Campus privacy area multi-sensor hierarchical cooperative behavior recognition method and system

Abstract

The application discloses a kind of campus privacy area multi-sensor layered cooperative behavior identification method and system, belong to multi-sensor fusion identification field.The method includes: infrared pyroelectric sensor low power value guard, personnel confirmation and point cloud generation are carried out after awakening millimeter wave radar by triggering;Stable target trajectory and number estimation are obtained by clustering and tracking to point cloud;According to target quantity, space close degree determines multi-person gathering;In the condition that personnel exist or gather, start smoke sensing unit, identify smoking behavior based on dynamic background increment;Overactive behavior is identified based on the micro-doppler energy, relative motion and trajectory intersection characteristics of millimeter wave radar;Establish time sequence memory decay accumulation to each type of candidate decision, output structured alarm.The system includes infrared pyroelectric sensor, millimeter wave radar, smoke sensing unit, edge processing unit, communication unit and power management unit.The application realizes privacy area multi-person gathering, smoking and overactive behavior identification under the premise of not collecting image, and considers low power consumption and high reliability.

Description

Technical Field

[0001] This invention belongs to the field of multi-sensor fusion recognition technology, and in particular relates to a multi-sensor hierarchical collaborative behavior recognition method and system for campus privacy areas. Background Technology

[0002] Campus restrooms, dormitory corridor corners, and areas around changing rooms are characterized by their high degree of concealment, unpredictable personnel flow, and strong privacy requirements. If incidents such as illegal smoking, abnormal gatherings of multiple people, or physical altercations occur, relying solely on manual patrols can easily lead to delayed detection, untimely response, and difficulty in tracing the cause of the incident.

[0003] Traditional cameras can provide rich image information, but they are not suitable for deployment in private areas such as bathrooms and dormitories. Using a single non-visual sensor also struggles to simultaneously achieve low power consumption, people counting, smoking detection, and the identification of aggressive behavior. For example, infrared pyroelectric sensors are suitable for sensing changes in human body heat radiation, but they are insensitive to stationary individuals, the number of people, and specific behavioral semantics, unable to distinguish between normal passage and abnormal gatherings, let alone identify smoking or aggressive behavior. Smoke detectors or particulate sensors are suitable for sensing changes in air composition, but they are easily affected by moisture, cleaning agent sprays, ventilation conditions, and background drift, resulting in a high false alarm rate when used alone, and they cannot provide evidence of the presence or nature of people's behavior. Millimeter-wave radar can provide human motion information such as distance, speed, angle, and micro-Doppler, which can be used for people tracking and behavior analysis, but long-term, continuous operation increases power consumption and edge computing load, making it unsuitable for battery-powered or low-power deployment scenarios.

[0004] Existing technologies lack a layered, collaborative solution that can comprehensively address low-power monitoring, human-triggered wake-up, people counting, multi-person gathering detection, smoking detection, and aggressive behavior recognition without image acquisition. Single-sensor solutions often only solve part of the problem: infrared pyroelectric sensors can only detect moving people and cannot meet the needs of detecting stationary people and recognizing behavior; smoke detectors can only detect changes in air composition and cannot correlate them with human behavior; millimeter-wave radar, while powerful, has high power consumption and is not suitable for continuous operation. In addition, existing solutions typically lack a time-series fusion mechanism for non-visual multi-source information in privacy areas, leading to frequent instantaneous false alarms or missed alarms and difficulty in outputting stable and reliable event alarms.

[0005] Therefore, a hierarchical collaborative method suitable for campus privacy areas is needed, which enables low-power sensors to perform wake-up gating, millimeter-wave radar to perform human target tracking and motion interaction analysis when necessary, smoke detectors to provide environmental evidence in the presence or gathering of people, and outputs stable event results through end-side time-series fusion, thereby achieving the identification of multiple people gathering, smoking and excessive behavior without collecting images. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a multi-sensor layered collaborative behavior recognition method and system for campus privacy areas. This method solves the problem that a single sensor is difficult to simultaneously handle low-power monitoring, multi-person gatherings, smoking, and excessive behavior recognition when cameras cannot be used in privacy areas by using a layered collaborative approach of infrared pyroelectric sensors, millimeter-wave radar, and smoke detectors.

[0007] To achieve the above objectives, the present invention provides the following technical solution: The multi-sensor hierarchical collaborative behavior recognition method for campus privacy areas provided by this invention includes the following steps: S1. Low-power monitoring and personnel confirmation trigger: The system collects the output signal of the infrared pyroelectric sensor. When no one is present or no confirmation is received, the infrared pyroelectric sensor remains on standby and the edge processing unit is in a low-power state. When the infrared pyroelectric signal is triggered, the control system changes from the low-power monitoring state to the personnel confirmation state. S2. Millimeter-wave radar startup and point cloud generation: The millimeter-wave radar is started under personnel confirmation. Static background suppression, range and velocity domain target detection, angle estimation and coordinate transformation are performed on the radar echo to form a target point cloud containing spatial position, motion speed and reflection intensity. S3. Target Clustering and Tracking: Perform target clustering and tracking on the target point cloud to obtain stable target trajectories and population estimation results; S4. Determination of Multiple People Gathering: Determine whether multiple people are gathering based on the number of stable targets, the spatial density between targets, and the duration. S5. Smoking behavior recognition: When there are people or multiple people gathered, the smoke detection unit enters the corresponding sampling state and identifies smoking behavior based on the dynamic background increment, rate of change and continuity of the smoke detection output. S6. Overreaction Recognition: Overreaction is identified by utilizing the high-frequency micro-Doppler energy characteristics, relative motion characteristics, acceleration fluctuation characteristics, and trajectory intersection characteristics generated by millimeter-wave radar. S7. Temporal Memory and Structured Alarms: Temporal memory is established for candidate decisions of multiple people gathering, smoking and excessive behavior. Instantaneous decisions are transformed into stable event outputs, and structured alarms are generated according to preset risk priorities, outputting the corresponding event type, number of people summary and alarm level.

[0008] Furthermore, the low-power state mentioned in step S1 includes: the edge processing unit is in a low-power or periodic health check state; when the infrared pyroelectric signal shows a change that matches the characteristics of human entry, the edge processing unit initiates the personnel confirmation process; the infrared pyroelectric sensor, as a first-level gate sensor, uses its judgment result to wake up the main control or high-power sensing module, and is not output separately as the final abnormal behavior result.

[0009] Furthermore, step S3, which involves target clustering and tracking of the target point cloud, specifically includes: constraining the monitoring area of ​​the radar point cloud, removing outliers, and clustering targets; combining data association, filtering prediction, and trajectory lifecycle management to form a stable target trajectory; and counting the number of currently valid targets.

[0010] Furthermore, the determination of multiple people gathering in step S4 is based on at least two of the following: the number of stable target trajectories, the average distance between targets, the area occupied by the target set, and the duration of the state; the candidate results of multiple people gathering serve as one of the conditional information for subsequent smoking identification and aggressive behavior identification.

[0011] Furthermore, in step S5, the smoke sensing unit includes one or more of a particulate matter sensor, a smoke sensor, or a volatile gas sensor; the smoking behavior recognition calculates the concentration change based on the dynamic background and outputs the smoking confidence level in combination with the person's status. Let the concentration or equivalent output of the smoke sensor at time t be... The historical background baseline is as follows: ; ; In the formula, The dynamic background baseline representing the smoke detector output; Indicates the length of the smoke detector background window; Indicates the current smoke concentration or equivalent output; Indicates historical smoke sensor output; Indicates the relative background increment; This indicates the rate of change between adjacent time points.

[0012] Furthermore, the overreaction recognition described in step S6 does not use image or video information, but constructs a behavior feature sequence based on the temporal point cloud and velocity spectrum statistical features of millimeter-wave radar, and outputs the overreaction confidence level through rule discrimination or lightweight classification model; The behavioral feature sequence includes one or more of the following: high-frequency micro-Doppler energy, relative proximity velocity, acceleration variance, and number of rapid trajectory crossings; Among them, the high-frequency micro-Doppler energy is obtained by integrating the slow-time radar echo signal in the high-frequency band after performing a short-time Fourier transform, and the relative proximity velocity vrel is calculated based on the position vectors and velocity vectors of the two targets: ; ; In the formula, This represents the relative approach velocity characteristics between the i-th and j-th stable targets; , Represents the target position vector; , Represents the target velocity vector; This indicates the avoidance of tiny positive quantities with a denominator of zero; Indicates the intensity of the target interaction; The target pair weights can be determined by target distance, signal-to-noise ratio, or trajectory confidence.

[0013] Furthermore, in step S7, establishing temporal memories for candidate decisions regarding multi-person gatherings, smoking, and aggressive behavior includes: setting an independent memory window for each event type e, accumulating candidate results according to time decay weights, with the accumulated value being: ; ; In the formula, e represents the event type. For large gatherings, For smoking, This is considered an excessive act. This represents the time decay weight of event e at the i-th historical position; Indicates the attenuation factor; This indicates the length of the memory window for event e; This represents the candidate outcome of event e at a given historical moment. This represents the accumulated memory value for event e; This indicates the stable output of event e; This represents the stable output threshold for event e; j is the index for weighted normalized summation.

[0014] Furthermore, in step S7, generating structured alarms according to preset risk priorities specifically involves classifying alarm levels from high to low based on the risk priorities of aggressive behavior, smoking behavior, and large gatherings; when multiple events occur simultaneously, the alarm with the highest priority is the main alarm, along with summary information of other events; the edge processing unit only uploads the event type, timestamp, number of people summary, sensor summary, and alarm level, without uploading the original image data.

[0015] The campus privacy area multi-sensor hierarchical collaborative behavior recognition system provided by this invention includes: Infrared pyroelectric sensor, millimeter-wave radar sensor, smoke detection unit, edge processing unit, communication unit, and power management unit; Among them, the infrared pyroelectric sensor is used for low-power monitoring and detection of human body entry trigger, and its output is connected to the wake-up input of the edge processing unit; Millimeter-wave radar sensors are used to activate under personnel confirmation, generate target point clouds, and perform multi-target tracking and motion interaction analysis; The smoke detection unit is used to sample and output smoke or particulate matter concentration data in the presence of people or when multiple people are gathered. The power management unit controls the enabling of high-power modules based on the state machine of the edge processing unit; The edge processing unit is configured to perform the method according to any one of claims 1 to 8.

[0016] Furthermore, the millimeter-wave radar sensor adopts a frequency-modulated continuous wave radar system or an equivalent radar system, with an operating frequency band of 24-80 GHz; The smoke detection unit employs one or more of a particulate matter sensor, a smoke sensor, or a volatile gas sensor; The edge processing unit includes a main control chip and an optional coprocessor. The main control chip is responsible for the state machine, sensor scheduling, data caching and communication reporting, while the coprocessor is responsible for radar feature classification or lightweight model inference. The communication unit wirelessly uploads the structured alarms output by the edge processing unit to the management platform.

[0017] The beneficial effects of this invention are as follows: The present invention provides a multi-sensor hierarchical collaborative behavior recognition method and system for campus privacy areas. This invention discloses a multi-sensor hierarchical collaborative technology for identifying multiple gatherings, smoking, and aggressive behaviors in campus privacy areas, integrating infrared pyroelectric sensors, smoke detectors, and millimeter-wave radar. This technology belongs to the field of non-visual security perception in campus privacy areas. The technology uses an infrared pyroelectric sensor for low-power personnel entry triggering, millimeter-wave radar for generating human target point clouds, tracking multiple people, estimating the number of people, and analyzing motion interactions under image-less conditions, and a smoke detector for confirming changes in particulate matter or gas related to smoking. An edge processing unit outputs event results according to the "personnel triggering - radar confirmation - gathering determination - specific identification - temporal fusion" chain. This method does not collect identifiable images. It performs hierarchical scheduling and confidence fusion of radar motion characteristics, environmental change characteristics, and sensor status characteristics at the edge, enabling graded early warning of risks related to multiple gatherings, smoking, and aggressive behaviors in campus restrooms, dormitories, and other privacy areas, while balancing privacy protection, low-power operation, and stability in abnormal event identification. Compared with existing technologies, the beneficial effects of this method specifically include the following: 1. Privacy protection: The system does not collect optical images or videos, but only processes pyroelectric trigger signals, smoke detection environmental quantities, millimeter-wave radar point clouds, and motion statistics, making it suitable for privacy-sensitive areas such as campus restrooms.

[0018] 2. Low power consumption operation: By using infrared pyroelectric primary gating, millimeter-wave radar on-demand wake-up, and smoke detection conditional sampling, the high-power sensing and inference process is limited to the presence of personnel or abnormal candidate scenarios, thereby reducing the overall energy consumption of the system.

[0019] 3. Multimodal complementarity: Human motion information and smoke-sensing environmental information are mutually verified, which can distinguish between multiple people passing by normally, short-term water vapor disturbance, smoking and risk of excessive behavior, and improve recognition stability; based on the dynamic background of smoke sensors, smoking behavior is identified incrementally, reducing false alarms due to environmental interference.

[0020] 4. Edge Deployable: The algorithm takes point cloud, trajectory, and statistical features as its main inputs and can be executed on edge processing units or coprocessing units, reducing the transmission of raw sensitive data; it achieves stable people estimation under image-less conditions by generating and clustering point clouds using millimeter-wave radar; it determines multi-person gatherings based on the number of targets and their spatial density, providing conditions for subsequent behavior recognition; and it uses micro-Doppler energy and trajectory interaction features to identify aggressive behavior, avoiding privacy collection.

[0021] 5. Clear event classification: The system can output warnings for multiple people gathering, smoking, and high-risk excessive behavior, which facilitates the campus management platform to handle events in a coordinated manner according to event type; the establishment of time-series memory decay accumulation transforms instantaneous judgments into stable events, improves output reliability, and ultimately achieves multi-behavior hierarchical collaborative identification while protecting privacy.

[0022] The above and other objects, advantages, and features of the present invention will be more fully set forth and demonstrated through the following detailed description of specific embodiments in conjunction with the accompanying drawings. Those skilled in the art, upon referring to the following detailed description and the accompanying drawings, will be able to better understand and realize the above advantages of the present invention. Other objects, features, and advantages of the present invention will become clearer after being described in detail in the detailed description section in conjunction with the accompanying drawings. Attached Figure Description

[0023] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following drawings are provided for illustration.

[0024] Figure 1 Flowchart of a multi-sensor hierarchical collaborative behavior recognition method for campus privacy areas; Figure 2 This is a schematic diagram of the system architecture; Figure 3 This is a schematic diagram of the hardware structure; Figure 4 This is a schematic diagram of the installation scenario; Figure 5 A flowchart for multi-sensor hierarchical collaboration; Figure 6 This is a schematic diagram of the on / off output algorithm for an infrared pyroelectric sensor. Figure 7 Point cloud results for dual-task identification of millimeter-wave radar focusing and over-aggressive behavior; Figure 8 This is a graph showing the time series and dynamic threshold alarm results of PM2.5 concentration under smoking and interference scenarios; Figure 9 This is a schematic diagram showing the results of the monitoring of personnel gathering on the host computer. Figure 10 A schematic diagram showing the results of smoking detection by the host computer. Figure 11 This is a schematic diagram showing the detection results of excessive behavior by the host computer. Detailed Implementation

[0025] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0026] Example 1

[0027] like Figure 1 As shown, Figure 1 This document presents a flowchart of a multi-sensor hierarchical collaborative behavior recognition method for campus privacy areas. It integrates infrared pyroelectric sensors, smoke detectors, and millimeter-wave radar to collaboratively identify gatherings of multiple people, smoking, and aggressive behaviors within campus privacy areas. The main principle is to use an infrared pyroelectric sensor as a low-power gating mechanism to activate the millimeter-wave radar for personnel confirmation and behavior analysis. Based on this, the smoke detector unit is conditionally activated. Through end-side multi-source data fusion and temporal memory, non-visual, low-power, and highly reliable identification of specific abnormal behaviors (gatherings of multiple people, smoking, and aggressive behaviors) within the privacy area is achieved.

[0028] This embodiment uses an edge computing node deployed in private areas such as campus restrooms and dormitory corridors as an example. The node integrates an infrared pyroelectric sensor, a millimeter-wave radar sensor, a smoke detection unit, an edge processing unit (e.g., an embedded system containing an ARM Cortex-M core main control chip and an optional AI coprocessor), a communication unit (such as a Wi-Fi or 4G Cat.1 module), and a power management unit.

[0029] The multi-sensor hierarchical collaborative behavior recognition method for campus privacy areas provided in this embodiment specifically includes the following steps: Step S1: Low-power monitoring and personnel confirmation triggering In the default state of no human activity, the system maintains an extremely low-power, passive mode. In this mode, only the infrared pyroelectric sensor is active, while the main control chip of the edge processing unit enters a low-power state of deep sleep or periodic health checks (such as waking up every 10 seconds to send a heartbeat packet), or enters a periodic health check state.

[0030] The infrared pyroelectric sensor continuously monitors the environment. When a person enters its detection range, the sensor outputs a signal with a level change consistent with the characteristics of human entry. This signal serves as a wake-up interrupt, directly connected to the wake-up input of the edge processing unit, triggering the main control chip to recover from a low-power state and initiate a complete personnel verification process.

[0031] It is important to emphasize that the infrared pyroelectric sensor plays only the role of a first-level gated sensor in this invention. Its judgment result is only used to wake up a higher power consumption sensing module (such as millimeter-wave radar), and will not be output separately as the basis for judging any final abnormal behavior (such as gathering, smoking, or excessive behavior). This fundamentally avoids false alarms caused by heat source interference from a single infrared sensor.

[0032] Step S2: Millimeter-wave radar startup and point cloud generation After the personnel confirmation status is activated, the edge processing unit immediately starts the millimeter-wave radar sensor through the power management unit. In this embodiment, the millimeter-wave radar adopts a frequency-modulated continuous wave (FMCW) system and operates at a frequency of 60 GHz to obtain high-resolution distance and velocity information.

[0033] After the radar is activated, the edge processing unit performs a series of end-side signal processing steps on the received radar echo data: Static background suppression: By subtracting the long-term averaged clutter map, strong reflection interference from stationary objects such as walls and pipes is filtered out.

[0034] Distance-Doppler processing: Perform distance-dimensional FFT (Fast Fourier Transform) and Doppler-dimensional FFT on each chirp sequence to generate a distance-Doppler map (RD Map).

[0035] Target detection: A constant false alarm rate (CFAR) detector is applied to the RD map to adaptively extract the distance and velocity cells of moving targets.

[0036] Angle estimation and coordinate transformation: For detected target points, angle estimation is performed using the phase difference of multiple receiving antennas (e.g., using amplitude-ratio monopulse or Capon beamforming algorithms) to calculate the azimuth and elevation angles of each target point. Finally, each detected point is transformed from the radar polar coordinate system (range, angle) to a three-dimensional Cartesian coordinate system (x, y, z), and combined with its Doppler velocity (v) and reflection intensity (amp) to form a structured target point cloud containing spatial position, motion velocity, and reflection intensity.

[0037] Step S3: Target Clustering and Tracking After obtaining the raw point cloud data, the edge processing unit executes target clustering and tracking algorithms to obtain stable and reliable personnel target trajectories and headcount estimates. The specific process is as follows: Region filtering and outlier removal: First, the point cloud range is constrained according to a preset monitoring area (e.g., a semi-circular area with a radius of 5 meters centered on the radar), and points outside the area are filtered out. Then, isolated noise points are removed through statistical filtering (e.g., outlier removal based on point cloud density).

[0038] Target clustering: The density-based spatial clustering algorithm (DBSCAN) is used on the remaining point cloud to aggregate points that are close to each other in spatial location into a candidate target cluster, and each cluster represents a potential human body.

[0039] Data association and filtering tracking: The state of each tracked target in the next frame is predicted using a Kalman filter, and then the candidate target cluster detected in the current frame is optimally matched with the tracking trajectory of the historical frames using the Hungarian algorithm (data association).

[0040] Track management: Each track maintains a lifecycle, including three states: "Tentative," "Confirmed," and "Terminated." Only targets that can be successfully associated with each other for multiple consecutive frames (e.g., 3 frames) will have their tracks promoted to "Confirmed" and included in the final count of valid targets. New clusters that appear briefly and cannot be associated with any existing tracks are considered clutter and their tracks are terminated if they disappear within several consecutive frames. This strict track management mechanism effectively suppresses erroneous counts caused by transient clutter.

[0041] Step S4: Determine if multiple people gather Based on the stable target trajectory obtained in step S3, the edge processing unit performs a multi-person clustering determination. This determination is not based on a single threshold comparison, but rather integrates information from multiple dimensions: Number of stable targets: The number N of "steady-state" targets currently being tracked.

[0042] Spatial compactness between targets: Calculate the average Euclidean distance Davg between all steady-state targets and the area A occupied by the convex hull formed by all target point clouds.

[0043] Duration: The above states of "N is greater than the preset threshold" (e.g., N≥3) and "Davg is less than the distance threshold" (e.g., 1.5 meters) or "A is less than the area threshold" (e.g., 8 square meters) need to last for a certain period of time (e.g., more than 5 seconds).

[0044] When all the above conditions are met, the edge processing unit determines that a "multi-person gathering candidate event" has been generated. This candidate event itself can be reported, and more importantly, it will serve as one of the conditions for initiating more refined behavior analysis (smoking, excessive behavior) to achieve hierarchical and on-demand scheduling of sensor resources.

[0045] Step S5: Smoking Behavior Recognition To further reduce power consumption and false alarms, the smoke detection unit is not always operational. Instead, when the presence of people (or multiple people gathering) is confirmed in step S1 or S4, the power management unit puts it into a high-frequency sampling state.

[0046] In this embodiment, the smoke detection unit employs a module integrating a particulate matter sensor (such as a laser PM2.5 sensor) and a volatile gas sensor (such as a TVOC sensor). The smoking behavior recognition algorithm is based on a dynamic background baseline. First, the edge processing unit outputs the smoke sensor value. Establish an average value for a sliding window (e.g., window size of 30 seconds) as a dynamic background baseline. The average value. Then, two key features are calculated in real time: Relative background increment: Used to detect sudden increases in smoke or particulate matter concentration.

[0047] Rate of change: , which is the difference between adjacent sampling times, is used to characterize the steepness of the concentration increase.

[0048] when and If a person continuously exceeds their preset threshold within a short period of time (e.g., within 10 seconds), and the person's status is "present" or "gathered", the system will output a high-confidence candidate judgment for smoking behavior.

[0049] Step S6: Identification of Overly Aggressive Behaviors The identification of aggressive behaviors (such as pushing, fighting, and sprinting) is entirely based on temporal point cloud and velocity spectrum information from millimeter-wave radar, without using any image or video information, thus fundamentally protecting campus privacy. The edge processing unit constructs a series of behavioral feature sequences from the tracking data: High-frequency micro-Doppler energy: The micro-Doppler spectrum is obtained by performing a short-time Fourier transform (STFT) on the slow-time radar echo signal of a selected target. High-frequency micro-Doppler energy is the energy integral of this spectrum within a high-frequency band (e.g., 50-200 Hz). Actions such as punching and kicking generate significant high-frequency micro-Doppler energy.

[0050] Relative approach velocity: To quantify the degree of conflict between two targets (such as two people), the relative approach velocity vrel is calculated. ; ; In the formula, This represents the relative approach velocity characteristics between the i-th and j-th stable targets; , Represents the target position vector; , Represents the target velocity vector; This indicates the avoidance of tiny positive quantities with a denominator of zero; Indicates the intensity of the target interaction; The target pair weights can be determined by target distance, signal-to-noise ratio, or trajectory confidence.

[0051] Acceleration variance: Calculates the variance of the acceleration of a single target within the tracking window. Overly aggressive behavior is usually accompanied by drastic changes in the direction and magnitude of motion, leading to a sharp increase in acceleration variance.

[0052] Rapid trajectory intersection: This counts the number of times two or more stable trajectories intersect within a short period (e.g., within 2 seconds). Frequent and rapid trajectory intersections are a typical characteristic of brawling or conflict among multiple people.

[0053] The edge processing unit combines these feature values ​​into a feature vector, which is then input into a pre-trained lightweight classification model (such as a random forest or a very small neural network). The model outputs an "overly aggressive behavior confidence score" between 0 and 1 in real time. When the confidence score exceeds a preset threshold, a candidate decision for the overly aggressive behavior is output.

[0054] Step S7: Temporal Memory and Structured Alarms In order to transform unstable decisions caused by single-frame or short-term fluctuations into reliable and smooth event outputs, this invention establishes independent temporal memory windows for the three event types of "multiple people gathering", "smoking" and "excessive behavior", and designs a confidence accumulation mechanism based on time decay.

[0055] Specifically, a memory window of length Me is set for each event type e. For each historical time ti within the window, the candidate decision result ye(ti) (with a value of 0 or 1) is assigned a weight ω{e,i} =exp(-λe i) that decays exponentially, where λe is the decay rate of that event type, and all weights are normalized. The cumulative value ηe(t) at time t is calculated as follows: ; ; In the formula, e represents the event type. For large gatherings, For smoking, This is considered an excessive act. This represents the time decay weight of event e at the i-th historical position; Indicates the attenuation factor; Indicates the length of the memory window for event e; This represents the candidate outcome of event e at a given historical moment. This represents the accumulated memory value for event e; This indicates the stable output of event e; This represents the stable output threshold for event e; j is the index for weighted normalized summation.

[0056] When the cumulative value Reach the preset stable output threshold for this event type Only then will the system finally output a stable event. =1. For example, even if a frame is misjudged as "excessive behavior" due to violent motion, no final alarm will be generated if the cumulative value of subsequent frames does not reach the threshold, thus effectively reducing false alarms caused by radar point cloud jitter, normal violent movements, or short-term environmental disturbances.

[0057] Finally, the edge processing unit generates structured alarm information based on preset risk priorities. In this invention, the risk priorities, from highest to lowest, are: aggressive behavior, smoking behavior, and large gatherings. When multiple stable events occur simultaneously, the alarm information prioritizes the highest priority event, accompanied by summary information of other concurrently occurring events. The generated structured alarm is a JSON data packet containing: event type (e.g., "fight"), timestamp, a summary of the number of people (e.g., "3 people"), and brief sensor data (e.g., "PM2.5: 350μg / m³"). 3 The data packet is displayed as "Radar: Overreaction Confidence 0.95" and an alarm level (e.g., "High Risk"). The packet is uploaded to the campus management platform via the communication unit, without uploading any original image or video data, fully complying with campus privacy regulations.

[0058] This embodiment also provides a multi-sensor hierarchical collaborative system for recognizing multiple gatherings, smoking, and aggressive behaviors in campus privacy areas, integrating infrared pyroelectric sensors, smoke detectors, and millimeter-wave radar to implement the above-mentioned methods. The system includes: Infrared pyroelectric sensor: used for low-power monitoring, its output is directly connected to the wake-up pin of the edge processing unit.

[0059] Millimeter-wave radar sensor: Employs FMCW technology and operates in the 60GHz band. It activates upon personnel confirmation and is used to generate target point clouds and perform multi-target tracking and motion interaction analysis.

[0060] Smoke Detector Unit: Integrates particulate matter sensor and TVOC sensor. Under the control of edge processing unit, it enters high sampling rate mode when people are present or multiple people gather.

[0061] Edge processing unit: Comprising a main control chip (such as the STM32 series Cortex-M MCU) and an optional coprocessor (such as GAP8 or K210 for running a lightweight radar behavior classification model). The main control chip is responsible for system state machine management, multi-sensor scheduling, data caching, and structured alarm reporting; the coprocessor undertakes computationally intensive tasks such as radar micro-Doppler feature extraction and classification model inference.

[0062] Communication unit: Uploads the structured alarms output by the edge processing unit to the cloud or a local management platform via Wi-Fi or 4G LTE Cat.1 wireless means.

[0063] Power management unit: Controlled by the state machine of the edge processing unit, it precisely controls the power supply and power-off of high-power consumption modules such as the millimeter-wave radar and the smoke sensing unit according to the current working phase (on-duty, personnel confirmation, behavior analysis), so as to realize the overall low-power consumption operation of the system.

[0064] Embodiment 2

[0065] The specific embodiments of the present invention are described below with reference to the accompanying drawings. The following embodiments are used to explain the technical principles and algorithm processes of the present invention, and are not intended to limit the protection scope of the present invention. All thresholds, window lengths, model coefficients and sampling periods can be calibrated or adaptively adjusted according to the spatial structure of the privacy area, ventilation conditions and management requirements, and the description does not limit specific numerical values.

[0066] As shown in Figure 2 the figure, the system of this embodiment includes a perception layer, an edge processing layer and an application alarm layer. The perception layer is composed of a pyroelectric infrared sensor, a millimeter-wave radar and a smoke sensing unit, which respectively provide personnel entry triggering, human body spatial movement information and smoking-related environment change information. The edge processing layer completes peripheral wake-up, timestamp alignment, task scheduling, aggregation determination, smoking determination, aggressive behavior recognition and time sequence fusion. The application alarm layer receives the structured events uploaded by the edge end, and executes prompting or linkage disposal according to the event type.

[0067] The data flow of the system is not simple parallel acquisition, but hierarchical collaboration. The pyroelectric infrared sensor is used to wake the system from the low-power on-duty state; the millimeter-wave radar is used to confirm personnel, estimate the number of people and analyze the motion relationship between targets; the smoke sensing unit provides smoking evidence under the condition of the presence or aggregation of personnel. Thus, a cross-modal recognition framework of "environmental quantity + human dynamic quantity" is formed.

[0068] As shown in Figure 3As shown, the hardware structure includes an edge processing unit, an infrared pyroelectric sensor, a millimeter-wave radar, a smoke detection unit, a power management circuit, and a communication unit. The output of the infrared pyroelectric sensor is connected to the wake-up input of the edge processing unit; the millimeter-wave radar provides radar measurement information to the edge processing unit through a digital interface; the smoke detection unit provides data related to particulate matter, smoke, or gas concentration through a serial interface; and the power management circuit controls the enabling of high-power modules according to a state machine.

[0069] The edge processing unit can consist of a main control chip and an optional coprocessor. The main control chip is responsible for the state machine, sensor scheduling, data caching, and communication reporting; the coprocessor can handle radar feature classification or lightweight model inference. This architecture enables the system to complete event identification locally, uploading only event summaries and preventing the leakage of private data.

[0070] like Figure 4 As shown, when deployed in campus corridors, infrared pyroelectric sensors can be set at entrances, top public areas, or locations that can cover the path of people entering, to detect changes in thermal radiation when people enter; millimeter-wave radar can be set above the side of the wall or above the public area outside the cubicle, so that it covers the handwashing area, passageway, or public activity space that management allows to be monitored; smoke detection units can be set in locations where the airflow is relatively stable and it is easy to detect the spread of local smoke.

[0071] To facilitate algorithm description, a radar coordinate system and a monitoring area coordinate system are established. Let the position of the i-th reflection point in the radar coordinate system be denoted as . The location in the monitoring area coordinate system is The two are converted through installation posture and positional relationship: ; In the formula, This represents the position vector of the i-th radar reflection point in the radar coordinate system; This represents the position vector of the reflection point in the coordinate system of the monitoring area; This represents the rotation matrix from the radar coordinate system to the monitoring area coordinate system; This represents the translation vector corresponding to the radar installation location; i is the reflection point index. This coordinate transformation is used to unify subsequent clustering, tracking, and region constraints.

[0072] like Figure 5As shown, the system's operating states can include low-power standby state, wake-up initialization state, personnel detection state, anomaly identification state, and alarm state. In low-power standby state, only infrared pyroelectric triggering and necessary health checks are retained. After infrared pyroelectric triggering, the system enters the wake-up initialization state, starts the millimeter-wave radar, and establishes a short-term background; if the radar confirms the presence of a human target within the monitoring area, it enters the personnel detection state; when the number of people, spatial concentration, or movement intensity meets the candidate conditions, it enters the anomaly identification state; when the time-series fusion result is valid, it enters the alarm state.

[0073] This process can be represented by a state transition function as follows: ; In the formula, This represents the system state at time t; Indicates the system state at the next moment; This represents the state transition function implemented by firmware or program; This indicates the result of personnel triggering the infrared pyroelectric sensor; This indicates the number of stable targets estimated by the radar. Indicates the confidence level for smoking; Indicates the confidence level of excessive behavior.

[0074] like Figure 6 As shown, the infrared pyroelectric sensor divides the monitoring range into multiple sensitive areas using a Fresnel lens. When a human body moves within the monitoring range, its infrared thermal radiation sequentially enters and leaves different sensitive areas, causing the intensity of the thermal radiation received by the two sensing units A and B of the dual pyroelectric chip to change sequentially. Because units A and B employ a differential output structure, when the human body heat source moves from one sensitive area to an adjacent sensitive area, the sensor output will generate alternating positive and negative pulse signals, such as... Figure 6 The waveform shown in the lower right corner indicates that there is human movement or heat source change within the monitoring area, but does not directly reflect the number of people, their posture, or specific behaviors.

[0075] Therefore, in this embodiment, the infrared pyroelectric sensor serves as a low-power "first-level wake-up / gating" unit: when there is no human movement, the subsequent detection modules remain in a low-power or sleep state; when the sensor detects a change in human body thermal radiation and generates a trigger signal, other detection modules are then activated, thereby reducing the overall power consumption of the system and improving the standby time and response efficiency of the device.

[0076] Let the preprocessed output of the infrared pyroelectric sensor at time t be... In length of Constructing trigger energy within the sliding window: ; In the formula, This represents the window energy of the infrared pyroelectric signal at time t; This represents the preprocessed output i sampling intervals from the current time. Indicates the window length; i is the sample number within the window.

[0077] To adapt to varying background thermal noise levels in bathrooms, the system sets adaptive trigger thresholds based on the background mean and fluctuation level. ; In the formula, This represents the infrared trigger threshold at time t; This represents the baseline mean of the infrared output within the background window; This represents the standard deviation of the infrared output within the background window; This represents the threshold adjustment coefficient, used to control trigger sensitivity.

[0078] Level 1 gated output is defined as: ; In the formula, This indicates the first-level gated output of infrared pyroelectric control; a value of 1 indicates that the personnel trigger has been established, and a value of 0 indicates that it has not been triggered. Indicates the most recent valid trigger time; The trigger hold time is used to prevent the system from immediately returning to sleep mode after a brief period of stillness following personnel entry; t represents the current time.

[0079] like Figure 7 As shown, after the millimeter-wave radar is activated by the infrared pyroelectric sensor, it analyzes the echo signal in terms of distance, velocity, and angle, generating point cloud data in a three-dimensional spatial coordinate system. The x, y, and z axes represent the target's lateral position, depth, and height distribution within the monitoring area, respectively. The blue and orange point clouds in the figure correspond to the echo clustering results of different human targets, namely "Target 1" and "Target 2." A small number of gray discrete points can be considered as environmental clutter or transient interference points, which can be removed through threshold filtering, clustering algorithms, or trajectory continuity judgment.

[0080] Based on this, the system can determine the number of people in the current area according to the number of point cloud clusters, determine the relative position and distribution state of each target point cloud according to the spatial coordinates of each target point cloud, and obtain the target movement trajectory, movement speed, acceleration, and changes in distance between targets through continuous frame point cloud changes. When multiple targets continuously approach each other in space, stay in the same area, or the distance between targets is less than a set threshold, it can be determined that there are multiple people gathered. When the target point cloud shows characteristics such as violent displacement, sudden speed change, reciprocating collision, or rapid approach and separation between multiple targets in a short period of time, it can be further used as a basis for identifying excessive behavior.

[0081] Let the complex echo of the m-th receiving channel at the fast time sampling point n and the slow time index l be... First, estimate the static background and then perform cancellation: ; In the formula, This represents the discrete complex echo of the m-th channel; n represents the fast-time sampling sequence number; l represents the slow-time or transmission period sequence number; This represents the estimated static background; This represents the number of slow-time samples used for background estimation; This represents the dynamic echo after background cancellation; m represents the receiving channel number.

[0082] The range-velocity spectrum is obtained by performing range and Doppler transformations on the dynamic echo. For the unit to be tested The noise is estimated using local reference cells, and a detection threshold is formed. ; In the formula, Represents distance and velocity units Nearby noise power estimation; This represents the set of reference cells surrounding the detection unit; Indicates the number of reference units; Reference unit The distance-velocity spectrum value; Indicates the adaptive detection threshold; This represents the threshold factor related to false alarm control; r and u represent distance cell indices, and v and w represent velocity cell indices.

[0083] ; In the formula, This represents the candidate target detection result. A value of 1 indicates that there is a candidate reflection point in the corresponding distance and velocity unit, while a value of 0 indicates that no target was detected.

[0084] For frequency-modulated continuous wave radar, the beat frequency, Doppler shift, and array phase difference of candidate reflection points can be converted into target range, radial velocity, and azimuth: ; In the formula, Indicates the target distance; c represents the electromagnetic wave propagation speed; Indicates beat frequency; Indicates the frequency modulation slope; Indicates the radial velocity of the target; Indicates the radar carrier wavelength; Indicates Doppler frequency shift; Indicates azimuth; This indicates the phase difference between the received signals of adjacent array elements; This represents the spacing between array elements. The above variables are used to illustrate the principle and do not limit specific radar parameters.

[0085] Converting polar coordinate measurements into point cloud representation yields the radar point cloud set at time t: ; In the formula, The set of point clouds representing time t; This represents the i-th point cloud reflection point; Represents the number of points in a point cloud; Represents spatial coordinates; Indicates radial velocity or equivalent motion velocity; Indicates reflected energy, signal-to-noise ratio, or intensity characteristics; This represents the transpose of a vector.

[0086] After region filtering of the point cloud, density clustering is used to obtain candidate object clusters. Then, cross-frame correlation and filtering tracking are performed on the centroids of each object cluster. Let the state vector of the k-th track be... The following prediction and update process can be adopted: ; ; ; In the formula, This represents the prior state estimate of the k-th track at time t; This represents the posterior state estimate; Represents the prior error covariance matrix; This represents the posterior error covariance matrix of the previous time step; Represents the state transition matrix; Represents the process noise covariance matrix; Indicates Kalman gain; Represents the observation matrix; Represents the radar measurement noise covariance matrix; This indicates the current measurement associated with the k-th track; the superscript "-" indicates the predicted measurement.

[0087] A target is considered stable when its tracks remain effectively correlated over a continuous period of time. Population estimates can be expressed as: ; In the formula, This represents the number of stable targets at time t; This indicates the total number of tracks currently being maintained; This represents the effective duration or confidence lifetime of the k-th track; Indicates stable flight path conditions; This indicates an indicator function that takes the value 1 if the condition is true and 0 otherwise.

[0088] Let the position vector of the i-th stable target be... The system calculates the average distance and occupied area between targets to determine whether the targets are clustered together. ; ; In the formula, This represents the average distance between stable targets at time t; , Let represent the position vectors of the i-th and j-th stable targets, respectively; Represents the Euclidean norm; Indicates candidate results for multiple people gathering; Indicates the number of people; Indicates the distance condition between targets; This represents the area occupied by the target set or the area of ​​its convex hull. Indicates the area occupied; Indicates the continuous duration of the aggregation state; This indicates a continuous condition. The above conditions may be adjusted based on the size of different campus restrooms and management strategies.

[0089] Aggressive behavior is typically accompanied by rapid approach between targets, velocity broadening, high-frequency limb movements, acceleration fluctuations, and increased trajectory intersections. The system extracts micro-Doppler and point cloud interaction features under conditions of multiple people present or gathered together. Let the radar slow-time signal be... The micro-Doppler representation was obtained using short-time analysis: ; ; In the formula, Indicates time position Micro-Doppler spectrum at the location; This represents a slow-time radar echo sequence; n represents the slow-time sample number. Indicated by The window function centered on; Indicates the short-time analysis window length; j represents the imaginary unit; f represents the Doppler frequency index; Indicates high-frequency motion energy; This represents the set of high-frequency motion bands.

[0090] For any two stable targets i and j, define the relative proximity motion between the targets: ; ; In the formula, This represents the relative approach velocity characteristics between the i-th and j-th stable targets; , Represents the target position vector; , Represents the target velocity vector; This indicates the avoidance of tiny positive quantities with a denominator of zero; Indicates the intensity of the target interaction; The target pair weights can be determined by target distance, signal-to-noise ratio, or trajectory confidence.

[0091] Furthermore, high-frequency motion energy, interaction intensity, acceleration fluctuations, and trajectory cross-statistics are used to construct an overreaction confidence level: ; In the formula, Indicates the confidence level of excessive behavior; This represents the Sigmoid function or its equivalent normalized mapping; This represents the normalized high-frequency motion energy; This represents the normalized interaction strength. This represents the normalized variance of the target acceleration. This represents the number of fast crossovers in the normalized trajectory; to Indicate feature weights; This indicates the bias term.

[0092] The aforementioned weights can be obtained from rule experience, training models, or field calibration.

[0093] ; In the formula, Indicates candidate results for excessive behavior; This represents the confidence threshold for excessive behavior.

[0094] As an alternative implementation of pure rule-based discrimination, the weighting coefficients in the above-described excessive behavior recognition process to and bias terms Instead of requiring training, the algorithm is empirically set based on the correlation between behavioral characteristics and overexcitation (e.g., high-frequency micro-Doppler energy and interaction intensity are given higher weights). In this case, the entire recognition process does not involve training or inference of any classification model; it is completed solely through fixed linear weighted sums, sigmoid mapping, and threshold comparison, achieving completely rule-based discrimination. This embodiment has lower computational requirements and is easier for on-site personnel to intuitively understand and adjust parameters.

[0095] To prevent misjudgments in a single frame caused by normal brisk walking, waving, or opening a door, candidate results also need to undergo sequential memory as described later.

[0096] like Figure 8 As shown, the smoke detection unit enters the sampling state when personnel confirm the presence of smoke or when multiple people gather.

[0097] Let the concentration or equivalent output of the smoke sensor at time t be... The historical background baseline is as follows: ; ; In the formula, The dynamic background baseline representing the smoke detector output; Indicates the length of the smoke detector background window; Indicates the current smoke concentration or equivalent output; Indicates historical smoke sensor output; Indicates the relative background increment; This indicates the rate of change between adjacent time points.

[0098] Combining the individual's condition with changes in smoke perception to form a smoking confidence level: ; ; In the formula, Indicates the confidence level for smoking; This represents the Sigmoid function or its equivalent normalized mapping; This represents the normalized increase in smoke sensation; This represents the normalized rate of change in smoke sensation; This indicates the result of infrared pyroelectric triggering by personnel; Indicates candidate results for multiple people gathering; to Indicates the weight of smoking feature; Indicates the bias term; Indicates the candidate results for smoking; This represents the confidence threshold for smoking.

[0099] When interference such as water vapor or cleaning agent spray only causes short-term changes without the support of personnel status or continuous monitoring, candidate smoking results will not directly form a final alarm, thereby reducing false alarms.

[0100] Specifically, such as Figure 8As shown, after the smoke detection unit is activated, it continuously samples the PM2.5 concentration within the monitoring area and compares the real-time sampled values ​​with the dynamic background threshold. In the figure, the blue curve represents the PM2.5 concentration change under smoking conditions, the orange curve represents the PM2.5 concentration change under water vapor interference conditions, and the green dashed line represents the dynamic alarm threshold updated in real time by the system based on the environmental baseline concentration. It can be seen that under normal conditions, the PM2.5 concentration remains at a low level; when smoking occurs, the PM2.5 concentration rises rapidly in a short period and continuously exceeds the dynamic threshold; while water vapor interference may also cause an increase in PM2.5 values, its peak value is lower, its duration is shorter, and it usually does not form a stable threshold-exceeding process.

[0101] This embodiment does not use a fixed absolute threshold because the baseline PM2.5 concentration in different locations is affected by factors such as ventilation conditions, environmental cleanliness, human activity, air conditioning ventilation, and outside air quality. Setting a fixed threshold directly could easily lead to false alarms when the baseline concentration is high, or missed alarms when ventilation is good and smoke is diluted. Therefore, the system establishes a dynamic background using historical sampling data over a period of time, and sets a relatively high threshold based on this background, allowing smoke detection to adapt to different environmental conditions.

[0102] like Figure 9 As shown, the host computer or management platform can display the personnel gathering detection results. The upper left corner of the interface displays the device's operating status, including information such as "Device Online Status," "Is There Anyone Here," "Number of Persons," "Wi-Fi RSSI," and "Device Running Time." This information is used to determine whether the front-end device is connected to the network normally, whether personnel have been detected entering, and the number of targets in the current monitoring area. In the figure, the number of people is displayed as 3, indicating that the system has identified three human targets in the current area using millimeter-wave radar.

[0103] The central part of the interface displays environmental monitoring data, including parameters such as PM1.0, PM2.5, PM10, CO2, and TVOC, which are used to help determine the on-site air quality and whether there is smoking or abnormal gas accumulation. The alarm status area on the left shows "Personnel Gathering Status" as "Alarm," while the smoke status and excessive behavior status are shown as "Normal," indicating that the main abnormal event currently identified by the system is a gathering of multiple people, and no smoking alarm or excessive behavior alarm has been triggered simultaneously.

[0104] The target coordinate graph on the right side of the interface displays the spatial location of human targets identified by the millimeter-wave radar. Different targets are displayed as points in the radar coordinate system, intuitively reflecting the distance, direction, and distribution of personnel relative to the equipment. The target list below further provides the target number, lateral distance, and longitudinal distance, facilitating accurate identification of personnel locations and their clustering positions by management personnel. The historical data table at the bottom records information such as detection time, upload sequence number, equipment runtime, PIR trigger status, radar status, number of personnel, environmental parameters, and alarm status, for subsequent traceability and statistical analysis.

[0105] In this embodiment, the aggregation events output from the edge do not include images, but only the event type, region number, timestamp, summary of stable participants, aggregation persistence status, and confidence level. To ensure stable display results, the aggregation candidate results need to undergo time-series memory processing. ; In the formula, The amount of time-series memory representing events involving large gatherings; Indicates the length of the clustered memory window; This represents the weight of the i-th historical result in the window; Indicates historical clustering candidate results; This indicates the output of stable clustered events; This represents the stable output threshold for clustered events.

[0106] like Figure 10 As shown, when the smoking candidate result remains valid under the support of personnel status, the system outputs a smoking alarm. The figure shows that the device is currently online, there are personnel in the detection area, and the number of personnel is 2. The values ​​of PM1.0, PM2.5, and PM10 in the figure are all significantly increased, with PM2.5 reaching 131 μg / m³. 3 This indicates a significant change in particulate matter concentration at the scene. As seen in the right-hand image, two people are present within the monitoring area, one of whom is smoking, corresponding to the abnormal particulate matter changes detected by the platform. In the alarm status area, "Personnel Gathering Status" is displayed as normal, "Smoking Monitoring Alarm" is displayed as an alarm, and "Excessive Behavior Alarm" is displayed as normal. This result shows that the system does not trigger alarms solely based on the number of people, but rather, after confirming the presence of people, further combines PM concentration changes collected by the smoke detection unit, dynamic threshold judgment, and a continuous confirmation mechanism to determine that suspected smoking behavior exists in the current scene. Simultaneously, the system did not detect multiple gatherings or excessive behavior, therefore the corresponding statuses remain normal.

[0107] Smoking alarms can include a summary of smoke detection changes, the presence status of individuals, whether it is accompanied by a large gathering of people, and the alarm level. Its stable output can be represented as: ; In the formula, The amount of time-series memory representing the smoking incident; Indicates the length of the smoking memory window; Indicates the weight of smoking memory; This indicates historical smoking candidate results; This indicates a stable smoking event output; This represents the stable output threshold for smoking events.

[0108] like Figure 11 As shown, the host computer or management platform is used to display the real-time detection results in the scenario of identifying aggressive behavior. The device in the image is online, and the system has detected the presence of three people in the current area, indicating that the front-end sensor has completed the identification of the on-site personnel. No significant abnormal increase in particulate matter and gas concentrations is observed in the image; therefore, the system has not classified the current event as smoking or abnormal air quality. In the alarm status area on the left, "Personnel Gathering Status" is displayed as an alarm, "Smoking Monitoring Alarm" is displayed as normal, and "Aggressive Behavior Alarm" is displayed as an alarm. This indicates that, based on the identification of multiple people present simultaneously, the system further determines the risk of abnormal aggressive behavior on-site based on the motion changes of the millimeter-wave radar point cloud and the interaction relationship between targets. Combined with... Figure 11 The scene on the right shows that there were multiple people in close contact, exhibiting abnormal body postures, or making large movements within the monitoring area, corresponding to the "abnormal bullying behavior" and "excessive behavior alarms" displayed on the platform. This result indicates that the system does not rely solely on a single threshold of the number of people, but rather comprehensively utilizes information such as the number of people, the distance between targets, movement speed, posture change trends, and interaction relationships between targets to make a comprehensive judgment on abnormal behavior.

[0109] When outputting alarms for excessive behavior, the edge processing unit can simultaneously include the number of targets, an interaction intensity summary, a high-frequency motion energy summary, and event confidence. To avoid false alarms caused by a single violent action, the candidate results for excessive behavior also need to undergo time decay accumulation. ; ; In the formula, e represents the event type. For large gatherings, For smoking, This is considered an excessive act. This represents the time decay weight of event e at the i-th historical position; Indicates the attenuation factor; Indicates the length of the memory window for event e; This represents the candidate outcome of event e at a given historical moment. This represents the accumulated memory value for event e; This indicates the stable output of event e; This represents the stable output threshold for event e; j is the index for weighted normalized summation.

[0110] When multiple stable events occur simultaneously, this embodiment can output alarms according to risk priority: aggressive behavior takes precedence over smoking behavior, and smoking behavior takes precedence over simple gathering of multiple people. If the aggressive behavior is resolved and the radar motion characteristics decrease, the system can continue to detect people; if the radar fails to detect any valid targets and infrared pyroelectricity is not triggered, the high-power peripherals are shut down and the system returns to a low-power standby state.

[0111] Without departing from the spirit of the present invention, the smoke detection unit may be a particulate sensor, a smoke sensor, a volatile gas sensor or a combination thereof; the millimeter-wave radar may be a radar module with different frequency bands, different antenna array forms or equivalent capabilities to output distance, speed, angle and micro-Doppler information; the overreaction behavior recognition may employ rule fusion, lightweight neural networks, support vector machines, decision trees or other edge-deployable classifiers.

[0112] In the above embodiments, the windows, thresholds, weights, and model coefficients listed in the formulas are all configurable variables used to illustrate the processing flow and the relationships between the variables, and do not limit specific parameter values. All equivalent substitutions made based on the hierarchical wake-up, multi-sensor collaboration, radar non-visual tracking, smoke-detector dynamic background confirmation, and event time-series fusion ideas of this invention should fall within the protection scope of this invention.

Claims

1. A multi-sensor hierarchical collaborative behavior recognition method for campus privacy areas, characterized in that: Includes the following steps: S1. Low-power monitoring and personnel confirmation trigger: The system collects the output signal of the infrared pyroelectric sensor. When no one is present or no confirmation is received, the infrared pyroelectric sensor remains on standby and the edge processing unit is in a low-power state. When the infrared pyroelectric signal is triggered, the control system changes from the low-power monitoring state to the personnel confirmation state. S2. Millimeter-wave radar startup and point cloud generation: The millimeter-wave radar is started under personnel confirmation. Static background suppression, range and velocity domain target detection, angle estimation and coordinate transformation are performed on the radar echo to form a target point cloud containing spatial position, motion speed and reflection intensity. S3. Target Clustering and Tracking: Perform target clustering and tracking on the target point cloud to obtain stable target trajectories and population estimation results; S4. Determination of Multiple People Gathering: Determine whether multiple people are gathering based on the number of stable targets, the spatial density between targets, and the duration. S5. Smoking behavior recognition: When there are people or multiple people gathered, the smoke detection unit enters the corresponding sampling state and identifies smoking behavior based on the dynamic background increment, rate of change and continuity of the smoke detection output. S6. Overreaction Recognition: Overreaction is identified by utilizing the high-frequency micro-Doppler energy characteristics, relative motion characteristics, acceleration fluctuation characteristics, and trajectory intersection characteristics generated by millimeter-wave radar. S7. Temporal Memory and Structured Alarms: Temporal memory is established for candidate decisions of multiple people gathering, smoking and excessive behavior. Instantaneous decisions are transformed into stable event outputs, and structured alarms are generated according to preset risk priorities, outputting the corresponding event type, number of people summary and alarm level.

2. The multi-sensor hierarchical collaborative behavior recognition method for campus privacy areas according to claim 1, characterized in that, The low-power state mentioned in step S1 includes: the edge processing unit is in a low-power or periodic health check state; when the infrared pyroelectric signal shows a change that matches the characteristics of human entry, the edge processing unit initiates the personnel confirmation process; the infrared pyroelectric sensor, as a first-level gate sensor, uses its judgment result to wake up the main control or high-power sensing module, and is not output separately as the final abnormal behavior result.

3. The multi-sensor hierarchical collaborative behavior recognition method for campus privacy areas according to claim 1, characterized in that, Step S3, which involves target clustering and tracking of the target point cloud, specifically includes: constraining the monitoring area of ​​the radar point cloud, removing outliers, and clustering targets. It also involves combining data association, filtering prediction, and track lifecycle management to form a stable target trajectory and counting the number of currently valid targets.

4. The multi-sensor hierarchical collaborative behavior recognition method for campus privacy areas according to claim 1, characterized in that, The determination of multiple people gathering in step S4 is based on at least two of the following: the number of stable target trajectories, the average distance between targets, the area occupied by the target set, and the duration of the state. The candidate results of multiple people gathering serve as one of the conditional information for subsequent smoking identification and aggressive behavior identification.

5. The multi-sensor hierarchical collaborative behavior recognition method for campus privacy areas according to claim 1, characterized in that, In step S5, the smoke detection unit includes one or more of a particulate matter sensor, a smoke sensor, or a volatile gas sensor; the smoking behavior recognition calculates the concentration change based on the dynamic background and outputs the smoking confidence level in combination with the person's status. Let the concentration or equivalent output of the smoke sensor at time t be... The historical background baseline is as follows: ； ； In the formula, The dynamic background baseline representing the smoke detector output; Indicates the length of the smoke detector background window; Indicates the current smoke concentration or equivalent output; Indicates historical smoke sensor output; Indicates the relative background increment; This indicates the rate of change between adjacent time points.

6. The multi-sensor hierarchical collaborative behavior recognition method for campus privacy areas according to claim 1, characterized in that, The overreaction recognition described in step S6 does not use image or video information, but constructs a behavior feature sequence based on the temporal point cloud and velocity spectrum statistical features of millimeter-wave radar, and outputs the overreaction confidence level through rule discrimination or lightweight classification model. The behavioral feature sequence includes one or more of the following: high-frequency micro-Doppler energy, relative proximity velocity, acceleration variance, and number of rapid trajectory crossings; Among them, the high-frequency micro-Doppler energy is obtained by integrating the slow-time radar echo signal in the high-frequency band after performing a short-time Fourier transform, and the relative proximity velocity vrel is calculated based on the position vectors and velocity vectors of the two targets: ； ； In the formula, This represents the relative approach velocity characteristics between the i-th and j-th stable targets; , Represents the target position vector; , Represents the target velocity vector; This indicates the avoidance of tiny positive quantities with a denominator of zero; Indicates the intensity of the target interaction; The target pair weights can be determined by target distance, signal-to-noise ratio, or trajectory confidence.

7. The multi-sensor hierarchical collaborative behavior recognition method for campus privacy areas according to claim 1, characterized in that, Step S7 involves establishing temporal memories for candidate decisions regarding group gatherings, smoking, and aggressive behavior, including: setting an independent memory window for each event type e, accumulating candidate results according to time decay weights, and accumulating the accumulated value as follows: ； ； In the formula, e represents the event type. For large gatherings, For smoking, This is considered an excessive act. This represents the time decay weight of event e at the i-th historical position; Indicates the attenuation factor; Indicates the length of the memory window for event e; This represents the candidate outcome of event e at a given historical moment. This represents the accumulated memory value for event e; This indicates the stable output of event e; This represents the stable output threshold for event e; j is the index for weighted normalized summation.

8. The multi-sensor hierarchical collaborative behavior recognition method for campus privacy areas according to claim 1, characterized in that, In step S7, generating structured alarms according to preset risk priorities specifically involves classifying alarm levels from high to low based on risk priorities such as aggressive behavior, smoking behavior, and large gatherings. When multiple events occur simultaneously, the alarm with the highest priority is the main alarm, along with summary information of other events. The edge processing unit only uploads the event type, timestamp, number of people summary, sensor summary, and alarm level, without uploading the original image data.

9. A multi-sensor hierarchical collaborative behavior recognition system for campus privacy areas, characterized in that: include: Infrared pyroelectric sensor, millimeter-wave radar sensor, smoke detection unit, edge processing unit, communication unit, and power management unit; Among them, the infrared pyroelectric sensor is used for low-power monitoring and detection of human body entry trigger, and its output is connected to the wake-up input of the edge processing unit; Millimeter-wave radar sensors are used to activate under personnel confirmation, generate target point clouds, and perform multi-target tracking and motion interaction analysis. The smoke detection unit is used to sample and output smoke or particulate matter concentration data in the presence of people or when multiple people are gathered. The power management unit controls the enabling of high-power modules based on the state machine of the edge processing unit; The edge processing unit is configured to perform the method according to any one of claims 1 to 8.

10. The campus privacy area multi-sensor hierarchical collaborative behavior recognition system according to claim 9, characterized in that, The millimeter-wave radar sensor adopts a frequency-modulated continuous wave radar system or an equivalent radar system, and operates in the frequency band of 24-80GHz. The smoke detection unit employs one or more of a particulate matter sensor, a smoke sensor, or a volatile gas sensor; The edge processing unit includes a main control chip and an optional coprocessor. The main control chip is responsible for the state machine, sensor scheduling, data caching and communication reporting, while the coprocessor is responsible for radar feature classification or lightweight model inference. The communication unit wirelessly uploads the structured alarms output by the edge processing unit to the management platform.

Images