A non-contact heart rate recognition and abnormality monitoring method and system

Abstract

The present application relates to the technical field of rehabilitation and sports monitoring, and discloses a non-contact heart rate recognition and abnormal monitoring method and system, which comprises the following steps: collecting a face image and binding an identity label during identity registration, and synchronously collecting a resting heart rate to establish a personalized initial heart rate benchmark database; a plurality of cameras deployed in multiple regions are cross-camera identity associated according to spatial topological relations, time continuity and face feature similarity, so as to realize identity continuous tracking during multi-region motion switching; initial rPPG signals under each perspective are extracted based on the tracked identity label, and a fusion rPPG signal is generated after quality evaluation and high-quality signal weighting fusion compensation of adjacent cameras; real-time physiological indicators are extracted according to a sliding time window, and compared with historical benchmark data in the personalized initial heart rate benchmark database, and an alarm is triggered when the deviation exceeds a preset threshold. The present application realizes continuous and stable heart rate monitoring and personalized recovery state evaluation during the multi-region motion switching process of rehabilitation personnel.

Description

Technical Field

[0001] This invention relates to the field of rehabilitation and exercise monitoring technology, specifically to a non-contact heart rate recognition and abnormality monitoring method and system. Background Technology

[0002] With the increasing health awareness of the public and the growing demand for exercise and fitness, more and more people who exercise want to monitor their heart rate in real time throughout the entire exercise process and assess their physical fatigue and recovery status based on heart rate changes. This helps them avoid risks such as palpitations, physical exhaustion, and recurrence of old injuries caused by overexertion or poor recovery. These people's activity scenarios include gyms, rehabilitation centers, and sports venues, and each type of scenario usually contains multiple independent functional areas (such as the aerobics area, strength training area, and group exercise area in a gym; and the training area, balance area, and recovery area in a rehabilitation center). Exercisers switch between different areas to complete their predetermined training programs, which brings unique scenario requirements for continuous cross-area tracking in continuous heart rate monitoring throughout the entire process.

[0003] Currently, heart rate monitoring in sports scenarios mainly falls into two categories: one is contact wearable devices, which rely on wristbands, chest straps, heart rate patches, etc., to collect signals through contact with the skin. During exercise, sweating, large-scale limb movements, and device displacement can all cause signal interruption and data distortion. At the same time, wearing foreign objects can affect limb movement and is not conducive to controlling the standard of training movements. The other is the traditional single-camera non-contact heart rate monitoring solution. This solution is feasible to a certain extent in static scenarios, but during dynamic movements such as running, limb training, and changing body positions, problems such as posture shift, image obstruction, and changes in lighting are prone to occur. A single camera cannot continuously track the same target in multiple areas, resulting in poor stability of heart rate signal monitoring, making it difficult to accurately determine whether the recovery state is abnormal.

[0004] In summary, existing monitoring methods, in dynamic multi-regional sports scenarios, cannot continuously track the same target moving across multiple regions using a single camera. This makes it difficult to obtain complete and continuous heart rate signals to establish individualized recovery assessment benchmarks, thus hindering the accurate determination of the athlete's recovery status. Summary of the Invention

[0005] The purpose of this invention is to provide a non-contact heart rate recognition and abnormality monitoring method and system to solve the technical problem in the prior art that it is difficult to obtain complete and continuous heart rate signals to establish individualized recovery assessment benchmarks because a single camera cannot continuously track the same target moving in multiple areas, thus making it difficult to accurately determine the exercise recovery status of rehabilitation personnel.

[0006] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution: A non-contact heart rate recognition and abnormality monitoring method includes the following steps: Step S1, Identity Registration and Benchmark Collection: Collect facial images of the target personnel, extract facial feature vectors and bind them with unique identity identifiers to establish an individual identity database; simultaneously collect the target personnel's resting heart rate, heart rate recovery baseline and safe heart rate range to establish a personalized initial heart rate benchmark database corresponding to the individual identity database; Step S2, Multi-camera collaborative tracking: By deploying multiple cameras in various areas to collect video streams in real time, face detection and feature extraction are performed on each frame of the image, and based on the spatial topology of the cameras, temporal continuity and facial feature similarity, the same target person under different camera views is associated with the unique identity identifier, thereby realizing continuous identity tracking across cameras. Step S3, Cross-camera rPPG signal fusion and compensation: Based on the face regions of the same identity that have been tracked, the initial rPPG signal from each camera viewpoint is extracted; the quality of each initial rPPG signal is evaluated, and the low-quality signal is weighted, fused and compensated using the high-quality signal collected by the neighboring camera to generate the fused rPPG signal; Step S4, Personalized Anomaly Determination: Based on the fused rPPG signal, extract real-time physiological indicators by sliding time window; compare the real-time physiological indicators with the corresponding historical benchmark data in the personalized initial heart rate benchmark database; when the indicator deviation exceeds the preset threshold, it is determined to be abnormal and an alarm is triggered.

[0007] Furthermore, in step S1, the method for extracting the facial feature vector is as follows: Facial landmark localization and face alignment are performed on the acquired raw face images to obtain standardized face images; The standardized face image is input into a lightweight face feature extraction network to obtain a deep feature vector; The deep feature vector is bound to the unique identifier and then stored locally in an encrypted manner.

[0008] Furthermore, in step S2, face detection, face alignment, and feature extraction are performed sequentially in each video frame captured by the camera, specifically as follows: Let the first Each camera is constantly The captured video frames are Face detection is performed on the image to obtain face bounding boxes and key point information: ; in, This represents the detected face bounding box; Indicates the corresponding facial key points; A network for extracting facial landmarks; Geometric alignment and normalization of the face region are performed based on key points to obtain a standardized face image. : ; The aligned face image is then input into the face feature extraction network. Feature extraction is performed to obtain the corresponding facial feature vector. : ; in, Given a facial feature vector, combine this feature vector with a timestamp. Output them together for subsequent identity matching and cross-camera tracking.

[0009] Furthermore, in step S2, the cross-camera identity association specifically includes: Before performing cross-camera facial feature vector comparison, illumination compensation preprocessing is performed on the acquired facial images to eliminate feature drift interference caused by differences in ambient color temperature and illumination intensity between different cameras. Install camera At any moment The obtained facial feature vector is ,Camera At any moment The obtained facial feature vector is Calculate cosine similarity : ; Under the conditions of satisfying the preset time continuity constraints and camera spatial topology constraints, the target with the highest similarity is selected as the same identity identifier.

[0010] Furthermore, in step S3, the specific method for extracting the initial rPPG signal from each camera's viewpoint is as follows: Based on facial landmarks, the forehead and cheek areas, which are less affected by facial expressions and occlusion, are selected as a local skin pixel set. ; For the facial area captured by the camera, extract a stable set of local skin pixels. and based on Calculate the time for each channel Average light intensity on: ; ; ; in, This represents the number of pixels in the skin region. For pixels The RGB values; based on the average light intensity, the rPPG signal of a single camera is calculated using a blind source separation algorithm. : ; ; ; in, It represents the standard deviation of the signal and is used for normalization to suppress amplitude differences.

[0011] Further, in step S3, the quality assessment is performed by calculating the signal-to-noise ratio. Identifying low-quality signals: ; in, This indicates the calculation of variance. For the initial rPPG signal The signal after bandpass filtering in the 0.5-4.0Hz range; The higher the value, the stronger the effective physiological signal.

[0012] Furthermore, in step S3, for signals determined to be of low quality... Utilizing the high-quality signals from nearby cameras Perform weighted fusion compensation: ; in, For the compensated signal, For nearby cameras, The weights are dynamically calculated based on each signal quality index, and satisfy the following conditions: .

[0013] Furthermore, in step S4, the extracted physiological indicators include heart rate. Heart rate variability and pulse amplitude The calculation method is as follows: Let the interval between consecutive heartbeats within the detection time window be _____. ,but: ; ; in, The total number of beats within the time window; Heart rate variability The calculation method uses the standard deviation of the heartbeat interval (SDNN): ; pulse amplitude By analyzing each peak heart rate within the time window The amplitude was statistically calculated and the median was taken; The corresponding values ​​are output in real time after each time window ends, so as to carry out continuous and real-time physiological monitoring.

[0014] Furthermore, in step S4, the specific method for determining an anomaly when the index deviation exceeds a preset threshold is as follows: Let the real-time indicator value be The historical mean of the corresponding individual is The historical standard deviation is The preset sensitivity coefficient is ,when It was determined to be a transient anomaly. Only in consecutive If the same anomaly is detected within a time window, it is confirmed as a valid anomaly event and an alarm is triggered.

[0015] To address the aforementioned technical problems, the present invention further provides the following technical solution: A monitoring system for performing the above-described non-contact heart rate recognition and abnormality monitoring method includes: Multiple cameras are deployed in various activity areas to capture video streams in real time; A host computer is electrically connected to the plurality of cameras. The host computer includes a processor and a memory. The memory stores a computer program that, when executed by the processor, causes the processor to perform the following operations: The identity registration module is used to collect facial images of target personnel, extract facial feature vectors and bind them with unique identity identifiers and store them, and simultaneously collect resting heart rate, heart rate recovery baseline and safe heart rate zone to establish a personalized initial heart rate benchmark database. The multi-camera tracking module is used to detect and extract features from faces in the video stream, and to perform cross-camera identity association based on spatiotemporal information and feature similarity to maintain the continuity of identity identification. The rPPG signal processing module is used to extract the initial rPPG signal from the associated facial skin region, evaluate the quality of each signal, and perform weighted fusion and compensation of low-quality signals using high-quality signals from nearby cameras. The anomaly analysis and alarm module is used to calculate real-time physiological indicators based on the fused rPPG signal, compare the real-time physiological indicators with the historical baseline data of the corresponding individual to determine anomalies, and trigger an alarm when an anomaly is determined. The database module is used to securely store identity information, facial feature vectors, personalized heart rate baseline data, historical physiological signal data, and abnormal event records.

[0016] Compared with the prior art, the present invention has the following advantages: This invention establishes a personalized initial heart rate baseline database by binding a unique identity identifier during identity registration and simultaneously collecting resting heart rate, heart rate recovery baseline, and safe heart rate range. This database serves as the benchmark for individualized recovery assessment. It utilizes a cross-field identity recognition architecture composed of multiple cameras based on spatial topological relationships, temporal continuity, and facial feature similarity. This recognition architecture is then used to continuously track the identity of the same target during multi-regional movement switching, enabling continuous and stable heart rate monitoring and personalized recovery status assessment for rehabilitation personnel during multi-regional movement switching. Attached Figure Description

[0017] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the monitoring process provided in an embodiment of this application.

[0019] Figure 2 This is a schematic diagram of the identity registration process provided in an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] like Figure 1 and Figure 2 As shown, the present invention provides a non-contact heart rate recognition and abnormality monitoring method, which is mainly applied to designated functional areas such as training areas, balance areas, and recovery areas in rehabilitation centers. Specifically, it includes the following steps: Step S1, Identity Registration and Baseline Collection: Collect facial images of the target personnel, extract facial feature vectors and bind them with unique identity identifiers to establish an individual identity database; simultaneously collect the target personnel's resting heart rate, heart rate recovery baseline and safe heart rate range to establish a personalized initial heart rate baseline database corresponding to the individual identity database; Step S2, Multi-camera collaborative tracking: By deploying multiple cameras in various areas to collect video streams in real time, face detection and feature extraction are performed on each frame of the image. Based on the spatial topology of the cameras, temporal continuity, and facial feature similarity, the same target person under different camera views is associated with a unique identity identifier, thereby achieving continuous identity tracking across cameras. Step S3, Cross-camera rPPG signal fusion and compensation: Based on the face regions of the same identity that have been tracked, the initial rPPG signal from each camera view is extracted; the quality of each initial rPPG signal is evaluated, and the low-quality signal is weighted, fused and compensated using the high-quality signal collected by the neighboring cameras to generate the fused rPPG signal; Step S4, Personalized Anomaly Detection: Based on the fused rPPG signal, extract real-time physiological indicators by sliding time window; compare the real-time physiological indicators with the corresponding historical benchmark data in the personalized initial heart rate benchmark database; when the indicator deviation exceeds the preset threshold, it is judged as abnormal and an alarm is triggered.

[0022] This invention aims to achieve continuous, contactless monitoring of a member's cardiac health by using multiple cameras for collaborative video acquisition in a rehabilitation environment and combining it with a POS extraction algorithm. This solves the wearing limitations of traditional wearable devices and improves the convenience of long-term health monitoring in rehabilitation scenarios.

[0023] Meanwhile, by introducing a signal quality assessment and multi-camera signal compensation fusion mechanism, this invention dynamically compensates for low-quality remote photoplethysmography (rPPG) signals using high-quality signals from nearby cameras, effectively solving the signal instability problem caused by changes in lighting, human movement, or partial obstruction, and improving the stability and accuracy of continuous monitoring.

[0024] Furthermore, this invention also calculates physiological indicators (heart rate, heart rate variability, and pulse amplitude) in real time through a preset time window and compares them with individual benchmarks to achieve automatic identification of abnormal physiological states. When an abnormality is detected, the system can automatically issue an alarm and record relevant data, thereby realizing continuous health monitoring and risk warning in the rehabilitation environment.

[0025] Specifically, when the method of the present invention is executed: First, the target person completes identity registration upon entering the center, that is, facial images are captured by a camera, facial feature vectors are extracted and bound to a unique identity identifier to establish an individual identity database. At the same time, their resting heart rate, heart rate recovery baseline and safe heart rate range are collected to establish a corresponding personalized initial heart rate benchmark database.

[0026] Then, multiple cameras deployed in various areas of the rehabilitation center (such as the training area, balance area, recovery area, etc.) collect video streams in real time. The system performs face detection and feature extraction on each frame of the image. Based on the spatial topology of the cameras, temporal continuity, and facial feature similarity, the system associates the same target person under different camera views with a unique identity identifier, thereby achieving continuous identity tracking across cameras.

[0027] Subsequently, based on the face regions of the tracked individuals with the same identity, the system extracts the initial rPPG signal from each camera's perspective, evaluates the quality of each signal, and uses high-quality signals collected by neighboring cameras to perform weighted fusion and compensation on the low-quality signals to generate the fused rPPG signal.

[0028] Finally, the system extracts real-time physiological indicators based on the fused rPPG signal using a sliding time window, compares the real-time indicators with the corresponding individual's historical baseline data, and determines an abnormality and triggers an alarm when the indicator deviation exceeds a preset threshold, thereby realizing continuous heart rate monitoring and recovery status assessment throughout the exercise process.

[0029] Compared to existing single-camera non-contact heart rate monitoring solutions, which suffer from technical defects in multi-area dynamic scenarios in rehabilitation centers, this solution addresses the issue of a single camera being unable to continuously track the same moving target across multiple areas, making it difficult to obtain complete and continuous heart rate signals to establish individualized recovery assessment benchmarks.

[0030] This invention utilizes multi-camera collaborative tracking to maintain the continuity of the target's identity across different camera viewpoints, ensuring continuous tracking of rehabilitation personnel as they switch between training, balance, and recovery zones. Through cross-camera rPPG signal quality assessment and weighted fusion compensation of neighboring camera signals, it achieves stable fused rPPG signals even when single-viewpoint signal quality deteriorates, resolving interference issues such as posture shifts, image occlusion, and lighting changes during exercise. Furthermore, by simultaneously collecting individual resting heart rate, heart rate recovery baseline, and safe heart rate range during identity registration to establish a personalized initial heart rate benchmark database, and dynamically comparing real-time indicators with individual historical benchmarks during anomaly detection, the assessment criteria align with individual physiological differences, achieving non-contact continuous heart rate monitoring and personalized recovery status assessment throughout the entire exercise process.

[0031] Based on the overall process described above, since the quality of facial feature vector extraction directly affects the accuracy of subsequent cross-camera identity association, and the collection of personalized heart rate baseline data needs to be completed synchronously during the identity registration stage, it is necessary to further define the facial feature vector extraction method and the baseline data collection method in step S1.

[0032] As a preferred embodiment, such as Figure 2 As shown, in step S1, the method for extracting the facial feature vector is as follows: Facial landmark localization and face alignment are performed on the acquired raw face images to obtain standardized face images; Standardized face images are input into a lightweight face feature extraction network to obtain deep feature vectors; The deep feature vector is bound to a unique identifier and then stored locally in an encrypted manner.

[0033] In this embodiment, a pre-trained convolutional neural network model is used to detect the acquired face images (left side, frontal, and right side), accurately identifying key facial feature points such as eyes, nose tip, and corners of the mouth, and performing face alignment and standardization based on these key points. Multiple face-aligned images are then input into the MobileFaceNet lightweight face feature extraction network. This model extracts local and global facial features through a lightweight convolutional structure and finally generates a compact depth feature vector.

[0034] Encoding and aggregating multiple images of the same member can effectively enhance the stability and robustness of features and reduce interference from changes in lighting, expression, and pose.

[0035] All identity information and feature data are stored locally in encrypted form to ensure the privacy and security of rehabilitation members. This step provides a stable and comparable feature vector foundation for cross-camera identity association in subsequent multi-camera tracking.

[0036] Among them, the extraction of facial feature vectors: The MobileFaceNet lightweight facial feature extraction network is trained using the ArcFace loss function and outputs a 512-dimensional normalized deep feature vector.

[0037] This network contains depthwise separable convolutional layers and global depthwise convolutional layers, with a computational cost of about 1 / 10 that of a regular FaceNet, making it suitable for real-time operation on edge computing devices in rehabilitation centers.

[0038] After normalization, the eigenvectors satisfy the following conditions: This eliminates feature amplitude drift caused by changes in illumination and contrast. Local encrypted storage uses the AES-256 symmetric encryption algorithm, with the key held by the central administrator to ensure the security of identity data during storage.

[0039] Since multi-camera collaborative tracking requires face detection and feature extraction to be performed in the video frames captured by each camera, it is necessary to further explain the single-frame processing method for each camera and the cross-camera identity association method in step S2.

[0040] In a preferred embodiment, in step S2, face detection, face alignment, and feature extraction are performed sequentially in each video frame captured by the camera, specifically as follows: Let the first Each camera is constantly The captured video frames are Face detection is performed on the image to obtain face bounding boxes and key point information: ; in, This represents the detected face bounding box; Indicates the corresponding facial key points; A network for extracting facial landmarks; Geometric alignment and normalization of the face region are performed based on key points to obtain a standardized face image. : ; The aligned face image is then input into the face feature extraction network. Feature extraction is performed to obtain the corresponding facial feature vector. : ; in, Given a facial feature vector, combine this feature vector with a timestamp. Output them together for subsequent identity matching and cross-camera tracking.

[0041] In this embodiment, the system uses MTCNN to perform face detection on the image, obtaining face bounding boxes and information on 68 facial key points. Based on these key points, the face region is geometrically aligned and normalized to eliminate the effects of pose variations.

[0042] The aligned face images are input into a face feature extraction network to extract deep feature vectors, providing a feature basis for subsequent cross-camera identity matching. This processing is executed independently in each frame of each camera to ensure real-time performance.

[0043] Among them, the face detection and key point extraction network It is implemented using MTCNN (Multi-task Cascaded Convolutional Neural Network).

[0044] MTCNN consists of three cascaded lightweight convolutional neural networks: The first-level P-Net (Proposal Network) quickly generates face candidate boxes on the image pyramid; The second-level R-Net (Refine Network) refines and filters the candidate boxes; The third-level O-Net (Output Network) outputs the final face bounding box. and 68 facial key points .

[0045] The 68 key points include: 17 facial contour points, 10 eyebrow points, 9 bridge of the nose points, 12 eye points, and 20 mouth points.

[0046] When the above processing is performed on each frame of image, the detection speed is no less than 30fps. Image alignment operation. A similarity transformation is used to normalize the face region to a 112×112 pixel standardized image based on the center points of the eyes and the tip of the nose, and to eliminate in-plane rotation.

[0047] Differences in ambient color temperature and light intensity between different cameras can cause the feature vector of the same target to drift under different cameras, affecting the accuracy of identity association. To solve this problem, the following preferred embodiments are provided.

[0048] In a preferred embodiment, step S2, cross-camera identity association specifically includes: Before performing cross-camera facial feature vector comparison, illumination compensation preprocessing is performed on the acquired facial images to eliminate feature drift interference caused by differences in ambient color temperature and illumination intensity between different cameras. Install camera At any moment The obtained facial feature vector is ,Camera At any moment The obtained facial feature vector is Calculate cosine similarity : ; Under the conditions of satisfying the preset time continuity constraints and camera spatial topology constraints, the target with the highest similarity is selected as the same identity identifier.

[0049] In this embodiment, before extracting and comparing facial feature vectors, the system performs illumination compensation preprocessing such as white balance correction or histogram specification on the acquired facial images to eliminate feature drift interference caused by differences in ambient color temperature and illumination intensity between different cameras.

[0050] Subsequently, under the conditions of satisfying the time continuity constraint (i.e., the time interval between two observations is small, usually less than 2 seconds) and the spatial topology relationship of the cameras (i.e., there is a direct or reachable path between the two cameras in the rehabilitation space structure), the similarity of multiple candidate targets is compared, and the matching object with the highest similarity is selected as the corresponding target of the same member in different cameras.

[0051] If a match is successful, the system will inherit the original identity ID of the target and update its current camera position, thereby forming a continuous target trajectory and realizing cross-camera identity tracking and status maintenance.

[0052] The camera spatial topology diagram is constructed using the following method: During the initial deployment phase of the system, users can mark the installation location and viewing direction of each camera on the floor plan of the rehabilitation space through the host management interface, generating an initial version of the topology diagram.

[0053] Topological graphs are undirected graphs It means that, among them A set of camera nodes. Let be the set of spatially adjacent edges, with edge weights. Indicates camera and cameras The probability of target transfer between them.

[0054] edge weight The initial value is based on physical distance. Set, satisfy .

[0055] After the system is running, when a cross-camera identity association is successfully established, the system records the transfer event and calculates the transfer frequency within the sliding time window. Dynamically update edge weights: ; in, Learning rate (default value) ), To the camera The total number of all transition events that started.

[0056] Through the above mechanism, the system continuously learns the actual movement patterns of rehabilitation personnel, achieving adaptive optimization of the topology graph.

[0057] In the time continuity constraint, a maximum allowable time interval threshold is defined. seconds, if the time difference between two observations If so, then the associated candidate is abandoned.

[0058] Since the quality of rPPG signal extraction directly affects the accuracy of subsequent heart rate calculation, and the quality assessment and fusion compensation of multi-channel signals is one of the core aspects of this invention, it is necessary to further explain the initial rPPG signal extraction method, signal quality assessment method, and signal fusion compensation method in step S3.

[0059] As a preferred embodiment, in step S3, the specific method for extracting the initial rPPG signal from each camera's viewpoint is as follows: Based on facial landmarks, the forehead and cheek areas, which are less affected by facial expressions and occlusion, are selected as a local skin pixel set. ; For the facial area captured by the camera, extract a stable set of local skin pixels. and based on Calculate the time for each channel Average light intensity on: ; ; ; in, This represents the number of pixels in the skin region. For pixels The RGB values ​​are calculated; based on the average light intensity, the rPPG signal of a single camera is calculated using a blind source separation algorithm. : ; ; ; in, It represents the standard deviation of the signal and is used for normalization to suppress amplitude differences.

[0060] In this embodiment, the system filters out the forehead and cheek areas that are less affected by facial expression changes and occlusion based on facial key points, and extracts a stable set of local skin pixels.

[0061] By calculating the average light intensity of the RGB channels in the skin area, the pulse wave signal is separated from the RGB signal using a blind source separation algorithm (i.e., the POS algorithm, corresponding to the above formula).

[0062] The advantage of this method is that it does not require additional hardware and can achieve non-contact acquisition of physiological signals using only a regular RGB camera. Furthermore, the skin pixel set in the forehead and cheek areas has good stability relative to facial expression changes and occlusion, ensuring the signal-to-noise ratio of the initial rPPG signal.

[0063] The aforementioned rPPG signal extraction algorithm is based on the POS (Plane Orthogonal to Skin) principle in blind source separation. The core assumption of this principle is that the RGB three-channel signal in skin reflected light contains two independent components: a pulse wave signal and a noise signal (including illumination variations, motion artifacts, etc.). By constructing... and Two orthogonal projection directions: This combination is sensitive to pulse waves and also has a certain degree of robustness to noise.

[0064] This combination mainly contains noise components and is relatively insensitive to pulse waves.

[0065] By from Subtract from the middle By separating proportional components, a pure pulse wave signal can be obtained. The proportionality coefficient This minimizes the variance of the noise component in the separated signal.

[0066] This algorithm does not require the establishment of a skin color model, but only relies on the statistical characteristics of the three-channel signal, making it suitable for robust rPPG signal extraction under different skin colors and lighting conditions.

[0067] Because the quality of each signal is inconsistent, it is necessary to evaluate the signal quality to determine whether fusion compensation is needed.

[0068] In a preferred embodiment, in step S3, the quality assessment is performed by calculating the signal-to-noise ratio. Identifying low-quality signals: ; in, This indicates the calculation of variance. For the initial rPPG signal The signal after bandpass filtering in the 0.5-4.0Hz range; The higher the value, the stronger the effective physiological signal.

[0069] In this embodiment, the system performs quality assessment on the acquired rPPG signals and identifies low-quality signals by analyzing the signal-to-noise ratio. The bandpass filtering range of 0.5-4.0Hz covers the frequency band of normal adult heart rate (corresponding to 30-240 beats / min), and the filtered signal contains effective physiological signals, while the residual contains noise.

[0070] A higher signal-to-noise ratio (SNR) indicates a stronger effective physiological signal and higher signal quality. The system classifies signals with an SNR below a preset threshold as low-quality signals, triggering subsequent fusion compensation processes.

[0071] Among them, bandpass filtering A fourth-order zero-phase Butterworth digital filter is used, with a passband range of 0.5–4.0 Hz. The filter's design parameters are: sampling rate… Hz, passband ripple dB, stopband attenuation dB.

[0072] Zero-phase filtering is achieved through a forward-backward filtering method, eliminating the phase delay of traditional IIR filters and ensuring that the filtered signal is aligned with the original signal on the time axis. The signal-to-noise ratio threshold is set to... dB, when A signal is considered high-quality if it is high-quality, and low-quality if it is low-quality.

[0073] This threshold, calibrated based on a large amount of experimental data, can effectively distinguish between valid physiological signals and noise-dominated signals.

[0074] Based on the signal quality assessment method, for signals determined to be of low quality, it is necessary to compensate using high-quality signals from nearby cameras.

[0075] In a preferred embodiment, in step S3, for signals determined to be of low quality... Utilizing the high-quality signals from nearby cameras Perform weighted fusion compensation: ; in, For the compensated signal, For nearby cameras, The weights are dynamically calculated based on each signal quality index, and satisfy the following conditions: .

[0076] In this embodiment, the system normalizes the signal quality score of each high-quality signal in the set of neighboring cameras, calculated in the signal quality assessment, using the Softmax function to obtain the final fusion weight. The higher the signal quality score, the greater the fusion weight assigned to it.

[0077] For signals captured by a single camera and determined to be of high quality, the original values ​​are kept unchanged. When multiple high-quality signals exist, they can be weighted and fused to further enhance signal stability and anti-interference capabilities.

[0078] This compensation mechanism ensures that even when the quality of a camera signal deteriorates due to changes in target posture, occlusion, or lighting interference, the system can still obtain a stable rPPG signal from a nearby camera with high-quality signals, thus guaranteeing the continuity of heart rate monitoring.

[0079] Among them, weight The specific calculation method is as follows: The system will use the signal-to-noise ratio calculated in step S3. As a signal mass fraction First, the set of nearby cameras. The quality scores of each signal are subjected to an exponential transformation, and then normalized using the Softmax function. ; in, Temperature parameter (default value) ), used to adjust the smoothness of weight allocation.

[0080] when When the value approaches 0, the weight allocation tends to assign primary weights to only the highest quality signal; when... As the value approaches infinity, the weight distribution tends to become uniform.

[0081] Preferably, A value of 1.0 ensures that for every 1dB increase in signal-to-noise ratio, the weight increases exponentially by approximately [value missing]. times.

[0082] Through the above mechanism, high-quality signals gain greater weight in the fusion process, while the contribution of low-quality signals to the fusion result is effectively suppressed, thereby ensuring the quality of the compensated signal.

[0083] Nearby camera collection The determination method is as follows: based on the camera spatial topology diagram constructed in step S2, select the camera with the current camera Spatial nearest and topological edge weight One to three cameras are considered as members of a neighborhood set.

[0084] Based on the rPPG signal fusion compensation method, since anomaly detection requires extracting physiological indicators from the fused rPPG signal and comparing them with personalized benchmarks.

[0085] In a preferred embodiment, the physiological indicators extracted in step S4 include heart rate. Heart rate variability and pulse amplitude The calculation method is as follows: Let the interval between consecutive heartbeats within the detection time window be _____. ,but: ; ; in, The total number of beats within the time window; Heart rate variability The calculation method uses the standard deviation of the heartbeat interval (SDNN): ; pulse amplitude By analyzing each peak heart rate within the time window The amplitude was statistically calculated and the median was taken; The corresponding values ​​are output in real time after each time window ends, so as to carry out continuous and real-time physiological monitoring.

[0086] In this embodiment, the system extracts key physiological indicators from the fused rPPG signal in 10-second time windows.

[0087] Heart rate is calculated from the continuous heartbeat intervals within the detection time window, reflecting the current heartbeat frequency; heart rate variability is calculated using the standard deviation of heartbeat intervals, reflecting the state of autonomic nervous system regulation of the heart; pulse amplitude is obtained by taking the median of the peak heartbeat amplitude, reflecting the intensity of peripheral vascular pulsation.

[0088] The above three indicators describe the physiological state of the heart from different dimensions. Using them in combination can improve the comprehensiveness and accuracy of abnormality detection.

[0089] Among them, the intercardiac interval The detection is performed by analyzing the fused rPPG signal. Peak detection is achieved. The specific method is as follows: First of all Perform 0.5-4.0Hz bandpass filtering to remove baseline drift and high-frequency noise; Then, an adaptive thresholding method is used to detect the peak, with 1.5 times the average signal amplitude within the current time window as the dynamic threshold. When the rising edge of the signal crosses the threshold and is followed by a falling edge, a peak heart rate moment is recorded. ; The time difference between two consecutive peak heartbeats is the heartbeat interval. .

[0090] For false positive peaks (i.e., abnormal values ​​where the interval between adjacent heartbeats is less than the lower physiological limit by 0.25 seconds or greater than the upper physiological limit by 2.5 seconds), the system uses median filtering to remove them: The current cardiac interval is compared with the median of the three cardiac intervals before and after it. Intervals with a deviation of more than 50% from the median are identified as false positives and removed. Missing intervals are filled in by linear interpolation.

[0091] At least 5 valid cardiac intervals must be detected within each 10-second time window. If the number is less than this, the physiological indicators of that time window are marked as invalid and skipped.

[0092] Based on the real-time indicator calculation method, the anomaly detection requires dynamic threshold comparison by combining individual historical benchmark data.

[0093] In a preferred embodiment, the specific method for determining an abnormality when the index deviation exceeds a preset threshold in step S4 is as follows: Let the real-time indicator value be The historical mean of the corresponding individual is The historical standard deviation is The preset sensitivity coefficient is ,when It was determined to be a transient anomaly. Only in consecutive If the same anomaly is detected within a time window, it is confirmed as a valid anomaly event and an alarm is triggered.

[0094] In this embodiment, anomaly detection employs dynamic threshold logic. During the initial learning period, physiological data of the user under exercise conditions are collected, and the historical mean and standard deviation of heart rate, heart rate variability, and pulse amplitude are calculated.

[0095] The real-time indicator value is compared with the individual's historical average. If the absolute deviation exceeds kk times the individual's historical standard deviation (kk is the preset sensitivity coefficient, with a default value of 2.0), it is judged as an instantaneous anomaly.

[0096] To prevent false alarms, the system only recognizes an event as valid if the same anomaly is detected within multiple consecutive time windows (e.g., 10 time windows). Once confirmed, the system will send an alert via HTTP push notification to the app and record the event details in the database.

[0097] This dynamic threshold mechanism enables the anomaly detection threshold to be adaptively adjusted according to individual physiological characteristics, avoiding misjudgment or missed judgment for people with special physical conditions when using a fixed threshold.

[0098] Among them, the number of continuous time windows The default value is That is, the system only confirms a valid abnormal event when the same abnormality is detected in 10 consecutive time windows (corresponding to 10 consecutive seconds, since each time window is 10 seconds and the sliding step size is 1 second, the 10 time windows correspond to a continuous time period of 10 seconds).

[0099] This value balances the false alarm rate and response latency: Value too small (e.g.) This can lead to false alarms triggered by a single transient disturbance. Value too large (e.g.) This will delay the alarm and may cause you to miss the best time to intervene.

[0100] Preset sensitivity coefficient The default value is This corresponds to the boundary of a 95% confidence interval in statistics, meaning that under normal circumstances, only about 5% of the indicator values ​​will exceed this threshold. scope.

[0101] Once an alarm is triggered, the system will push structured JSON data containing user identity, alarm level, measured value of the indicator, trigger threshold and timestamp to the cloud server and the guardian's mobile APP via HTTPS protocol within 5 seconds.

[0102] Based on the above method, the present invention also provides a monitoring system for performing the above-described non-contact heart rate recognition and abnormality monitoring method, comprising: Multiple cameras are deployed in various activity areas to capture video streams in real time; The host unit is electrically connected to multiple cameras. The host unit includes a processor and a memory. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the following operations: The identity registration module is used to collect facial images of target personnel, extract facial feature vectors and bind them with unique identity identifiers and store them, and simultaneously collect resting heart rate, heart rate recovery baseline and safe heart rate zone to establish a personalized initial heart rate benchmark database. The multi-camera tracking module is used to detect and extract features from faces in the video stream, and to perform cross-camera identity association based on spatiotemporal information and feature similarity to maintain the continuity of identity identification. The rPPG signal processing module is used to extract the initial rPPG signal from the associated facial skin region, evaluate the quality of each signal, and perform weighted fusion and compensation of low-quality signals using high-quality signals from nearby cameras. The anomaly analysis and alarm module is used to calculate real-time physiological indicators based on the fused rPPG signal, compare the real-time physiological indicators with the historical baseline data of the corresponding individual to determine anomalies, and trigger an alarm when an anomaly is determined. The database module is used to securely store identity information, facial feature vectors, personalized heart rate baseline data, historical physiological signal data, and abnormal event records.

[0103] When the system is in operation, cameras are deployed in the main activity areas of the rehabilitation center to capture video streams in real time and transmit them to the host computer. The host computer then sequentially performs operations such as identity registration, multi-camera tracking, rPPG signal processing, and anomaly analysis and alarm. Specifically: The identity registration module establishes an individual identity database and a personalized initial heart rate benchmark database.

[0104] The multi-camera tracking module enables continuous identity tracking across cameras, ensuring the continuity of the target's identity identifier when switching between different functional areas.

[0105] The rPPG signal processing module extracts and fuses multiple rPPG signals to ensure that a stable fused signal can still be obtained even when the signal quality of a single viewpoint deteriorates.

[0106] The anomaly analysis and alarm module compares real-time physiological indicators with an individual's historical baseline and triggers tiered alarms when anomalies are detected.

[0107] The modular architecture of this system enables each function to be implemented independently and work together, providing a complete integrated hardware and software solution for heart rate monitoring and recovery status assessment throughout the entire exercise process in rehabilitation scenarios.

[0108] The data flow relationships between the modules in the system are as follows: The video stream captured by the cameras is transmitted via a data cable to the multi-camera tracking module on the host computer. The multi-camera tracking module then outputs the identification identifier for each target person. and the location of the face bounding box in the current frame. and timestamp .

[0109] rPPG signal processing module As input, extract the rPPG signal of the face region corresponding to the identity identifier, and output the fused rPPG signal after quality assessment and fusion compensation. .

[0110] The anomaly analysis and alarm module is based on As input, calculate physiological indicators using a sliding time window and output abnormality judgment results.

[0111] The database module receives identity information and baseline data from the identity registration module, historical physiological signal data from the rPPG signal processing module, and abnormal event records from the anomaly analysis and alarm module, and stores them in a unified encrypted manner.

[0112] Data is transmitted between modules through an asynchronous communication mechanism based on message queues, ensuring that video stream processing (30fps) and physiological signal analysis (one time window per second) are matched in time scale.

[0113] The invention will be further illustrated by a specific application scenario below.

[0114] Scenario Setting: A sports rehabilitation center has cameras 1, 2, and 3 deployed in the training area, balance area, and recovery area, respectively, covering all activity areas of the rehabilitation personnel. Person Zhang is undergoing post-operative rehabilitation training.

[0115] 1. Identity registration stage: When Zhang first visited the rehabilitation center, the system captured facial images of his frontal and half-profile faces using any of the cameras. MTCNN was used to detect facial landmarks in the images, obtaining 68 facial key points. Face alignment was performed based on these key points, and the image size was normalized to 112×112 pixels.

[0116] Standardized images are input into the MobileFaceNet network to extract 512-dimensional facial feature vectors, which are then bound to a unique identifier (such as "P001") and stored encrypted. Simultaneously, Zhang's resting heart rate of 72 beats / min, heart rate recovery baseline (the decrease in heart rate 3 minutes after exercise cessation), and safe heart rate range (80–140 beats / min) are collected to establish a personalized initial heart rate baseline database corresponding to the identifier.

[0117] 2. Multi-camera tracking stage: Mr. Zhang began rehabilitation training, performing limb movements in the training area. Camera 1 captured the video stream in real time, performing face detection and extracting feature vectors for each frame.

[0118] When Zhang moved from the training area to the balance area, the face in the field of view of camera 1 gradually disappeared, and a new target was detected in the field of view of camera 2.

[0119] The system associates the target in camera 2 with the identity identifier P001 based on the spatial topology of the cameras (camera 1 and camera 2 are adjacent cameras), temporal continuity (the time difference between the disappearance of the target in camera 1 and the appearance of the target in camera 2 is <2 seconds) and the cosine similarity of facial features (similarity >0.85), thus maintaining the continuity of the identity identifier.

[0120] 3. rPPG signal extraction and fusion stage: Based on the tracked facial region of Zhang, the system extracts the initial rPPG signal from each camera's perspective.

[0121] When Zhang was running in the training area, the signal quality from the perspective of camera 1 decreased due to changes in his movement posture. The calculated SNR value was 5.2dB, which was determined to be a low-quality signal.

[0122] The system detected an SNR of 18.7dB from the nearby camera 2, which was determined to be a high-quality signal.

[0123] The system uses the high-quality signal from camera 2 to perform weighted fusion compensation on the low-quality signal from camera 1, generating a stable fused rPPG signal from which heart rate information is extracted.

[0124] 4. Anomaly detection stage: The system calculates Zhang's heart rate, HRV, and pulse amplitude in real time using a 10-second sliding time window.

[0125] After completing a set of high-intensity training, Mr. Zhang sat quietly in the recovery area to recover. The system monitored Mr. Zhang's heart rate in real time: for the first 30 seconds, his heart rate was 138 beats / minute, which was near the upper limit of the safe range; at the 40th second, his heart rate was 145 beats / minute.

[0126] The system compared this value with Zhang's personal historical benchmark (mean 128 times / min, standard deviation 8 times / min). The deviation was 17 times / min, which is more than twice the standard deviation (16 times / min), and was therefore judged as a transient anomaly.

[0127] After continuous monitoring for 10 time windows, Zhang's heart rate remained between 142 and 148 beats per minute, which the system confirmed as a valid abnormal event. The system pushed an alarm to the guardian's mobile app via HTTPS protocol, with the message: "Zhang's heart rate is 145 beats per minute, exceeding the individual baseline limit. Please pay attention immediately."

[0128] Meanwhile, the event details (User ID: P001, Alarm Level: Level 1, Measured Heart Rate: 145 bpm, Personal Benchmark Average: 128 bpm, Standard Deviation: 8 bpm, Trigger Threshold: 16 bpm, Timestamp) are encrypted and stored in the database.

[0129] It should be noted that the cameras involved in this invention are deployed in designated functional areas of the rehabilitation center, such as the training area, balance area, and recovery area, which are controlled environments known to the rehabilitation personnel. The camera's data collection range is limited to rehabilitation training activities within each functional area, and does not involve indiscriminate monitoring of public places. Furthermore, identity information and physiological data are stored locally with encryption and are used solely for individualized assessment of motor recovery status. Rehabilitation personnel have completed informed consent and identity registration procedures upon entering the center, and the data usage is strictly limited to rehabilitation assessment purposes.

[0130] Furthermore, when this solution is applied to public exercise areas in non-rehabilitation institutions such as homes or gyms, the system only collects facial images and heart rate signals of volunteers. The scope of data use is strictly limited to individualized assessment of exercise recovery status and does not involve indiscriminate monitoring in public places.

[0131] In addition, although the technical solution provided by this invention is mainly applied to dynamic sports scenarios in multiple areas such as rehabilitation centers and gyms, the technical architecture of "identity registration → multi-camera collaborative tracking → cross-camera rPPG signal fusion → personalized anomaly judgment" is also applicable to static monitoring scenarios in a single area.

[0132] When applied to static heart rate monitoring of the elderly or children in a home environment, the mechanism of multi-camera collaborative tracking and cross-regional identity association is simplified to continuous observation under a single camera or fixed viewpoint, while the core logic of rPPG signal extraction and personalized benchmark comparison remains unchanged.

[0133] This solution features technical redundancy designed for dynamic multi-region scenarios, ensuring higher monitoring stability and reliability when backward compatible with simpler scenarios such as static and single-region environments. In other words, a solution that works effectively in dynamic multi-region scenarios will also exhibit good heart rate recognition and anomaly detection capabilities when applied to static scenarios.

[0134] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and not to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. All such modifications or substitutions should be covered within the protection scope of this application, and should not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A non-contact heart rate recognition and abnormality monitoring method, characterized in that, Includes the following steps: Step S1, Identity Registration and Benchmark Collection: Collect facial images of the target personnel, extract facial feature vectors and bind them with unique identity identifiers to establish an individual identity database; simultaneously collect the target personnel's resting heart rate, heart rate recovery baseline and safe heart rate range to establish a personalized initial heart rate benchmark database corresponding to the individual identity database; Step S2, Multi-camera collaborative tracking: By deploying multiple cameras in various areas to collect video streams in real time, face detection and feature extraction are performed on each frame of the image, and based on the spatial topology of the cameras, temporal continuity and facial feature similarity, the same target person under different camera views is associated with the unique identity identifier, thereby realizing continuous identity tracking across cameras. Step S3, Cross-camera rPPG signal fusion and compensation: Based on the face regions of the same identity that have been tracked, the initial rPPG signal from each camera viewpoint is extracted; the quality of each initial rPPG signal is evaluated, and the low-quality signal is weighted, fused and compensated using the high-quality signal collected by the neighboring camera to generate the fused rPPG signal; Step S4, Personalized Anomaly Detection: Based on the fused rPPG signal, extract real-time physiological indicators by sliding time window; compare the real-time physiological indicators with the corresponding historical benchmark data in the personalized initial heart rate benchmark database; when the indicator deviation exceeds the preset threshold, it is determined to be abnormal and an alarm is triggered; at the same time, the data within the deviation is added to the benchmark database and the benchmark database is updated in real time.

2. The non-contact heart rate recognition and abnormality monitoring method according to claim 1, characterized in that, In step S1, the facial feature vector is extracted in the following way: Facial landmark localization and face alignment are performed on the acquired raw face images to obtain standardized face images; The standardized face image is input into a lightweight face feature extraction network to obtain a deep feature vector; The deep feature vector is bound to the unique identifier and then stored locally in an encrypted manner.

3. The non-contact heart rate recognition and abnormality monitoring method according to claim 1, characterized in that, In step S2, face detection, face alignment, and feature extraction are performed sequentially in each video frame captured by the camera, specifically as follows: Let the first Each camera is constantly The captured video frames are Face detection is performed on the image to obtain face bounding boxes and key point information: ； in, This represents the detected face bounding box; Indicates the corresponding facial key points; A network for extracting facial landmarks; Geometric alignment and normalization of the face region are performed based on key points to obtain a standardized face image. : ； The aligned face image is then input into the face feature extraction network. Feature extraction is performed to obtain the corresponding facial feature vector. : ； in, Given a facial feature vector, combine this feature vector with a timestamp. Output them together for subsequent identity matching and cross-camera tracking.

4. The non-contact heart rate recognition and abnormality monitoring method according to claim 3, characterized in that, In step S2, the cross-camera identity association specifically includes: Before performing cross-camera facial feature vector comparison, illumination compensation preprocessing is performed on the acquired facial images to eliminate feature drift interference caused by differences in ambient color temperature and illumination intensity between different cameras. Install camera At any moment The obtained facial feature vector is ,Camera At any moment The obtained facial feature vector is Calculate cosine similarity : ； Under the conditions of satisfying the preset time continuity constraints and camera spatial topology constraints, the target with the highest similarity is selected as the same identity identifier.

5. The non-contact heart rate recognition and abnormality monitoring method according to claim 1, characterized in that, In step S3, the specific method for extracting the initial rPPG signal from each camera's viewpoint is as follows: Based on facial landmarks, the forehead and cheek areas, which are less affected by facial expressions and occlusion, are selected as a local skin pixel set. ; For the facial area captured by the camera, extract a stable set of local skin pixels. and based on Calculate the time for each channel Average light intensity on: ； ； ； in, This represents the number of pixels in the skin region. For pixels RGB values; Based on the average light intensity, the rPPG signal of a single camera is calculated using a blind source separation algorithm. : ； ； ； in, It represents the standard deviation of the signal and is used for normalization to suppress amplitude differences.

6. The non-contact heart rate recognition and abnormality monitoring method according to claim 5, characterized in that, In step S3, the quality assessment is performed by calculating the signal-to-noise ratio. Identifying low-quality signals: ； in, This indicates the calculation of variance. For the initial rPPG signal The signal after bandpass filtering in the 0.5-4.0Hz range; The higher the value, the stronger the effective physiological signal.

7. The non-contact heart rate recognition and abnormality monitoring method according to claim 6, characterized in that, In step S3, for signals determined to be of low quality Utilizing the high-quality signals from nearby cameras Perform weighted fusion compensation: ； in, For the compensated signal, For nearby cameras, The weights are dynamically calculated based on each signal quality index, and satisfy the following conditions: .

8. The non-contact heart rate recognition and abnormality monitoring method according to claim 1, characterized in that, In step S4, the extracted physiological indicators include heart rate. Heart rate variability and pulse amplitude The calculation method is as follows: Let the interval between consecutive heartbeats within the detection time window be _____. ,but: ； ； in, The total number of heartbeats within the time window; Heart rate variability The calculation method uses the standard deviation of the heartbeat interval (SDNN): ； pulse amplitude By analyzing each peak heart rate within the time window The amplitude was statistically calculated and the median was taken; The corresponding values ​​are output in real time after each time window ends, so as to carry out continuous and real-time physiological monitoring.

9. A non-contact heart rate recognition and abnormality monitoring method according to claim 8, characterized in that, In step S4, the specific method for determining an abnormality when the indicator deviation exceeds a preset threshold is as follows: Let the real-time indicator value be The historical mean of the corresponding individual is The historical standard deviation is The preset sensitivity coefficient is ,when It was determined to be a transient anomaly. Only in consecutive If the same anomaly is detected within a time window, it is confirmed as a valid anomaly event and an alarm is triggered.

10. A monitoring system for performing the non-contact heart rate recognition and abnormality monitoring method according to any one of claims 1 to 9, characterized in that, include: Multiple cameras are deployed in various activity areas to capture video streams in real time; A host computer is electrically connected to the plurality of cameras. The host computer includes a processor and a memory. The memory stores a computer program that, when executed by the processor, causes the processor to perform the following operations: The identity registration module is used to collect facial images of target personnel, extract facial feature vectors and bind them with unique identity identifiers and store them, and simultaneously collect resting heart rate, heart rate recovery baseline and safe heart rate range to establish a personalized initial heart rate benchmark database. The multi-camera tracking module is used to detect and extract features from faces in the video stream, and to perform cross-camera identity association based on spatiotemporal information and feature similarity to maintain the continuity of identity identification. The rPPG signal processing module is used to extract the initial rPPG signal from the associated facial skin region, evaluate the quality of each signal, and perform weighted fusion and compensation of low-quality signals using high-quality signals from nearby cameras. The anomaly analysis and alarm module is used to calculate real-time physiological indicators based on the fused rPPG signal, compare the real-time physiological indicators with the historical baseline data of the corresponding individual to determine anomalies, and trigger an alarm when an anomaly is determined. The database module is used to securely store identity information, facial feature vectors, personalized heart rate baseline data, historical physiological signal data, and abnormal event records.

Images