Fireproof electronic fence comprehensive safety management system based on multi-source perception and intelligent linkage

Abstract

The invention relates to the technical field of intelligent security and risk early warning, and particularly discloses a fireproof electronic fence comprehensive safety management system based on multi-source perception and intelligent linkage. Thermal imaging and visible light camera shooting technologies are fused to generate an initial spatio-temporal data stream, and behavior, environment and target attribute features are extracted in parallel; calculating a dynamic risk integral value; when the initial risk integral exceeds a threshold value, the system executes active stimulation intervention and synchronously adjusts a perception strategy, and captures a target response to generate a stimulation feedback data stream; a risk integral value is updated by analyzing micro expressions and postures in the feedback data, and closed-loop treatment of silence recovery, progressive inquiry or manual intervention is executed; according to the invention, the dynamic quantitative evaluation and differential disposal of the target risk are realized, the accuracy of risk identification and the disposal flexibility are improved, and the fire prevention capability of the natural area is effectively enhanced.

Description

Technical Field

[0001] This invention relates to the field of intelligent security and risk warning technology, and more specifically, to a comprehensive safety management system for fireproof electronic fences based on multi-source sensing and intelligent linkage. Background Technology

[0002] Currently, preventing human-caused fire risks in large natural areas such as forests and grasslands faces challenges including a wide monitoring range, dispersed potential risk points, and limited effectiveness of traditional monitoring methods. Since human patrols are limited by cost, terrain, and all-weather coverage, technological means play a crucial role. The core issue lies in how to identify high-potential-risk human activities early and accurately from massive amounts of monitoring information, and take appropriate intervention measures to prevent fire accidents.

[0003] Commonly adopted solutions in the industry include deploying fire awareness networks based on fixed smoke or temperature sensors, and building video surveillance systems covering key areas. The former is a reactive alarm technology, which only triggers when a fire has already formed and generated sufficient smoke or heat, thus limiting its effectiveness in prevention. The latter provides real-time monitoring through video images, and some systems integrate motion detection-based intrusion alarms, which can send alerts to the backend when a moving target is detected entering the protected area. Some more advanced systems have introduced thermal imaging technology, utilizing its sensitivity to heat sources to detect abnormal hotspots or biological targets within forest areas.

[0004] However, the aforementioned traditional methods have room for improvement when dealing with complex human-caused risk prevention scenarios. Fixed sensor-based solutions rely on passive alarm mechanisms, failing to provide early warnings before risky behavior occurs. Systems based on conventional video surveillance and simple motion detection are inadequate in distinguishing between normal pedestrians, tourists, and high-risk individuals with potential intentions such as arson. They are prone to generating numerous invalid alarms due to non-threatening events such as animal activity and swaying tree shadows, increasing the burden of manual verification. Even solutions incorporating thermal imaging technology, while addressing the detection of nighttime or camouflaged targets to some extent, typically only detect the presence of the target. They lack comprehensive analysis of multi-dimensional information such as target behavior, surrounding environment, and carried items, making it impossible to dynamically and quantitatively assess the target's true risk level. Intervention measures are often limited to single, open-ended alarms, lacking immediate verification of intervention effectiveness and closed-loop escalation of response strategies. Summary of the Invention

[0005] In view of this, in order to solve the problems mentioned in the background technology, a comprehensive safety management system for fireproof electronic fences based on multi-source sensing and intelligent linkage is proposed.

[0006] The objective of this invention can be achieved through the following technical solution: This invention provides a comprehensive safety management system for fireproof electronic fences based on multi-source sensing and intelligent linkage, including: an initial spatiotemporal data stream generation module, which controls a thermal imaging camera to acquire thermal sensing images and identify moving heat source signals, and links a visible light camera to lock onto the target and acquire continuous visible light images based on the moving heat source signals, and aligns the continuous visible light images with spatiotemporal trajectory data to generate an initial spatiotemporal data stream.

[0007] The initial dynamic risk integral value generation module extracts behavioral feature signals, contextual environment signals, and target attribute signals in parallel based on the initial spatiotemporal data stream, and generates the initial dynamic risk integral value through weighted summation and time-series accumulation calculation.

[0008] When the initial dynamic risk score exceeds the preset first-level warning threshold, the composite instruction data packet generation module calls the intervention-perception dual-modal instruction generation unit to generate a composite instruction data packet containing intervention content instructions and posterior perception strategy instructions.

[0009] The stimulus feedback data stream generation module drives the smart speaker to perform active auditory stimulation according to the composite instruction data packet, and simultaneously adjusts the parameters of the visible light camera to enter the feedback capture mode, and collects the target reaction video to generate a stimulus feedback data stream.

[0010] The feedback quantization feature signal generation module analyzes micro-expressions and instantaneous postures in the stimulus feedback data stream, extracts and quantifies the biological features representing the subtle movements of the target face and limbs, and generates feedback quantization feature signals.

[0011] The risk update and handling decision module incorporates the feedback quantitative characteristic signals into the calculation model of the initial dynamic risk score, generates the updated dynamic risk score, and performs closed-loop handling based on the updated dynamic risk score, such as restoring silence, progressive questioning, or manual intervention.

[0012] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: (1) The present invention can quantitatively assess the potential risk of a target by fusing multimodal features and establishing an initial dynamic risk integral value. The system does not rely on a single motion detection or heat source signal, but processes the behavior pattern reflected by the spatiotemporal trajectory of the target, the contextual environmental information combined with geographic information and meteorological data, and the target attribute information analyzed by image recognition in parallel. By weighted fusing and time-series accumulation of these heterogeneous data of different dimensions, the system can generate a continuously changing risk quantification index, thereby realizing a dynamic and comprehensive judgment of the target risk level. This mechanism enables the system to distinguish between low-risk normal passage behavior and abnormal activities with high-risk characteristics, guiding monitoring resources and human attention to targets that are truly potentially threatening, and improving the accuracy of risk identification.

[0013] (2) This invention, through the design of a mechanism that simultaneously executes active stimulus intervention and adaptive perception strategies, can deeply probe and verify the target's intentions. When the initial risk score reaches a preset threshold, the system not only performs intervention actions such as verbal dissuasion, but more importantly, it simultaneously adjusts the working mode of the front-end sensing device, such as switching to a high frame rate to capture the target's instantaneous reaction after being stimulated. By quantitatively analyzing the collected feedback information such as micro-expressions and posture changes, the system can obtain objective data about the target's psychological state. This constructs a closed loop from "perception-analysis-intervention-re-perception-re-analysis," enabling the system to correct risk judgments based on the target's actual feedback, providing data support for subsequent decisions on whether to downgrade the situation, escalate the situation, or intervene manually, thus avoiding the shortcomings of traditional open-loop interventions that cannot confirm the effect.

[0014] (3) This invention establishes a closed-loop progressive handling process based on feedback verification, enabling automated and differentiated upgrades of handling methods. The system incorporates the parsed feedback quantitative characteristic signals into the risk calculation model with high weighting coefficients to generate updated dynamic risk scores. Changes in these scores directly drive subsequent handling decisions. If the score decreases due to the target's compliant behavior, the system can automatically deactivate the alarm; if the score jumps to a higher threshold due to the target's provocative behavior, the system automatically triggers the next level of stronger intervention and perception strategies, forming a progressive questioning mechanism. This mechanism allows the handling process to be dynamically adjusted according to the actual evolution of the risk, achieving a smooth and automatic upgrade from gentle persuasion to strong deterrence and even manual intervention, enhancing the system's flexibility and appropriateness in dealing with targets of different risk levels. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the system module structure connection of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 This invention provides a comprehensive safety management system for fire-prevention electronic fences based on multi-source perception and intelligent linkage, including: an initial spatiotemporal data stream generation module, an initial dynamic risk score generation module, a composite instruction data packet generation module, a stimulus feedback data stream generation module, a feedback quantification feature signal generation module, and a risk update and disposal decision module.

[0019] The initial spatiotemporal data stream generation module is connected to the initial dynamic risk integral value generation module. The initial dynamic risk integral value generation module is connected to the composite instruction data packet generation module. The composite instruction data packet generation module is connected to the stimulus feedback data stream generation module. The stimulus feedback data stream generation module is connected to the feedback quantification feature signal generation module. Both the initial dynamic risk integral value generation module and the feedback quantification feature signal generation module are connected to the risk update and disposal decision module.

[0020] The initial spatiotemporal data stream generation module controls the thermal imaging camera to acquire thermal sensing images and identify moving heat source signals. Based on the moving heat source signals, it links the visible light camera to lock onto the target and acquire continuous visible light images. The continuous visible light images are aligned with the spatiotemporal trajectory data to generate the initial spatiotemporal data stream.

[0021] In a specific embodiment of the present invention, the process of aligning continuous visible light images with spatiotemporal trajectory data to generate an initial spatiotemporal data stream includes: applying an inter-frame difference algorithm to the original thermal sensing image frames acquired by the thermal imaging camera, identifying regions of pixel value change, and performing connected component analysis.

[0022] When the average temperature of the connected domain is higher than the preset moving heat source detection threshold and the displacement exceeds the preset motion judgment distance, it is confirmed as a moving heat source signal.

[0023] Using a pre-calibrated coordinate mapping model, the two-dimensional coordinates of the moving heat source signal are converted into the pan-tilt control parameters of the visible light camera, the target is locked and a real-time target tracking algorithm is started, and continuous visible light images are acquired during the tracking process.

[0024] Geometric projection operations are used to convert the position of the target in the continuous visible light image into latitude and longitude coordinates in the geodetic coordinate system to generate spatiotemporal trajectory data. The continuous visible light image and the spatiotemporal trajectory data are time-stamped and the image matrix, latitude and longitude coordinates and motion vectors are encapsulated into an initial spatiotemporal data stream.

[0025] Specifically, this process is executed by a multi-source data acquisition and processing module deployed on edge computing nodes or a backend central server within the fire prevention area. First, the module periodically acquires raw thermal imaging frames from a wide-angle thermal imaging camera deployed at the forest boundary via the RTSP protocol. The module applies an inter-frame difference algorithm to at least two consecutively acquired raw thermal imaging frames to identify regions where pixel values ​​change, and performs connected component analysis on the pixel blocks within these regions. When the temperature corresponding to the average pixel value of a connected component is higher than a preset moving heat source detection threshold, and its pixel area is larger than a preset minimum target filtering area, the system marks the centroid coordinates of that connected component as an initial hotspot. By comparing the positional changes of the initial hotspot in consecutive frames, if its displacement exceeds a preset motion judgment distance, the system identifies it as a moving heat source signal. Subsequently, the module encapsulates the two-dimensional coordinates of the moving heat source signal in the thermal imaging coordinate system into a target indication message. Next, the module calls a pre-calibrated coordinate mapping model, which stores the spatial transformation matrix between the wide-angle thermal imaging camera and the specified high-definition visible light PTZ camera. Through a coordinate mapping model, the two-dimensional coordinates in the target indication message are converted into translation, tilt, and initial zoom parameters required by the HD Visible Light PTZ camera's PTZ control protocol. The module encapsulates these parameters into ONVIF commands and sends them to the HD Visible Light PTZ camera via the network. The HD Visible Light PTZ camera executes the ONVIF commands, aligns its optical lens with the physical location of the target and magnifies it. After locking onto the intruding target, it initiates a real-time target tracking algorithm based on a kernel correlation filter. During real-time target tracking, the Visible Light PTZ camera continuously adjusts its PTZ attitude to keep the target centered in the field of view and acquires and outputs continuous visible light images at a preset frame rate. The module synchronously records the timestamp of each frame of continuous visible light images and reads the current PTZ angle and zoom ratio of the HD Visible Light PTZ camera in real time. Combining the preset geospatial location and attitude data of the HD Visible Light PTZ camera, geometric projection calculations are used to convert the target's position in each frame of the image into latitude and longitude coordinates in the geodetic coordinate system, generating spatiotemporal trajectory data containing timestamps and location information. Subsequently, the module performs data alignment operations to strictly synchronize the continuous visible light image sequence and the spatiotemporal trajectory data according to the timestamps. Finally, the multi-source data acquisition and processing module creates a data queue as a cache, encapsulating the RGB matrix of each frame of continuous visible light image, the latitude and longitude coordinates in the spatiotemporal trajectory data at the corresponding time, and the instantaneous velocity and direction vectors calculated from the continuous coordinate points into a structured data unit, and pushing it into the data queue in chronological order, thereby generating an initial spatiotemporal data stream containing target appearance, position and motion information for subsequent processing steps.

[0026] The mobile heat source detection threshold is a critical temperature value used to distinguish biological heat sources from environmental background thermal radiation. In this embodiment, assuming the application scenario is the spring fire prevention season in a temperate forest, this threshold is set to 33-35℃ to effectively identify human targets and exclude localized high-temperature areas on the ground caused by sunlight. The minimum target filtering area is a lower limit of pixel area set to avoid misjudging small animals such as rabbits or birds as intrusive targets. Its value is calibrated based on the resolution of the wide-angle thermal imaging camera and its installation height, and is usually set to the number of pixels that can cover the cross-sectional projection of a human body at an equivalent distance, for example, no less than 30 pixels at a distance of 50m. The spatiotemporal trajectory data is a time series, and its data nodes consist of timestamps, longitude, latitude, and altitude. The initial spatiotemporal data stream is a composite data structure, and its basic unit is a data frame, which contains at least a timestamp, an RGB image matrix, three-dimensional geographic coordinates, and a two-dimensional motion vector.

[0027] For example, at a certain time T0, a wide-angle thermal imaging camera deployed at the entrance of a forest area detects a moving heat source signal with an average temperature of 34.5℃ within its 640x512 resolution field of view. The centroid coordinates of this moving heat source shift from (250, 400) at time T0 to (255, 402) at time T0+0.5s. Since its temperature exceeds the 33℃ moving heat source detection threshold and its displacement is greater than a motion detection distance of 2 pixels, the system determines it to be a valid intrusion target. The multi-source data acquisition and processing module immediately converts the coordinates (250, 400) into control commands for a high-definition visible light PTZ camera using a preset coordinate mapping model, such as a translation angle of 15.8 degrees, a tilt angle of -10.2 degrees, and a zoom ratio of 5x. After executing the commands, the high-definition visible light PTZ camera locks onto a person wearing a blue shirt and begins tracking and acquiring data at a frame rate of 30fps. At time T1 = T0 + 3.0s, the system records the first frame of visible light image I1 (1920x1080 pixels) and analyzes the PTZ camera's attitude data, calculating the target's geographic coordinates as (118.521°E, 32.088°N). Subsequently, at time T2 = T0 + 4.0s, the system records the thirty-first frame of visible light image I2. At this point, the target has moved, and the system calculates its geographic coordinates as (118.523°E, 32.089°N) based on the new PTZ camera attitude. Finally, the system generates two consecutive data units in the initial spatiotemporal data stream. The first data unit contains the data matrix of T1 and image I1, the coordinates (118.521°E, 32.088°N), and an initialized zero vector. The second data unit contains the data matrix of image I2 (T2), coordinates (118.523°E, 32.089°N), and a motion vector with a velocity of approximately 2.5 m / s and a northeast direction, calculated based on the positional change from T1 to T2. This data stream is then fed into the subsequent feature extraction module for processing.

[0028] The initial dynamic risk integral value generation module extracts behavioral feature signals, contextual environment signals, and target attribute signals in parallel based on the initial spatiotemporal data stream, and generates the initial dynamic risk integral value through weighted summation and time-series accumulation calculation.

[0029] In a specific embodiment of the present invention, the specific process of extracting behavioral feature signals, contextual environment signals and target attribute signals in parallel based on the initial spatiotemporal data stream includes: calculating the velocity, acceleration and heading angle change rate of continuous coordinate points in the initial spatiotemporal data stream, and matching them with a preset behavioral pattern rule base to generate behavioral feature signals.

[0030] Geographic coordinates are submitted to a geographic information system to obtain land use types, timestamps are compared with meteorological databases to obtain wind and humidity data, and contextual environmental signals are generated by mapping through a multidimensional environmental risk lookup table.

[0031] Continuous visible light images are input into the target detection model to identify the category labels of the items carried by the target, and the data is quantified according to a preset item risk level table to generate target attribute signals.

[0032] Specifically, this is performed by the multimodal feature extraction and risk assessment module deployed in the backend central server. Upon receiving the generated initial spatiotemporal data stream, the module creates a unique identifier for the target associated with the initial spatiotemporal data stream and starts three parallel processing threads to extract multimodal features. The first thread handles behavioral features, continuously extracting spatiotemporal trajectory data from the initial spatiotemporal data stream. It calculates the velocity, acceleration, and rate of change of heading angle between continuous coordinate points and matches them with behavioral templates preset in the behavioral pattern rule base to quantify and generate behavioral feature signals. The second thread handles contextual features, extracting geographic coordinates and timestamps from the spatiotemporal trajectory data. It submits the geographic coordinates to a Geographic Information System (GIS) for spatial querying to obtain land use type, vegetation cover, and sensitive point information for the target area. Simultaneously, it compares the timestamps with a meteorological database and holiday calendar to obtain the current wind speed, air humidity, and special time period markers. Then, based on a multidimensional environmental risk lookup table, it maps these discrete environmental factors into a comprehensive contextual environmental signal. The third thread is responsible for processing target attribute features. It extracts the RGB matrix of continuous visible light images from the initial spatiotemporal data stream and inputs it into a pre-trained YOLOv7 object detection model. This model can identify and output the category label and confidence score of the items carried by the target in the image, such as a shovel, backpack, or incense sticks. The module converts the identified item categories into a quantified target attribute signal based on a pre-defined item risk level table. At each time step, the multimodal feature extraction and risk assessment module aggregates the behavioral feature signals, contextual environment signals, and target attribute signals generated by the three threads.

[0033] It should also be noted that the specific process of quantifying and generating behavioral feature signals by calculating the velocity, acceleration, and rate of change of heading angle between continuous coordinate points and matching them with behavioral templates preset in the behavioral pattern rule base includes: using a sliding window difference algorithm to select the current moment... and the two preceding sampling times , Given the latitude and longitude coordinates, the ground displacement distance between adjacent time points is first calculated using the semi-versus formula, and then the instantaneous velocity scalar is obtained by combining the timestamp difference. The calculation formula is as follows: , in, For the Earth's radius, and These are the radian values ​​for latitude and longitude, respectively. For timestamps; Secondly, the quotient of the difference between two adjacent instantaneous velocity scalars is calculated to obtain the instantaneous acceleration. The calculation formula is: , At the same time, the arctangent function of the coordinate points is used to calculate the direction angle of the trajectory tangent. ( Then, the rate of change of heading angle is obtained by calculating the ratio of the differences in heading angles between adjacent time points. The calculation formula is: .

[0034] The system performs Z-score normalization on the calculated instantaneous velocity, instantaneous acceleration, and rate of change of heading angle to construct a three-dimensional behavioral feature vector for the current moment. And compare it with multiple standard behavior template vectors preset in the behavior pattern rule base. Perform Euclidean distance matching calculations on templates such as normal passage, loitering, and rapid advance. , The system selects the calculation results Minimum target template And the normalized risk weight value preset by the target template. As output, behavioral feature signals are generated through quantization.

[0035] In a specific embodiment of the present invention, the specific process of generating an initial dynamic risk score through weighted summation and time-series cumulative calculation includes: establishing a risk score buffer for the target, and performing weighted summation on the behavioral feature signal, contextual environment signal and target attribute signal according to a preset weight matrix to obtain an instantaneous risk score.

[0036] An exponential moving average algorithm is used to update the instantaneous risk score over time, thereby generating an initial dynamic risk score that changes continuously over time.

[0037] Specifically, the module maintains a risk score buffer in memory for the current target. This buffer is a first-in, first-out queue used to store the risk calculation results within the most recent time window. Next, it performs a weighted summation of the three types of feature signals based on a preset weight matrix to obtain an instantaneous risk score. Finally, to smooth short-term fluctuations and reflect the cumulative effect of risk, the module uses an exponential moving average algorithm to perform time-series cumulative calculations on the instantaneous risk score sequence, ultimately generating an initial dynamic risk score that changes continuously over time. This score is then associated with the target's unique identifier and output. At any sampling time... Instantaneous risk score The calculation method is as follows: , Subsequently, the initial dynamic risk score Updated using the exponential moving average algorithm: , All signals and integral values ​​in the formula are dimensionless normalized values, ensuring that features from different physical sources can be calculated within a unified framework.

[0038] Among them, behavioral feature signals Context signals With target attribute signal At any moment Numerical values ​​are extracted and normalized to the 0-1 range from three dimensions: behavior, environment, and target attributes. For example, for behavioral feature signals, in the behavioral pattern rule base, the behavioral template "traveling at a constant speed along a fixed path" corresponds to a score of 0.1, while the behavioral template "repeatedly wandering in a non-open area" corresponds to a score of 0.7. The preset weight matrix in this embodiment is represented by a set of weight coefficients. , , Their sum is 1, used to adjust the contribution of different feature dimensions to the final risk. Based on statistical analysis of historical fire causation data, carrying flammable materials poses the highest risk, therefore the weight is set to 1. , , The preset time window is set to 30 seconds. This duration is based on the statistical average of the time required for normal personnel to cross the monitored area, and is used to distinguish between normal passage and abnormal lingering. For a moment The initial dynamic risk score, This is the integral value from the previous time step. Smoothing coefficient. It is a constant between 0 and 1, and in this embodiment it is set to 0.3. It is used to control the update speed. The larger the value, the faster the system reacts to the current instantaneous risk.

[0039] For example, the system continues to process the generated initial spatiotemporal data stream. At time T2, the multimodal feature extraction and risk assessment module receives a data unit containing image I2, coordinates (118.523°E, 32.089°N), and motion vectors. First, the behavior feature processing thread analyzes its passage speed of 2.5 m / s, determines its behavior pattern as "normal passage," and retrieves the corresponding behavior feature signal from the behavior pattern rule base. The value is 0.1. In parallel, the context environment feature processing thread queries the GIS system for coordinates, returning the "cemetery area" label. Simultaneously, it queries the meteorological and calendar database to determine that the current period is "Qingming Festival" and is characterized by "strong winds and dryness." High-risk context environment signals are then matched against the environmental risk lookup table. The value is 0.9. Simultaneously, the target attribute feature processing thread inputs image I2 into the YOLOv7 model. The model identifies that the target is holding an object resembling "incense sticks" and generates a high-risk target attribute signal based on the item risk level table. The value is 0.8. Next, the system uses preset weights. Calculate the instantaneous risk score: .

[0040] Finally, the system uses a smoothing coefficient. Update the risk score, assuming the initial dynamic risk score value from the previous time step. If the value is 0.1, then the initial dynamic risk integral value at the current moment is: .

[0041] When the initial dynamic risk score exceeds the preset first-level warning threshold, the composite instruction data packet generation module calls the intervention-perception dual-modal instruction generation unit to generate a composite instruction data packet containing intervention content instructions and posterior perception strategy instructions.

[0042] In a specific embodiment of the present invention, the specific process of calling the intervention-perception dual-modal instruction generation unit to generate a composite instruction data packet containing intervention content instructions and posterior perception strategy instructions includes: the intervention content instructions encapsulate the device identifier of the smart speaker near the target and the flexible dissuasion voice content to be played.

[0043] The posterior perception strategy instruction encapsulates the sensor parameter adjustment plan for the visible light camera. The sensor parameter adjustment plan includes a trigger time stamp synchronized with the end of voice playback, a frame rate parameter for switching to a high frame rate acquisition mode, and zoom parameters for zooming in on specific areas of the target.

[0044] Specifically, this is executed by the risk level determination and instruction generation module in the backend central server. This module continuously reads the generated initial dynamic risk score value associated with a specific target at a preset polling cycle. Internally, the module maintains a state machine for each tracked target, recording the currently triggered risk level. Within a processing cycle, the module compares the latest initial dynamic risk score value with a preset Level 1 warning threshold. If the initial dynamic risk score value is greater than the Level 1 warning threshold, and the target's state machine indicates that a Level 1 warning has not yet been triggered, the system determines that the risk level has increased and locks the event to trigger an instruction generation process. At this time, the system calls an intervention-perception bimodal instruction generation unit. The intervention-perception bimodal instruction generation unit first accesses a tiered intervention strategy library, and based on the currently reached Level 1 risk level, retrieves the corresponding gentle persuasion voice content and posterior perception strategy template. Subsequently, the intervention-perception bimodal instruction generation unit generates a structured composite instruction data packet. The composite instruction data packet contains two main parts. The first part is the intervention content instruction, which encapsulates the device identifier of the smart speaker near the target and the text or audio file path of the gentle persuasion voice content to be played. The second part is the posterior perception strategy instruction, which encapsulates the device identifier of the high-definition visible light PTZ camera responsible for tracking the target and encodes a detailed set of sensor parameter adjustment plans. Specifically, the sensor parameter adjustment plan includes a trigger time stamp precisely synchronized with the end of voice playback, a frame rate parameter for switching to a high frame rate acquisition mode, and a set of zoom and focus parameters for close-up shooting of key areas such as the target's face or hands. Finally, the risk level determination and instruction generation module serializes the generated composite instruction data packet into a JSON-formatted string, ready to distribute it to the front-end execution device via a message queue.

[0045] The Level 1 warning threshold is the critical dividing line that distinguishes the target behavior from "suspicious" to "having a clear potential risk." The Level 1 warning threshold is set based on statistical analysis of the risk score evolution curves in historical fire prevention early warning cases, aiming to ensure timely responses to behaviors with obvious risk tendencies while avoiding excessive intervention in low-risk normal activities. In this embodiment, the Level 1 warning threshold is set to 0.25. The intervention-perception bimodal instruction generation unit is a software functional module that converts the abstract risk level input into a specific set of instructions that can be directly executed by physical devices. The trigger time stamp in the posterior perception strategy instruction is a Unix timestamp in milliseconds, calculated by adding the playback duration of the selected audio content to the current time, ensuring that the perception recognition action begins immediately after the intervention stimulus ends. The parameters of the high frame rate acquisition mode are typically set to 100fps to 150fps, far higher than the 30fps of conventional monitoring, in order to capture micro-expressions or subconscious rapid movements that the target may exhibit after being stimulated.

[0046] For example, the system continuously monitors the initial dynamic risk score generated by the target in the example. The initial dynamic risk score has reached 0.256 at time T2. The risk level determination and instruction generation module compares this value with the preset level 1 warning threshold of 0.25. Since 0.256 is greater than 0.25, and the target's state machine shows that the level 1 warning has not yet been triggered, the system immediately calls the intervention-perception dual-modal instruction generation unit. This unit retrieves the soft dissuasion voice corresponding to level 1 risk from the hierarchical intervention strategy library, which is "You have entered a high-risk forest fire zone, please pay attention to fire safety," and finds that the playback duration of this voice is 4.5 seconds. Then, the unit begins to construct a composite instruction data package. Its intervention content instruction part specifies the smart speaker with ID SPK-03 located near the target to play the above voice content. Its posterior perception strategy instruction part specifies the high-definition visible light PTZ camera with ID PTZ-02 and sets the trigger time stamp to a timestamp 4.5 seconds after the current time T2. The instructions also included camera parameters, specifying that the frame rate should be adjusted from 30fps to 120fps and the zoom should be increased to 12x for close-up shots of the target's face. Finally, a composite instruction data packet containing these instructions was generated, serialized, and prepared for transmission to the corresponding smart sound column and visible light PTZ camera.

[0047] The stimulus feedback data stream generation module drives the smart speaker to perform active auditory stimulation according to the composite instruction data packet, and simultaneously adjusts the parameters of the visible light camera to enter the feedback capture mode, and collects the target reaction video to generate a stimulus feedback data stream.

[0048] In a specific embodiment of the present invention, the specific process of acquiring target response video to generate stimulus feedback data stream includes: intelligent sound column parsing intervention content instructions and executing voice playback to provide active auditory stimulation to the target.

[0049] The visible light camera interprets the a posteriori perception strategy instructions and instantaneously modifies the register values ​​of the image signal processor when the internal clock reaches the trigger time marker.

[0050] Configure the frame rate parameter to a preset high frame rate value and drive the lens to zoom, enter the feedback capture mode to capture high frame rate close-up video, and encapsulate the captured high frame rate close-up video into a stimulus feedback data stream with a specific feedback identifier.

[0051] Specifically, the process is executed collaboratively by intelligent devices deployed at the front end of the fire protection zone and the command dispatch and execution module in the back-end central server. First, upon receiving the composite command data packet, the command dispatch and execution module parses it. The module extracts the intervention content command and, based on the device identifier contained within, publishes the playback command to a designated topic via the MQTT message protocol. This topic is subscribed to by intelligent speakers within the target area. Upon receiving the command, the intelligent speakers decode the audio content and, through their built-in digital-to-analog converter and power amplifier, drive the speakers to play the audio, thus providing an active auditory stimulus to the target. Simultaneously, the command dispatch and execution module extracts the posterior perception policy command from the same composite command data packet. Based on the device identifier contained in the command, the module sends the command to the visible light PTZ camera tracking the target via the ONVIF protocol or the device's private SDK. The embedded processor within the visible light PTZ camera receives and parses the posterior perception policy command, loading its parameter configuration into a task queue. This task queue is driven by an internal clock synchronized with the Network Time Protocol (NTP). The camera firmware continuously monitors the current system time. When the system time reaches the trigger time marker specified in the instruction, the firmware immediately modifies the relevant register values ​​of the image signal processor (ISP) with high priority, adjusts the sensor's frame rate and exposure time to preset values, and drives the lens motor to complete zoom and focus operations, instantly putting the camera into the preset feedback capture mode. In this mode, the visible light PTZ camera begins to acquire high frame rate close-up video. The acquired video frames are encapsulated into a new data stream with specific feedback identifiers, generating a stimulus feedback data stream specifically for analyzing the target's response, and transmitted back to the backend analysis server via the RTSP protocol.

[0052] Active auditory stimulation is an acoustic intervention with explicit informational content. Its purpose is to observe and assess the target's response patterns by conveying warnings or dissuasion messages, rather than simply using noise to drive away the target. Feedback capture mode is a special operating state of a visible light PTZ camera, characterized by a frame rate several times higher than conventional monitoring modes and a magnified view of a specific area. This mode is designed to capture subtle changes in facial expressions or body movements that occur within a fraction of a second after stimulation. The stimulus feedback data stream is a logically independent data channel. Specific metadata tags are added to the header of its data packets to distinguish it from the generated initial spatiotemporal data stream, ensuring that the backend processing module can accurately route it to a specialized feedback behavior analysis model.

[0053] For example, the aforementioned generated composite instruction data packet is received by the command dispatch and execution module. The module first issues a playback instruction to the smart speaker SPK-03, causing it to begin playing a 4.5-second voice message, "You have entered a high-risk forest fire zone; please be careful with fire," at time T2. Simultaneously, the module sends a post-hoc perception strategy instruction containing a trigger time stamp T2+4.5s, a frame rate of 120fps, and a zoom of 12x to the visible light PTZ-02. The PTZ-02's internal clock is synchronized with the server, and it enters a waiting state upon receiving the instruction. At time T2+4.5s, the instant the voice playback ends, the PTZ-02's firmware is triggered, immediately increasing the sensor's frame rate from 30fps to 120fps and zooming the lens to 12x, aiming at the target's face. The camera then begins acquiring and encoding high-frame-rate video, labeling this video stream as "stimulus_feedback_stream_PTZ02," which is the stimulus feedback data stream.

[0054] The feedback quantization feature signal generation module analyzes micro-expressions and instantaneous postures in the stimulus feedback data stream, extracts and quantifies the biological features representing the subtle movements of the target face and limbs, and generates feedback quantization feature signals.

[0055] In a specific embodiment of the present invention, the specific process of parsing micro-expressions and instantaneous postures in the stimulus feedback data stream includes: calling a face detection algorithm to extract facial image sequences from the stimulus feedback data stream, inputting them into a micro-expression recognition model to analyze the activation combinations of facial action units, and obtaining micro-expression features.

[0056] The instantaneous posture analysis model is called to extract key skeletal points of the human body, and the position and velocity changes of the skeletal points within a short time window after the stimulus occurs are calculated to obtain the body posture characteristics.

[0057] The intention mapping rule engine is used to make logical judgments on micro-expression features and body posture features to evaluate the subconscious reaction of the target after receiving active auditory stimulation.

[0058] Specifically, this is primarily performed by the feedback behavior parsing module deployed within the backend central server. Upon receiving the generated stimulus feedback data stream with specific feedback identifiers, this module immediately processes it. First, it invokes a face detection and tracking algorithm to locate and extract the target's facial image sequence from consecutive frames of a high-frame-rate close-up video. Subsequently, the module inputs the facial image sequence into a pre-trained micro-expression recognition model. This model, based on a three-dimensional convolutional neural network structure, can analyze pixel and texture changes in facial regions over a short period, thereby identifying specific activation combinations of facial action units. Simultaneously, another parallel processing thread invokes a graph neural network-based instantaneous posture analysis model. This model first constructs a spatiotemporal graph based on the human body's topology, defining extracted key skeletal points as nodes in the graph. Within a preset short time window after the stimulus occurs, for each skeletal point, the first-order difference method is used to calculate the coordinate difference between the current moment and the previous moment, and then divided by the sampling time interval to obtain the instantaneous velocity vector. Then, the second-order difference method is used to calculate the difference between the instantaneous velocity vectors of adjacent moments and divided by the time interval to obtain the instantaneous acceleration vector. Simultaneously, the head's position is calculated based on the relative position changes of the head and neck nodes. The normal vector of the plane is rotated angularly. Then, the above position, velocity, acceleration and angular velocity data are used as dynamic features of the nodes and input into the spatiotemporal graph convolutional network. The spatial dimension convolution aggregates limb coordination features and the temporal dimension convolution extracts action evolution features. Finally, the fully connected layer analyzes and outputs the discrimination results: if the sum of the magnitudes of the acceleration vectors of all nodes is detected to quickly return to zero after a brief peak and remain stationary, it is judged as body stiffness; if the cumulative displacement vector of the center point of the torso points in a direction away from the warning area, it is judged as backward movement; if the peak value of the head rotation angular velocity exceeds the preset physiological threshold, it is judged as rapid head rotation.

[0059] In a specific embodiment of the present invention, the biometric features representing the subtle movements of the target's face and limbs are extracted and quantified to generate a feedback quantified feature signal. The specific method for generating the feedback quantified feature signal is as follows: the generated feedback quantified feature signal is a vector containing compliance, alertness, and provocation. Among them, compliance is positively correlated with the target's body turning away from the restricted area and no negative emotional action unit is detected; alertness is positively correlated with the activation intensity of action units representing surprise or fear and the sudden stiffening of body posture; and provocation is positively correlated with the activation intensity of action units representing contempt or anger and the maintenance of a confrontational posture.

[0060] Specifically, a multi-dimensional quantification function integrating micro-expression action units and physical motion features is constructed to numerically calculate various indicators: For compliance, the system constructs a unit safety vector pointing from the target's current position away from the restricted area. The dot product of the instantaneous velocity vector of the target's torso center of mass and this safety vector is calculated and divided by the product of their magnitudes to obtain a cosine similarity, which is used to geometrically represent the directional consistency of "body turning away". At the same time, the normalized activation intensity of the action units corresponding to negative emotions is statistically analyzed, and the difference between this and the value of 1 is calculated as the emotional calmness coefficient (i.e., the probability of no negative emotion being detected). The cosine similarity and the emotional calmness coefficient are linearly weighted and summed according to preset weights. For alertness, the system extracts the average activation intensity of action units representing surprise or fear in real time, and monitors the acceleration vector magnitude of key skeletal points throughout the body. When the acceleration of all key points... When the sum of the degree modulus values ​​is lower than the preset static noise threshold, time integration is initiated to accumulate the duration of the body's "sudden stiffness". This duration is normalized and then superimposed on the above-mentioned emotion intensity values. For the provocation degree, the system extracts the activation intensity of the action unit representing contempt or anger, and constructs a risk unit vector from the target to the warning line or law enforcement officer. The projection component of the target's instantaneous acceleration vector on the risk vector is calculated (i.e., dot product operation). Only when this component is positive is it included in the "maintaining confrontation" approximation intensity. The derivative of the acceleration vector with respect to time is obtained to obtain the modulus of jerk (quickness) to quantify the sudden intensity of the action. The micro-expression intensity, the positive approximation projection component, and the quickness modulus are multi-factor weighted and fused, and finally mapped to the standard range of 0 to 1 through a nonlinear activation function to output a three-dimensional quantitative rating vector.

[0061] The micro-expression and instantaneous posture analysis model is a composite model, consisting of a micro-expression recognition model for facial expression analysis and an instantaneous posture analysis model for body dynamics analysis. Compliance is a normalized value between 0 and 1, positively correlated with changes in the target's body orientation and the calmness of their facial expression; when the target's body smoothly turns away from the restricted area, and no activation of negative emotional motor units such as frowning or downturned corners of the mouth is detected, the compliance value approaches 1. Alertness is also a normalized value between 0 and 1, positively correlated with the intensity of activation of motor units representing surprise or fear and sudden changes in body posture; when the target exhibits wide eyes, raised eyebrows, and sudden stiffening or rapid head turning, the alertness value is high. The provocation level is also a normalized value between 0 and 1. Its calculation is positively correlated with the activation intensity of action units representing contempt or anger and whether a confrontational posture is maintained. When the target is detected to have expressions such as one corner of the mouth turned up, frowning, or glaring, and the body orientation does not change or even leans forward, the provocation level is relatively high. The feedback quantization feature signal is a three-dimensional feature vector in the form of [compliance value, alertness value, provocation value].

[0062] For example, the feedback behavior analysis module receives a stimulus feedback data stream that began recording at time T2+4.5s. Analyzing this 120fps video stream, the module finds that at time T2+4.7s, within 200ms after the speech stimulus ends, the micro-expression recognition model detects a brief expression of surprise on the target's face, characterized by raised eyebrows and widened eyes; the activation intensities of the corresponding action units AU1+AU2+AU5 all exceed 0.7. Immediately following, the instantaneous posture analysis model detects that the target's head rapidly turns towards the sound source, with a peak angular velocity of 40 degrees / second in the neck. Subsequently, at time T2+5.2s, the micro-expression model detects that the target's expression changes to a frown, with the activation intensity of the corresponding action unit AU4 reaching 0.8 and remaining thereafter. The intent mapping rule engine comprehensively analyzes these features: due to the absence of a compliant departure action and the presence of a negative facial expression, the compliance level is determined to be 0.1; due to the obvious surprise reaction and rapid posture change, the alertness level is determined to be 0.8; and due to the subsequent continuous frowning and confrontational expression, the provocation level is determined to be 0.4. Finally, the system generates a set of feedback quantified feature signals, with specific values ​​as a vector [0.1, 0.8, 0.4], and passes them to the risk update module.

[0063] The risk update and handling decision module incorporates the feedback quantitative characteristic signals into the calculation model of the initial dynamic risk score, generates the updated dynamic risk score, and performs closed-loop handling based on the updated dynamic risk score, such as restoring silence, progressive questioning, or manual intervention.

[0064] In a specific embodiment of the present invention, the specific process of generating the updated dynamic risk score includes: weighting the compliance, alertness and provocation in the feedback quantization feature signal according to the internal feedback weight configuration to generate a scalar feedback score.

[0065] The scalar feedback score is multiplied by a preset high-weight coefficient as an incremental term and added to the instantaneous risk score calculation model.

[0066] Based on the instantaneous risk score including feedback increments, the updated dynamic risk score, which has been verified by feedback, is calculated using the exponential moving average algorithm.

[0067] Specifically, this is executed by the risk update and handling decision-making module in the backend central server. This module first receives the generated feedback quantitative feature signal. To integrate this multi-dimensional feedback signal into the existing risk calculation model, the module, based on an internal feedback weight configuration, weights and sums the compliance, alertness, and provocation indices in the feedback quantitative feature signal to generate a single-dimensional scalar feedback score. Subsequently, the module multiplies this scalar feedback score by a preset high-weight coefficient and adds it as an increment to the model used to calculate the instantaneous risk score. In this way, the system allows the real-time feedback of the target to directly and significantly influence the risk assessment results. The module uses the updated instantaneous risk score and, following the exponential moving average algorithm, calculates and generates an updated dynamic risk score that has been validated by the feedback.

[0068] The formula explains that, firstly, at time... Scalar feedback score The calculation method is as follows: .

[0069] Subsequently, the dynamic risk score was updated. Updated by introducing scalar feedback scores: , All signals, scores, and integrals in the formula are dimensionless normalized values, ensuring that information from all dimensions can be calculated on the same scale.

[0070] in, , , They represent the time intervals. The values ​​of compliance, alertness, and provocation are obtained from the feedback quantization feature signals. , , These are their respective weights, used to adjust the degree of influence of different feedback emotions on risk assessment. In this embodiment, based on the highest risk indicativeness of provocative behavior, the weights are set as follows: , , . For a moment The updated dynamic risk score value, This is the updated score from the previous moment. This is the instantaneous risk score without feedback information; its calculation method has been defined in the initial dynamic risk score generation module. Preset high-weight coefficients. This value, set to 0.6, amplifies the overall impact of feedback behavior on the risk score, reflecting the system's emphasis on direct observational evidence.

[0071] In a specific embodiment of the present invention, the specific process of performing closed-loop processing of restoring silence, progressive questioning, or manual intervention based on the updated dynamic risk score includes: if the updated dynamic risk score drops below the preset risk relief threshold, then the silent tracking mode is restored.

[0072] If the updated dynamic risk score exceeds the preset deterrence threshold, the system returns to the composite instruction data packet generation module to generate a composite instruction data packet containing a more aggressive level of voice and targeted perception strategies, forming a progressive interrogation.

[0073] If the updated dynamic risk score reaches the preset maximum threshold for human intervention, the initial spatiotemporal data stream and stimulus feedback data stream are aggregated to generate a comprehensive evidence file, which is then pushed to the human agent, and the two-way intercom module of the smart speaker is activated.

[0074] Specifically, the risk update and handling decision module then compares the updated dynamic risk score with multiple risk thresholds stored in the handling strategy library to execute differentiated handling. If the updated dynamic risk score drops below the preset risk de-escalation threshold due to high compliance feedback, the system downgrades the target's threat level and reverts to silent tracking mode. If the updated dynamic risk score jumps due to high alertness or high provocation feedback and exceeds a higher-level deterrence threshold, the system determines that a progressive interrogation is required and immediately sends an event message containing the current risk level back to the input queue of the composite command packet generation module to trigger the generation and execution of a more aggressive composite command packet. If the updated dynamic risk score reaches the highest preset threshold for human intervention, the system will activate the highest level alarm procedure. This procedure automatically collects and packages the complete risk evolution path of the target, the initial spatiotemporal data stream, the stimulus feedback data stream, and the characteristic signals generated at each stage to form a comprehensive evidence file. This file is then pushed to the monitoring terminal of the human agent via WebSocket. At the same time, an API request is sent to the communication gateway to unlock and activate the two-way intercom module of the smart speaker deployed near the target, authorizing the human agent to perform remote real-time intervention.

[0075] In one specific embodiment of the present invention, the risk mitigation threshold is set to 0.1, the deterrence threshold is set to 0.5, and the maximum human intervention threshold is set to 0.8. These thresholds are set based on statistical analysis of risk score evolution data from over 1000 real and simulated fire prevention scenarios. Progressive questioning is a closed-loop escalation intervention strategy, meaning the system's intervention intensity increases progressively with the accumulation of negative feedback from the target.

[0076] For example, at time T3, after receiving the feedback quantized feature signal vector [0.1, 0.8, 0.4], the risk update and disposal decision module first calculates the scalar feedback score. The module then uses this score to update the dynamic risk score. Assuming that at time T3, other characteristics of the target remain unchanged, the instantaneous risk score is without feedback. It remains at 0.62, while the integral value at the previous time step... That is, the value calculated in the initial dynamic risk score generation module. Its value is 0.256. Therefore, the updated dynamic risk score is: .

[0077] The system compared the value of 0.5434 with various thresholds and found that it exceeded the deterrence threshold of 0.5, but had not yet reached the highest human intervention threshold of 0.8. Therefore, the system determined to execute a progressive interrogation, resubmitting an event indicating a risk level of "Level 2 Deterrence" to the composite command data packet generation module. This generates and executes a composite command data packet with a more severe warning voice and more targeted perception and recognition strategies, such as capturing close-up shots of the target's hand movements, thus forming a continuously escalating closed-loop response process.

[0078] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A comprehensive safety management system for fire-prevention electronic fences based on multi-source sensing and intelligent linkage, characterized in that: include: The initial spatiotemporal data stream generation module controls the thermal imaging camera to acquire thermal sensing images and identify moving heat source signals. Based on the moving heat source signals, it links the visible light camera to lock onto the target and acquire continuous visible light images. The continuous visible light images are aligned with the spatiotemporal trajectory data to generate the initial spatiotemporal data stream. The initial dynamic risk score generation module extracts behavioral feature signals, contextual environment signals, and target attribute signals in parallel based on the initial spatiotemporal data stream, and generates the initial dynamic risk score through weighted summation and time-series accumulation calculation. When the initial dynamic risk score exceeds the preset first-level warning threshold, the composite instruction data packet generation module calls the intervention-perception dual-modal instruction generation unit to generate a composite instruction data packet containing intervention content instructions and posterior perception strategy instructions. The stimulus feedback data stream generation module drives the smart speaker to perform active auditory stimulation according to the composite instruction data packet, and simultaneously adjusts the parameters of the visible light camera to enter the feedback capture mode, and collects the target reaction video to generate the stimulus feedback data stream. The feedback quantization feature signal generation module analyzes the micro-expressions and instantaneous postures in the stimulus feedback data stream, extracts and quantifies the biological features that characterize the subtle movements of the target face and limbs, and generates feedback quantization feature signals. The risk update and handling decision module incorporates the feedback quantitative characteristic signals into the calculation model of the initial dynamic risk score, generates the updated dynamic risk score, and performs closed-loop handling based on the updated dynamic risk score, such as restoring silence, progressive questioning, or manual intervention.

2. The fireproof electronic fence integrated safety management system based on multi-source sensing and intelligent linkage according to claim 1, characterized in that: The specific process of aligning continuous visible light images with spatiotemporal trajectory data to generate an initial spatiotemporal data stream includes: The inter-frame difference algorithm is applied to the raw thermal image frames acquired by the thermal imaging camera to identify regions of pixel value variation and perform connected component analysis. When the average temperature of the connected domain is higher than the preset moving heat source detection threshold and the displacement exceeds the preset motion judgment distance, it is confirmed as a moving heat source signal. Using a pre-calibrated coordinate mapping model, the two-dimensional coordinates of the moving heat source signal are converted into the pan-tilt control parameters of the visible light camera, the target is locked and a real-time target tracking algorithm is started, and continuous visible light images are acquired during the tracking process; Geometric projection operations are used to convert the position of the target in the continuous visible light image into latitude and longitude coordinates in the geodetic coordinate system to generate spatiotemporal trajectory data. The continuous visible light image and the spatiotemporal trajectory data are time-stamped and the image matrix, latitude and longitude coordinates and motion vectors are encapsulated into an initial spatiotemporal data stream.

3. The fireproof electronic fence integrated safety management system based on multi-source sensing and intelligent linkage according to claim 1, characterized in that: The specific process of extracting behavioral feature signals, contextual environment signals, and target attribute signals in parallel based on the initial spatiotemporal data stream includes: The velocity, acceleration, and rate of change of heading angle of continuous coordinate points in the initial spatiotemporal data stream are calculated and matched with a preset behavior pattern rule base to generate behavior feature signals. The geographic coordinates are submitted to the geographic information system to obtain the land use type, the timestamp is compared with the meteorological database to obtain wind and humidity data, and the contextual environmental signal is generated by mapping through a multidimensional environmental risk lookup table. Continuous visible light images are input into the target detection model to identify the category labels of the items carried by the target, and the data is quantified according to a preset item risk level table to generate target attribute signals.

4. The fireproof electronic fence integrated safety management system based on multi-source sensing and intelligent linkage according to claim 3, characterized in that: The specific process of generating the initial dynamic risk integral value through weighted summation and time-series accumulation includes: A risk score buffer is established for the target. Based on a preset weight matrix, the behavioral feature signal, the contextual environment signal, and the target attribute signal are weighted and summed to obtain the instantaneous risk score. An exponential moving average algorithm is used to update the instantaneous risk score over time, thereby generating an initial dynamic risk score that changes continuously over time.

5. The fireproof electronic fence integrated safety management system based on multi-source sensing and intelligent linkage according to claim 1, characterized in that: The specific process of invoking the intervention-perception bimodal instruction generation unit to generate a composite instruction data packet containing intervention content instructions and posterior perception strategy instructions includes: The intervention content instruction encapsulates the device identifier of the smart speaker near the target and the soft dissuasive voice content to be played. The posterior perception strategy instruction encapsulates the sensor parameter adjustment plan for the visible light camera. The sensor parameter adjustment plan includes a trigger time stamp synchronized with the end of voice playback, a frame rate parameter for switching to a high frame rate acquisition mode, and zoom parameters for zooming in on specific areas of the target.

6. The fireproof electronic fence integrated safety management system based on multi-source sensing and intelligent linkage according to claim 5, characterized in that: The specific process of acquiring target response video to generate stimulus feedback data stream includes: The intelligent sound column analyzes the intervention instructions and executes voice playback, providing active auditory stimulation to the target. The visible light camera interprets the a posteriori perception strategy instructions and instantaneously modifies the register value of the image signal processor when the internal clock reaches the trigger time mark; Configure the frame rate parameter to a preset high frame rate value and drive the lens to zoom, enter the feedback capture mode to capture high frame rate close-up video, and encapsulate the captured high frame rate close-up video into a stimulus feedback data stream with a specific feedback identifier.

7. The fireproof electronic fence integrated safety management system based on multi-source sensing and intelligent linkage according to claim 4, characterized in that: The specific process of analyzing micro-expressions and instantaneous postures in the stimulus feedback data stream includes: The facial detection algorithm is called to extract facial image sequences from the stimulus feedback data stream, and the micro-expression recognition model is input to analyze the activation combination of facial action units to obtain micro-expression features; The instantaneous posture analysis model is called to extract key human skeletal points, and the position and velocity changes of the skeletal points within a short time window after the stimulus occurs are calculated to obtain the body posture characteristics. The intention mapping rule engine is used to make logical judgments on micro-expression features and body posture features to evaluate the subconscious reaction of the target after receiving active auditory stimulation.

8. The fireproof electronic fence integrated safety management system based on multi-source sensing and intelligent linkage according to claim 7, characterized in that: The specific method for extracting and quantifying the biometric features representing the subtle facial and limb movements of the target, and generating feedback quantified feature signals is as follows: the generated feedback quantified feature signals are vectors containing compliance, alertness, and provocation. Among them, compliance is positively correlated with the target body turning away from the restricted area and no negative emotional action units are detected; alertness is positively correlated with the activation intensity of action units representing surprise or fear and the sudden stiffening of body posture; and provocation is positively correlated with the activation intensity of action units representing contempt or anger and the maintenance of a confrontational posture.

9. The fireproof electronic fence integrated safety management system based on multi-source sensing and intelligent linkage according to claim 8, characterized in that: The specific process for generating the updated dynamic risk score includes: Based on the internal feedback weight configuration, the compliance, alertness and provocation in the feedback quantization feature signal are weighted and calculated to generate a scalar feedback score. The scalar feedback score is multiplied by a preset high-weight coefficient as an incremental term and added to the instantaneous risk score calculation model; Based on the instantaneous risk score including feedback increments, the updated dynamic risk score, which has been verified by feedback, is calculated using the exponential moving average algorithm.

10. The fireproof electronic fence integrated safety management system based on multi-source sensing and intelligent linkage according to claim 9, characterized in that: The specific process of performing closed-loop handling based on the updated dynamic risk score, including restoring silence, progressive questioning, or manual intervention, includes: If the updated dynamic risk score drops below the preset risk resolution threshold, the silent tracking mode will be restored. If the updated dynamic risk score exceeds the preset deterrence threshold, return to the compound instruction data packet generation module to generate a compound instruction data packet containing a stronger level of voice and targeted perception strategies, forming a progressive interrogation. If the updated dynamic risk score reaches the preset maximum threshold for human intervention, the initial spatiotemporal data stream and stimulus feedback data stream are aggregated to generate a comprehensive evidence file, which is then pushed to the human agent, and the two-way intercom module of the smart speaker is activated.

Images