with grid coordinates as the core and time authority, function label, and risk coefficient as attributes.

7. The multi-angle campus panoramic intelligent monitoring system according to claim 6 is characterized in that: The implementation steps of the global risk causal chain deduction in the dormitory fire scenario include: When the current sensor detects that the load on the power strip continues to exceed the limit and the infrared sensor identifies that the density of people exceeds the limit, a composite abnormality mark is marked; Retrieve the escape route topology and flammable material distribution matrix associated with the dormitory grid; The campus event knowledge base was used to simulate the fire evolution chain: a circuit overload caused sparks to ignite the bed curtains, which in turn caused smoke to spread along the ventilation ducts; A dynamic risk vector field is generated based on smoke diffusion particle swarm simulation, and the evacuation path is planned using the tangential component of the dynamic risk vector field as a constraint. Push vector maps with real-time route updates to teachers and students' mobile phones, and overlay a semi-transparent warning layer of smoke-covered areas on corridor displays; Automatically remove privacy blur from escape route cameras based on risk levels and generate a heat map of occupant density to assist in rescue dispatch.

8. The multi-angle campus panoramic intelligent monitoring system according to claim 1 is characterized in that: The dynamic task allocation of the sensor network includes: The node value coefficient is calculated through a bionic optimization algorithm, which is determined based on the target distance, remaining power and network delay; When the density of people gathering on the playground exceeds the set density threshold, the neighboring node group starts the face tracking mode and compresses non-critical background pixels; The remote node group dispatches drones to capture a bird's-eye view of crowd distribution, and the edge node generates a crowd density heat map that is projected onto the security center's circular screen. High-density areas in the crowd density heat map are automatically marked as stampede risk areas and trigger evacuation broadcasts.

9. The multi-angle campus panoramic intelligent monitoring system according to claim 4 is characterized in that: The implementation of the hierarchical privacy protection includes: Deploy a piezoelectric sensor array at the entrance of the psychological counseling room to analyze the periodic characteristics of pedaling duration; When the environmental noise sensor detects an abnormal fundamental frequency of the sobbing sound pattern, it generates an encrypted warning message and sends it to the school doctor terminal; A dynamic pixel reconstruction strategy is used in real time for indoor video streams to transform human forms into abstract geometric blocks while preserving motion vectors.

10. The multi-angle campus panoramic intelligent monitoring system according to claim 5 is characterized in that: The implementation of cross-campus collaborative actions includes: Compress the millimeter-wave radar waveform features of climbing walls into a 128-bit fingerprint code; After receiving the 128-bit fingerprint code, the mountain campus matches similar waveform patterns in the local feature library through Hamming distance; If the matching difference value is lower than the dynamic threshold, the lightweight detection model is downloaded to the edge device and the risk factor of the semantic grid is updated synchronously.

Images