An automated biometric feature attendance verification system

Abstract

This invention relates to the fields of intelligent facility management and education technology, and discloses a non-biometric automated attendance verification system. This system aims to address the technical shortcomings of existing attendance methods, such as privacy leaks, susceptibility to cheating, high false alarm rates, and poor hardware compatibility. By integrating computer vision, dynamic threshold adjustment, time-smoothing filtering, occlusion compensation, and spatial density verification technologies, it achieves accurate, privacy-protected, and hardware-compatible automated attendance verification. This system does not require the storage of biometric data or video data, is compatible with existing high-angle CCTV infrastructure, and effectively improves the accuracy, privacy security, and practical applicability of attendance verification, making it suitable for classroom attendance scenarios in various educational institutions.

Description

Technical Field

[0001] This invention relates to the fields of intelligent facility management and education technology, specifically to an automated attendance verification system based on non-biological characteristics. Background Technology

[0002] In educational institutions, attendance tracking is an essential administrative process for verifying credits and monitoring safety. Traditionally, this process is carried out through manual roll call, which is prone to human error and consumes teaching time.

[0003] Existing technological solutions have significant drawbacks: Hardware token systems (RFID / Bluetooth): These systems verify the presence of the device (token) rather than the student, which makes it easy for "proxy attendance" (one student carrying multiple cards) to occur.

[0004] Biometric identification (facial recognition): These systems face significant regulatory hurdles (e.g., GDPR, PIPL) in storing biometric data of minors. Furthermore, they typically require expensive, line-of-sight hardware, which is incompatible with standard high-angle classroom closed-circuit television (CCTV) infrastructure.

[0005] Simple Occupancy Counters: Standard headcount algorithms lack "temporal context." They cannot distinguish between brief absences (such as going to the restroom) and genuine absences, leading to false negative alerts. Furthermore, they frequently fail in high-density classroom environments where students often shield themselves from view.

[0006] In summary, existing attendance technology solutions are inadequate in terms of privacy protection, anti-cheating capabilities, accuracy, and hardware compatibility, and cannot meet the needs of educational institutions for efficient, accurate, and compliant attendance verification. Therefore, a system is needed that can utilize existing hardware to provide accurate and privacy-protected attendance verification by compensating for signal noise, obstruction, and authorized schedule changes. Summary of the Invention

[0007] To achieve the above-mentioned objectives, this invention provides a non-biological automated attendance verification system. This system, through a design concept of "decoupling identity and presence," does not identify the specific individuals in the classroom, but only verifies the presence of a required number of human entities and a reasonable spatial distribution pattern. The specific technical solution is as follows: The system includes the following core functional modules, which work together to achieve a complete attendance verification process: (a) Computer Vision Module As the core data acquisition and processing unit of the system, this module is coupled to the image acquisition interface of the existing classroom monitoring camera to receive real-time video streams and perform non-biological target detection and counting.

[0008] The specific workflow is as follows: After the video stream is transmitted to the AI ​​processing unit, it is processed in real time in volatile memory without permanently storing any video data; the AI ​​processing unit uses non-biological pose estimation algorithms (such as YOLO algorithm, OpenPose algorithm, etc.) to identify individual targets by detecting human skeleton features in the video stream, avoiding the extraction of biometric information such as facial features, and ensuring privacy protection; based on the human skeleton detection results, a real-time occupancy count (Nreal) is generated in the classroom. This count is anonymized data, which only reflects the number of human entities currently present and is not associated with any personal identity information.

[0009] (ii) Dynamic threshold unit This unit is used to dynamically determine the expected attendance (Nexpected). Its core lies in real-time integration with external administrative databases, rather than relying on a fixed registration list, to adapt to dynamic changes in authorized absences such as sick leave.

[0010] The specific implementation method is as follows: Data source: The system controller periodically queries the external school management system (SMS) through a preset polling protocol to retrieve approved "sick leave" or "other authorized absence" tickets during the course session, ensuring that the obtained authorized absence information is real-time and accurate; Calculation logic: The expected attendance is calculated using the following formula: Nexpected(t) = Nenrolled - Nsick_leave(t), where Nenrolled is the total number of registered students for the course, and Nsick_leave(t) is the number of authorized absences approved at time t; Polling mechanism: To cope with temporary administrative approval delays, the polling operation is performed at preset intervals during the course session. For example, polling is performed 15 minutes before the course starts (T-15 minutes) and 30 minutes after the course starts (T+30 minutes) to ensure that Nexpected(t) can be updated in a timely manner and accurately reflect the actual number of attendees.

[0011] (III) Time Smoothing Engine This engine is used to filter out instantaneous fluctuations in occupancy counts, distinguish noise signals such as brief absences from actual absences, and avoid false alarms.

[0012] The specific working principle is as follows: Sliding time window setting: The system presets a fixed-length sliding time window W (e.g., 300 seconds, or 5 minutes) to calculate the moving average of the real-time count Nreal within this window, so as to smooth out the counting deviation caused by instantaneous fluctuations; Gap timer mechanism: When the moving average of the real-time count Nreal is lower than the expected count Nexpected(t), the system starts the "Gap Timer" to begin accumulating the duration of the count mismatch; Alarm triggering conditions: The system generates an alarm flag and triggers an attendance anomaly alarm only when the cumulative duration of the gap timer exceeds the duration of the sliding time window W. If the moving average of Nreal recovers to Nexpected(t) or higher within the window W, the gap timer is reset and no alarm is triggered. This mechanism effectively filters out instantaneous count mismatches caused by brief absences such as students going to the toilet or teachers temporarily entering and leaving, reducing the false alarm rate.

[0013] (iv) Occlusion Compensation Logic To address the blind spot problem caused by camera shooting angle limitations and obstacles such as pillars in the classroom, this logic uses a "vector memory" mechanism to retain the "virtual presence" state of objects in the blind spot, preventing omissions.

[0014] The specific implementation process is as follows: Zone division: The system pre-maps the physical space of the classroom, dividing it into "active sea area" and "blind / inference zone". The active sea area is the area where the camera can clearly capture and effectively detect the human skeleton, while the blind / inference zone is the area that cannot be effectively detected due to occlusion or shooting angle. Motion vector tracking: The system tracks the motion vector of each detected object in real time. When the object's trajectory intersects the boundary of the blind zone / inferred area, and the visual tracking of the object is subsequently lost in the visible area, it is determined that the object has entered the blind zone. Virtual Presence Token: If an object enters the blind zone and its intersection with the classroom exit boundary is not detected (i.e., it has not left the classroom), the system creates a "Virtual Presence Token" for the object, maintaining the object's presence status in the count (count + 1). Marker Maintenance and Elimination: The duration of the virtual presence marker is set to a preset decay period, or until the object reappears in the visible area. If the object does not reappear in the visible area after the decay period, or if the object is detected to intersect with the exit boundary (having left the classroom), the virtual presence marker is eliminated, and the count is decremented by 1. This mechanism simulates the principle of "object constancy," effectively solving the problem of missed recordings caused by high-angle camera occlusion.

[0015] (v) Spatial density verification module This module is used to detect situations where the total count is correct but the personnel distribution is abnormal, thereby preventing violations such as substitute attendance and unauthorized auditing, and further improving the effectiveness of attendance verification.

[0016] The specific working method is as follows: Reference model establishment: The system pre-collects spatial distribution data of people in the classroom under normal attendance conditions and generates a reference heat map model. This model reflects the reasonable occupancy density range of each sub-area (such as front row, middle row, and back row) in a normal class scenario. Real-time distribution comparison: The system generates a heat map of the spatial distribution of occupants in the classroom in real time, compares it with a reference heat map model, and analyzes the occupancy density of each sub-region. Anomaly Alert Generation: If the real-time total count Nreal ≥ Nexpected(t), but the occupancy density of a specific sub-region (such as the back row) exceeds the threshold set by the reference model, while other key sub-regions (such as the front row) are vacant, the system determines this as an abnormal spatial distribution, generates a "regional anomaly alert," and sends a secondary "verification prompt" to the instructor, clearly indicating the specific sub-region of the anomaly. This facilitates manual verification by the instructor to investigate cases of substitute attendance, unauthorized auditing, etc. Compared with existing technologies, this invention provides a non-biological automated attendance verification system with the following beneficial effects: This non-biological automated attendance verification system offers a high level of privacy protection: the system employs non-biological gesture recognition technology, detecting only human skeletal features without extracting or storing facial features or other biological information. Furthermore, all video data is processed in real-time only in volatile memory and is not permanently stored, fundamentally avoiding the risk of biometric data leakage. It fully complies with the privacy protection requirements of relevant laws and regulations such as GDPR and PIPL, making it particularly suitable for underage students.

[0017] High attendance accuracy: Through multi-dimensional compensation and filtering mechanisms, the core defects of existing technologies are effectively solved: the time smoothing engine filters out instantaneous noise caused by brief absences, the occlusion compensation logic solves the problem of missed recording in camera blind spots, the dynamic threshold unit adapts to authorized absences in real time, the spatial density verification module identifies violation scenarios where the total count is correct but the distribution is abnormal, and the collaborative effect of multiple modules greatly reduces attendance counting errors and significantly improves alarm accuracy.

[0018] Good hardware compatibility: The system directly couples with existing optical hardware such as classroom monitoring cameras, without the need to deploy expensive hardware such as RFID card readers or dedicated biometric collection equipment, and without the need to modify existing CCTV infrastructure, which greatly reduces the deployment cost and implementation difficulty for educational institutions and facilitates large-scale promotion and application.

[0019] Flexible administrative adaptability: The system integrates with the school management system (SMS) in real time and dynamically updates authorized absence information through a timed polling mechanism. It can quickly respond to changes in administrative processes such as temporary sick leave approval, avoid attendance counting errors caused by administrative delays, and adapt to the flexibility needs of educational institutions' administrative work.

[0020] Comprehensive violation detection: It not only verifies whether the total number of attendees matches, but also uses spatial density analysis to identify instances where the total count is correct but the actual violation is not, such as substitute attendance or unauthorized auditing. Compared to existing technologies that only focus on matching the number of attendees, the depth and comprehensiveness of attendance verification are significantly improved, better meeting the core needs of educational institutions for attendance management. Attached Figure Description

[0021] Figure 1 is a high-level system architecture diagram of the present invention, showing the data flow relationship between the camera, the AI ​​processing unit, and the administrative database. The real-time video stream captured by the camera is transmitted to the AI ​​processing unit. The real-time count (Nreal) output by the AI ​​processing unit and the authorized absence data provided by the administrative database are used together to calculate the dynamic expected count (Nexpected). The system generates an alarm signal based on the comparison result of Nreal and Nexpected and related verification logic.

[0022] Figure 2 is a flowchart of the time smoothing algorithm, showing the process of distinguishing between instantaneous noise and true absence. After the process starts, the system receives Nreal and Nexpected. First, it determines whether the gap timer duration exceeds the sliding time window W. If not, it determines whether Nreal is less than Nexpected. If so, it starts or increments the gap timer. Otherwise, it resets the gap timer. If the gap timer duration exceeds the window W, an alarm flag is generated and the process ends.

[0023] Figure 3 is a region-based heuristic mapping diagram showing the division of the physical space of the classroom. The diagram clearly distinguishes between the active seating zone and the blind / inference zones. The active seating zone is the area where the camera can effectively detect the active seats, while the blind / inference zones are areas that are obscured or cannot be effectively detected.

[0024] Figure 4 shows the state transition diagram for entering vector analysis, illustrating the state transition process of the object in the occlusion compensation logic. It includes the transition relationships of the object's normal detection state in the visible area, the virtual presence state when entering the blind zone, the state recovery when returning to the visible area, and the state elimination after leaving the classroom, clarifying the conditions for the creation, maintenance, and elimination of the virtual presence marker. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be described in detail and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0026] I. Overall System Deployment The deployment of the non-biometric automated attendance verification system of the present invention is based on the existing classroom infrastructure of educational institutions, without the need for additional dedicated hardware. The specific deployment method is as follows: Hardware coupling: The system's image acquisition interface is physically coupled with the surveillance cameras already installed in the classroom. It supports various types of CCTV devices, such as mainstream analog cameras and network cameras (IP cameras). It receives real-time video streams output by the cameras through video interfaces (such as HDMI, CVBS, RTSP protocols, etc.). No hardware modification to the cameras is required. It is only necessary to ensure that the cameras are working properly and that the shooting angle covers the main activity areas of the classroom.

[0027] Processing Unit Deployment: The AI ​​processing unit can be deployed locally or on edge computing nodes, with local deployment being preferred to reduce network transmission latency. The AI ​​processing unit needs to have sufficient computing power to support the operation of real-time attitude estimation algorithms. The recommended configuration is: CPU clock speed ≥ 2.8GHz, GPU memory ≥ 4GB, RAM ≥ 8GB, and volatile memory (such as DDR4 memory) for storage to ensure that video data is only temporarily stored during real-time processing and is immediately deleted after processing, without retaining any original video data.

[0028] Database interface integration: The system controller establishes a communication connection with the school management system (SMS) through a network interface (such as Ethernet or Wi-Fi), and configures the corresponding interface protocols (such as RESTful API or SOAP protocol) to ensure that sick leave approval data and course registration data can be obtained normally through polling. The interface communication adopts an encrypted transmission method (such as HTTPS protocol) to ensure the security of data transmission and prevent information leakage.

[0029] Alarm output configuration: The system supports multiple alarm output methods, which can be configured according to the needs of educational institutions, including: sending push notifications to the mobile terminals of teachers (such as mobile APP), displaying alarm prompts on the front-end display devices in the classroom (such as electronic blackboard), and sending emails or system messages to the office terminals of school administrators; the alarm information should include key information such as course name, time, and anomaly type (such as insufficient number of students, abnormal regional distribution) to facilitate rapid response.

[0030] II. Detailed Implementation of Core Modules (a) Implementation of the computer vision module Video stream acquisition: The image acquisition interface acquires real-time video streams at a fixed frame rate (15-30fps recommended) according to the camera's output format, and converts the video stream into a digital image format that the AI ​​processing unit can recognize (such as RGB format, with a resolution adapted to the camera output, 1920×1080 pixels recommended); during the acquisition process, the video stream is pre-processed, including noise reduction (using Gaussian filtering algorithm) and image enhancement (adjusting brightness and contrast) to improve the accuracy of subsequent pose detection.

[0031] Pose estimation algorithm selection and configuration: In this embodiment, the YOLOv8 algorithm is combined with the OpenPose pose detection framework. The YOLOv8 algorithm is used for fast target detection and localization of human body regions, while the OpenPose algorithm is used to extract key points of the human skeleton (such as 18 key nodes including the head, shoulders, elbows, and waist). The algorithm parameters are configured as follows: confidence threshold ≥ 0.6, IOU threshold ≥ 0.5, and the maximum number of targets to be detected is set to the maximum capacity of the classroom (e.g., 60 people) to ensure accurate detection of all human targets in high-density scenes.

[0032] Real-time counting generation: For each frame image, the system first detects all human candidate regions using the YOLOv8 algorithm, and then extracts the human skeleton key points of each candidate region using the OpenPose algorithm. If the number of key points is ≥10 (ensuring a complete human body), it is determined to be a valid human target and counted. The system performs consistency verification on the counting results of multiple consecutive frames (5 frames recommended). If the counting difference between consecutive frames is ≤1, the average value is taken as the real-time count Nreal for that time period. If the difference exceeds 1, the detection is repeated to avoid counting fluctuations caused by false detections in a single frame image.

[0033] (ii) Implementation of dynamic threshold units Basic data acquisition: 24 hours before the course starts, the system retrieves the total number of registered students (Nenrolled) for the course from the school management system (SMS) and stores it in local volatile memory; the registration data must include information such as course number, class, and a list of registered students (only used for counting the number of students, and not storing students' personal identification details) to ensure the accuracy of Nenrolled.

[0034] Polling Mechanism Execution: The polling operation is executed at preset time intervals, specifically: the first poll is executed 15 minutes before the start of the course (T-15 minutes) to obtain the number of sick leave approved up to that time, Nsick_leave1; the second poll is executed 30 minutes after the start of the course (T+30 minutes) to obtain the number of newly added sick leave, Nsick_leave2, at which point Nexpected(t) = Nenrolled - (Nsick_leave1 + Nsick_leave2); if the course duration exceeds 2 hours, an additional poll can be added in the middle of the course (e.g., T+90 minutes) to further adapt to the situation of temporary approval.

[0035] Dynamic threshold adjustment: If a change in the number of sick leave applicants is detected during the polling process (i.e., Nsick_leave(t) is updated), the system immediately recalculates Nexpected(t) and applies the new expected count to subsequent attendance verifications. At the same time, the system records the time of the threshold adjustment, the values ​​before and after the adjustment, and the reason (such as adding new sick leave approvals) to facilitate subsequent traceability and verification.

[0036] (III) Implementation of the Time Smoothing Engine Sliding time window configuration: The duration of the sliding time window W can be flexibly configured according to the course type. For regular theoretical courses, it is recommended to be set to 300 seconds (5 minutes); for courses such as experimental courses and training courses where students may leave more frequently and briefly, the window duration can be adjusted to 600 seconds (10 minutes); the window duration is set through system parameters, and educational institutions are supported to customize according to actual needs.

[0037] Moving average calculation: The system calculates the moving average of Nreal in real time according to the sliding time window W. The calculation method is: for the current moment t, the moving average = (the sum of all Nreal values from the moment t - W to the moment t) / the number of samples within the window; the number of samples = window duration × video capture frame rate, ensuring that the calculation result can accurately reflect the average number of people present within the window.

[0038] Gap timer control: The timing unit of the gap timer is seconds, and the initial value is 0; when the moving average < Nexpected (t), the gap timer starts to accumulate (add 1 per second); when the moving average ≥ Nexpected (t), the gap timer is immediately reset to 0; if the value of the gap timer reaches W (such as 300 seconds), the system generates a "personnel shortage alarm" and continues to count until the moving average resumes to Nexpected (t) and above, at which point the alarm stops and the timer is reset.

[0039] (4) Implementation of occlusion compensation logic Area mapping calibration: During the system deployment phase, the boundary coordinates of the visible area and the blind area / inferred area are manually calibrated through the panoramic image of the classroom captured by the camera; the boundary coordinates adopt the pixel coordinate system, with the upper left corner of the image as the origin, the horizontal direction as the x-axis, and the vertical direction as the y-axis; for example, the coordinate range of the visible area can be calibrated as (x1 = 100, y1 = 100) to (x2 = 1820, y2 = 980), and the coordinate range of the blind area (such as the area blocked by the pillar in the corner of the classroom) is (x3 = 1600, y3 = 700) to (x4 = 1820, y4 = 980); after calibration, the system stores the area coordinate data for subsequent motion vector judgment.

[0040] Motion vector analysis: For each detected human target, the system calculates its motion vector (including direction and speed) based on its position coordinates in multiple consecutive frames of images. The direction of motion is determined by the difference in position change between adjacent frames, and the speed of motion is calculated by the ratio of the position change to the time interval (unit: pixels / second). When the position coordinates of the human target are close to the boundary of the blind zone, and the direction of motion is pointing into the blind zone, and the speed of motion is within a reasonable range (e.g., 10-50 pixels / second, corresponding to the normal walking speed of a human), it is determined that the target is about to enter the blind zone.

[0041] Virtual Presence Marker Management: When a target enters the blind zone and visual tracking is lost, the system creates a virtual presence marker. The marker contains information such as the time the target entered the blind zone and the blind zone number. The decay period of the virtual presence marker is preset to 180 seconds (3 minutes). If the target reappears in the visible area within the decay period, the system eliminates the virtual presence marker and resumes normal counting. If the target still does not appear after the decay period ends and no intersection between the target and the exit boundary is detected, the virtual presence marker is eliminated and the count is decremented by 1. If an intersection between the target and the exit boundary (such as the coordinate range of the classroom door) is detected, the virtual presence marker is immediately eliminated and the count is decremented by 1.

[0042] (v) Implementation of the spatial density verification module Reference Heatmap Model Establishment: In the initial stage after system deployment, spatial distribution data for at least 10 normal class scenarios is collected, with each collection lasting 30 minutes, generating multi-frame spatial distribution heatmaps. The heatmaps are generated as follows: the visible area of ​​the classroom is divided into several sub-regions (a 10×10 grid is recommended, with each sub-region being 192×108 pixels in size), the number of human targets in each sub-region is counted, and different colors are assigned according to the number (e.g., blue = 0 people, green = 1 person, yellow = 2 people, red ≥ 3 people); the heatmaps collected from multiple times are averaged to determine the reasonable occupancy density range for each sub-region (e.g., the reasonable density for the front row sub-region is 1-2 people, and the reasonable density for the back row sub-region is 0-2 people), forming a reference heatmap model and storing it.

[0043] Real-time heat map generation and comparison: The system generates a spatial distribution heat map of the current moment in real time, and the generation method is consistent with the reference model; the real-time heat map is compared with the reference heat map model sub-region by sub-region, and the density deviation value of each sub-region is calculated (deviation value = real-time density - reference density upper limit); if there are at least 3 consecutive sub-regions with deviation values ​​> 0 (i.e., density exceeds the standard), and these sub-regions are concentrated in the same area (such as the back row), and at least 2 consecutive sub-regions in the front row have a density of 0, then it is judged as an abnormal spatial distribution.

[0044] Secondary verification prompt sending: When abnormal spatial distribution is detected, the system immediately sends a secondary "verification prompt" to the teaching teacher. The prompt information includes the location description of the abnormal area (such as "the density in the left rear area exceeds the standard, and the front middle area is vacant") and a screenshot of the real-time heat map. After receiving the prompt, the teacher can verify the abnormal reason through classroom observation or taking attendance. If it is confirmed as a situation such as a substitute class or unauthorized attendance, the teacher can feedback the verification result through the system, and the administrative staff will conduct subsequent processing based on the feedback result.

[0045] III. System operation process The complete operation process of the non-biometric automated attendance verification system of the present invention is as follows: Preprocessing stage (15 minutes before the start of the course): The system starts, the image acquisition interface establishes a connection with the camera, starts preheating to ensure normal video stream acquisition; The AI processing unit loads the pose estimation algorithm model and completes the initialization configuration; The system controller performs the first polling, obtains the course enrollment number Nenrolled and the number of approved sick leave Nsick_leave1 from the school management system, and calculates the initial expected count Nexpected_initial = Nenrolled - Nsick_leave1; The system loads the pre-calibrated area mapping data and the reference heat map model to complete the preparations before operation.

[0046] Real-time attendance stage (from the start of the course to 30 minutes before the end): The image acquisition interface continuously acquires the video stream, transmits it to the AI processing unit for real-time processing, and generates an anonymous real-time count Nreal; The time smoothing engine calculates the moving average of Nreal within the sliding time window W and compares it with the current Nexpected (t); If the moving average < Nexpected (t), start the gap timer to accumulate the mismatch duration. If the accumulated duration exceeds W, generate a "number shortage alarm"; The occlusion compensation logic tracks the motion vector of the human target in real time, creates a virtual presence mark for the target entering the blind area, and maintains the counting accuracy; The system controller performs the second polling 30 minutes after the start of the course, updates Nsick_leave (t), and recalculates Nexpected (t); The spatial density verification module generates a heat map in real time and compares it with the reference model. If abnormal distribution is detected, it sends a "regional anomaly verification prompt".

[0047] Concluding phase (30 minutes before the end of the course until the end): The system continuously verifies attendance, but the alarm sensitivity is adjusted appropriately, such as shortening the sliding time window W to 120 seconds (2 minutes) to avoid unnecessary alarms caused by students leaving before the end of the class. The system records complete attendance data for this course, including real-time count change curves, Nexpected(t) adjustment records, alarm trigger records (if any), and teacher feedback results (if any). After the course ends, the system will summarize the attendance data, generate an attendance report, and upload it to the school management system for administrative staff to view and archive. Clear all temporary data (including video stream data, count data, etc.) from the volatile memory, and the system enters standby mode, waiting for the next class attendance task.

[0048] In summary, this non-biometric automated attendance verification system aims to address the technical shortcomings of existing attendance methods, such as privacy leaks, susceptibility to cheating, high false alarm rates, and poor hardware compatibility. By integrating computer vision, dynamic threshold adjustment, temporal smoothing filtering, occlusion compensation, and spatial density verification, it achieves accurate, privacy-preserving, and hardware-compatible automated attendance verification. The system utilizes existing classroom surveillance cameras to capture real-time video streams and generates anonymous real-time occupancy counts using a non-biometric pose estimation algorithm. It dynamically adjusts the expected attendance based on an external sick leave database; employs a sliding time window to filter instantaneous fluctuations, avoiding false alarms caused by brief absences; compensates for missed entries due to camera blind spots using a virtual presence marking mechanism; and combines spatial distribution heatmap verification to identify situations where the total count is correct but constitutes an anomaly. This system does not require storing biometric or video data, is compatible with existing high-angle CCTV infrastructure, effectively improves the accuracy, privacy security, and practical applicability of attendance verification, and is suitable for classroom attendance scenarios in various educational institutions.

Claims

1. A non-biological automated attendance verification system, characterized in that: Including a computer vision module, a dynamic thresholding unit, a temporal smoothing engine, occlusion compensation logic, and a spatial density verification module, this system implements attendance verification through the following steps: (a) The computer vision module acquires real-time video streams within a limited physical space, applies non-biological target detection algorithms to identify individual targets and generate anonymous real-time counts (Nreal). (b) The dynamic threshold unit retrieves the dynamic expected count (Nexpected) by cross-referencing the pre-defined class schedule registration database with the real-time sick leave application database. (c) The time smoothing engine calculates the time stability index of Nreal within a predetermined sliding time window to filter out instantaneous fluctuations; (d) The occlusion compensation logic performs virtual presence marking on objects entering the predefined blind zone to compensate for omissions caused by occlusion; (e) The spatial density verification module compares the real-time spatial distribution of occupants with the reference heat map model to identify distribution anomalies; (f) The system generates corresponding alarms based on the stability comparison results of Nreal and Nexpected and the spatial distribution verification results.

2. The non-biological automated attendance verification system according to claim 1, characterized in that: The computer vision module employs a non-biological pose estimation algorithm, specifically a combination of the YOLO algorithm and the OpenPose pose detection framework. It identifies individual targets by detecting key points of the human skeleton, without extracting biometric information such as facial features. Furthermore, all video data is processed in real time only in volatile memory and is not permanently stored.

3. The non-biological automated attendance verification system according to claim 1, characterized in that: The dynamic expected count of the dynamic threshold unit is calculated using the formula Nexpected(t) = Nenrolled - Nsick_leave(t), where Nenrolled is the total number of registered students for the course, and Nsick_leave(t) is the number of authorized absences approved at time t. The dynamic threshold unit updates Nexpected(t) at preset intervals via a polling protocol, with the polling intervals including 15 minutes before the course starts and 30 minutes after the course starts.

4. The non-biological automated attendance verification system according to claim 1, characterized in that: The duration of the sliding time window W of the time smoothing engine is 60-600 seconds, with a default setting of 300 seconds; the time stability index is the moving average of Nreal within the sliding time window. When the moving average is lower than Nexpected(t), a gap timer is started. An "insufficient number of people alarm" is generated only when the cumulative duration of the gap timer exceeds W.

5. The non-biological automated attendance verification system according to claim 1, characterized in that: The specific implementation of the occlusion compensation logic includes: (a) Divide the physical space into visible and blind / inferred areas in advance, and mark the boundary coordinates of each area; (b) Track the motion vector of the object and create a "virtual presence marker" when the object's motion trajectory intersects the blind zone boundary and visual tracking is lost, and no intersection between the object and the exit boundary is detected; (c) The virtual presence marker is maintained for a preset decay period (60-300 seconds, default 180 seconds), or until the object reappears in the visible area, during which time the presence count of the object is maintained.

6. The non-biological automated attendance verification system according to claim 1, characterized in that: The reference heatmap model of the spatial density verification module is generated by collecting spatial distribution data from at least 10 normal class scenarios. It divides the physical space into several sub-regions and determines the reasonable occupancy density range of each sub-region. When the real-time total count Nreal ≥ Nexpected(t), but there are at least 3 consecutive sub-regions whose occupancy density exceeds the reference model threshold and the corresponding key sub-regions are vacant, an "area anomaly alarm" is generated.

7. The non-biological automated attendance verification system according to claim 1, characterized in that: The occlusion compensation logic uses a pixel coordinate system to define the boundaries of the region division. The visible region is the active seating area where the camera can effectively detect the human skeleton, while the blind / inferred region is the area that cannot be effectively detected due to occlusion or shooting angle issues.

8. The non-biological automated attendance verification system according to claim 1, characterized in that: The computer vision module performs consistency verification on the real-time count Nreal. Specifically, it compares the count results of 5 consecutive frames. If the count difference of consecutive frames is ≤1, the average value is taken as the Nreal for that time period. If the difference exceeds 1, the detection is repeated.

9. The non-biological automated attendance verification system according to claim 1, characterized in that: The system supports multiple alarm output methods, including sending push notifications to the instructor's mobile terminal, displaying alarm prompts on the front-end display device in the classroom, and sending emails or system messages to the office terminals of administrative personnel. The alarm information includes the course name, time, type of anomaly, and specific location information.

10. The non-biological automated attendance verification system according to claim 1, characterized in that: The system's core parameters support custom configuration, including the duration of the sliding time window W, the polling interval, the virtual presence marker decay period, the attitude detection confidence threshold, the spatial density deviation threshold, and the video acquisition frame rate.

Images