A non-contact multimodal respiratory monitoring method and system based on infrared sensing

Abstract

This invention relates to the field of respiratory monitoring technology and discloses a non-contact multimodal respiratory monitoring method and system based on infrared sensing. The method includes: acquiring an infrared temperature matrix and constructing a preprocessed temperature field; calculating the thermal profile gradient field and centroid displacement to generate a rigid motion vector; using the rigid motion vector to perform dynamic position compensation on the oral and nasal sampling areas, and extracting the weighted average temperature of the area using an inverse gradient weighting algorithm; demodulating the AC thermal signal through time-domain differential operation; quantifying the energy characteristics of the rigid motion vector and the AC thermal signal, and determining normal breathing, obstructive apnea, central apnea, and body movement interference states through dual-channel logic. This invention solves the problems of target loss and edge noise interference caused by body movement in non-contact monitoring through motion tracking and inverse gradient weighted sampling, and achieves accurate differentiation of pathological types of apnea by utilizing mechanical-thermal flow dual-modal characteristics.

Description

Technical Field

[0001] This invention relates to the field of respiratory monitoring technology, specifically to a non-contact multimodal respiratory monitoring method and system based on infrared sensing. Background Technology

[0002] Respiratory monitoring is of great significance for the diagnosis and care of diseases such as sleep apnea syndrome. Traditional monitoring methods, such as polysomnography (PSG), usually rely on contact sensors such as nasal cannulas or chest-abdominal breathing belts. Prolonged wear can easily cause discomfort to the subject and even interfere with normal sleep, affecting the accuracy of the monitoring data. Infrared thermal imaging technology, as a non-contact detection method, extracts respiratory signals by sensing the temperature difference between exhaled airflow and the environmental background and the human face. Due to its advantages such as passive detection, no radiation, no dependence on ambient light, and high user comfort, it is gradually becoming an important technological direction in the field of respiratory monitoring.

[0003] However, existing infrared monitoring technologies still face challenges in practical applications. During natural sleep, head movements or body shifts often cause the airflow area around the mouth and nose to deviate from the preset range, resulting in the loss of the monitoring target. Furthermore, to balance cost and privacy, low-resolution sensors are often used, and even slight body movements can produce significant non-physiological temperature jumps at the edges of the thermal image. This edge noise easily drowns out weak airflow thermal signals, leading to a decrease in the system's signal-to-noise ratio.

[0004] Accurately distinguishing between obstructive and central sleep apnea is crucial for treatment planning. The core difference lies in whether thoracic and abdominal respiratory drive is preserved when airflow ceases. Existing infrared monitoring methods mostly focus on single-modal extraction of facial airflow thermal signals, making it difficult to simultaneously quantify the body's mechanical movement characteristics that characterize respiratory effort. Due to the lack of simultaneous monitoring of respiratory drive, single thermal flow detection cannot determine the presence of respiratory drive when airflow disappears, thus making it difficult to effectively differentiate between the two pathological types. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a non-contact multimodal respiratory monitoring method and system based on infrared sensing. This solves the problems of existing non-contact respiratory monitoring technologies, such as the easy loss of sampling area when the subject moves, low signal-to-noise ratio due to large edge noise in low-resolution infrared images, and difficulty in distinguishing between obstructive and central sleep apnea pathological types using a single monitoring modality.

[0006] To achieve the above objectives, the present invention provides a non-contact multimodal respiratory monitoring method based on infrared sensing, comprising the following steps:

[0007] The original temperature matrix of the target area is acquired using an infrared thermopile array sensor, and a preprocessed temperature field is constructed by performing spatial smoothing filtering.

[0008] The spatial gradient magnitude of the preprocessed temperature field is calculated to construct the human body thermal profile gradient field. The centroid of the thermal profile is calculated based on the magnitude weighting. Rigid motion vectors representing the rigid displacement of the human body are generated based on the centroid displacement of consecutive frames.

[0009] The center of the airflow sampling area of ​​the mouth and nose is dynamically compensated by using rigid motion vectors, and the weighted average temperature of the area is extracted by using inverse gradient weighting calculation based on the inverse of gradient magnitude.

[0010] The DC component is filtered out by performing time-domain differential operation on the regional weighted average temperature, and the AC heat signal is demodulated.

[0011] The energy characteristics of rigid motion vectors and AC thermal signals are quantified, and normal breathing, obstructive apnea, central apnea, and body movement disturbances are determined based on dual-signal logic.

[0012] Furthermore, the process of performing spatial smoothing filtering to construct the preprocessed temperature field includes: constructing an expanded temperature matrix using the boundary copying and padding method; constructing a normalized Gaussian convolution kernel; and performing sliding convolution on the expanded temperature matrix using the Gaussian convolution kernel to generate the preprocessed temperature field.

[0013] Furthermore, the process of calculating the spatial gradient magnitude to construct the human body thermal contour gradient field includes: selecting the Sobel operator to convolve with the preprocessed temperature field respectively, calculating the horizontal and vertical gradient components of the pixel; calculating the Euclidean norm of the two components to obtain the gradient magnitude, and constructing the human body thermal contour gradient field.

[0014] Furthermore, the process of generating a rigid motion vector based on the centroid displacement of consecutive frames includes: calculating the ratio of the first moment to the zero moment with the gradient magnitude as the weight to obtain the weighted centroid coordinates of the human body thermal profile; and obtaining the rigid motion vector by calculating the difference between the weighted centroid coordinates of the current frame and the previous frame.

[0015] Furthermore, the process of dynamically compensating the center of the nasal and oral airflow sampling area using rigid motion vectors includes: initializing a rectangular sampling window; superimposing the rigid motion vector onto the center of the previous frame window to calculate the theoretical coordinates; determining the current frame window position through rounding and boundary clamping, so that the window follows the thermal contour of the human body for translation.

[0016] Furthermore, the process of using inverse gradient weighting based on the inverse of gradient magnitude includes: calculating the inverse of the pixel gradient magnitude within the sampling window as the inverse gradient weight coefficient (including the zero constant); calculating the weighted average of the pixel temperature values ​​within the window to obtain the regional weighted average temperature.

[0017] Furthermore, the process of performing time-domain differential operation on the regional weighted average temperature includes: maintaining a first-in-first-out queue of temperature data, calculating its arithmetic mean as a dynamic temperature baseline; and subtracting the dynamic temperature baseline from the current regional weighted average temperature to obtain the original respiratory differential signal.

[0018] Furthermore, it also includes signal gain amplification and normalization: the standard deviation of the original respiratory differential signal within the statistical sliding time window is calculated; the gain coefficient is obtained by dividing the target amplitude constant by the standard deviation, the original signal is amplified, and a standardized respiratory waveform signal is output.

[0019] Furthermore, the process of determining the respiratory state of the monitored object based on dual-signal logic includes: calculating the sum of the magnitudes of rigid motion vectors within the evaluation window as mechanical activity energy, and the sum of the absolute values ​​of standardized respiratory waveforms as thermal flow respiratory energy; if the mechanical energy is greater than the body motion determination threshold, it is determined to be body motion interference; if the mechanical energy is low and the thermal flow energy is consistently lower than the airflow cutoff threshold, it is determined to be suspected apnea; in suspected apnea, if the mechanical energy is greater than the respiratory effort threshold, it is determined to be obstructive apnea, otherwise it is determined to be central apnea.

[0020] A second aspect of the present invention provides a non-contact multimodal respiratory monitoring system based on infrared sensing, comprising:

[0021] The infrared acquisition and preprocessing module is configured to acquire the original temperature matrix and perform spatial smoothing filtering, and output the preprocessed temperature field.

[0022] The thermal gradient breathing tracking module is configured to construct a thermal profile gradient field and calculate a weighted centroid to generate a rigid motion vector representing respiratory fluctuations or body movements.

[0023] The anti-vibration airflow sampling module is configured to use a rigid motion vector to correct the coordinates of the sampling area and combine an inverse gradient weighting algorithm to output the weighted average temperature of the area.

[0024] The breathing airflow extraction module is configured to remove the DC component through time-domain differential operation and demodulate the AC thermal signal.

[0025] The dual-channel pathological judgment module is configured to calculate the energy characteristics of rigid motion vectors and alternating thermal signals, and output the states of normal breathing, obstructive apnea, central apnea, and body movement disturbance through logical judgment.

[0026] This invention provides a non-contact multimodal respiratory monitoring method and system based on infrared sensing. It has the following beneficial effects:

[0027] 1. This invention generates a rigid motion vector representing the overall displacement of the human body by constructing a human body thermal profile gradient field and calculating a weighted centroid. This vector is then used to perform real-time dynamic coordinate compensation for the airflow sampling area of ​​the mouth and nose. This mechanism allows the monitoring window to actively follow natural body movements such as head rotation or body descent, ensuring continuous locking onto the airflow generation area of ​​the mouth and nose. This solves the problem of lost or deviated regions of interest due to changes in the position of the measured object in non-contact monitoring, and guarantees signal continuity during long-term monitoring.

[0028] 2. This invention employs an inverse gradient weighting algorithm based on the inverse of the gradient magnitude within the sampling area, assigning low weights to edge contour pixels with high gradients and high weights to flat airflow regions with low gradients. This algorithm effectively suppresses non-physiological large temperature jump noise caused by minute jitter of edge pixels in low-resolution infrared images, significantly improving the signal-to-noise ratio and anti-interference capability of respiratory thermal signals while preserving weak respiratory airflow heat exchange signals.

[0029] 3. This invention constructs a mechanical-thermal flow dual-channel physiological assessment logic by simultaneously extracting the rigid motion vector characterizing the driving force of thoracic and abdominal breathing and the alternating thermal signal characterizing airflow heat exchange. The system utilizes a monocular infrared sensor to achieve synchronous monitoring of respiratory effort and actual ventilation. By determining whether mechanical respiratory driving force is retained when airflow stops, it can accurately distinguish between obstructive apnea and central apnea, overcoming the technical limitations of single thermal imaging modalities in subdividing respiratory pathological types. Attached Figure Description

[0030] Figure 1 This is a system framework diagram of an embodiment of the present invention;

[0031] Figure 2 This is a flowchart of a method according to an embodiment of the present invention;

[0032] Figure 3 This is a schematic diagram of the thermal profile construction and motion tracking logic based on thermal gradient in an embodiment of the present invention;

[0033] Figure 4 This is a schematic diagram of dynamic compensation of the sampling area and anti-jitter sampling logic based on inverse gradient in an embodiment of the present invention;

[0034] Figure 5 This is a schematic diagram of the differential demodulation logic for the respiratory airflow thermal signal in an embodiment of the present invention;

[0035] Figure 6 This is a schematic diagram of the pathological state judgment logic based on mechanical-thermal flow dual-channel energy according to an embodiment of the present invention;

[0036] Figure 7This is a waveform diagram showing the synchronous comparison between infrared non-contact monitoring data and standard polysomnography data in an embodiment of the present invention.

[0037] Among them, 10 is the infrared thermal imaging acquisition and preprocessing module; 20 is the thermal gradient respiratory drive tracking module; 30 is the anti-shaking airflow targeted sampling module; 40 is the respiratory airflow differential signal extraction module; and 50 is the dual-channel respiratory pathology judgment module. Detailed Implementation

[0038] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] See attached document Figure 1 , Figure 1 This is a structural block diagram of a non-contact multimodal respiratory monitoring system based on infrared sensing according to an embodiment of the present invention. The present invention provides a non-contact multimodal respiratory monitoring system based on infrared sensing, comprising:

[0040] The infrared thermal image acquisition and preprocessing module 10 is configured to drive the infrared sensor to acquire the original temperature matrix covering the human chest, abdomen and face, and perform spatial smoothing filtering to output a low-noise preprocessed temperature field suitable for feature extraction.

[0041] The thermal gradient breathing-driven tracking module 20 is configured to construct a human thermal profile gradient field and calculate a weighted centroid, and generate a rigid motion vector representing breathing fluctuations or body movements based on the centroid displacement.

[0042] The anti-shaking airflow targeted sampling module 30 is configured to dynamically correct the coordinates of the airflow sampling area of ​​the mouth and nose using a rigid motion vector, and combine it with an inverse gradient weighting algorithm to suppress edge noise and output a region-weighted average temperature with a high signal-to-noise ratio.

[0043] The breathing airflow differential signal extraction module 40 is configured to perform time-domain differential operation on the regional weighted average temperature to remove the DC component and demodulate the AC heat signal reflecting the airflow heat exchange.

[0044] The dual-channel respiratory pathology judgment module 50 is configured to calculate the energy characteristics of rigid motion vectors and alternating thermal signals, and then output the normal breathing, obstructive apnea, central apnea, and body movement disturbance states through logical judgment.

[0045] See attached document Figure 2 , Figure 2This is a flowchart of an infrared non-contact multimodal respiratory monitoring method based on thermal gradient motion vector compensation according to an embodiment of the present invention. The present invention provides an infrared non-contact multimodal respiratory monitoring method based on thermal gradient motion vector compensation, comprising the following steps:

[0046] S1. The infrared thermopile array sensor is used to acquire the original temperature matrix of the target area at a fixed frame rate, and spatial Gaussian smoothing filtering is performed to suppress random thermal noise, thereby constructing preprocessed temperature field data with high signal-to-noise ratio.

[0047] S2 calculates the spatial gradient magnitude of the temperature field data to construct the human body thermal profile gradient field, calculates the thermal profile centroid based on the gradient magnitude weighting, and generates a motion vector representing the overall rigid displacement of the human body based on the centroid displacement of consecutive frames.

[0048] S3 locks the airflow sampling area of ​​the mouth and nose, uses motion vectors to perform real-time dynamic compensation of its center coordinates to follow the movement of the human body, and uses the inverse of the gradient magnitude to perform inverse gradient weighting calculation on the pixels in the area to suppress edge jitter noise.

[0049] S4 extracts the weighted average temperature of the sampling area, filters out the DC component of the basal body temperature through time-domain difference operation, and demodulates the AC heat signal that reflects the changes in heat exchange of respiratory airflow.

[0050] S5 quantifies the short-time energy characteristics of the motion vector magnitude and the AC thermal signal, respectively, and determines the respiratory state based on dual-signal logic: excessive magnitude is determined to be body motion interference, both signals are high energy and they are determined to be normal breathing, only high motion energy is determined to be obstructive apnea, and both signals are low energy and they are determined to be central apnea.

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the specific embodiments of this invention will be further described in detail below in conjunction with the above-described process steps.

[0052] See attached document Figure 2 In step S1, the infrared thermal image acquisition and preprocessing module 10 mainly performs the acquisition and spatial preprocessing of infrared thermal image data, specifically including the following sub-steps:

[0053] S11, Non-contact acquisition of infrared thermal image data. The infrared thermal image acquisition and preprocessing module 10 drives the low-resolution infrared thermopile array sensor to perform data acquisition. Considering the limited sensor resolution (typically 32×32 to 120×120 pixels), in order to ensure that the contours of human facial organs and the undulations of the chest and abdomen can be distinguished simultaneously within the field of view, the sensor is fixedly installed on a bracket directly above or diagonally above the bed to be monitored. The installation height is maintained within 0.5 meters to 1.5 meters from the human body surface, so that the sensor's field of view (FOV) can completely cover the area from the head to the lower abdomen of the supine human body, and avoid the mouth and nose area being covered by the blanket.

[0054] During the acquisition process, the infrared thermal image acquisition and preprocessing module 10 controls the sensor to work continuously at a fixed sampling frame rate (e.g., 5Hz to 20Hz), converting the received infrared radiation energy of the target area into electrical signals and quantizing them to generate the original two-dimensional temperature matrix. Original two-dimensional temperature matrix The size is ,in Indicates the number of rows in the matrix. Indicates the number of columns in the matrix. Coordinates in the original two-dimensional temperature matrix. The value at the location The absolute temperature value representing the corresponding spatial location, where 1 .

[0055] S12, Spatial domain filtering preprocessing of raw temperature data. Because infrared thermopile sensors not only have internal electronic thermal noise, but are also highly susceptible to random environmental thermal noise caused by thermal radiation reflection from surrounding objects and air convection, the raw data contains non-physiological high-frequency noise. Therefore, the infrared thermal image acquisition and preprocessing module 10 employs a spatial Gaussian smoothing filter algorithm to reduce noise in the data.

[0056] First, to address the boundary effect problem where edge pixels of the matrix cannot be completely covered by the convolution kernel, the infrared thermal image acquisition and preprocessing module 10 employs a boundary copying and filling method to process the original two-dimensional temperature matrix. To extend the temperature matrix, construct an extended temperature matrix. Expanding the temperature matrix The size is ,in The filter radius is [value]. The fill rule is to [complicate] the original two-dimensional temperature matrix. The pixel values ​​at the edge are copied outwards and extended. One pixel width.

[0057] Subsequently, the infrared thermal image acquisition and preprocessing module 10 constructs a size of Gaussian convolution kernel Any position within the convolution kernel The formula for calculating the weighting coefficient is as follows:

[0058] ;

[0059] In the formula, The standard deviation of the Gaussian distribution is used to control the degree of smoothing. and These represent the horizontal and vertical coordinate offsets of elements within the convolution kernel relative to the center point, respectively, with values ​​ranging from [value range missing]. to To maintain energy conservation in the temperature field and avoid changes in overall image brightness caused by filtering, the calculated weight coefficients are normalized to ensure that the sum of all weight coefficients within the convolution kernel is equal. .

[0060] Finally, using the normalized Gaussian convolution kernel For the extended temperature matrix Perform sliding convolution operation to generate preprocessed temperature field data. For the original coordinates The formula for calculating the smoothed temperature value of a pixel at a given location is as follows:

[0061] ;

[0062] In the formula, Indicates the first [item] in the preprocessed temperature field data line, number Filtered temperature values ​​of each column pixel; These represent the row and column indices of the pixel in the original two-dimensional temperature matrix, respectively, with values ​​ranging from... , ; The radius of the Gaussian filter is a positive integer. These represent the row offset and column offset variables within the convolution kernel, respectively, with values ​​ranging from 0 to 1. to ; This represents the expanded temperature matrix after boundary filling; This indicates that the temperature matrix is ​​expanded during the convolution operation. The coordinate index of the neighboring pixels involved in the calculation; This indicates that the Gaussian convolution kernel is located at a relatively central position. The normalized weighting coefficient at the location.

[0063] Through the above calculations, the infrared thermal image acquisition and preprocessing module 10 outputs preprocessed temperature field data. It effectively suppresses high-frequency random noise while preserving the gradient contour features of human body thermal distribution.

[0064] See attached document Figure 3 , Figure 3 This is a schematic diagram of thermal contour construction and motion tracking logic based on a thermal gradient according to an embodiment of the present invention. In step S2, the thermal gradient breathing-driven tracking module 20 performs edge feature extraction and motion quantization based on the preprocessed temperature field data output in step S1, specifically including the following sub-steps:

[0065] S21, Constructing the human body thermal contour gradient field. The thermal gradient breathing-driven tracking module 20 uses a discrete spatial differential operator to perform convolution operations on the preprocessed temperature field data to extract human body edge features. The thermal gradient breathing-driven tracking module 20 selects a (3×3) Sobel operator as the gradient calculation kernel. The Sobel operator includes a horizontal convolution kernel. Convolution kernel in the vertical direction The specific matrix values ​​of these two convolution kernels are defined as follows:

[0066] ;

[0067] The thermal gradient breathing-driven tracking module 20 uses the above-mentioned horizontal convolution kernels and vertical convolution kernel Compared with preprocessed temperature field data Perform sliding convolution to calculate all pixels. horizontal gradient components at [location] and vertical gradient components Specifically, the sliding convolution operation process is as follows: The thermal gradient breathing-driven tracking module 20 aligns the center of the 3×3 convolution kernel with the target pixel in the preprocessed temperature field data, and multiplies each weight coefficient in the convolution kernel with the corresponding temperature values ​​of the target pixel and its surrounding neighborhood. Then, all multiplication results are summed, and the sum is the gradient component of the target pixel in the direction of the convolution kernel. After completing the calculation for one pixel, the thermal gradient breathing-driven tracking module 20 controls the convolution kernel to move pixel by pixel on the temperature field data in a left-to-right, top-to-bottom order until all pixels have been traversed, thereby generating complete horizontal and vertical gradient matrices. Based on this, the thermal gradient breathing-driven tracking module 20 synthesizes the gradient vectors and calculates their Euclidean norms to obtain the gradient magnitude of the pixel. Gradient magnitude The calculation formula is as follows:

[0068] ;

[0069] In the formula, This represents the rate of temperature change of a pixel in the horizontal direction. This represents the rate of temperature change of a pixel in the vertical direction. This represents the gradient magnitude intensity of a pixel. The above calculation process traverses all pixels within the field of view, and the resulting set of gradient magnitudes constitutes the human body thermal contour gradient field. In the human body thermal contour gradient field, regions with larger values ​​correspond to the boundaries where there is a significant temperature difference between the human body and the background environment, namely the head contour, shoulder contour, and chest and abdomen edges; regions with smaller values ​​correspond to flat areas where the temperature distribution is uniform in the background or inside the human body.

[0070] S22, Calculate the centroid of the thermal profile based on gradient magnitude weighting. The thermal gradient breathing-driven tracking module 20 uses the thermal profile gradient field to calculate the weighted centroid coordinates of the human body to quantify the overall position of the human body. This embodiment uses gradient magnitude... As weighting coefficients for the corresponding pixel spatial coordinates, the ratio of the first moment to the zeroth moment is calculated. Assume the height of the field of view image is... Width is The thermal gradient breathing-driven tracking module 20 calculates the current frame according to the following formula. Weighted centroid coordinates :

[0071] ;

[0072] In the formula, Indicates the first Vertical coordinates of the weighted centroid of the human body thermal profile in the frame image; Indicates the first The horizontal coordinate position of the weighted centroid of the human body thermal contour in the frame image; Represents the row index of the image matrix, corresponding to the vertical spatial position; Represents the column index of the image matrix, corresponding to the horizontal spatial position; Indicates the first Frame image in coordinates Gradient magnitude at; Indicates the total number of rows in the image; This represents the total number of columns in the image. Since edge pixels with high gradient magnitudes contribute more to the calculation of the first moment of the molecule, the weighted centroid coordinate calculation method makes the resulting centroid coordinates essentially the energy centroid of the human body thermal contour edge.

[0073] S23 generates a rigid motion vector representing body movement or respiratory fluctuations. The thermal gradient respiratory-driven tracking module 20 has an internal frame buffer unit for storing the weighted centroid coordinates calculated from the previous frame. The thermal gradient breathing-driven tracking module 20 reads the weighted centroid coordinates of the current frame. Weighted centroid coordinates of the previous frame The rigid motion vector representing the overall spatial movement trend of the human body thermal profile at adjacent sampling times is calculated through vector subtraction. Rigid motion vector The calculation formula is as follows:

[0074] ;

[0075] In the formula, Indicates the first The rigid motion vector of the frame; It represents the vertical displacement component of a rigid motion vector, reflecting the slight movement of the human body along the body axis; The horizontal displacement component represents the rigid motion vector, reflecting the slight movement of the human body in the left and right directions; and These are the centroid coordinates of the current frame; and These are the centroid coordinates of the previous frame. Rigid motion vector. The physical significance lies in the following: when the human body is in a static supine position, the rigid motion vector mainly reflects the minute centroidal oscillations caused by the periodic expansion and contraction of the chest and abdomen edges due to respiratory movements; when the human body turns over or moves its limbs significantly, the rigid motion vector reflects a large mechanical displacement. The rigid motion vector not only serves as the basis for subsequent judgment of respiratory driving force, but also as a feedback quantity transmitted to subsequent modules for dynamic coordinate compensation of the sampling area.

[0076] See attached document Figure 4 , Figure 4 This is a schematic diagram of dynamic compensation and anti-jitter sampling logic for the sampling area according to an embodiment of the present invention. In step S3, the dynamic compensation and anti-jitter sampling module 30 performs real-time position correction on the breathing sampling area based on the rigid motion vector generated in step S2, and extracts a high signal-to-noise ratio temperature signal using an inverse gradient weighted algorithm, specifically including the following sub-steps:

[0077] S31, Locking and Dynamic Coordinate Compensation of the Respiratory Sampling Area. In the initial monitoring phase, the dynamic compensation and anti-jitter sampling module 30 initializes a rectangular sampling window at the geometric center of the preprocessed temperature field data or a preset anatomical feature location. The rectangular sampling window is structurally represented as a sub-matrix extracted from the original two-dimensional temperature matrix, with its size set as follows: Pixels (e.g., 1 / 4 to 1 / 3 of the total sensor resolution). The pixel size is physically large enough to cover the mouth and nose of a person in a supine position and the surrounding airflow diffusion area.

[0078] To address the issue that the rectangular sampling window deviates from the actual mouth and nose area due to head rotation or body sliding during natural sleep, the dynamic compensation and anti-shake sampling module 30 employs a dynamic compensation mechanism based on feedforward motion vectors.

[0079] The dynamic compensation and anti-jitter sampling module 30 reads the rigid motion vector of the current frame output by the thermal gradient breathing driven tracking module 20. This is superimposed onto the center coordinates of the rectangular sampling window from the previous frame. Assume the center coordinates of the previous frame window are... Then the theoretical center coordinates of the current frame window The update formula is as follows:

[0080] ;

[0081] In the formula, They represent the first The horizontal and vertical coordinates of the center of the frame rectangular sampling window; They represent the first The horizontal and vertical coordinates of the center of the frame rectangular sampling window; They represent the first The horizontal and vertical displacement components of the frame rigid motion vector.

[0082] Considering that the index of the image matrix must be an integer, the dynamic compensation and anti-shake sampling module 30 calculates... and Rounding is performed to obtain the discretized center pixel coordinates. Simultaneously, to prevent the rectangular sampling window from moving outside the image boundary, the dynamic compensation and anti-jitter sampling module 30 performs boundary clamping operations on the center coordinates to ensure that the calculated window range always remains within the image boundaries. Within the field of view. Through the above frame-by-frame accumulation and correction, the rectangular sampling window can synchronously translate following the rigid displacement of the human body's thermal contour, thereby continuously locking onto the area where airflow occurs from the mouth and nose.

[0083] S32, construct an inverse gradient weighting coefficient matrix based on the inverse of the gradient magnitude. Within the locked rectangular sampling window, it includes high-gradient pixels of edge contours (such as facial contour edges) and low-gradient pixels of flat areas (such as skin surfaces and airflow areas). Since edge pixels in infrared images can produce large non-physiological temperature jumps when there is slight jitter, while temperature changes in flat areas are mainly caused by heat exchange from breathing airflow and are not sensitive to slight movements, the dynamic compensation and anti-jitter sampling module 30 adopts an inverse gradient weighting strategy to suppress edge noise.

[0084] The dynamic compensation and anti-shake sampling module 30 extracts all pixels within the rectangular sampling window. Corresponding gradient magnitude The gradient magnitude data originates from the gradient field generated in step S2. The dynamic compensation and anti-jitter sampling module 30 calculates the inverse gradient weight coefficient for each pixel based on the principle that larger gradients result in smaller weights, and smaller gradients result in larger weights. The calculation formula is as follows:

[0085] ;

[0086] In the formula, The local coordinates within the rectangular sampling window are: The weighting coefficients of the pixels; This indicates the gradient magnitude intensity of the pixel in the human body thermal contour gradient field; To prevent tiny positive numbers with a denominator of zero (e.g.) ), used to ensure the stability of numerical calculations and limit the maximum value of weights.

[0087] In step S32, the dynamic compensation and anti-jitter sampling module 30 generates a weight matrix that is consistent with the size of the rectangular sampling window, in which the edge texture area is given a lower weight, while the flat airflow heat exchange area is given a higher weight.

[0088] S33, Demodulation of the regional average temperature based on inverse gradient weighting. The dynamic compensation and anti-jitter sampling module 30 uses the weight matrix generated above to perform weighted average calculation on the temperature data within the rectangular sampling window to extract the respiratory thermal signal at the current moment. The dynamic compensation and anti-jitter sampling module 30 preprocesses the temperature value of each pixel within the window. Its corresponding weighting coefficient Multiply the results, sum all the products, and divide by the sum of the weighting coefficients to obtain the weighted average temperature of the current frame. The calculation formula is as follows:

[0089] ;

[0090] In the formula, Indicates the first The respiratory thermal signal output in the frame is a scalar value that represents the average temperature change caused by airflow in the mouth and nose area. Indicates the height of the rectangular sampling window (number of pixel rows); Indicates the width (number of pixel columns) of the rectangular sampling window; This represents the local row index within the rectangular sampling window, with a value ranging from 1 to... ; This represents a local column index within the rectangular sampling window, with values ​​ranging from 1 to... ; This represents the temperature value of the corresponding pixel within the rectangular sampling window in the preprocessed temperature field data. This represents the inverse gradient weight coefficient for the corresponding pixel.

[0091] Through the above weighted calculation, the final output is It suppresses motion artifacts introduced by high gradient edge pixels and mainly reflects temperature fluctuations in low gradient flat areas (i.e., facial skin and exhaled airflow areas), thus enabling the demodulation of high signal-to-noise ratio respiratory thermal waveforms even when the subject has slight body movements.

[0092] See attached document Figure 5 , Figure 5 This is a schematic diagram of differential demodulation logic for respiratory airflow thermal signals according to an embodiment of the present invention. In step S4, the respiratory signal differential demodulation module 40 performs time-domain baseline drift removal and respiratory waveform extraction on the weighted average temperature signal output in step S3, specifically including the following sub-steps:

[0093] S41, Establish a dynamic temperature baseline. Because the infrared thermopile sensor is affected by changes in ambient temperature and slow fluctuations in human body surface temperature, the weighted average temperature output in step S3... It includes DC components and low-frequency drift components. The respiratory signal differential demodulation module 40 uses a sliding time window algorithm to calculate the current dynamic temperature baseline to separate the AC component caused by respiratory airflow.

[0094] The respiratory signal differential modulation module 40 is internally configured with a length of A first-in, first-out (FIFO) queue for temperature data is used to store the most recent data. Weighted average temperature data of frames, queue length The value is based on the system sampling frame rate. Determined by a preset minimum human respiratory rate (e.g., 6 breaths / minute) to ensure the time window covers at least one complete respiratory cycle (i.e., The value is approximately equal to the sampling frame rate multiplied by 10 seconds. The respiratory signal differential demodulation module traverses the temperature data first-in-first-out queue 40 times, processing all data stored in the queue. The historical temperature data are summed up, and the sum is divided by the queue length. This yields the arithmetic mean at the current moment. The arithmetic mean is defined as the current dynamic temperature baseline. It reflects the background and average body surface temperature trend after excluding instantaneous respiratory fluctuations.

[0095] S42, Demodulation and extraction of respiratory differential signal. The respiratory signal differential demodulation module 40 performs differential operation, and calculates the weighted average temperature of the current frame. Subtract the dynamic temperature baseline calculated in step S41 Obtain the raw respiratory differential signal of the current frame. Through subtraction, the respiratory signal differential demodulation module 40 filters out low-frequency interference such as ambient temperature drift and slow changes in body temperature, while retaining the high-frequency respiratory airflow heat exchange signal.

[0096] Specifically, when the subject exhales, the hot airflow causes the value to be higher than the baseline, resulting in a positive value for the original respiratory differential signal; when the subject inhales, the cold air causes the temperature of the nasal area to be lower than the baseline, resulting in a negative value for the original respiratory differential signal.

[0097] S43, Signal gain amplification and normalization output. Because the response amplitude of non-contact infrared temperature measurement to changes in airflow is small and affected by the measurement distance, the respiratory signal differential demodulation module 40 performs adaptive gain amplification on the original respiratory differential signal.

[0098] The respiratory signal differential demodulation module 40 first calculates the fluctuation energy of the original respiratory differential signal within the current sliding time window. Specifically, the calculation method is as follows: calculate the fluctuation energy of the most recent signal within the window... Standard deviation of the raw respiratory differential signal Standard deviation The calculation follows the statistical definition, and the formula is as follows:

[0099] ;

[0100] In the formula, Indicates the first time within the sliding time window The original respiratory differential signal values ​​at each historical moment; Indicates the current sliding time window The arithmetic mean of the original respiratory differential signals of the frame; Indicates the length of the sliding time window; The summation index variable takes values ​​ranging from 0 to... .

[0101] Based on this, the respiratory signal differential demodulation module 40 dynamically normalizes the signal using a preset target amplitude constant and outputs a standardized respiratory waveform signal. The calculation formula is as follows:

[0102] ;

[0103] In the formula, Indicates the first The frame output is a standardized breathing waveform signal, which is used for subsequent frequency counting and pause detection; This indicates the preset normalization target amplitude (e.g., set to 1.0), used to unify the amplitude range of the waveform; This represents the standard deviation of the original respiratory differential signal within the current sliding time window, and characterizes the average intensity of the current respiratory signal; To prevent small protection constants where the denominator is zero (e.g.) ); This represents the original respiratory differential signal calculated in step S42.

[0104] Through the above processing, the standardized respiratory waveform signal output by the respiratory signal differential demodulation module 40 is obtained. It is a quasi-sine waveform that fluctuates around the zero axis. The quasi-sine waveform eliminates the influence of ambient temperature drift and unifies the waveform amplitude at different measurement distances.

[0105] See attached document Figure 6 , Figure 6 This is a schematic diagram of the pathological state determination logic based on a dual-channel energy system of mechanical and thermal flux according to an embodiment of the present invention. In step S5, the pathological state analysis module 50 comprehensively utilizes the rigid motion vector (mechanical channel data) output in step S2 and the standardized respiratory waveform signal (thermal flux channel data) output in step S4 to determine the current physiological state of the monitored object through a dual-channel energy assessment mechanism, specifically including the following sub-steps:

[0106] S51, constructs a dual-channel energy index for mechanical and thermal flow. The pathological state analysis module 50 sets a duration of... The evaluation time window (e.g., 10 seconds) covers the evaluation time window. Each sampling frame. The pathological state analysis module 50 performs energy quantification calculations on the mechanical channel data and the heat flow channel data respectively within the evaluation time window to obtain quantitative indicators for subsequent judgment.

[0107] For the mechanical channel, the pathological state analysis module 50 acquires the rigid motion vector for each frame within the evaluation time window. Calculate its Euclidean norm (i.e., modulus), and sum the modulus of all frames within the window to obtain the mechanical activity energy value. The mechanical activity energy value physically represents the total cumulative displacement of the human body's thermal profile on the imaging plane within the assessment period. The calculation formula is as follows:

[0108] ;

[0109] In the formula, This indicates the energy value of mechanical activity within the current evaluation window; This indicates the total number of frames within the evaluation time window; The first in the window The horizontal and vertical components of the frame rigid motion vector.

[0110] For the heat flow channel, the pathological state analysis module 50 acquires standardized respiratory waveform signals within the assessment time window. The heat flow respiration energy value is obtained by calculating the sum of their absolute values. Mechanical activity energy value Physically, it represents the overall intensity of airflow heat exchange at the mouth and nose during the assessment period. The calculation formula is as follows:

[0111] ;

[0112] In the formula, This indicates the heat flow respiratory energy value within the current assessment window; Indicates the first in the window The standardized respiratory waveform signal value of the frame.

[0113] S52, motion disturbance elimination and awake state identification. The pathological state analysis module 50 calculates the mechanical activity energy value. The value is compared with a preset body movement detection threshold. The body movement detection threshold is typically set to 5 to 10 times the microkinetic energy during resting respiration, used to distinguish large-amplitude non-respiratory movements. Specifically, the body movement detection threshold is not fixed; it is obtained by identifying a period of time during the system's initialization phase or real-time monitoring. During the stable low-energy period, the arithmetic mean of the rigid motion vector magnitude during this period is used as the baseline respiratory microkinetic energy. The calculation formula is:

[0114] ;

[0115] in, The number of sampling frames within a stable period. For the first The rigid motion vector magnitude of the frame, then according to the formula The body motion determination threshold is calculated, where the gain coefficient is... The value range is set to [15, 10]. This range is based on a large amount of experimental data and aims to construct the energy difference boundary between resting breathing micro-movements and large-amplitude macroscopic movements such as turning over and turning the head, so as to eliminate the interference of monitoring distance, target size difference and environmental thermal background fluctuation on the decision criterion and ensure the consistency of the system under different working conditions.

[0116] when When the body movement threshold is exceeded, it indicates that the monitored subject is currently turning over, moving limbs significantly, or is awake and active. In this case, the pathological state analysis module 50 marks the current state as the body movement period. To prevent misjudgment due to drastic fluctuations in the respiratory waveform baseline caused by body movement, the pathological state analysis module 50 stops executing the detection logic for apnea events during the body movement period, and the data during this period is not included in the statistical duration of the sleep apnea-hypopnea index.

[0117] S53, capture of suspected apnea events based on heat flow loss. When the value is below the body movement threshold, indicating that the monitored subject is in a relatively static state, the pathological state analysis module 50 performs apnea detection. The module will then use the heat flow respiratory energy value... Compare with the preset airflow cutoff threshold.

[0118] If the energy value of heat flow respiration If the airflow remains below the airflow cutoff threshold for an extended period exceeding the preset pathological assessment time limit (typically 10 seconds), it is determined that there has been a significant decrease or cessation of airflow through the mouth and nose. The pathological state analysis module 50 initially marks this event as an apnea event and records the start time and duration of the event.

[0119] S54, classification of apnea types based on residual mechanical micromovements. For the apnea events captured in step S53, this embodiment utilizes the differences in dual-channel data to further distinguish between central apnea and obstructive apnea. The pathological state analysis module 50 re-analyzes the mechanical activity energy values ​​during the apnea event. This is then compared with a preset breathing effort threshold. It should be noted that the breathing effort threshold is much smaller than the body motion determination threshold in step S52, its magnitude only slightly higher than the system's noise floor. It is specifically designed to identify minute vibrations caused by chest and abdominal breathing. Its acquisition is based on statistical calibration of the system's background noise. Specifically, within a calibration time window where there is no monitoring target or the monitoring object is confirmed to be absolutely stationary, the cumulative sum of the rigid motion vector magnitudes is calculated as the system's zero-input noise floor. This reflects the non-physiological readings introduced by sensor electronic thermal noise and environmental micro-airflow disturbances, and is then calculated according to the formula:

[0120] ;

[0121] in, Describe the respiratory effort threshold. The signal-to-noise ratio gain coefficient is set to a value range of [1.2, 1.5]. The signal-to-noise ratio gain value is intended to construct a sensitivity limit that closely matches the noise floor, so as to accurately capture the extremely weak mechanical vibration signal generated by abnormal breathing of the chest and abdomen while filtering background clutter.

[0122] The decision logic is as follows:

[0123] If the energy value of mechanical activity A reading above the respiratory effort threshold indicates that although airflow through the mouth and nose has ceased (thermal flow channels are silent), there are still small, regular shifts in the body's thermal profile. This corresponds to the physiological phenomenon of upper airway obstruction but the respiratory center's driving force still exists, i.e., paradoxical chest-abdominal breathing. Based on this, the pathological state analysis module 50 specifically classifies the currently captured suspected apnea event as obstructive apnea (OSA).

[0124] If the energy value of mechanical activity A reading below the respiratory effort threshold indicates that airflow through the mouth and nose has ceased, and the body's mechanical movements are also at a state of extreme stillness. This corresponds to the cessation of chest and abdominal movements due to the disappearance of respiratory center driving force. Based on this, the pathological state analysis module 50 specifically classifies the currently captured suspected apnea event as central apnea (CSA).

[0125] Through the aforementioned dual-channel decision logic, using monocular infrared thermal imaging data, accurate capture of sleep apnea events and automatic classification of their pathological types are achieved without contact with the human chest and abdomen belt.

[0126] See attached document Figure 7 , Figure 7 This is a synchronous comparison waveform diagram of infrared non-contact monitoring data and standard polysomnography data according to an embodiment of the present invention. In a specific clinical application embodiment, the monitoring system of the present invention is deployed in the laboratory of a sleep apnea monitoring center for non-contact monitoring of volunteers suspected of having sleep apnea syndrome. The hardware device includes a 32×32 resolution infrared thermopile sensor, mounted on a universal bracket 1.0 meter above the bed, with the lens vertically downwards aimed at the subject's head and chest area. To verify the accuracy of this scheme, the subject also wore a medical-grade polysomnography system as a gold standard control, which includes a nasal cannula airflow sensor and a chest and abdomen breathing sensing strap. The system acquires infrared thermal image data in real time at a frame rate of 10Hz and processes it in a pipeline according to the aforementioned five steps. The entire monitoring process lasts for approximately 8 hours overnight.

[0127] During monitoring, the system first acquires the subject's thermal field distribution through an infrared thermal imaging acquisition and preprocessing module, and removes background electronic noise using Gaussian smoothing filtering. Subsequently, a thermal gradient breathing-driven tracking module locks the subject's position based on the gradient field of the human body's thermal profile and generates a rigid motion vector. Experimental data shows that when the subject makes slight head movements or adjusts their sleeping posture, the anti-shaking airflow targeted sampling module uses this motion vector to correct the coordinates of the sampling window in real time, ensuring that the window always follows the subject's mouth and nose. Compared to fixed-point monitoring schemes without dynamic compensation, this invention reduces the signal loss rate when the subject naturally turns over and can continuously output high signal-to-noise ratio respiratory thermal waveforms.

[0128] like Figure 7 As shown in the figure, a time-series data segment containing an obstructive sleep apnea event is presented. The upper curve in the figure represents the standard airflow waveform collected by a medical nasal cannula, the middle curve represents the standardized respiratory thermal waveform demodulated by this invention, and the lower curve represents the trend of mechanical activity energy changes calculated by this invention. In the middle segment of the time axis, both the standard airflow waveform and the respiratory thermal waveform of this invention exhibit a sharp drop in amplitude, approaching zero, and the duration exceeds 10 seconds, indicating the occurrence of an airflow cessation event. At the same time, the lower mechanical activity energy curve does not return to zero but remains in a continuous oscillating state above the noise floor. This characteristic of airflow cessation with residual mechanical energy reflects the pathological mechanism of obstructive sleep apnea, namely, although the upper airway is blocked, the central nervous system is still driving the chest and abdominal muscles to make respiratory efforts.

[0129] Based on the aforementioned data characteristics, the dual-channel respiratory pathology judgment module of this invention outputs the classification results of obstructive sleep apnea within this time period, which is consistent with the conclusions of doctors manually interpreting polysomnography data. If relying solely on a single heat flow signal, the system can only identify respiratory arrest but cannot distinguish between central and obstructive apnea; if relying solely on a single radar or accelerometer mechanical signal, it may misjudge apnea as normal breathing due to the detection of chest and abdominal fluctuations. Statistical analysis of 240 hours of clinical data from 30 subjects demonstrates that this invention achieves a level of sensitivity and specificity highly correlated with contact devices in apnea event detection. Furthermore, by removing the facial contact device, the subjects' sleep comfort scores are improved, verifying the practicality of this solution in home and clinical monitoring scenarios.

[0130] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A non-contact multimodal respiratory monitoring method based on infrared sensing, characterized in that, Includes the following steps: The raw temperature matrix of the target area is acquired using an infrared thermopile array sensor, and spatial smoothing filtering is performed to construct preprocessed temperature field data. The spatial gradient magnitude is calculated on the preprocessed temperature field data to construct the human body thermal profile gradient field. The thermal profile centroid is calculated based on the gradient magnitude and a rigid motion vector representing the overall rigid displacement of the human body is generated according to the centroid displacement of consecutive frames. The sampling area of ​​the airflow from the mouth and nose is locked, and the center coordinates of the sampling area are dynamically compensated using the rigid motion vector. The weighted average temperature of the sampling area of ​​the airflow from the mouth and nose is extracted by using inverse gradient weighting based on the inverse of the gradient magnitude. The weighted average temperature of the region is subjected to time-domain difference operation to filter out the DC component and demodulate the AC thermal signal. The energy characteristics of the rigid motion vector and the AC thermal signal are quantified respectively, and the respiratory state of the monitored object is determined based on the dual-signal logic. The respiratory state includes normal breathing, obstructive apnea, central apnea, and body movement disturbance.

2. The non-contact multimodal respiratory monitoring method based on infrared sensing according to claim 1, characterized in that, The execution of spatial smoothing filtering to construct preprocessed temperature field data includes: The original temperature matrix is ​​expanded using the boundary copying and filling method to construct an expanded temperature matrix; Construct a Gaussian convolution kernel, wherein the weight coefficients within the Gaussian convolution kernel are calculated based on a Gaussian distribution and then normalized. The preprocessed temperature field data is generated by performing a sliding convolution operation on the expanded temperature matrix using the Gaussian convolution kernel.

3. The non-contact multimodal respiratory monitoring method based on infrared sensing according to claim 1, characterized in that, The step of calculating the spatial gradient magnitude of the preprocessed temperature field data to construct the human body thermal profile gradient field includes: The Sobel operator, which includes horizontal and vertical convolution kernels, is selected. The horizontal and vertical convolution kernels are respectively subjected to sliding convolution with the preprocessed temperature field data to calculate the horizontal and vertical gradient components of all pixels. Calculate the Euclidean norm of the horizontal gradient component and the vertical gradient component to obtain the gradient magnitude of all pixels, and construct the human body thermal contour gradient field from the gradient magnitudes of all pixels.

4. The non-contact multimodal respiratory monitoring method based on infrared sensing according to claim 1, characterized in that, The step of calculating the centroid of the thermal profile based on the gradient magnitude weighting, and generating a rigid motion vector representing the overall rigid displacement of the human body based on the centroid displacement of consecutive frames, includes: Using the gradient magnitude as the weighting coefficient of the corresponding pixel spatial coordinates, the ratio of the first moment to the zero moment is calculated to obtain the weighted centroid coordinates of the human body thermal contour in the current frame. Read the weighted centroid coordinates of the human body thermal outline in the current frame and the weighted centroid coordinates of the human body thermal outline in the previous frame; The rigid motion vector is obtained by calculating the difference between the weighted centroid coordinates of the human body thermal profile between the current frame and the previous frame using vector subtraction.

5. The non-contact multimodal respiratory monitoring method based on infrared sensing according to claim 1, characterized in that, The dynamic compensation of the center coordinates of the nasal and oral airflow sampling area using the rigid motion vector includes: Initialize a rectangular sampling window as the oral and nasal airflow sampling area; The rigid motion vector is superimposed onto the center coordinates of the rectangular sampling window in the previous frame to calculate the theoretical center coordinates of the rectangular sampling window in the current frame; The theoretical center coordinates are rounded and the boundary clamping operation is performed to determine the position of the rectangular sampling window in the current frame, so that the rectangular sampling window is translated following the rigid displacement of the human body thermal contour.

6. The non-contact multimodal respiratory monitoring method based on infrared sensing according to claim 5, characterized in that, The step of extracting the region-weighted average temperature of the nasal and oral airflow sampling area by using inverse gradient weighting based on the reciprocal of the gradient magnitude includes: Extract the gradient magnitude corresponding to all pixels within the rectangular sampling window; Calculate the inverse gradient weight coefficients for all the pixels, where the inverse gradient weight coefficients are the gradient magnitude plus the reciprocal of a constant to prevent the denominator from being zero; The temperature value of each pixel within the rectangular sampling window is multiplied by its corresponding inverse gradient weight coefficient and summed. This sum is then divided by the sum of all inverse gradient weight coefficients to obtain the weighted average temperature of the region.

7. The non-contact multimodal respiratory monitoring method based on infrared sensing according to claim 1, characterized in that, The step of performing time-domain difference calculation on the weighted average temperature of the region to filter out the DC component and demodulate the AC thermal signal includes: Configure a first-in-first-out queue for temperature data to store historical weighted average temperature data of a preset length; Calculate the arithmetic mean of all data in the first-in-first-out queue of the temperature data, and define it as the current dynamic temperature baseline; Subtracting the dynamic temperature baseline from the weighted average temperature of the region in the current frame yields the original respiratory differential signal, which is the basic form of the AC heat signal.

8. The non-contact multimodal respiratory monitoring method based on infrared sensing according to claim 7, characterized in that, After obtaining the original respiratory differential signal, the process also includes gain amplification and normalization of the signal: Calculate the standard deviation of the original respiratory differential signal within the current sliding time window; The gain coefficient is obtained by dividing the preset target amplitude constant by the standard deviation, and the gain coefficient is applied to the original respiratory differential signal to output a standardized respiratory waveform signal.

9. The non-contact multimodal respiratory monitoring method based on infrared sensing according to claim 8, characterized in that, The process of quantifying the energy characteristics of the rigid motion vector and the alternating thermal signal, and determining the respiratory state of the monitored object based on dual-signal logic, includes: Within the evaluation time window, the sum of the magnitudes of the rigid motion vectors is calculated as the mechanical activity energy value, and the sum of the absolute values ​​of the standardized respiratory waveform signals is calculated as the thermal flow respiratory energy value. If the mechanical activity energy value is greater than the body movement determination threshold, it is determined to be the body movement interference; If the mechanical activity energy value is less than the body movement determination threshold, and the thermal respiratory energy value remains below the airflow cutoff threshold for more than the pathological determination time limit, it is determined to be a suspected sleep apnea event. In the suspected apnea event, if the mechanical activity energy value is greater than the breathing effort threshold, it is determined to be obstructive apnea. In the suspected sleep apnea event, if the mechanical activity energy value is less than the respiratory effort threshold, it is determined to be central sleep apnea. Wherein, the breathing effort threshold is less than the body movement determination threshold.

10. A non-contact multimodal respiratory monitoring system based on infrared sensing, wherein the method is described according to any one of claims 1-9, characterized in that, include: The infrared thermal image acquisition and preprocessing module is configured to drive the infrared sensor to acquire the raw temperature matrix covering the human chest, abdomen and face, and perform spatial smoothing filtering to output preprocessed temperature field data. The thermal gradient breathing-driven tracking module is configured to construct a human thermal profile gradient field and calculate a weighted centroid, and generate a rigid motion vector representing breathing fluctuations or body movements based on the centroid displacement. The anti-vibration airflow targeted sampling module is configured to dynamically correct the coordinates of the mouth and nose airflow sampling area using the rigid motion vector, and combine it with the inverse gradient weighting algorithm to suppress edge noise and output the weighted average temperature of the area. The breathing airflow differential signal extraction module is configured to perform time-domain differential operation on the weighted average temperature of the region to remove the DC component and demodulate the AC heat signal reflecting the heat exchange of the airflow. The dual-channel respiratory pathology judgment module is configured to calculate the energy characteristics of the rigid motion vector and the AC thermal signal, and then logically distinguish and output normal breathing, obstructive apnea, central apnea, and body movement disturbance states.

Images