Multi-angle campus panoramic intelligent monitoring system

Abstract

The application discloses a multi-angle campus panoramic intelligent monitoring system and relates to the technical field of security monitoring systems.The system comprises a sensing module, a processing module, a risk deduction module and a response execution module.The sensing module is used for obtaining sensing data based on a sensor network deployed in each region of a campus and dynamically distributing sensing tasks according to regional functions.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 sensing data in real time through a reconfigurable hardware node to generate compliance features and abnormal identifiers.The risk deduction module is used for constructing a semantic grid based on the compliance features and abnormal identifiers output by the processing module, associating regional function attributes, time permissions and historical risk heat maps and performing global risk causal chain deduction.The response execution module is used for triggering dynamic responses according to deduction results, including hierarchical privacy protection and cross-campus collaborative actions.

Description

TECHNICAL FIELD

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

[0002] With the complication and diversification of campus safety incidents, building an intelligent security system covering all scenarios has become a core requirement for educational institutions.

[0003] The existing campus security system usually deploys cameras, sensors, access control systems and other multi-modal sensing devices, trying to achieve risk early warning by integrating video monitoring, environmental monitoring and personnel trajectory data. However, the current technical architecture has significant bottlenecks in the depth fusion of multi-modal data and scene adaptation: traditional multi-modal fusion schemes mostly stay at the level of simple splicing of data, such as time-synchronous display of video pictures and sensor alarm signals, but lack of deep correlation and collaborative analysis mechanism.

[0004] In view of the above problems, no effective solution has been proposed so far. SUMMARY

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

[0006] The present application provides a multi-angle campus panoramic intelligent monitoring system, comprising:

[0007] A perception module is configured to obtain perception data based on a sensor network deployed in each area of the campus, and dynamically allocate perception tasks according to the area function; wherein the sensor network includes cameras, millimeter wave radars, pressure sensors, environmental noise sensors and Bluetooth devices;

[0008] A processing module is configured to process the perception data in real time through a reconfigurable hardware node to generate compliance features and abnormal identifiers, wherein the processing method comprises: using a time sequence rule embedded causal reasoning network to encode the campus rules into a neural network differentiable rule synapse; using an asymmetric modal contrast learning network to fuse the millimeter wave gait entropy features and the Bluetooth connection entropy features in the privacy area;

[0009] A risk deduction module is configured to construct a semantic grid based on the compliance features and abnormal identifiers output by the processing module, associate the area function attributes, time permissions and historical risk heat maps, and perform global risk causal chain deduction;

[0010] A response execution module is configured to trigger dynamic response according to the deduction result, including hierarchical privacy protection and cross-campus collaborative action.

[0011] Further, in the time sequence rule embedded causal reasoning network, the generation of the neural network differentiable rule synapse comprises:

[0012] Decompose the laboratory two-person authorization rule into spatial conditions and time conditions, and encode them as weight matrices and bias vectors of neural networks, respectively;

[0013] When the millimeter wave radar detects a single person sign in the laboratory area and the timestamp is in the late night period, activate the convolution kernel operation of the rule synapse;

[0014] Check whether the signal features of the power consumption equipment match the dangerous chemical operation mode through the differentiable knowledge graph, and if they match, generate the abnormal identifier and trigger the access control locking instruction.

[0015] Further, after triggering the access control locking instruction, the method further comprises:

[0016] Call the laboratory area camera to capture the personnel face image, and extract the non-sensitive contour features through the edge node;

[0017] Match the non-sensitive contour features with the authorized personnel model in the compliance feature library through dynamic time warping;

[0018] If the matching similarity is lower than the safety threshold, send a two-person verification request and personnel motion trajectory heat map to the security terminal;

[0019] Synchronously cut off the power supply of high-risk equipment and record the operation log to the risk heat map of the semantic grid.

[0020] Further, the operation of the asymmetric modal contrast learning network in the dormitory building privacy area includes:

[0021] Collect gait waveform through millimeter wave radar, and extract sliding window variance of step length coefficient of variation as millimeter wave gait entropy feature;

[0022] Synchronously acquire the Bluetooth device connection sequence, and calculate the spectral entropy value of signal strength fluctuation as the Bluetooth connection entropy feature;

[0023] Construct a double-channel attention mechanism on the edge node: the gait entropy channel focuses on the joint swing frequency, and the device entropy channel enhances the connection interval feature;

[0024] Fuse the double-channel output into an anonymous compliance feature vector, and only when the vector Euclidean distance is lower than the adaptive threshold of the dormitory member template library, the verification passes;

[0025] If the continuous verification fails, activate the corridor camera to generate a low-resolution thermal imaging map and push it to the dormitory manager terminal.

[0026] Further, the template updating mechanism of the anonymous compliance feature vector includes:

[0027] Collecting winter thick shoe gait data, calculating seasonal shift of joint swing frequency entropy through Bayesian probability model;

[0028] When the seasonal shift exceeds the preset threshold, the gait entropy weight coefficient is reduced and the decision proportion of Bluetooth connection entropy is increased;

[0029] The updated template library is distributed to each campus edge node after encryption, and the permission attribute in the semantic grid is updated synchronously.

[0030] Further, the construction of the semantic grid comprises:

[0031] The three-dimensional geographic model of the campus is discretized into cubic grid units with a side length of 30 cm;

[0032] Bind the test week mobile phone signal shielding rule to the teaching building grid. If the mobile device radio frequency signal is detected during the rule effective period, an abnormal identifier is generated;

[0033] Load the heat weight value of the historical falling event into the stair grid. The heat weight value increases in real time with the snow weather;

[0034] Form a four-dimensional index structure with grid coordinates as the core and time permissions, function labels and risk coefficients as attributes.

[0035] Further, the implementation steps of the global risk causal chain deduction in the dormitory fire scene comprise:

[0036] When the current sensor detects that the load of the patch panel is continuously over-limit and the infrared sensor identifies that the personnel density is over-standard, a composite abnormal identifier is marked;

[0037] Retrieve the escape passage topology graph and flammable material distribution matrix associated with the dormitory grid;

[0038] Call the campus event knowledge base to simulate the fire evolution chain: circuit overload causes sparks to ignite the bed curtain, which in turn causes smoke to spread along the ventilation duct;

[0039] Based on the smoke diffusion particle swarm simulation, a dynamic risk vector field is generated, and the tangent component of the dynamic risk vector field is used as a constraint condition to plan the evacuation path;

[0040] Push the vector map including real-time path update to the mobile phones of teachers and students, and superimpose a translucent warning layer of the smoke coverage area on the corridor display screen;

[0041] According to the risk level, automatically remove the privacy blur of the escape path camera, and generate a personnel density heat map to assist rescue dispatch.

[0042] Further, the dynamic task allocation of the sensor network comprises:

[0043] The node value coefficient is calculated by a biomimetic optimization algorithm, and the node value coefficient is determined based on target distance, remaining power and network delay;

[0044] When the playground gathering personnel density exceeds the set density threshold, the adjacent node group starts the face tracking mode and compresses the non-key background pixels;

[0045] The remote node group schedules a drone to take a bird's-eye view of the crowd distribution, and the crowd density heat map generated by the edge node is projected to the security center ring screen;

[0046] The high-density area in the crowd density heat map is automatically marked as a stampede risk area and triggers the evacuation broadcast.

[0047] Further, the implementation of hierarchical privacy protection includes:

[0048] A piezoelectric sensor array is deployed at the entrance of the psychological counseling room to analyze the periodic characteristics of the duration of the stampede;

[0049] When the environmental noise sensor identifies the abnormal fundamental frequency of the sipping soundprint, it generates an encrypted warning message and sends it to the school doctor terminal;

[0050] The indoor video stream enables a dynamic pixel reorganization strategy in real time, converts human body shapes into abstract geometric color blocks and retains motion vectors.

[0051] Further, the implementation of cross-campus collaborative action includes:

[0052] The climbing wall's millimeter wave radar waveform features are compressed into 128-bit fingerprint encoding;

[0053] After receiving the 128-bit fingerprint encoding, the mountainous school receives the 128-bit fingerprint encoding, and matches the similar waveform mode of the local feature library through Hamming distance;

[0054] If the matching difference value is lower than the dynamic threshold, download the lightweight detection model to the edge device, and update the risk coefficient of the semantic grid synchronously.

[0055] Based on the embodiments provided in the present application, the perception module obtains perception data based on the sensor network deployed in each area of the campus and dynamically allocates perception tasks according to the area function, 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 diversified security needs of each area of the campus. The processing module can realize rapid processing of perception data by using reconfigurable hardware nodes to process perception data in real time, and timely generate compliance features and abnormal identifiers; using the time sequence rule embedded causal reasoning network, the campus rules are encoded into the neural network differentiable rule synapse, so that the campus rules are deeply integrated into the neural network processing process, realizing the depth analysis and reasoning of the compliance of the perception data, rather than simply splicing at the data level; in the privacy area, the asymmetric modal contrast learning network is used to fuse the millimeter wave gait entropy feature and the Bluetooth connection entropy feature, which not only realizes the deep fusion analysis of the multi-modal data in the privacy area, but also avoids directly using video and other methods that may leak privacy, thereby improving the ability of abnormal detection in the privacy area while ensuring privacy. The risk deduction module constructs a semantic grid based on the output of the processing module, associates the area function attribute, time authority and historical risk heat map, and executes global risk causal chain deduction, which can perform global correlation analysis of risks from multiple dimensions such as area function, time authority, historical risk, etc., mine the causal relationship between risks, realize more comprehensive and in-depth deduction of campus risks, and change the situation that the prior art lacks depth correlation and collaborative analysis. The response execution module triggers dynamic response according to the deduction result, including hierarchical privacy protection and cross-campus collaborative action, which can realize hierarchical privacy protection according to the risk situation, and better balance safety and privacy when responding to risks; the cross-campus collaborative action can make each campus respond to security incidents in a coordinated manner, improving the coordination and effectiveness of the overall security response of the campus. BRIEF DESCRIPTION OF DRAWINGS

[0056] The drawings described herein are used to provide further understanding of the embodiments of the present application, and constitute a part of the present application. The schematic embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings:

[0057] Figure 1 It is a structure diagram of an optional multi-angle campus panoramic intelligent monitoring system according to an embodiment of the present application;

[0058] Figure 2 It is a flowchart of generation of an optional neural network differentiable rule synapse according to an embodiment of the present application;

[0059] Figure 3 It is a flowchart of operation of an optional asymmetric modal contrast learning network in a privacy area of a dormitory building according to an embodiment of the present application.

[0060] The objectives, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0061] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0062] Optionally, as shown in the present application, a multi-angle campus panoramic intelligent monitoring system is provided, comprising: Figure 1

[0063] The perception module 101 is configured to acquire perception data based on a sensor network deployed in each area of the campus, and dynamically allocate perception tasks according to area functions; wherein the sensor network comprises a camera, a millimeter wave radar, a pressure sensor, an environmental noise sensor and a Bluetooth device.

[0064] In the embodiment, the campus is divided into functionally differentiated area spaces according to the use of the campus buildings, and the sensors dynamically adjust the perception tasks according to the area functions. For example, the laboratory area (3rd floor of the chemical building): millimeter wave radar + current sensor are deployed to focus on monitoring the compliance of hazardous chemical operations; the privacy area of the dormitory building (4th floor corridor of No. 1 building): millimeter wave radar + Bluetooth device are combined to consider identity verification and privacy protection; the playground open area: cameras + environmental noise sensors are arranged for crowd density monitoring and sudden noise warning.

[0065] The processing module 102 is configured to process the perception data in real time through a reconfigurable hardware node to generate compliance features and abnormality identification, wherein the processing method comprises: using a time sequence rule embedded causal reasoning network to encode the campus rules into a neural network differentiable rule synapse; using an asymmetric modal contrast learning network to fuse the millimeter wave gait entropy feature and the Bluetooth connection entropy feature in the privacy area.

[0066] In the embodiment, the reconfigurable hardware node is a hardware module based on FPGA / edge computing unit. For example, when the teaching building enters the examination week, the edge node automatically switches the computing resources from face recognition to mobile phone signal spectrum analysis, and realizes the dynamic loading of the radio frequency signal detection module through the reconfiguration of the hardware logic.

[0067] ​The timing rule embedded causal inference network disassembles campus safety rules (such as double authorization and time period permission) into spatio-temporal conditions that can be calculated by a neural network, and realizes rule triggering and causal inference through a differentiable synapse. For example, the library closing period (22:00-6:00) rule is encoded as: the time condition corresponds to the neural network bias vector (the bias value after 22 o'clock is +1), and the space condition corresponds to the entrance area weight matrix. When the infrared sensor detects that a single person enters, the abnormality identifier is output through synapse convolution operation.

[0068] The asymmetric modal contrast learning network adopts an asymmetric strategy to fuse different sensor features in the privacy area: the millimeter wave radar obtains physical features (gait), and the Bluetooth device obtains behavior features (device connection rules), and realizes anonymous identity verification through contrast learning. For example, at the entrance of the psychological counseling room, the pressure waveform (gait) collected by the piezoelectric sensor array and the change sequence of the visitor's Bluetooth MAC address are processed by the contrast learning network to generate a 128-dimensional anonymous feature vector, which only retains the time correlation features and removes the biological identification information

[0069] The risk deduction module 103 is used for constructing a semantic grid based on the compliance features and abnormality identifiers output by the processing module, associating the area function attributes, time permissions and historical risk heat maps, and performing global risk causal chain deduction;

[0070] The compliance features include feature vectors extracted from sensor data that comply with safety rules, for example, in the laboratory double authorization scenario, the millimeter wave radar simultaneously detects the spatial distribution features of 2 signs.

[0071] The response execution module 104 is used for triggering dynamic response according to the deduction result, including hierarchical privacy protection and cross-campus collaborative action.

[0072] Based on the embodiments provided in the present application, the perception module obtains perception data based on the sensor network deployed in each area of the campus and dynamically allocates perception tasks according to the area function, can flexibly configure perception resources according to the function requirements of different areas, improve the pertinence and effectiveness of perception, avoid resource waste, and better adapt to the diversified security needs of each area of the campus. The processing module processes the perception data in real time through the reconfigurable hardware node, can realize fast processing of the perception data, and timely generate compliance features and abnormal identifiers; the time sequence rule embedded causal reasoning network encodes the campus rules into the neural network differentiable rule synapse, so that the campus rules are deeply integrated into the neural network processing process, realizing the depth analysis and reasoning of the compliance of the perception data, rather than simply splicing at the data level; the asymmetric modal contrast learning network is used in the privacy area to fuse the millimeter wave gait entropy feature and the Bluetooth connection entropy feature, which not only realizes the deep fusion analysis of the multi-modal data in the privacy area, but also avoids directly using video and other methods that may leak privacy, improving the ability of abnormal detection in the privacy area under the premise of protecting privacy. The risk deduction module constructs a semantic grid based on the output of the processing module, associates the area function attribute, time authority and historical risk heat map, and executes global risk causal chain deduction, which can perform global correlation analysis on risks from multiple dimensions such as area function, time authority and historical risk, mine the causal relationship between risks, realize more comprehensive and in-depth deduction of campus risks, and change the situation that the prior art lacks depth correlation and collaborative analysis. The response execution module triggers dynamic response according to the deduction result, including hierarchical privacy protection and cross-campus collaborative action, which can realize hierarchical privacy protection according to the risk situation, and better balance safety and privacy when responding to risks; the cross-campus collaborative action can make each campus respond to security incidents in a coordinated manner, improving the coordination and effectiveness of the overall security response of the campus.

[0073] Further, as shown in Figure 2 the generation of the neural network differentiable rule synapse in the time sequence rule embedded causal reasoning network includes:

[0074] S201, the laboratory double-person authorization rule is disassembled into space conditions and time conditions, and is respectively coded into a weight matrix and a bias vector of a neural network;

[0075] In the present embodiment, the rule disassembly logic includes: space condition: laboratory area coordinates (such as X=123, Y=456, Z=0) correspond to the spatial feature extraction kernel of the neural network weight matrix; time condition: the late night period (23:00-5:00) is converted into the timestamp weighting coefficient of the bias vector. Exemplarily, when the millimeter wave radar detects a single person sign 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 verification, and the two conditions are met at the same time. The power equipment signal analysis is started.

[0076] S202, when the millimeter wave radar detects a single human sign in the laboratory area and the timestamp is in the late night period, activate the convolution kernel operation of the rule synapse;

[0077] S203, check whether the power equipment signal feature matches the dangerous chemical operation mode through the differentiable knowledge graph, and if it matches, generate an abnormal identification and trigger the access control locking instruction.

[0078] In this embodiment, the differentiable knowledge graph verification includes: constructing the dangerous chemical operation mode (such as ether distillation requiring continuous heating of 300W) as a knowledge graph node, and verifying the current feature through a differentiable graph neural network. Dangerous chemicals: flammable and explosive: ether (anti-static grounding is required during operation, and the heating power is less than or equal to 300W) toxic: potassium cyanide (two people and two locks are managed, and two RFID permissions need to be activated at the same time for taking and using).

[0079] In a specific implementation, the timing rule synapse activation function is:

[0080]

[0081] wherein, is a spatial condition factor (laboratory area single human sign detection result, 1 for detecting a single person, otherwise 0); is a time condition factor (late night period encoding, 1 for 23:00-05:00 period, otherwise 0); is a spatial condition weight matrix (according to the double authorization rule, the single human sign risk weight is 2.0); is a time condition bias vector (the risk weight of the late night period is 1.5); is a knowledge graph embedding vector of the power equipment signal feature (dimension 128, generated by a graph convolution network); is a reference graph vector of the compliant dangerous chemical operation mode (dimension 128); is a knowledge graph verification coefficient (dynamically valued according to the dangerous chemical grade, 1.2 for toxic storage area, and 0.8 for ordinary laboratory); is a Sigmoid activation function (mapping the input to an activation probability in the [0, 1] interval); is the 2-norm of the vector.

[0082] Based on the above formula, through the linear combination of the spatial condition (single human sign) and the time condition (late night period), combined with the knowledge graph verification item (the Euclidean distance normalized value of the device signal and the compliant mode), the traditional Boolean rule is converted into a differentiable neural network operation, solving the problem that static rules cannot adapt to complex scenarios.

[0083] Based on the embodiments provided in the present application, by decomposing the laboratory two-person authorization rule into space conditions and time conditions and encoding it into the weight matrix and bias vector of the neural network, the system can accurately identify the abnormal situation of single person sign in the laboratory area at night. When the millimeter wave radar detects a single person sign and the timestamp is in the late night period, the convolution kernel operation of the rule synapse is activated, combined with the matching verification of the signal characteristics of the electrical equipment and the operation mode of the hazardous chemicals in the differentiable knowledge graph, the abnormal identification can be generated in a short time and the access control locking instruction is triggered, which realizes the accurate detection and real-time control of the laboratory illegal operation, effectively improves the safety protection ability of the laboratory high-risk scene, and avoids safety accidents caused by single person operation of hazardous chemicals.

[0084] Further, after triggering the access control locking instruction, the method further comprises:

[0085] Calling the laboratory area camera to capture the personnel face image, extracting the non-sensitive contour features through the edge node;

[0086] In the present embodiment, the feature extraction mechanism of the non-sensitive contour feature recognition includes: the edge node extracts the geometric parameters (16 non-biological feature points such as jaw line curvature and nose bridge slope) of the face contour through CNN, and removes sensitive information such as texture and skin color. For example, after the laboratory access control is locked, the camera captures a person wearing a mask, and the system only extracts the contour curve from the eyebrows to the jaw (ignoring the mask covered area), and generates an 8-dimensional contour vector (such as [1.2, 0.8, 1.5,...])

[0087] Dynamically time warping matching the non-sensitive contour features with the authorized personnel models in the compliance feature library;

[0088] Among them, the compliance feature library stores the non-sensitive contour templates of authorized personnel, for example, the contour vector set {V1, V2, V3} of laboratory administrator A corresponds to different angles.

[0089] If the matching similarity is lower than the safety threshold, a two-person verification request and a personnel motion trajectory heat map are sent to the security terminal;

[0090] Among them, the safety threshold example: normal working period threshold: 0.7 (Euclidean distance <0.7 is determined as an authorized personnel); late night abnormal period threshold: 0.9 (to improve the verification accuracy and reduce the risk of false release).

[0091] 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 embodiments provided in the present application, after triggering the access lock instruction, the laboratory area camera is called to capture the personnel face image and the non-sensitive contour features are extracted through the edge node, which not only ensures the privacy protection of the personnel face information, but also obtains the key features for identity verification. The non-sensitive contour features are matched with the authorized personnel model in the compliance feature library through dynamic time warping, and if the matching similarity is lower than the safety threshold, a two-person verification request and a personnel motion trajectory heat map are quickly sent to the security terminal, and 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 accurate verification information and personnel trajectory for security personnel, further strengthens the subsequent processing capability of abnormal conditions in the laboratory, and forms a complete safety closed loop from detection, response to verification.

[0092] Further, as shown in Figure 3 the running of the asymmetric modal contrast learning network in the dormitory building privacy area includes:

[0093] S301, collect gait waveform through millimeter wave radar, and extract sliding window variance of step length variation coefficient as millimeter wave gait entropy feature;

[0094] In the present embodiment, the calculation of the sliding window variance of the step length variation coefficient includes calculating the standard deviation of 10 consecutive steps, and then calculating the variance of the 30-second sliding window (for example, normal gait variance <0.5, and variance >0.8 when wearing thick shoes). Exemplarily, in winter, a student wearing snow boots walks in the dormitory building, and the step length variation coefficient variance rises from 0.4 to 1.1, triggering the gait entropy feature deviation warning.

[0095] S302, synchronously acquire the Bluetooth device connection sequence, and calculate the spectral entropy value of signal strength fluctuation as the Bluetooth connection entropy feature;

[0096] S303, construct a double-channel attention mechanism on the edge node: the gait entropy channel focuses on the joint swing frequency, and the device entropy channel strengthens the connection interval feature; S304, fuse the double-channel output into an anonymized compliance feature vector, and only when the vector Euclidean distance is lower than the adaptive threshold of the dormitory member template library, the verification passes;

[0097] The adaptive threshold of the dormitory member template library includes but is not limited to 1.2, 1.5, etc.

[0098] In this embodiment, the gait entropy channel outputs a joint swing frequency feature vector (16 dimensions), and the device entropy channel outputs a Bluetooth connection interval feature vector (8 dimensions). The two are fused into a 24-dimensional anonymous vector through an attention mechanism. When the Euclidean distance between the fused vector and the dormitory member template library is <1.2, it is determined to be a legal member, otherwise the corridor camera is activated to generate a heat map (for example, the Bluetooth connection interval of an outsider is irregular, resulting in a distance calculation value of 2.5)

[0099] S305, if the continuous verification fails, activate the corridor camera to generate a low-resolution heat map and push it to the dormitory manager terminal.

[0100] In a specific embodiment, the asymmetric modal contrast learning fusion function is:

[0101]

[0102] wherein, is the millimeter wave gait entropy feature vector (including the sliding window variance of the step variation coefficient); is the Bluetooth connection entropy feature vector (including the spectral entropy value of signal strength fluctuation); is the gait entropy channel attention mechanism (focusing on 0.8-1.2Hz joint swing frequency features); is the device entropy channel attention mechanism (strengthening 10-30 second Bluetooth connection interval features); is the gait entropy weight coefficient (dynamic range [0.2, 0.8], automatically reduced to 0.3 in winter).

[0103] Based on the above formula, the key features of gait entropy (joint swing frequency) and device entropy (connection interval stability) are extracted through a double-channel attention mechanism, and the temperature-sensitive weight coefficient α is used to dynamically adjust the modal fusion ratio: when the winter temperature is lower than the reference value, α automatically decreases (such as to 0.3), reducing the impact of thick shoe gait changes on verification, while increasing the decision-making proportion of Bluetooth connection entropy. Compared with fixed weight fusion (such as the traditional method of gait entropy accounting for 50%), this mechanism reduces the verification failure rate in winter scenarios, and protects user privacy through anonymous feature vectors (fusion dimension 128) (only abstract features are transmitted, not raw signals). In addition, the attention mechanism focuses on specific frequency / interval features (such as gait entropy focusing on 0.8-1.2Hz joint swing, and device entropy strengthening connection interval of more than 10 seconds), effectively distinguishing the behavior patterns of dormitory members and outsiders, and improving the security accuracy of private areas.

[0104] Based on the embodiments provided in the present application, in the privacy area of the dormitory building, the gait waveform is collected by the millimeter wave radar to extract the millimeter wave gait entropy feature, the Bluetooth device connection sequence is synchronously acquired to calculate the Bluetooth connection entropy feature, and a double-channel attention mechanism is constructed on the edge node to focus on the joint swing frequency and strengthen the connection interval feature respectively, and the double-channel output is fused into an anonymized compliance feature vector. This multi-modal fusion method can accurately identify the identity of the person while protecting the privacy of the dormitory members. Only when the vector Euclidean distance is lower than the adaptive threshold of the dormitory member template library, the verification passes, and if the continuous verification fails, the corridor camera can generate a low-resolution thermal imaging map and push it to the dormitory terminal. This scheme effectively balances the needs of privacy protection and security verification, improves the accuracy and security of personnel identity verification in the dormitory area, and reduces the risk of illegal entry of outsiders into the dormitory.

[0105] Further, the template updating mechanism of the anonymized compliance feature vector includes:

[0106] The winter thick shoe gait data is collected, the seasonal offset of the joint swing frequency entropy is calculated by the Bayesian probability model, and when the seasonal offset exceeds the preset threshold, the gait entropy weight coefficient is reduced and the decision proportion of the Bluetooth connection entropy is increased;

[0107] In the present embodiment, the weight dynamic adjustment mechanism includes: when the winter offset is greater than 15%, the gait entropy weight is reduced from 60% to 40%, and the Bluetooth connection entropy is increased from 30% to 50%, and the thick shoe gait changes greatly and depends on the device connection rule. Threshold examples: winter mode: 15% (triggered when the temperature is less than 5°C); rainy season mode: 10% (triggered when the air humidity is greater than 80%).

[0108] The updated template library is distributed to each campus edge node after encryption, and the permission attribute in the semantic grid is updated synchronously.

[0109] Based on the embodiments provided in the present application, for the template updating mechanism of the anonymized compliance feature vector, the winter thick shoe gait data is collected, the seasonal offset of the joint swing frequency entropy is calculated by the Bayesian probability model, when the offset exceeds the preset threshold, the weight coefficients of the gait entropy and the Bluetooth connection entropy are automatically adjusted, and the updated template library is distributed to each campus edge node after encryption, and the permission attribute in the semantic grid is updated synchronously. This mechanism can adaptively cope with the influence of seasonal changes on personnel gait features, ensuring that the accuracy of dormitory member identity verification is not significantly affected in different seasonal environments. By dynamically adjusting the weight and distributed template updating, the system can maintain the ability to accurately identify the identity of the personnel in the dormitory area for a long time, avoid verification errors caused by seasonal factors, and improve the environmental adaptability and stability of the system.

[0110] Further, the construction of the semantic grid includes:

[0111] The campus three-dimensional geographic model is discretized into cubic grid units with a side length of 30 cm; the examination week mobile phone signal shielding rule is bound to the teaching building grid, and if the mobile device radio frequency signal is detected during the effective period of the rule, an abnormality identifier is generated;

[0112] In this embodiment, the campus buildings, roads, greenery, and other entity spaces are converted into a campus three-dimensional geographic model, which is discretized into cubic grid units with a precision of 30 cm for spatial positioning and rule binding. For example, the main building of the library is modeled as a 100x80x20 (units: grid) three-dimensional structure, where the bookshelf area grid is bound to the "RFID book positioning rule", and the reading area grid is loaded with the "personnel density monitoring task".

[0113] The rule effective period includes a shielding period set according to the course schedule, usually 8:00-12:00 and 14:00-17:00 on exam days. For example, on May 30, 2025 (the third day of the exam week), all grids on the third floor of the teaching building activate the shielding rule during 8:30-11:30, and if the mobile phone signal strength is detected to be >-70 dBm during this period, an abnormality identifier is generated.

[0114] The stair grid is loaded with a heat weight value of historical fall incidents, and the heat weight value is increased in real time with snow and rain weather;

[0115] Based on historical accident data, the stair grid is assigned a heat weight value. For example, the east stair of the administrative building has had 2 falls in the winter of 2024, and the initial weight is set to 0.7; during a moderate rain, the weight x 1.8 → 1.26, triggering a "non-slip mat installation" reminder. The dormitory stair has a weight x 1.5 (due to insufficient light) after the evening study ends (22:00-23:00), and if it is raining at the same time, it is x 1.8, the maximum weight = initial 0.5 x 1.5 x 1.8 = 1.35.

[0116] A four-dimensional index structure is formed with grid coordinates as the core and time authority, function label, and risk coefficient as attributes.

[0117] Based on the embodiments provided in the present application, when constructing the semantic grid, the campus three-dimensional geographic model is discretized into cubic grid units with a side length of 30 cm, the grid of different functional areas is bound with corresponding rules and historical risk heat weight values, and a four-dimensional index structure with grid coordinates as the core is formed. This fine grid division and attribute binding enables the system to accurately associate campus rules (such as mobile phone signal shielding rules during exam week) and historical risk data (such as real-time increase of stair fall event heat weight value with snow weather) to specific spatial locations and time dimensions. During real-time monitoring, it can be judged whether to generate an abnormal identifier according to the attribute information of the grid unit, providing accurate and efficient basic data support for global risk causal chain deduction, and realizing fine modeling and dynamic management of risks in each area of the campus.

[0118] Further, the implementation steps of global risk causal chain deduction in the dormitory fire scenario include:

[0119] When the current sensor detects that the load of the power strip is continuously over-limit and the infrared sensor identifies that the personnel density is over-standard, a composite abnormal identifier is marked;

[0120] In this embodiment, the continuous over-limit of the power strip load includes: ordinary dormitory power strip: continuous > 10A (such as simultaneous use of hair dryer 1600W + hand warmer 500W, current ≈ (1600+500) / 220 ≈ 9.5A, close to over-limit). Laboratory high-risk equipment power strip: continuous > 5A (such as illegal use of electric cooker 1200W, current ≈ 1200 / 220 ≈ 5.45A, triggering over-limit).

[0121] The over-standard personnel density includes: standard dormitory (20㎡): > 8 people (per capita area < 2.5㎡, such as 6 people living in 10 people in a room); corridor (2m wide): > 3 people / ㎡ (such as 60 people gathering in a 10m corridor, density 60 / (10x2)=3 people / ㎡).

[0122] The composite abnormality needs to meet ≥2 independent abnormal conditions, and the ordinary abnormality is triggered by a single condition. For example, composite abnormality: power strip current 12A (over-limit) + 10 people in the dormitory (density over-standard) → generate 0x1101 identifier. Ordinary abnormality: single current 12A → generate 0x1001 identifier, single 10 people → generate 0x0101 identifier.

[0123] Retrieve the escape passage topology graph and flammable material distribution matrix associated with the dormitory grid;

[0124] Exemplarily, the escape passage topology map associated with the dormitory grid shows that the nearest safety exit is 15 meters to the east, and the standby passage needs to be detoured for 20 meters through the stairwell; the flammable material distribution matrix marks: bed curtain (burning speed 0.2 m / s), bedding (heat value 15 MJ / kg), books (burning smoke toxicity coefficient 0.8).

[0125] The campus event knowledge base is called to simulate the fire evolution chain: circuit overload causes sparks to ignite the bed curtain, which in turn causes smoke to spread along the ventilation duct;

[0126] Exemplarily, the fire evolution chain simulation includes: circuit overload (15A for 5 minutes) → sparks generated by insulation layer melting (temperature > 300°C) → ignition of bed curtain (0.5 meters from the power strip) → smoke spreads along the ventilation duct (diameter 20 cm) at a speed of 0.8 m / s to the corridor.

[0127] Based on the smoke diffusion particle swarm simulation, a dynamic risk vector field is generated, and the tangent component of the dynamic risk vector field is used as a constraint condition to plan an evacuation path;

[0128] Exemplarily, the smoke diffusion particle swarm simulation includes: simulate the dormitory fire, release 1000 smoke particles, each particle diffuses according to the Brownian motion law, adjust the direction according to the wind speed of 1.2 m / s when encountering a ventilation opening, and generate a dynamic risk vector field (red area represents particle concentration > 500 particles / m3).

[0129] Push the vector map including real-time path updates to the teachers' and students' mobile phones, and superimpose a semi-transparent warning layer of the smoke coverage area on the corridor display screen;

[0130] According to the risk level, automatically remove the privacy blur of the escape path camera, and generate a personnel density heat map to assist in rescue dispatch.

[0131] In this embodiment, the risk level is divided into: first level (low): smoke concentration < 200 ppm, temperature < 60°C; second level (medium): 200-500 ppm, 60-80°C; third level (high): > 500 ppm, > 80°C. The removal conditions include: when the risk level reaches level two and above, automatically remove the privacy blur of the escape passage camera (such as changing the face mosaic to outline display).

[0132] Personnel density heat map: visually display the area crowd density with color gradient, generated based on infrared sensor or camera data. For example, in the dormitory fire scene, the heat map shows that the corridor corner is red (density > 4 people / ㎡), and the safety exit is green (density < 1.5 people / ㎡), which assists in rescue dispatch.

[0133] In a specific embodiment, the dynamic risk vector field generation model is:

[0134]

[0135] Based on the above formula, is the risk vector field at time t (unit: person / (s·m³), pointing to the direction of risk reduction) at three-dimensional space coordinates (x, y, z); is the smoke diffusion speed (unit: m / s, initial value 0.3 m / s, increase by 0.1 m / s for every 5°C increase in temperature); is the personnel density function (unit: person / m³, generated by the UAV heat map and edge node calculation); is the Euclidean distance from the current position to the nearest escape route (unit: m, calculated based on the three-dimensional geographical model of the campus); is the channel resistance coefficient (dynamically adjusted according to the historical fall event heat weight, 0.5 for stairs and 1.2 for corridors); is the escape route length (unit: m, pre-stored in the four-dimensional index structure of the semantic grid); is the gradient operator (derivative of spatial coordinates, unit: 1 / m).

[0136] Based on the above formula, the risk flow per unit time per unit area is quantified by the product of the smoke diffusion speed (positively correlated with temperature, simulating the thermal buoyancy effect) and the personnel density; the exponential term describes the attenuation effect of escape route distance and resistance on risk (e.g., stairs have a high historical fall risk, 0.5, increasing the attractiveness of escape routes). The negative gradient direction ensures that the vector field points to the place with the smallest risk, and in combination with the smoke diffusion path generated by the particle swarm simulation, it realizes intelligent evacuation path planning. Dynamically avoid the smoke coverage area (e.g., the vector field sets a repulsive component in the area where the smoke concentration >500ppm). In addition, by removing the privacy blur of the escape route cameras, the personnel density heat map (based on the tangential component of the risk vector field) generated synchronously can assist the security terminal in adjusting the rescue strategy in real time, improving the evacuation efficiency, and solving the problems of "risk deduction lag" and "rough path planning" of traditional security systems.

[0137] Based on the embodiments provided in the present application, in a dormitory fire scene, when the current sensor detects that the load of the patch panel continuously exceeds the limit and the infrared sensor identifies that the personnel density exceeds the limit, a composite abnormal identifier is marked, then the escape passage topological graph and the combustible 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 an evacuation path is planned with the tangential component of the dynamic risk vector field as the constraint condition. The scheme can multi-modal fusion detect abnormal signals in the early stage of fire, generate a scientific evacuation path, push a vector map containing real-time path updates to the mobile phones of teachers and students, superimpose a translucent warning layer of the smoke coverage area on the corridor display screen, automatically remove the privacy blur of the escape path camera according to the risk level, and generate a personnel density heat map to assist rescue dispatch. This series of operations realizes early warning, dynamic risk assessment and efficient evacuation guidance for dormitory fires, provides strong support for the safe escape and rescue dispatch of teachers and students in a fire scene, and maximizes the reduction of losses caused by fires.

[0138] Further, the dynamic task allocation of the sensor network comprises:

[0139] The node value coefficient is calculated by a bionic optimization algorithm, and the node value coefficient is determined based on the target distance, the remaining power and the network delay;

[0140] In the embodiment, the bionic optimization algorithm includes simulating biological behaviors such as ant colony foraging and bee colony division of labor to dynamically optimize the sensor task allocation strategy. For example, when a playground meeting occurs, the algorithm preferentially selects a node group with the minimum path loss of the “sensing node-personnel center” path, similar to the ant colony searching for the shortest path, while avoiding overwork of nodes with low power.

[0141] Illustratively, the node value coefficient and the target distance: distance to target < 50 meters: value + 0.4 (such as a camera 30 meters from the center of the playground); 50-100 meters: value + 0.2 (such as a camera 80 meters from the edge of the playground); > 100 meters: value - 0.3 (such as a camera 150 meters from the playground).

[0142] The node value coefficient and the 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).

[0143] The node value coefficient and the 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).

[0144] When the playground gathering personnel density exceeds the set density threshold, the adjacent node group starts the face tracking mode and compresses the non-key background pixels;

[0145] For example, the density threshold is set: yellow warning: > 2 people / ㎡ (gather at the playground entrance); red warning: > 4 people / ㎡ (crowded in front of the stage). When the playground gathers, the adjacent node group includes: fixed cameras within 50 meters around the playground (30 meters away from the crowd center); millimeter wave radar installed on the stand rail (20 meters away from the crowd).

[0146] Non-key background pixels: When tracking the crowd, the pixels of the background area such as trees and fences in the video are compressed into 16x16 pixel blocks (original resolution 1920x1080), and only the high resolution of the area where the person is located (such as 512x512) is reserved.

[0147] The remote node group dispatches a drone to take an aerial view of the crowd distribution, and the edge node generates a crowd density heat map and projects it to the security center ring screen.

[0148] For example, when the playground gathers, the remote node group includes: long-focus cameras deployed on the top of the teaching building (200 meters away from the playground); and dispatched drones (300 meters high above the playground).

[0149] The high-density area in the crowd density heat map is automatically marked as a stampede risk area and triggers the evacuation broadcast.

[0150] Wherein, if the density of a certain area is > 2.5 people / ㎡ or > 5 people / ㎡, the area is a high-density area.

[0151] Based on the embodiments provided in the present application, for dynamic task allocation of sensor networks, the node value coefficient is calculated by a bionic optimization algorithm, and the sensing task is reasonably allocated based on factors such as target distance, remaining power, and network delay. When the playground gathering personnel density exceeds the set threshold, the adjacent node group can start the face tracking mode and compress the non-key background pixels, the remote node group dispatches a drone to take an aerial view of the crowd distribution, and the edge node generates a crowd density heat map and projects it to the security center ring screen. The high-density area in the heat map is automatically marked as a stampede risk area and triggers the evacuation broadcast. This dynamic task allocation mechanism realizes efficient use of sensor network resources, can monitor the crowd distribution and density change in personnel-intensive scenes such as playground gatherings in real time and comprehensively, quickly identifies the stampede risk area and triggers the evacuation measure in time, effectively improves the safety protection capability of large-scale gathering activities in the campus, and reduces the possibility of stampede accidents.

[0152] Further, the implementation of hierarchical privacy protection includes:

[0153] A piezoelectric sensor array is deployed at the entrance of the psychological counseling room to analyze the periodic characteristics of the stampede duration.

[0154] For example, a 3x3 piezoelectric sensor array is embedded in the ground at the entrance of the psychological counseling room, each sensor is spaced 20 cm apart, arranged in a matrix, which can locate the stepping position and pressure distribution. The analysis of periodic characteristics includes:

[0155] The stepping interval is >1.5 seconds (such as "step-step" interval 2 seconds) when entering normally, and the time interval is <0.8 seconds (such as "step-step-step" interval 0.5 seconds) in emergency situations. The system analyzes the frequency periodicity of the stepping signal by FFT.

[0156] When the environmental noise sensor identifies the abnormal base frequency of the sobbing soundprint, it generates an encrypted warning information and sends it to the school doctor terminal;

[0157] For example, normal conversation: female base frequency 180-250Hz, male 100-150Hz; Sobbing abnormality: female >280Hz (sharp crying), male >180Hz (suppressed crying).

[0158] The indoor video stream enables real-time dynamic pixel reorganization strategy to convert human body shape into abstract geometric color blocks and retain motion vectors.

[0159] The dynamic pixel reorganization strategy includes: converting indoor personnel into abstract geometric color blocks: standing posture → blue cylindrical body (1.7m high, 0.3m in diameter); waving action → adding red fan-shaped dynamic vector (radius 0.5m, direction changes with action); walking trajectory → green dotted line retains the motion path (such as a straight line from point A to point B).

[0160] Based on the embodiments provided in the present 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 stepping duration, combined with the identification of the abnormal base frequency of the sobbing soundprint by the environmental noise sensor, which can accurately detect possible abnormal situations in the room, generate encrypted warning information and send it to the school doctor terminal, ensuring the security and timeliness of information transmission. At the same time, the indoor video stream enables real-time dynamic pixel reorganization strategy to convert human body shape into abstract geometric color blocks and retain motion vectors, which can monitor the motion state of indoor personnel in real time without revealing sensitive information such as personnel's face, balancing the demand for privacy protection and security monitoring. The scheme 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 warning and intervention in abnormal situations, improving the safety management level of privacy areas in the campus.

[0161] Further, the implementation of cross-campus collaborative action includes:

[0162] The millimeter wave radar waveform features of climbing the fence are compressed into 128-bit fingerprint encoding;

[0163] Exemplarily, for the millimeter wave radar waveform feature, normal climbing: the waveform presents high frequency jitter (frequency 10-20 Hz), large amplitude change (peak > 0.5V), and duration of 5-10 seconds; animal climbing (such as cat): the waveform frequency is lower (5-8 Hz), the amplitude change is small (peak < 0.3V), and the duration is irregular.

[0164] After receiving the 128-bit fingerprint code, the mountain school receives the similar waveform mode in the local feature library through Hamming distance matching;

[0165] Exemplarily, the climbing code sent by the main school is 10110101... (128 bits), and the "person climbing" code in the local feature library of the mountain school is 10100101..., and the Hamming distance between the two is 2 (only the 3rd and 6th bits are different), which is determined to be highly similar.

[0166] If the matching difference value is lower than the dynamic threshold, a lightweight detection model is downloaded to the edge device, and the risk coefficient of the semantic grid is updated synchronously.

[0167] Exemplarily, the dynamic threshold includes: school difference threshold: plain school (no vegetation interference): matching difference value ≤ 10 bits → determined as climbing; mountain school (tree shelter): difference value ≤ 15 bits → determined as climbing. Time period dynamic threshold: day (good visibility): threshold 10 bits; night (much noise): threshold 15 bits.

[0168] In this embodiment, after receiving the climbing warning, the mountain school downloads a 500KB CNN detection model (containing a climbing action recognition layer) from the main school, deploys it to the edge node, and the model can analyze the millimeter wave radar waveform in real time, and the recognition rate is improved. The risk coefficient of the fence area grid of the mountain school is increased from 0.6 to 0.9, the time permission attribute is updated (such as increasing the millimeter wave scanning frequency to 5 seconds / time at night 23:00-5:00), and the "recent climbing risk high" label is marked.

[0169] Based on the embodiments provided in the present application, in the cross-school collaborative action implementation, the millimeter wave radar waveform feature of climbing the fence is compressed into a 128-bit fingerprint code, the mountain school receives the code, and matches the similar waveform mode in the local feature library through Hamming distance. If the matching difference value is lower than the dynamic threshold, a lightweight detection model is downloaded to the edge device, and the risk coefficient of the semantic grid is updated synchronously. This cross-school collaborative mechanism realizes the rapid sharing of security information and the collaborative update of detection models between different schools, so that the mountain school and other remote areas can obtain the abnormal event feature information of other schools in time, and quickly improve the detection ability of climbing the fence and other abnormal behaviors. Form a collaborative defense system in the whole campus, effectively improve the overall response ability and security protection level of the campus security system.

[0170] In one embodiment of the present application, the multi-algorithm collaborative prevention and control of laboratory dangerous chemical operation includes:

[0171] Sensor deployment and task dynamic allocation: millimeter wave radar is deployed in the laboratory workbench area (precision 10 cm), which scans the human body thermal imaging profile every 200 ms, combined with pressure sensors (precision ± 5 kg) to determine whether the standing position of personnel is in front of the dangerous chemical storage cabinet. Bionic optimization algorithm dynamically schedules sensors: when detecting the period of 22:00-6:00, the camera adjacent to the experimental table automatically switches to the 120fps high-speed camera mode, and the remote corridor camera reduces the sampling rate to 15fps to save bandwidth (node value coefficient calculation: distance < 30 meters + 0.4, time period risk level + 0.3).

[0172] Time sequence rule embedded causal reasoning: rule coding: the "two-person authorization" rule is disassembled into spatial feature matrix and time gating parameters: spatial conditions: the laboratory coordinates (X=120-130, Y=40-50, Z=0-2) correspond to the high weight area of the 3D convolution kernel, and the single body sign signal of the millimeter wave radar produces specific response; Time condition: through the LSTM time gating unit, the time stamp of 22:00-6:00 is converted into a bias vector, so that the activation threshold of the rule synapse in this period is reduced by 30%. Causal verification: the differentiable knowledge graph constructs a dangerous chemical operation state machine, when the current sensor detects that the heating device power curve exceeds the rated value by 15% for 5 minutes (such as 345W for a 300W device), the graph automatically traverses the "power overrun→temperature anomaly→dangerous chemical volatilization" state transition path, and the state transition probability at each step is obtained by training historical accident data (such as the transition probability from power overrun to temperature anomaly 0.85).

[0173] Non-sensitive feature extraction and dynamic matching: the edge node uses an improved HOG algorithm to extract the gradient direction histogram of the face contour, and only retains 6 geometric feature points (non-sensitive features) such as the lower jaw line and brow bone, which are compressed to 16 dimensions. Dynamic time warping algorithm introduces attention mechanism: the historical profile data of authorized personnel in the laboratory is sorted by time to form a feature sequence with timestamp, when the angle deviation between the to-be-matched feature and the template is > 25°, the time axis is stretched by ± 10% for elastic matching, and the matching time is controlled within 80 ms.

[0174] Semantic grid driven risk evolution simulation: build a 30cm precision lab semantic grid, each grid binds attributes such as device power threshold, dangerous chemical storage location, etc. When detecting single human sign + power overrun, retrieve the "dangerous chemical operation risk map" of the target grid, simulate the possible accident chain: primary risk: single person operation does not trigger double authorization; secondary risk: power overrun may cause container overheating (combined with device temperature sensor data verification); ultimate risk: dangerous chemicals volatilize to critical concentration (link gas sensor data). Particle swarm simulation model simulates smoke diffusion, introduces real-time data of ventilation system (such as exhaust duct air speed 1.2m / s) as the guiding vector of particle motion, so that the simulation error of diffusion path is less than 15%.

[0175] Hierarchical privacy protection and device linkage: after triggering the access control lock, the laboratory video stream enables the "dynamic pixel mask" technology: on the premise of preserving the human joint motion trajectory, the pixels of the face area are randomly rearranged, the rearranged video stream can identify limb movements (such as reaching for reagents) but cannot restore facial features. High-risk device power cut-off uses double-loop control: first cut off the main power supply through a relay, and within 50ms, cut off the backup power supply through a hardware logic circuit, ensuring that the cut-off action is completed within 100ms, and the operation log is written into the blockchain for storage.

[0176] In another embodiment of the present application, multi-modal identity verification and risk response in the privacy area of the dormitory building includes:

[0177] Gait and Bluetooth feature real-time collection: millimeter wave radar collects gait waveform at a frequency of 50Hz, calculates the sliding window variance of step length coefficient of variation (window size 10 steps, sliding interval 2 steps), normal variance range 0.6-1.0 in winter thick shoe scene, variance of outsiders can reach more than 1.3 due to different shoe sole materials. Bluetooth device listens to 2.4G / 5G frequency band, records the signal strength fluctuation of device connection sequence, calculates the spectral entropy value through fast Fourier transform, the device connection entropy value of dormitory members is between 0.5-0.7 (regular interval), and the entropy value of foreign devices is usually >0.8 (irregular connection).

[0178] 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 (main frequency interval of human walking), and through the self-attention mechanism, weights are assigned to each frequency component. In winter, the ankle swing feature weight is automatically reduced (due to the influence of thick shoes), and the knee swing feature weight is increased. Device entropy channel: LSTM network learns the time sequence pattern of Bluetooth connection intervals. When detecting 3 consecutive connection intervals < 2 minutes (high-frequency connection behavior of non-dormitory members), trigger abnormal weight gain. Fusion strategy: Through dynamic routing algorithm, automatically adjust the dual-channel weight according to the time period (Bluetooth weight accounts for 60% at night, gait weight accounts for 50% during the day), the dimension of the fused anonymous feature vector is 24.

[0179] Bayesian model-driven seasonal adaptation: Establish a prior distribution model of winter gait, determine the mean μ = 1.0Hz and standard deviation σ = 0.2Hz of joint swing frequency based on historical data. When the mean of real-time collected data deviates from μ by more than 1.5σ (i.e. < 0.7Hz or > 1.3Hz), it is determined to be affected by thick winter shoes. Incremental Bayesian update algorithm is used, and each time valid winter gait data is detected, μ and σ of the prior distribution are updated to adapt the model to different years of winter shoe changes (e.g. if the snow boots in 2025 are 1cm thicker than the shoe sole in 2024, the model automatically adjusts the threshold).

[0180] Multi-level verification and risk deduction: First verification: If the Euclidean distance between the fused feature vector and the template library is less than 1.2 (dormitory member threshold), it is passed, and 1.2-1.5 is started secondary verification (call the low-resolution thermal imaging of the corridor camera). Three consecutive verification failures (distance > 1.5): activate the "intrusion of outsiders" risk chain in the semantic grid, retrieve the historical intrusion records of the dormitory grid, link the millimeter wave radar on adjacent floors to expand the scanning range (from 3 meters to 5 meters), and at the same time push the thermal map of the personnel gait trajectory to the dormitory terminal.

[0181] The above is only a preferred embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A multi-angle campus panoramic intelligent monitoring system, characterized in that, The method comprises the following steps of: a perception module is used to obtain perception data based on a sensor network deployed in each area of the campus, and dynamically allocate perception tasks according to the area functions; wherein the sensor network comprises a camera, a millimeter wave radar, a pressure sensor, an environmental noise sensor and a Bluetooth device; a processing module is used to process the perception data in real time through a reconfigurable hardware node to generate compliance features and abnormality identification, wherein the processing method comprises: using a time sequence rule embedded causal reasoning network to encode the campus rules into a neural network differentiable rule synapse; wherein in the time sequence rule embedded causal reasoning network, the generation of the neural network differentiable rule synapse comprises: decomposing the laboratory double-person authorization rule into spatial conditions and time conditions, and respectively encoding them into the weight matrix and bias vector of the neural network; when the millimeter wave radar detects a single person sign in the laboratory area and the time stamp is in the late night period, the convolution kernel operation of the rule synapse is activated; through the differentiable knowledge graph, it is verified whether the signal features of the electrical equipment match the dangerous chemical operation mode, if they match, the abnormality identification is generated and the access control locking instruction is triggered; in the privacy area, the millimeter wave gait entropy feature and the Bluetooth connection entropy feature are fused using an asymmetric modal contrast learning network; wherein the operation of the asymmetric modal contrast learning network in the dormitory building privacy area comprises: collecting the gait waveform through the millimeter wave radar, extracting the sliding window variance of the step length variation coefficient as the millimeter wave gait entropy feature; synchronously obtaining the Bluetooth device connection sequence, calculating the spectral entropy value of the signal strength fluctuation as the Bluetooth connection entropy feature; constructing a double-channel attention mechanism on the edge node: the gait entropy channel focuses on the joint swing frequency, and the device entropy channel strengthens the connection interval feature; the double-channel output is fused into an anonymized compliance feature vector, which is only verified when the vector Euclidean distance is lower than the adaptive threshold of the dormitory member template library; if the continuous verification fails, the corridor camera generates a low-resolution thermal imaging map and pushes it to the dormitory terminal; a risk deduction module is used to construct a semantic grid based on the compliance features and abnormality identification output by the processing module, associate the area function attributes, time permissions and historical risk thermal maps, and perform global risk causal chain deduction; a response execution module is used to trigger dynamic response according to the deduction result, including hierarchical privacy protection and cross-campus collaborative action. 2.The multi-angle campus panoramic intelligent monitoring system according to claim 1, characterized in that, After triggering the access control locking instruction, the method further comprises: calling the laboratory area camera to capture the personnel face image, extracting the non-sensitive contour feature through the edge node; dynamically time warping the non-sensitive contour feature and the authorized personnel model in the compliance feature library for matching; if the matching similarity is lower than the safety threshold, a double-person verification request and a personnel motion trajectory thermal map are sent to the security terminal; synchronously cutting off the power supply of the high-risk equipment and recording the operation log to the risk thermal map of the semantic grid. 3.The multi-angle campus panoramic intelligent monitoring system according to claim 1, characterized in that, The template updating mechanism of the anonymized compliance feature vector comprises: collecting winter thick shoe gait data, calculating the seasonal offset of joint swing frequency entropy through a Bayesian probability model; when the seasonal offset exceeds the preset critical value, the gait entropy weight coefficient is reduced and the decision proportion of the Bluetooth connection entropy is increased. The updated template library is encrypted and distributed to each campus edge node, synchronously updating the permission attributes in the semantic grid.

4. The multi-angle campus panoramic intelligent monitoring system according to claim 1, characterized in that, The construction of the semantic grid includes: Discretizing the three-dimensional geographic model of the campus into a cubic grid unit with a side length of 30 cm; Binding the examination week mobile phone signal shielding rules to the teaching building grid, and generating an abnormal identifier if a mobile device radio frequency signal is detected during the effective period of the rules; Loading the heat weight value of historical falling events for the staircase grid, which increases in real time with snow and rain weather; Forming a four-dimensional index structure with grid coordinates as the core and time permissions, function labels, and risk coefficients as attributes.

5. The multi-angle campus panoramic intelligent monitoring system according to claim 4, 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 of the patch panel continues to exceed the limit and the infrared sensor identifies that the personnel density exceeds the standard, a composite abnormal identifier is marked; Retrieving the escape passage topology graph and flammable material distribution matrix associated with the dormitory grid; Calling the campus event knowledge base to simulate the fire evolution chain: circuit overload causes sparks to ignite bed curtains, which in turn causes smoke to spread along the ventilation duct; Based on the smoke diffusion particle swarm simulation, a dynamic risk vector field is generated, and the tangent component of the dynamic risk vector field is used as a constraint condition to plan the evacuation path; Pushing a vector map including real-time path updates to the mobile phones of teachers and students, and superimposing a translucent warning layer of the smoke coverage area on the corridor display screen; According to the risk level, automatically remove the privacy blur of the escape path camera, and generate a personnel density heat map to assist in rescue dispatch.

6. The multi-angle campus panoramic intelligent monitoring system according to claim 1, characterized in that, The dynamic task allocation of the sensor network includes: Calculating the node value coefficient based on the target distance, remaining power, and network delay through a bionic optimization algorithm; When the playground gathering personnel density exceeds the set density threshold, the adjacent node group starts the face tracking mode and compresses the non-key background pixels; The remote node group schedules a drone to take an aerial view of the crowd distribution, and the edge node generates a crowd density heat map and projects it to the security center ring screen; The high-density area in the crowd density heat map is automatically marked as a stampede risk area and triggers an evacuation broadcast.

7. The multi-angle campus panoramic intelligent monitoring system according to claim 1, characterized in that, The implementation of hierarchical privacy protection includes: Deploying a piezoelectric sensor array at the entrance of the psychological counseling room to analyze the periodic characteristics of the duration of the stampede; When the environmental noise sensor identifies an abnormal base frequency of the sobbing sound, an encrypted warning message is generated and sent to the school doctor terminal; The indoor video stream real-time enables a dynamic pixel reorganization strategy, which converts human form into abstract geometric color blocks and preserves motion vectors.

8. The multi-angle campus panoramic intelligent monitoring system according to claim 1, characterized in that, The implementation of cross-campus collaborative action includes: Compressing the millimeter wave radar waveform features of climbing the fence into a 128-bit fingerprint code; After receiving the 128-bit fingerprint code, the mountainous campus matches similar waveform patterns in the local feature library through Hamming distance; If the matching difference value is below the dynamic threshold, download the lightweight detection model to the edge device, and synchronously update the risk coefficient of the semantic grid.

Images