Non-contact vital sign state detection and recognition system

Abstract

The application relates to the field of intelligent medical treatment and computer vision technology, and discloses a non-contact vital sign state detection and recognition system. The system comprises an embedded controller and a camera, is used for collecting video data containing a face region, and obtains a region of interest of the face region through a face detection algorithm. The system extracts a color channel signal from the region of interest, obtains a stable physiological signal sequence through signal fusion, detrending processing and adaptive filtering, and extracts vital sign parameters such as a heart rate and a respiratory rate through frequency domain analysis. Meanwhile, a model is established in combination with blood flow waveform characteristics to realize blood pressure estimation, and visual information such as facial expressions and skin states is fused to perform comprehensive state analysis. The system can realize non-contact vital sign detection and health state recognition under the condition of ordinary camera equipment, has the advantages of high real-time performance, low equipment cost and wide application scenarios, and can be applied to scenes such as remote medical treatment, intelligent old-age care and daily health monitoring.

Description

Technical Field

[0001] This disclosure relates to intelligent healthcare, and more particularly to a non-contact vital sign detection and recognition system. Background Technology

[0002] With the development of artificial intelligence and medical engineering, non-contact vital sign monitoring has gradually attracted attention. Traditional methods for detecting indicators such as heart rate, blood pressure, and respiratory rate mostly rely on contact sensors, such as ECG electrodes, finger-clip pulse oximeters, and inflatable cuffs. While these methods offer high accuracy, they have many limitations in continuous monitoring, public places, telemedicine, and daily health management for the elderly, such as inconvenience in wearing them, poor comfort, and insufficient applicability to various scenarios. Therefore, researchers have begun exploring video-based methods for detecting human physiological signals to improve ease of use and universality.

[0003] In existing research, remote photoplethysmography (rPPG) has become one of the mainstream methods. For example, existing literature has pointed out that rPPG uses a standard RGB camera to capture subtle color changes on the face, achieving non-contact heart rate monitoring, eliminating the need for physical contact, and its accuracy is comparable to traditional PPG. Early rPPG methods included signal processing techniques based on PCA, ICA, CHROM, Green channels, and POS, extracting heart rate signals through ROI localization and filtering algorithms. However, these methods are relatively sensitive to changes in lighting, motion artifacts, and facial occlusion, and are not robust enough. Although deep learning methods have made positive progress in improving rPPG accuracy, such as the PhysNet 3D convolutional network method, which can significantly reduce heart rate prediction errors on datasets such as UBFC; to cope with changes in lighting, existing research has proposed image enhancement models combined with rPPG extraction networks, which significantly improve the accuracy of heart rate estimation under different lighting conditions. Furthermore, review articles have pointed out that deep learning methods are generally more accurate than traditional methods, especially performing better in complex environments. On the other hand, while commercial services such as Face++, Azure, and Amazon Rekognition offer face-based recognition of gender, age, emotion, and skin condition, these are limited to facial attribute analysis and cannot output physiological indicators such as heart rate and respiratory rate. Furthermore, they rely on cloud services, which is not conducive to real-time applications on embedded devices.

[0004] In summary, existing technologies can either only detect a single physiological indicator (such as heart rate) or only provide facial attribute analysis (such as emotion), lacking an integrated solution that can realize facial recognition, non-contact multi-parameter vital sign monitoring (heart rate, respiration, etc.), and comprehensive emotion analysis on an embedded platform. Summary of the Invention

[0005] Based on this, the present invention addresses the above-mentioned shortcomings by proposing a non-contact vital sign detection and recognition system based on face recognition and rPPG technology, which has the advantages of strong real-time performance and high scalability.

[0006] Firstly, this disclosure proposes a non-contact vital sign status detection and recognition system, comprising: a face detection and ROI extraction module, configured to adaptively generate a region of interest (ROI) in the forehead and cheek regions based on the current detection box when the coverage ratio of the current detection box and the historical region of interest (ROI) is lower than a threshold or when face detection results are continuously lost for more than a set time; and to extract a region of interest (ROI) from the face detection box and the historical region of interest (ROI) in consecutive frames. When the overlap coverage ratio is not less than a preset threshold and the face retention time is not less than a preset duration, the historical Region of Interest (ROI) is used. The signal preprocessing module is configured to extract pixel intensity information frame by frame from the ROI, convert it into a multi-channel time series signal, fuse the multi-channel signal according to a preset weight, and preprocess the fused time series signal. The spatiotemporal blood flow image construction module is configured to use the displacement field information between adjacent frames to constrain the spatial consistency of the extracted pixel intensity information and generate a virtual blood flow propagation image reflecting the local blood flow change trend. The frequency domain analysis module is configured to obtain the spectral distribution in the heart rate band and respiratory band through spectral decomposition based on the preprocessed time series signal. The output module is configured to estimate and output vital sign parameters, including blood pressure, heart rate, and respiratory rate, based on the spectral distribution in the heart rate band and respiratory band and combined with the amplitude change characteristics of the blood flow waveform.

[0007] In one embodiment of the above technical solution, continuous video stream data containing the face region is acquired by a USB camera or CSI camera connected to an embedded controller. The acquired video stream is read frame by frame in chronological order by an independent video acquisition and processing thread, and a timestamp is added to each frame. At the same time, a jitter reduction and buffering mechanism is used to ensure the uniformity and continuity of the timeline.

[0008] In one embodiment of the above technical solution, the reading of the continuous video stream containing the face region and subsequent signal processing adopt a double-buffered circular queue structure.

[0009] In one embodiment of the above technical solution, after obtaining the face detection box, the position and size parameters of the detection box are smoothed by an exponential moving average time-smoothing process with a smoothing coefficient of 0.3–0.4.

[0010] In one embodiment of the above technical solution, the system includes a face detection module, which is configured to acquire a face detection box based on each frame of a video stream containing a face region.

[0011] In one embodiment of the above technical solution, the ROI generation module is configured to further perform subpixel-level displacement compensation on the generated region of interest (ROI).

[0012] In one embodiment of the above technical solution, the multi-channel signal includes at least the average intensity sequences of the red channel, green channel and blue channel, and the signals of each channel are sampled synchronously on a unified time axis, and the fusion is a linear combination or principal component mapping.

[0013] Secondly, this disclosure provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed by any of the systems described in this disclosure.

[0014] Thirdly, this disclosure proposes a non-contact integrated physiological-psychological health analysis system, characterized in that the non-contact integrated physiological-psychological health analysis system includes: a multimodal information acquisition module configured to acquire non-contact visual features, including facial expression states, skin brightness, and skin texture; and a psychological-physiological mutual information module configured to assess the user's comprehensive health status based on non-contact visual features and physiological signals synchronously aligned on a time axis, wherein the comprehensive health status includes at least stress level, fatigue level, and mood fluctuation trend; wherein the physiological signals are acquired through any of the aforementioned systems.

[0015] The beneficial technical effects of this solution are as follows: This solution enables real-time face recognition and non-contact multi-parameter vital sign monitoring on an embedded platform. Building upon this, the invention further integrates emotion recognition, skin condition analysis, and physiological signal detection to form a unified output of multimodal health status. This solution converts the ROI region into color channels and performs color channel fusion according to preset weights to respond to changes in ambient light, achieving synchronous adaptation between the signal physical layer and the feature layer. This fundamentally solves the accuracy attenuation problem of traditional rPPG algorithms under uneven lighting and skin color differences. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1This is a schematic diagram of the workflow of a non-contact vital signs detection and recognition system in one implementation method. Detailed Implementation

[0018] Most non-contact vital sign detection methods perform well under laboratory conditions, but they have significant shortcomings in real-world applications. Existing rPPG-based methods are easily affected by changes in ambient lighting, facial movement, and camera resolution, leading to large fluctuations in heart rate and respiratory rate detection results, and insufficient stability and robustness. Furthermore, most of these methods can only monitor a single physiological indicator, commonly heart rate, with some methods extending to respiratory rate, but there is still a lack of effective methods for detecting critical vital signs such as blood pressure. In addition, existing non-contact detection solutions often focus only on physiological parameters, and the analysis of attributes such as emotion and skin condition provided by facial recognition APIs remains disconnected from vital sign detection, failing to form a unified multimodal health assessment system. Moreover, while deep learning methods can achieve high accuracy on server-side devices, their real-time operation on embedded devices is still limited by computing power and power consumption, making it difficult to meet the needs of portable and low-cost applications.

[0019] To address the aforementioned issues, this invention proposes a non-contact vital sign detection and recognition system based on face recognition and rPPG technology. It combines optical signal modeling with the self-regulating mechanism of deep neural networks to construct an "optical-neural coupled learning" framework. The system achieves synchronous adaptation between the signal physical layer and the feature layer by adjusting the weights within the neural network in real time to respond to changes in ambient light, fundamentally solving the accuracy degradation problem of traditional rPPG algorithms under uneven lighting and skin color differences. Through face detection and stable tracking of ROI (Region of Interest) on an embedded platform, combined with signal filtering and spectral analysis, the system effectively improves the robustness and accuracy of heart rate and respiratory rate estimation. Simultaneously, it introduces an empirical model to estimate blood pressure, expanding the parameter range of non-contact detection. Furthermore, this invention integrates emotion recognition and skin condition analysis provided by the Face++ API with physiological signal detection to form a multimodal health status output.

[0020] Compared with existing methods, this invention not only improves the stability and accuracy of detection, but also enables a comprehensive assessment of physiological and psychological states. It can be run in real time on embedded devices such as Jetson Nano, taking into account low cost, portability and universality of application, and meeting the needs of various scenarios such as telemedicine, smart elderly care and daily health management.

[0021] The following description, in conjunction with the accompanying drawings, clearly and completely describes how the technical solution of this case is implemented. Obviously, the described embodiments are only a part of the embodiments of this case, and not all of them. Based on the embodiments in this case, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application.

[0022] See Figure 1 The workflow of a non-contact vital sign detection and recognition system includes: acquiring facial video streams; face detection and ROI extraction; signal preprocessing; frequency domain analysis and peak localization; parameter analysis and multimodal fusion. Each workflow step can be designed as a corresponding module. Details are as follows.

[0023] 1. Capture facial video streams The non-contact vital signs detection and recognition system includes an embedded controller, which is preferably an embedded computing platform such as Jetson Nano or Orin Nano.

[0024] The embedded controller is connected to a USB camera or a CSI camera for real-time acquisition of continuous video stream data containing facial regions. To ensure the stability of subsequent remote photoplethysmography (rPPG) signal extraction, the video acquisition end preferably uses a fixed frame rate, ranging from 25 to 30 fps, with a resolution of not less than 640×480, preferably 1280×720.

[0025] During system initialization, the update range of the camera's automatic exposure and automatic white balance parameters is limited to keep the adjustment variation within a preset range between consecutive frames, thus avoiding low-frequency brightness drift noise introduced by frequent parameter fluctuations. The update range characterizes the intensity of change in the camera's automatic adjustment parameters between adjacent frames.

[0026] The system employs a dedicated video acquisition and processing thread to read the acquired video stream frame by frame in chronological order, adding a timestamp to each frame. A jitter-reducing buffer mechanism is also used to ensure the uniformity and continuity of the timeline. To suppress random noise introduced by the image sensor, lightweight image denoising processing is performed on the acquired frames without significantly increasing the computational burden, preferably using 3×3 spatial smoothing filtering or Gaussian filtering.

[0027] To balance real-time performance and power consumption control, a double-buffered circular queue structure is used for video stream reading and subsequent signal processing. The data window length for time-domain and frequency-domain analysis is preferably 20 s, with a refresh interval of 2 s, to achieve a balance between frequency domain resolution and system latency.

[0028] 2. Face detection and ROI extraction In each frame of the video stream, the system uses a lightweight face detection model to obtain face detection boxes to determine the spatial location of the face in the current frame. The lightweight face detection model can be one of the existing publicly available face detection network models, such as a lightweight detection model based on deep learning, but this invention does not limit the specific model form.

[0029] After obtaining the face detection bounding box, the position and size parameters of the bounding box are smoothed over time using an exponential moving average (EMA). The smoothing coefficient is preferably 0.3–0.4 to reduce spatial fluctuations caused by detection jitter.

[0030] Based on a smoothed face detection bounding box, Regions of Interest (ROIs) are adaptively generated in the forehead and cheek regions. To avoid interference from non-skin areas such as the eyes and lips with the optical signal, non-skin sub-regions are removed within the ROI by incorporating a skin color threshold or a facial landmark mask.

[0031] To ensure the continuity of ROI over time, the system sets a face retention time threshold and a minimum coverage ratio threshold. When the overlap ratio between the face detection box and the historical ROI in consecutive frames is not less than a preset threshold (preferably preset threshold ≥ 0.7), and the face retention time is not less than a preset duration (preferably preset duration ≥ 2 s), the historical ROI is used. When the coverage ratio between the detection box and the historical ROI is lower than the threshold or when consecutive face detection results are lost for more than the set duration, the ROI is re-determined based on the current face detection box.

[0032] In scenarios with slight head movement, subpixel-level displacement compensation is performed on the ROI by combining optical flow information or correlation tracking results between adjacent frames, so that the ROI remains relatively stable in the camera coordinate system, thereby significantly improving the signal-to-noise ratio of subsequent rPPG signal extraction.

[0033] In one embodiment, a face detection and ROI extraction module is provided, which is configured to adaptively generate a region of interest (ROI) in the forehead and cheek regions based on the current detection box when the coverage ratio between the current detection box and the historical ROI is lower than the threshold or when the face detection results are lost continuously for more than a set time. When the overlap coverage ratio between the face detection box and the historical ROI in consecutive frames is not lower than a preset threshold and the face retention time is not lower than a preset time, the historical ROI is used.

[0034] 3. Signal preprocessing The system extracts pixel intensity information frame by frame from the ROI region and converts it into a multi-channel time series signal. The multi-channel signal includes at least the average intensity sequences of the red, green, and blue channels, and the signals of each channel are sampled synchronously on a unified time axis.

[0035] During the signal construction phase, the system uses the following units for preprocessing: (1) Color channel fusion unit, used to perform linear combination or principal component mapping of multi-channel signals according to preset weights, so as to enhance the effective components related to blood volume changes; (2) Trend removal unit, used to eliminate baseline drift of the fused signal. The detrending process preferably adopts polynomial fitting or moving average to suppress low-frequency interference caused by illumination changes. (3) Adaptive filtering unit, used to dynamically adjust the filtering parameters according to the amplitude of ambient light change; (4) Bandpass filter unit, used to retain signal components within the physiological frequency range and suppress motion artifacts.

[0036] In one embodiment, a signal preprocessing module is provided, which is configured to extract pixel intensity information frame by frame from the region of interest (ROI), convert it into a multi-channel time series signal, fuse the multi-channel signal according to a preset weight, and preprocess the fused time series signal.

[0037] Based on the above preprocessing, the system further includes a spatiotemporal blood flow image construction module, which consists of a time intensity change mapping subunit and a spatial motion correlation subunit, wherein: The temporal intensity change mapping subunit is used to extract the intensity change information of each pixel within the ROI in consecutive frames; the spatial motion association subunit is used to combine the displacement field information between adjacent frames to constrain the spatial consistency of the intensity change, thereby generating a virtual blood flow propagation image that reflects the local blood flow change trend.

[0038] The spatiotemporal blood flow image construction module maps local light intensity changes on the face into a visual result similar to a blood flow velocity field, enabling non-contact characterization of blood flow changes and providing multi-dimensional feature inputs for subsequent frequency domain analysis.

[0039] 4. Frequency Domain Analysis and Peak Location After the signal stabilizes, the system performs multi-scale frequency domain analysis on the preprocessed time series signal.

[0040] For frequency domain analysis, the preferred sliding window length is 20 s, and the preferred window refresh step size is 2 s. The Fast Fourier Transform or Welch power spectrum estimation method is used to perform spectral decomposition on the signal within each time window to obtain the spectral distribution within the heart rate and respiratory frequency bands as the frequency domain analysis results.

[0041] During the peak positioning stage, the system does not rely solely on a single maximum spectral peak for judgment, but further introduces the following joint judgment mechanism: (1) spectral energy distribution ratio criterion, used to measure whether the energy proportion in the target frequency band is higher than the preset threshold; (2) phase consistency criterion, used to evaluate the phase stability of the main frequency component in adjacent time windows; (3) frequency continuity criterion, used to determine whether the change of the main peak frequency in adjacent windows is within the physiologically reasonable range.

[0042] When the above criteria are met simultaneously, the corresponding frequency is determined to be a valid physiological peak value; when the criteria are not met, the system determines that the credibility of the current signal is reduced.

[0043] When signal reliability decreases, the system calls the signal reliability assessment module, which adaptively adjusts the filtering parameters and analysis window configuration based on the current signal-to-noise ratio and spectrum stability index, forming a closed-loop self-optimizing spectrum detection process.

[0044] In one embodiment, a frequency domain analysis module is provided, which is configured to obtain the spectral distribution within the heart rate and respiratory frequency bands through spectral decomposition based on the preprocessed time series signal.

[0045] In one embodiment, a peak localization module is provided, which is configured to determine the validity of the signal based on the spectral decomposition results. When the signal is determined to be invalid, the reliability of the signal is reduced.

[0046] In one embodiment, a signal reliability assessment module is provided, which adaptively adjusts the filtering parameters and analysis window configuration based on the current signal-to-noise ratio and spectral stability index when the signal reliability decreases.

[0047] 5. Parameter Analysis and Multimodal Fusion In one embodiment, an output module is provided, configured to estimate and output vital sign parameters, including blood pressure, heart rate, and respiratory rate, based on the spectral distribution within the heart rate and respiratory frequency bands and in conjunction with the amplitude variation characteristics of the blood flow waveform. The estimation model can be an empirical regression model or a lightweight neural network model.

[0048] The system also extracts features from the acquired facial images to obtain non-contact visual features related to the face, such as facial expression states, skin brightness, and skin texture. These non-contact visual features are synchronized with physiological signals on the time axis and sent as multimodal inputs to the psycho-physiological mutual information fusion module. In one embodiment, the system includes a multimodal information acquisition module configured to acquire non-contact visual features, including facial expression states, skin brightness, and skin texture.

[0049] The psycho-physiological information fusion module is configured to assess a user’s overall health status based on non-contact visual features and physiological signals that are synchronously aligned on the timeline. The overall health status includes at least stress level, fatigue level, and mood fluctuation trend.

[0050] The psychophysiological information fusion module includes a feature encoding unit, a graph structure construction unit, and a joint inference unit. The feature encoding unit encodes multi-source heterogeneous features from the physiological signal analysis module and the facial attribute analysis module and maps them to a unified feature space. These multi-source heterogeneous features include at least heart rate features, respiration features, pulse waveform features, blood pressure estimation-related features, facial expression features, skin condition features, and their temporal dynamic features. The graph structure construction unit constructs a mutual information graph structure representing the coupling relationship between emotional changes and physiological parameters based on node features and association weights in the unified feature space. The joint inference unit performs association modeling and fusion inference on the multimodal features based on the mutual information graph structure to obtain a comprehensive health status assessment result for the user.

[0051] The comprehensive health status assessment results include at least one or more of the following: stress level, fatigue level, emotional state, emotional fluctuation trend, cardiovascular status assessment, abnormal risk warning, and comprehensive health score. The system ultimately outputs these assessment results on an embedded terminal graphical interface, simultaneously displaying heart rate, respiratory rate, blood pressure estimates and their trends, thus achieving joint monitoring and health assessment of the user's physiological and psychological state.

[0052] 6. Embedded System Architecture and Self-Organizing Collaboration Mechanism This solution adopts a self-organizing embedded multi-task collaborative architecture at the system level. Algorithm tasks are automatically allocated to the CPU, GPU, and image signal processing unit according to priority, forming a parallel pipelined processing structure to achieve dynamic collaboration of real-time signal acquisition, filtering, and result display.

[0053] Experiments show that on the Jetson Orin Nano platform, the system frame rate can be stably maintained at over 20 frames per second, with power consumption below 10 watts, demonstrating high energy efficiency and low latency. Furthermore, the system can automatically optimize parameters based on different user skin color, lighting, and posture characteristics through an online learning mechanism, achieving cross-scene self-calibration and long-term learning capabilities.

[0054] In summary, this disclosure proposes a non-contact vital sign detection and recognition system, comprising: a face detection and ROI extraction module, configured to adaptively generate ROIs in the forehead and cheek regions based on the current detection box when the coverage ratio of the current detection box and the historical ROIs is lower than the threshold or when face detection results are continuously lost for more than a set time; and to extract ROIs from the face detection box and the historical ROIs in consecutive frames. When the overlap coverage ratio is not less than a preset threshold and the face retention time is not less than a preset duration, the historical Region of Interest (ROI) is used. The signal preprocessing module is configured to extract pixel intensity information frame by frame from the ROI, convert it into a multi-channel time series signal, fuse the multi-channel signal according to a preset weight, and preprocess the fused time series signal. The spatiotemporal blood flow image construction module is configured to use the displacement field information between adjacent frames to constrain the spatial consistency of the extracted pixel intensity information and generate a virtual blood flow propagation image reflecting the local blood flow change trend. The frequency domain analysis module is configured to obtain the spectral distribution in the heart rate band and respiratory band through spectral decomposition based on the preprocessed time series signal. The output module is configured to estimate and output vital sign parameters, including blood pressure, heart rate, and respiratory rate, based on the spectral distribution in the heart rate band and respiratory band and combined with the amplitude change characteristics of the blood flow waveform. This disclosure proposes a non-contact integrated physiological-psychological health analysis system, comprising: a multimodal information acquisition module configured to acquire non-contact visual features, including facial expression states, skin brightness, and skin texture; and a psychological-physiological mutual information module configured to assess a user's overall health status based on non-contact visual features and physiological signals synchronously aligned on a time axis, wherein the overall health status includes at least stress level, fatigue level, and mood fluctuation trend; wherein the physiological signals are acquired through any of the aforementioned systems.

[0055] The system proposed in this invention can achieve non-contact vital sign detection and health status recognition under ordinary camera equipment conditions. It has the advantages of high real-time performance, low equipment cost and wide application scenarios, and can be applied to scenarios such as telemedicine, smart elderly care and daily health monitoring.

[0056] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein. The scope of the invention is defined by the appended claims.

Claims

1. A non-contact vital sign detection and identification system, characterized in that, include: The face detection and ROI extraction module is configured to adaptively generate a region of interest (ROI) in the forehead and cheek regions based on the current detection box when the coverage ratio between the current detection box and the historical ROI is lower than the threshold or when face detection results are lost continuously for more than a set time. When the overlap coverage ratio between the face detection box and the historical ROI is not lower than a preset threshold and the face retention time is not lower than a preset time in consecutive frames, the historical ROI is used. The signal preprocessing module is configured to extract pixel intensity information frame by frame from the region of interest (ROI), convert it into a multi-channel time series signal, fuse the multi-channel signal according to a preset weight, and preprocess the fused time series signal. The spatiotemporal blood flow image construction module is configured to use displacement field information between adjacent frames to constrain the spatial consistency of extracted pixel intensity information, thereby generating a virtual blood flow propagation image that reflects the local blood flow change trend. The frequency domain analysis module is configured to obtain the spectral distribution within the heart rate and respiratory frequency bands through spectral decomposition based on the preprocessed time series signal. The output module is configured to estimate and output vital signs parameters, including blood pressure, heart rate, and respiratory rate, based on the spectral distribution within the heart rate and respiratory frequency bands and in combination with the amplitude variation characteristics of the blood flow waveform.

2. The system according to claim 1, characterized in that, The continuous video stream data containing the face region is acquired through a USB camera or CSI camera connected to the embedded controller. The acquired video stream is read frame by frame in chronological order through an independent video acquisition and processing thread, and a timestamp is added to each frame. At the same time, a jitter reduction and buffering mechanism is used to ensure the uniformity and continuity of the timeline.

3. The system according to claim 1, characterized in that, The reading of continuous video streams containing face regions and subsequent signal processing adopt a double-buffered circular queue structure.

4. The system according to claim 1, characterized in that, After obtaining the face detection bounding box, the position and size parameters of the bounding box are smoothed by exponential moving average row time with a smoothing coefficient of 0.3–0.

4.

5. The system according to claim 1, characterized in that, The system includes a face detection module, which is configured to acquire a face detection box based on each frame of a video stream containing a face region.

6. The system according to claim 1, characterized in that, The ROI generation module is configured to also perform subpixel-level displacement compensation on the generated region of interest (ROI).

7. The system according to claim 1, characterized in that, The multi-channel signal includes at least the average intensity sequences of the red, green, and blue channels, and the signals of each channel are sampled synchronously on a unified time axis. The fusion is a linear combination or principal component mapping.

8. A computer-readable storage medium, characterized in that: The system stores a computer program that can be loaded by a processor and executed by the system as described in any one of claims 1 to 7.

9. A non-contact integrated physiological-psychological health analysis system, characterized in that, The non-contact integrated physiological-psychological health analysis system includes: The multimodal information acquisition module is configured to acquire non-contact visual features, including facial expression states, skin brightness, and skin texture. The psycho-physiological mutual information module is configured to assess a user’s overall health status based on non-contact visual features and physiological signals that are synchronously aligned on a time axis. The overall health status includes at least stress level, fatigue level, and mood fluctuation trend. The physiological signal is acquired by the system described in any one of claims 1-7.

10. The integrated physiological-psychological health analysis system according to claim 9, characterized in that, The overall health status is displayed in a graphical interface.

Images