A video physiological signal dynamic extraction method and device

Abstract

The application discloses a kind of video physiological signal dynamic extraction method and device, collect video and carry out face detection and posture correction to each frame video;The corrected face image is divided into micro-grid, and the brightness of each micro-grid is response-delay coding, and the pulse sequence of presynaptic pulse neuron is generated;Through membrane potential model and mutual inhibition between pulse neurons, competition is carried out and combined with connection weight, and the pulse emission time and pulse frequency of postsynaptic pulse neuron are obtained;According to the time difference between presynaptic and postsynaptic pulse neuron, dynamically adjust connection weight;After sorting and screening the pulse frequency of all postsynaptic neurons, it is mapped to the pulse sequence formed by corresponding micro-grid and weighted fusion, and the final pulse wave signal is generated;Finally, the physiological parameters of practitioner are obtained by calculating pulse wave signal, and the connection weight is adjusted by signal quality evaluation, and the strategy optimization of pulse wave signal extraction is realized continuously.

Description

Technical Field

[0001] This invention relates to the field of occupational safety and health monitoring in high-risk industries (such as coal mines, chemical plants, and power plants), and specifically to a video-based dynamic extraction method and device for physiological signals. Background Technology

[0002] In high-risk industrial environments, real-time monitoring of the physiological state of employees is of great significance. Traditional contact-based physiological monitoring devices (such as heart rate monitors and pulse oximeter finger clips) have the following problems: 1. Poor comfort: Prolonged wear can cause discomfort and affect normal work. 2. Hygiene hazards: Sharing the device with multiple people poses a risk of cross-infection. 3. Low reliability: They are prone to failure in high-temperature, high-humidity, and dusty industrial environments.

[0003] To overcome the problems of contact-based physiological monitoring devices, non-contact video physiological monitoring devices have emerged. Existing non-contact video physiological monitoring devices primarily rely on traditional computer vision technology, which involves capturing video of the worker's face and then analyzing the captured facial video and images to achieve non-contact physiological monitoring. While this device and method overcomes the problems of contact-based methods, based on current user experience, it faces the following issues because existing non-contact video acquisition uses fixed facial areas (such as cheeks, forehead, and nose) for signal extraction: 1. Static area selection: Unable to adapt to dynamic scenarios such as head movements and facial expression changes.

[0004] 2. Lagging response: Area switching requires complete video frame processing, with a response delay of hundreds of milliseconds.

[0005] 3. Feature solidification: It relies on manually determining and extracting features for analysis, and lacks adaptive learning capabilities.

[0006] 4. High computational complexity: Existing deep learning models involve large amounts of computation and are difficult to deploy on embedded devices.

[0007] In summary, existing non-contact physiological monitoring equipment suffers from technical bottlenecks such as slow response, poor environmental adaptability, and high power consumption, making it difficult to meet the rapid monitoring needs of high-risk industries. Summary of the Invention

[0008] To address the problems existing in the prior art, the present invention provides a video-based dynamic extraction method and apparatus for physiological signals, which can effectively solve the problems existing in the prior art.

[0009] To achieve the above objectives, the technical solution adopted by this invention is: a video-based dynamic extraction method for physiological signals, comprising the following steps: Step 1: Collect facial videos of employees and perform facial detection and posture correction on each frame of the video.

[0010] Step 2: Divide the corrected facial image into microgrids and perform response-delay coding on the brightness of each microgrid to generate a pulse sequence of presynaptic spiking neurons, which includes the pulse firing time of each presynaptic spiking neuron.

[0011] Step 3: Receive the pulse sequence from Step 2, compete through the membrane potential model and mutual inhibition between spiking neurons, and combine the connection weights between presynaptic and postsynaptic spiking neurons to make each microgrid correspond to a spiking event of a postsynaptic spiking neuron, and output the spiking time and number of spiking events for each spiking event.

[0012] Step 4: Dynamically adjust the connection weights based on the time difference between the firing times of the presynaptic spiking neurons and the postsynaptic spiking neurons.

[0013] Step 5: Based on the number of pulses of all postsynaptic neurons in Step 3, count the pulse firing frequency, sort them from largest to smallest, select the microgrids corresponding to the top K postsynaptic spiking neurons, and perform weighted fusion of the pulse sequences formed by these microgrids to generate the final pulse wave signal.

[0014] Step Six: Calculate the physiological parameters of practitioners by analyzing the pulse wave signal, evaluate the signal quality by calculating the quality evaluation index, and use the quality evaluation index as a reward signal to adjust the connection weights in Step Four, continuously optimizing the strategy for pulse wave signal extraction.

[0015] Furthermore, in step one, face detection and pose correction are performed on the video, specifically as follows: S1.1. Predefine the target point set and extract the facial key points of practitioners from each frame of video to form the source point set.

[0016] S1.2. The least squares method is used to solve for the 6-parameter affine transformation matrix, which is used to minimize the error between the source point set and the target point set after the affine transformation matrix is ​​applied.

[0017] S1.3. Using the affine transformation matrix obtained in step S1.2, transform the source point set to obtain the corrected facial image.

[0018] Furthermore, step two specifically involves: S2.1 Divide the corrected facial image evenly into... The system consists of overlapping microgrids, each uniquely corresponding to a presynaptic spiking neuron. This one-to-one mapping ensures the accuracy of subsequent signal tracing and extracts the RGB three-channel brightness values ​​of each presynaptic spiking neuron. .

[0019] S2.2. Using response-delay coding, the RGB three-channel brightness values ​​are converted into a pulse moment to generate a presynaptic spiking neuron pulse sequence: (S=1 indicates pulse firing, S=0 indicates no pulse); and determine the pulse timing of each presynaptic spiking neuron.

[0020] Furthermore, in step three, postsynaptic spiking neurons compete through lateral inhibition, suppressing low-frequency postsynaptic spiking neurons and enhancing the output of high-frequency postsynaptic spiking neurons, thus achieving dynamic filtering of signal sources. Specifically, each microgrid is connected to a corresponding postsynaptic spiking neuron through connection weights. To establish connections, the membrane potential of each postsynaptic spiking neuron j... As it evolves over time, the potential updates in each processing cycle as follows: in: Indicates the current pulse input With current connection weight The weighted sum, It is the connection weight updated in the previous processing cycle; This indicates the frame period, which is 33.3ms in this embodiment; This represents the membrane potential of postsynaptic spiking neuron j; This represents the membrane time constant, set to 10 ms. This represents the lateral inhibition strength coefficient, set to 0.1; This represents the resting potential, typically -70mV.

[0021] Postsynaptic spiking neuron pulse firing rules: Let... The threshold for issuance, At that time, postsynaptic spiking neurons A pulse is delivered, and then the membrane potential is reset. (Restless potential), that is: .

[0022] Through the above process, the real-time membrane potential of the postsynaptic spiking neuron is obtained. Postsynaptic spiking neuron firing timing Postsynaptic spiking neuron pulse frequency .

[0023] Furthermore, the specific formula for dynamically adjusting the connection weights in step four is as follows: in: This indicates the amount of change in weight; These represent the pulse firing times of presynaptic spiking neuron i and postsynaptic spiking neuron j, respectively. This represents the exponentially decaying term, with a value between 0 and 1. : These are the learning rates for weight enhancement and weight reduction, respectively, with values ​​ranging from [0,1]. This indicates the upper limit of the weight, set to 1; This represents the time decay constant.

[0024] Furthermore, step five specifically includes: S5.1 Calculate the firing frequency of each postsynaptic spiking neuron (corresponding to the presynaptic spiking neuron in the face), using the following formula: In the formula, Postsynaptic spiking neurons within the time window T Number of distributions Postsynaptic spiking neurons The pulse firing frequency.

[0025] S5.2, according to The signals are sorted in descending order of size, and the presynaptic spiking neurons corresponding to the top K most frequent postsynaptic spiking neurons are selected to form the optimal presynaptic spiking neuron pulse sequence for the region; where K is a preset parameter to ensure a balance between the sparsity and effectiveness of the signal source.

[0026] S5.3. The pulse wave signal is obtained by weighted summation of the presynaptic spiking neuron pulse sequences in the optimal region. The specific formula is as follows: In the formula, Indicates pulse wave signal; Indicates the normalized weights; This represents the low-pass filtering result of the pulse sequence.

[0027] Furthermore, step six specifically includes: S6.1 Calculate the signal-to-noise ratio of the pulse wave signal. The specific formula is as follows: in: : Represents the variance (effective signal strength) of the pulse wave signal. , for The time mean; The variance (noise intensity) of the signal residual is represented by residual analysis.

[0028] S6.2. The physiological parameters of the current employees are calculated through pulse wave, including heart rate, blood oxygen saturation, and blood pressure.

[0029] S6.3. Construct a quality assessment index Q for stability by calculating the signal-to-noise ratio (SNR) and heart rate, blood oxygen, and blood pressure. The specific formula is as follows: in: Indicates the preset maximum reference value for the signal-to-noise ratio; This represents the preset maximum reference value for the standard deviation of heart rate; This represents the preset maximum reference value for the standard deviation of blood oxygen. This represents the preset maximum reference value for the standard deviation of blood pressure. , : both represent weighting coefficients, and ; Indicates the standard deviation of heart rate. , For the first Heart rate values ​​within a time window This represents the average heart rate. Indicates the standard deviation of blood oxygen. , For the first Blood oxygen levels within a specific time window. Mean blood oxygen; Indicates the standard deviation of blood pressure. , For the first Blood pressure values ​​within a time window, This represents the average blood pressure.

[0030] S6.4, Reward signal R=Q, using the reward signal to dynamically adjust the learning rate in the connection weights: in: , Indicates the initial learning rate; This represents the feedback strength coefficient, which controls the degree to which the reward signal adjusts the learning rate.

[0031] Furthermore, the heart rate determination process is as follows: the peak points of the pulse wave signal are extracted by peak detection method, and the average value of the interval between adjacent peaks is calculated to obtain the heart rate.

[0032] The process of determining blood oxygen saturation is as follows: based on the AC / DC ratio method of the rPPG signal of the R channel (red light) and G channel (using green light to approximate infrared light, corresponding to the 660nm / 940nm band) in the RGB three-channel brightness values, the blood oxygen saturation is calculated by using the linear relationship between blood oxygen and light absorption coefficient.

[0033] The process of determining blood pressure is as follows: based on the characteristic parameters of rPPG signal (peak value, rise slope, decay time) combined with a linear regression model, blood pressure is calculated by utilizing the correlation between blood pressure and rPPG waveform morphology.

[0034] A video-based dynamic physiological signal extraction device includes a memory, a processor, and a computer program stored in the memory, characterized in that the processor executes the computer program to implement the steps of the above-described video-based dynamic physiological signal extraction method.

[0035] Compared with the prior art, the present invention has the following advantages: 1. This invention employs a time-series-dependent weight dynamic adjustment mechanism and a signal frequency-based competitive selection mechanism to achieve real-time adaptive optimization of physiological signal extraction sources in complex dynamic environments. This enables automatic avoidance of noise regions based on environmental changes, thereby improving the accuracy and stability of signal extraction.

[0036] 2. This invention, through a progressive processing flow of "perception-encoding-evaluation-selection-aggregation-feedback", can maintain stable facial image extraction performance under complex environments such as changes in lighting and head movements, and has strong robustness.

[0037] 3. The entire processing of the method of the present invention meets the requirements of area switching delay ≤100ms and physiological indicator monitoring time ≤8 seconds, thus having good instantaneous response capability.

[0038] 4. This invention first divides the microgrid and maps presynaptic spiking neurons and postsynaptic spiking neurons one-to-one through response-delay coding. Then, postsynaptic spiking neurons compete and filter each other through lateral inhibition, thereby finally obtaining the pulse wave signal. The entire process effectively reduces computational complexity and enables efficient operation on resource-constrained devices (such as embedded devices). Attached Figure Description

[0039] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation

[0040] The present invention will be further described below.

[0041] like Figure 1 As shown, the present invention includes the following steps: Step 1: Collect facial videos of employees and perform facial detection and pose correction on each frame of the video. Specifically: S1.1, Predefined target point set: Collect a large-scale, diverse face database (e.g., CelebA, LFW, etc.), containing tens of thousands of frontal face images with pre-annotated key points. Then, calculate the average value of each key point position in all samples, and use it as the coordinates of the corresponding key points in the standard frontal face template to form the target point set; and extract the facial key points of practitioners (such as the center of the left eye, the center of the right eye, the tip of the nose, the corner of the mouth, etc.) from each frame of video to form the source point set.

[0042] S1.2. The least squares method is used to solve for the 6-parameter affine transformation matrix, which is used to minimize the error between the source point set and the target point set after the affine transformation matrix is ​​applied. Among them, 0.7 , -0.5 , -50 , (pixels).

[0043] S1.3. Using the affine transformation matrix obtained in step S1.2, transform the source point set to obtain the corrected facial image.

[0044] Step 2: Divide the corrected facial image into microgrids and perform response-delay coding on the brightness of each microgrid to generate a pulse sequence of presynaptic spiking neurons. This pulse sequence includes the firing time of each presynaptic spiking neuron, specifically: S2.1 Divide the corrected facial image evenly into... The system consists of overlapping microgrids, each uniquely corresponding to a presynaptic spiking neuron. This one-to-one mapping ensures the accuracy of subsequent signal tracing and extracts the RGB three-channel brightness values ​​of each presynaptic spiking neuron. .

[0045] S2.2. Using response-delay coding, the RGB three-channel brightness values ​​are converted into a pulse moment to generate a presynaptic spiking neuron pulse sequence: (S=1 indicates pulse firing, S=0 indicates no pulse); and determine the pulse timing of each presynaptic spiking neuron.

[0046] Pulse firing time Defined as: The average brightness value of each presynaptic spiking neuron is in each video frame. : in: Indicates the maximum delay threshold; Represents the coding coefficients (which adjust the mapping relationship between signal strength and delay). .

[0047] Presynaptic spiking neuron pulse sequence generation rules: When hour, ;otherwise .

[0048] Step 3: Receive the pulse sequence from Step 2. Through a membrane potential model and competition between spiking neurons based on mutual inhibition, combined with the connection weights between presynaptic and postsynaptic spiking neurons, each microgrid corresponds to a spiking event of a postsynaptic spiking neuron. Competition between postsynaptic spiking neurons occurs through lateral inhibition, suppressing low-frequency spiking neurons and enhancing the output of high-frequency spiking neurons, thus achieving dynamic filtering of the signal source. The pulse firing time and number of pulse firing events are output. Specifically, each microgrid (index i) is connected to a corresponding postsynaptic spiking neuron (index j) through connection weights. To establish connections, the membrane potential of each postsynaptic spiking neuron j... As it evolves over time, the potential updates in each processing cycle as follows: in: Indicates the current pulse input With current connection weight The weighted sum, It is the connection weight updated in the previous processing cycle; This indicates the frame period, which is 33.3ms in this embodiment; This represents the membrane potential of postsynaptic spiking neuron j; This represents the membrane time constant, set to 10 ms. This represents the lateral inhibition strength coefficient, set to 0.1; This represents the resting potential, typically -70mV.

[0049] Postsynaptic spiking neuron pulse firing rules: Let... The threshold for issuance, At that time, postsynaptic spiking neurons A pulse is delivered, and then the membrane potential is reset. (Restless potential), that is: .

[0050] Through the above process, the real-time membrane potential of the postsynaptic spiking neuron is obtained. Postsynaptic spiking neuron firing timing Postsynaptic spiking neuron pulse frequency .

[0051] Step 4: Dynamically adjust the connection weights based on the time difference between the firing times of the presynaptic spiking neurons and the postsynaptic spiking neurons. The specific formula is as follows: in: This indicates the amount of change in weight; These represent the pulse firing times of presynaptic spiking neuron i and postsynaptic spiking neuron j, respectively; the presynaptic pulse firing time... Generated by response-delay coding, it is a temporal mapping of microgrid brightness, recording the triggering time of physiological signals in this region; the presynaptic pulse firing time. Equal to pulse firing time Postsynaptic pulse timing : Generated after processing by the lateral inhibition competition network, it is the output response time of the presynaptic pulse sequence after neuronal competition; This represents the exponentially decaying term, with a value between 0 and 1. : These are the learning rates for weight enhancement and weight reduction, respectively, with a value range of [0,1]. During operation, they will be adjusted by the feedback of the reward signal R from step six. This indicates the upper limit of the weight, set to 1; This represents the time decay constant.

[0052] Step 5: Based on the number of pulses from all postsynaptic neurons in Step 3, count the pulse firing frequencies and sort them from largest to smallest. Select the microgrids corresponding to the top K postsynaptic spiking neurons, and perform weighted fusion of the pulse sequences formed by these microgrids to generate the final pulse wave signal. Specifically: S5.1 Calculate the firing frequency of each postsynaptic spiking neuron (corresponding to the presynaptic spiking neuron in the face), using the following formula: In the formula, Postsynaptic spiking neurons within the time window T Number of distributions Postsynaptic spiking neurons The pulse firing frequency.

[0053] S5.2, according to Sort the neurons in descending order of size, and select the presynaptic spiking neurons corresponding to the top K most frequent postsynaptic spiking neurons to form the optimal presynaptic spiking neuron pulse sequence for the region: Where K is a preset parameter to ensure a balance between the sparsity and effectiveness of the signal source.

[0054] S5.3. The pulse wave signal is obtained by weighted summation of the presynaptic spiking neuron pulse sequences in the optimal region. The specific formula is as follows: In the formula, Indicates pulse wave signal; Indicates the normalized weights; This represents the low-pass filtering result (smoothing noise) of a pulse sequence.

[0055] The specific calculation process for the above normalized weights (positively correlated with the pulse frequency of postsynaptic spiking neurons) is as follows: The low-pass filtering result of the pulse sequence, using a first-order low-pass filter, is calculated as follows: in Filter coefficients .

[0056] Step Six: Calculate the physiological parameters of the practitioners using the pulse wave signal, evaluate the signal quality by calculating a quality assessment index, and use this index as a reward signal to adjust the connection weights in Step Four. Continuous strategy optimization is performed on the pulse wave signal extraction process, specifically as follows: S6.1 Calculate the signal-to-noise ratio of the pulse wave signal. The specific formula is as follows: in: : Represents the variance (effective signal strength) of the pulse wave signal. , for The time mean; The variance (noise intensity) of the signal residual is represented by residual analysis.

[0057] S6.2. The physiological parameters of the current employees are calculated through pulse wave, including heart rate, blood oxygen saturation, and blood pressure. The heart rate determination process is as follows: pulse waves are extracted using peak detection method. peak point Calculate the average value of the interval between adjacent peaks: The heart rate (HR) is then: .

[0058] The process for determining blood oxygen saturation is as follows: Based on the AC / DC ratio method of the rPPG signal from the R channel (red light) and G channel (approximately infrared light, corresponding to the 660nm / 940nm band) of the RGB three-channel brightness values, blood oxygen saturation is calculated using the linear relationship between blood oxygen and light absorption coefficient. Specifically: First, extract the rPPG signal components from the R channel (red light) and infrared light (IR, corresponding to the G channel): Calculate the AC (alternating current component, pulsating part) and DC (direct current component, static part): Define the light absorption ratio R: Empirical formula for blood oxygen saturation (calibrated based on Lambert-Beer Law): Where: a: calibration coefficient; b: calibration constant.

[0059] The process for determining blood pressure is as follows: based on the characteristic parameters of the rPPG signal (peak value, rise slope, decay time) combined with a linear regression model, blood pressure is calculated using the correlation between blood pressure and the rPPG waveform morphology. Specifically: Step 1: Extract key feature parameters of the rPPG signal: Peak amplitude: ; Slope of ascent: decay time: (Time to decay from peak value to 50% of amplitude); Normalized heart rate value: (Reference heart rate 100 beats / minute).

[0060] Step 2: Regression model of systolic blood pressure (SBP) and diastolic blood pressure (DBP): in: : Regression coefficients, calculated using lasso regression with 500+ samples.

[0061] S6.3. Construct a quality assessment index Q for stability by calculating the signal-to-noise ratio (SNR) and heart rate, blood oxygen, and blood pressure. The specific formula is as follows: in: Indicates the preset maximum reference value for the signal-to-noise ratio; This represents the preset maximum reference value for the standard deviation of heart rate; This represents the preset maximum reference value for the standard deviation of blood oxygen. This represents the preset maximum reference value for the standard deviation of blood pressure. , : both represent weighting coefficients, and ; Indicates the standard deviation of heart rate. , For the first Heart rate values ​​within a time window This represents the average heart rate. Indicates the standard deviation of blood oxygen. , For the first Blood oxygen levels within a specific time window. Mean blood oxygen; Indicates the standard deviation of blood pressure. , For the first Blood pressure values ​​within a time window, This represents the average blood pressure.

[0062] S6.4, Reward signal R=Q, using the reward signal to dynamically adjust the learning rate in the connection weights: in: , Indicates the initial learning rate; This represents the feedback strength coefficient, which controls the degree to which the reward signal adjusts the learning rate.

[0063] A video-based dynamic physiological signal extraction device includes a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps of the aforementioned video-based dynamic physiological signal extraction method. By running the dynamic extraction steps of this invention through this device, it can adapt to complex dynamic environments (such as changes in lighting and facial movements) in real time, continuously optimize the signal source selection strategy, and ensure the accuracy and stability of pulse wave signal extraction.

[0064] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for dynamic extraction of video-based physiological signals, characterized in that, Includes the following steps: Step 1: Collect facial videos of employees and perform facial detection and posture correction on each frame of the video. Step 2: Divide the corrected facial image into microgrids and perform response-delay coding on the brightness of each microgrid to generate a pulse sequence of presynaptic spiking neurons, which includes the pulse firing time of each presynaptic spiking neuron. Step 3: Receive the pulse sequence from Step 2. Competition is generated through membrane potential model and mutual inhibition between spiking neurons. Combined with the connection weights between presynaptic and postsynaptic spiking neurons, each microgrid corresponds to a spiking event of a postsynaptic spiking neuron. The timing and number of spiking events for each event are output. This includes: competition between postsynaptic spiking neurons through lateral inhibition, suppressing low-frequency spiking neurons and enhancing the output of high-frequency spiking neurons, thus achieving dynamic filtering of the signal source. Specifically, each microgrid is connected to a corresponding postsynaptic spiking neuron through connection weights. To establish connections, the membrane potential of each postsynaptic spiking neuron j... As it evolves over time, the potential updates in each processing cycle as follows: in: Indicates the current pulse input With current connection weight The weighted sum, It is the connection weight updated in the previous processing cycle; Indicates the frame period; This represents the membrane potential of postsynaptic spiking neuron j; Indicates the membrane time constant; Indicates the lateral inhibition strength coefficient; Indicates resting potential; Postsynaptic spiking neuron pulse firing rules: Let... The threshold for issuance, At that time, postsynaptic spiking neurons A pulse is delivered, and then the membrane potential is reset. ; Through the above process, the real-time membrane potential of the postsynaptic spiking neuron is obtained. Postsynaptic spiking neuron firing timing Postsynaptic spiking neuron pulse frequency ; Step 4: Dynamically adjust the connection weights based on the time difference between the firing times of the presynaptic spiking neurons and the postsynaptic spiking neurons. Step 5: Based on the number of pulses from all postsynaptic spiking neurons in Step 3, count the pulse firing frequencies and sort them from largest to smallest. Select the microgrids corresponding to the top K postsynaptic spiking neurons, and perform weighted fusion of the pulse sequences formed by these microgrids to generate the final pulse wave signal. Specifically: S5.1 Calculate the firing frequency of each postsynaptic spiking neuron using the following formula: In the formula, Postsynaptic spiking neurons within the time window T Number of distributions Postsynaptic spiking neurons The pulse firing frequency; S5.2, according to Sort the neurons in descending order of size, select the K most frequent postsynaptic spiking neurons and their corresponding presynaptic spiking neurons to form the optimal presynaptic spiking neuron pulse sequence for the region. S5.

3. The pulse wave signal is obtained by weighted summation of the presynaptic spiking neuron pulse sequences in the optimal region. The specific formula is as follows: In the formula, Indicates pulse wave signal; Indicates the normalized weights; This represents the low-pass filtering result of the pulse sequence; Step Six: Calculate the physiological parameters of practitioners by analyzing the pulse wave signal, evaluate the signal quality by calculating the quality evaluation index, and use the quality evaluation index as a reward signal to adjust the connection weights in Step Four, continuously optimizing the strategy for pulse wave signal extraction.

2. The video-based dynamic extraction method for physiological signals according to claim 1, characterized in that, Step one involves performing face detection and pose correction on the video, specifically as follows: S1.1, Predefine the target point set, and extract the facial key points of practitioners from each frame of video to form the source point set; S1.

2. The least squares method is used to solve for the 6-parameter affine transformation matrix, so that the error between the source point set and the target point set is minimized after the source point set is transformed by the affine transformation matrix. S1.

3. Using the affine transformation matrix obtained in step S1.2, transform the source point set to obtain the corrected facial image.

3. The video-based dynamic extraction method for physiological signals according to claim 1, characterized in that, Step two specifically involves: S2.1 Divide the corrected facial image evenly into... The overlapping microgrids, each microgrid uniquely corresponding to a presynaptic spiking neuron, and the RGB three-channel brightness values ​​of each presynaptic spiking neuron are extracted; S2.

2. Using response-delay coding, the RGB three-channel brightness values ​​are converted into a pulse moment to generate a presynaptic spiking neuron pulse sequence; and the pulse moment of each presynaptic spiking neuron is determined.

4. The video-based dynamic extraction method for physiological signals according to claim 3, characterized in that, The specific formula for dynamically adjusting the connection weights in step four is as follows: in: This indicates the amount of change in weight; These represent the pulse firing times of presynaptic spiking neuron i and postsynaptic spiking neuron j, respectively. This represents the exponentially decaying term, with a value between 0 and 1. : These are the learning rates for weight enhancement and weight reduction, respectively, with values ​​ranging from [0,1]. This indicates the upper limit of the weight, set to 1; This represents the time decay constant.

5. The video-based dynamic extraction method for physiological signals according to claim 4, characterized in that, Step six specifically involves: S6.1 Calculate the signal-to-noise ratio of the pulse wave signal; S6.

2. The physiological parameters of the current employees are calculated through pulse wave, including heart rate, blood oxygen saturation, and blood pressure. S6.

3. Construct a quality assessment index Q for stability by calculating the signal-to-noise ratio (SNR) and heart rate, blood oxygen, and blood pressure. The specific formula is as follows: in: Indicates the preset maximum reference value for the signal-to-noise ratio; This represents the preset maximum reference value for the standard deviation of heart rate; This represents the preset maximum reference value for the standard deviation of blood oxygen. This represents the preset maximum reference value for the standard deviation of blood pressure. , : both represent weighting coefficients, and ; This represents the standard deviation of heart rate; Indicates the standard deviation of blood oxygen; Indicates the standard deviation of blood pressure; S6.4, Reward signal R=Q, using the reward signal to dynamically adjust the learning rate in the connection weights: in: , Indicates the initial learning rate; This represents the feedback strength coefficient, which controls the adjustment of the learning rate by the reward signal.

6. The video-based dynamic extraction method for physiological signals according to claim 5, characterized in that, The heart rate determination process is as follows: the peak points of the pulse wave signal are extracted by peak detection method, and the average value of the interval between adjacent peaks is calculated to obtain the heart rate; The process of determining blood oxygen saturation is as follows: based on the AC / DC ratio of the rPPG signals of the R and G channels in the RGB three-channel brightness values, blood oxygen saturation is calculated using the linear relationship between blood oxygen and light absorption coefficient. The process of determining blood pressure is as follows: based on the characteristic parameters of rPPG signal combined with a linear regression model, blood pressure is calculated by utilizing the correlation between blood pressure and rPPG waveform morphology.

7. A video-based dynamic physiological signal extraction device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

Images