A non-line-of-sight moving human body breathing waveform recovery method based on millimeter wave radar

Abstract

This invention provides a method for recovering respiratory waveforms of a moving human body from non-line-of-sight (NLS) observations based on millimeter-wave radar. The method includes: transmitting a signal from the millimeter-wave radar to a target object; acquiring the reflected or diffracted echo; processing the echo to generate a virtual channel distance-time response; performing multipath sensing processing on the virtual channel distance-time response to obtain multi-channel radio frequency (RF) observations; and processing the multi-channel RF observations using a BreathDiff-SSM model to recover the respiratory waveform. This invention eliminates the need for sensors worn by the human body, cameras, and multi-radar coordination. It can recover waveform-level respiratory information from weak, multipath, time-varying, and motion-contaminated NLS millimeter-wave observations, offering advantages such as simple deployment, privacy protection, and applicability to real indoor environments.

Description

Technical Field

[0001] This invention relates to the field of millimeter-wave radar signal processing technology, and in particular to a method for recovering the non-line-of-sight moving human breathing waveform based on millimeter-wave radar. Background Technology

[0002] Breathing is a vital sign reflecting the state of the human cardiopulmonary system. Continuous, non-contact respiratory monitoring can be used in scenarios such as sleep assessment, rehabilitation care, long-term health management, and indoor security monitoring. While existing wearable sensors can obtain high-precision respiratory signals, they require contact with the human body, leading to problems such as wearing burden, poor comfort, and insufficient compliance with long-term use. Although visual, thermal imaging, and acoustic methods can achieve non-contact detection, they are easily limited by lighting conditions, occlusion, posture, environmental noise, and privacy protection requirements.

[0003] Millimeter-wave radar offers advantages such as non-contact operation, privacy, insensitivity to light, miniaturization, and low-power deployment, making it suitable for indoor respiratory sensing. However, in real indoor environments, the human body is rarely a stationary target; it may walk, turn, sway, or shift relative to the radar. Furthermore, walls, corridor corners, furniture, cabinets, or the direction the body faces can obstruct the direct path of the radar to the chest cavity, making non-line-of-sight (NLOS) propagation a common occurrence.

[0004] Under non-line-of-sight conditions, the direct reflection path of radar to the chest cavity is blocked, and respiratory information can only be transmitted through indirect multipath components formed by reflection, diffraction, or multiple scattering. These components are typically weak in intensity, vary significantly with scene geometry, and are easily interfered with by static clutter and other human motion. For moving human bodies, the phase change introduced by macroscopic body motion is much greater than the micro-displacement of chest cavity breathing, resulting in strong non-stationary characteristics in the slow-time radar signal. Therefore, non-line-of-sight sensing of breathing in moving human bodies is not simply a noise-enhanced version of ordinary line-of-sight breathing detection, but rather a problem requiring the recovery of the underlying respiratory waveform from weak, multipath, time-varying observations contaminated by motion.

[0005] Existing radio frequency (RF) breathing sensing methods are mostly focused on line-of-sight or weak obstruction conditions, stationary targets, breathing rate estimation, or simple motion suppression. In completely non-line-of-sight scenarios with human movement, relying solely on the strongest peak, fixed path modeling, or reference channel cancellation makes it difficult to consistently obtain a stable breathing path and to recover a morphologically consistent breathing waveform. Therefore, a method is needed to recover high-fidelity breathing waveforms in non-line-of-sight moving scenarios using a single commercial millimeter-wave radar. Summary of the Invention

[0006] The purpose of this invention is to provide a non-line-of-sight moving human breathing waveform recovery method based on millimeter-wave radar. Through multipath sensing preprocessing and conditional diffusion state space recovery model, the human breathing waveform is recovered from dynamically changing non-line-of-sight indirect reflections.

[0007] To achieve the above objectives, the present invention provides a method for recovering the respiratory waveform of a moving human body without line of sight based on millimeter-wave radar, the method comprising: S11. The millimeter-wave radar transmits signals to the target object, obtains the echo after reflection or diffraction, processes the echo, and generates a virtual channel range-time response. S12. Perform multipath sensing processing on the virtual channel distance-time response to obtain multi-channel radio frequency observations; S13. The multi-channel radio frequency observations were processed using the BreathDiff-SSM model to restore the respiratory waveform.

[0008] Furthermore, a virtual channel distance-time response is formed, specifically including: S21. Deploy a single commercial FMCW millimeter-wave radar on one side of the target scene, so that the human body being tested is in a non-line-of-sight area where the direct path of the radar wave is blocked. S22. Control the millimeter-wave radar to transmit linear frequency modulated continuous wave signals and receive radio frequency echo signals after reflection or diffraction by human body and environmental structures. S23. Mix the radio frequency echo signal with the local oscillator signal of the millimeter-wave radar to obtain the intermediate frequency analog signal, and then perform analog-to-digital conversion sampling to generate the original ADC data. S24. Using MIMO virtual array technology, the multi-channel echo components in the original ADC data are separated and rearranged to obtain spatial dimension information. Then, the data is reassembled according to the dimensional order of frame, chirp, virtual antenna channel and fast time sampling to construct multi-dimensional radar data. S25. The fast time dimension of the multidimensional radar data is processed by range FFT, while the slow time dimension is processed by static clutter suppression and the slow time axis is phase-sorted to finally generate the virtual channel range-time complex response.

[0009] Furthermore, multipath sensing processing is performed on the virtual channel distance-time response, specifically including: S31. Perform minimum variance distortionless response beamforming on each antenna domain snapshot in the virtual channel range-time response to obtain the angle-range domain response spectrum; S32. After removing the DC component from the angle-distance domain response spectrum along the time dimension, extract multiple candidate peaks from the angle-distance domain response spectrum of each frame to form a candidate multipath peak set. S33. Calculate the weighted distance based on the angle difference and distance difference of each candidate peak, and construct the cross-frame association cost matrix; S34. The Hungarian matching algorithm is used to solve the global minimum cost of the cross-frame association cost matrix, obtain the one-to-one optimal association between candidate peaks, and form a time-continuous non-line-of-sight multipath trajectory. S35. Based on the distance position corresponding to the non-line-of-sight multipath trajectory, extract and expand the phase sequence from the virtual channel distance-time response; S36. Stack all virtual channel distance-time responses and phase sequences of all non-line-of-sight multipath trajectories to generate multi-channel radio frequency observations.

[0010] Furthermore, the angle-distance domain response spectrum is generated, specifically including: S41. Determine the set of scanning angles based on the detection field of view of the millimeter-wave radar and the preset angle step size; S42. Estimate the sample covariance matrix by taking a snapshot of the antenna domain at each range cell in the t-th frame of the virtual channel range-time response and introducing a diagonal loading term to obtain the loaded covariance matrix. S43. Based on each scanning angle in the set of scanning angles, construct the array steering vector corresponding to the current angle, combine the array steering vector with the covariance matrix after loading, and calculate the beam weight with minimum variance distortion-free response. S44. The antenna domain snapshots are weighted and summed using beam weights to obtain the array response in the current scanning angle direction; S45. By traversing the set of scanning angles, the array response of the corresponding distance unit under multiple scanning angles is obtained, forming the angle-dimensional response spectrum of the corresponding distance unit; S46. Traverse all distance cells in frame t, arrange the angular response spectra of all distance cells according to the distance dimension, and obtain the angular-distance domain response spectrum of frame t.

[0011] Furthermore, multi-peak extraction is performed on the angle-range spectrum of each frame, specifically including: S51. For any angle-distance unit, if the response amplitude is greater than the response amplitude of the adjacent preceding and following distance units in the distance dimension, and greater than the response amplitude of the adjacent preceding and following angles in the angle dimension, then the angle-distance unit is determined to be a local maximum. S52. At the same time, the amplitude of the candidate peak must also be greater than the product of the median amplitude of all units in the current frame angle-distance spectrum and the first preset coefficient. S53. Sort all candidate peaks that simultaneously satisfy the local maximum condition and the adaptive amplitude threshold condition in descending order of amplitude, and retain only the first preset number of candidate peaks.

[0012] Furthermore, the formation of time-continuous non-line-of-sight multipath trajectories specifically includes: S61. Obtain the set of candidate multipath peaks extracted in the t-th frame. Select several candidate peaks with the highest amplitude in the first frame to initialize the active trajectory. The number of active trajectories shall not exceed the preset maximum number of maintained trajectories. S62. For each activity trajectory at the current moment, calculate the reference center using the effective observation position of the most recent frame of the current trajectory. The reference center includes a reference angle and a reference distance. S63. Based on the angle difference and distance difference between the current candidate peak and the reference center of each activity trajectory, calculate the matching cost of all candidate peaks and all activity trajectories, and construct the cost matrix. S64. The Hungarian algorithm is used to solve for the global minimum cost of the cost matrix, and the one-to-one optimal match between the candidate peak and the activity trajectory is obtained. S65. For a successfully matched candidate peak, update the location information of the corresponding activity trajectory. For an activity trajectory that does not match a candidate peak, maintain its state for a preset time length and wait for subsequent matching. S66. When the number of active trajectories is less than the preset maximum number of maintained trajectories, one or more new trajectories are initialized using candidate peaks that have not been matched with active trajectories. S67. Traverse all frames and return the associated time-continuous non-line-of-sight multipath trajectory.

[0013] Furthermore, each trajectory in the non-line-of-sight multipath trajectory contains time-varying distance cell position information, generating multi-channel radio frequency observations, specifically including: S71. For each non-line-of-sight multipath trajectory, based on the distance position of the target trajectory in each frame, sample in the virtual channel distance-time response to obtain the complex value of each virtual channel at the corresponding distance position. S72. Extract the complex-valued phase principal value and perform phase expansion operation along the time dimension to obtain the continuous phase sequence under the combination of each virtual channel and each non-line-of-sight multipath trajectory. S73. Stack all virtual antenna channels and phase sequences under all multipath trajectory combinations according to the virtual channel dimension to form a multi-channel radio frequency observation.

[0014] Furthermore, the BreathDiff-SSM model includes: The conditional diffusion model is used to use multi-channel radio frequency observations as conditional features to guide the conditional diffusion model to gradually recover clean breathing waveforms from noisy breathing waveforms and generate a prior distribution. The selective state-space module is used to model the long-term dynamic changes of respiratory signals and changes in observation reliability based on prior distributions, and to recover respiratory waveforms.

[0015] Furthermore, the conditional diffusion model includes: Temporal conditional pyramids are used to map multi-channel radio frequency observations to different resolution levels of U-Net, providing aligned temporal conditional features for features at each scale. The frequency cue pyramid is used to decompose the conditional features of multi-channel RF observations into frequency cue features through filtering.

[0016] Furthermore, the selective state space module consists of conditional frequency injection, local morphological modeling, global state modeling, and feature fusion to recover the respiratory waveform, specifically including: S81. Obtain the diffusion characteristics of the current layer of the backbone network of the conditional diffusion model, as well as the temporal conditional characteristics and frequency cue characteristics at the same scale as the current layer. S82. Normalize, nonlinear activation and one-dimensional convolutional projection are performed on the features obtained from the backbone network in sequence to obtain the projected features. S83. After injecting the time-domain conditional features and frequency cue features into the projected features, the projected injected features are obtained. S84. Local morphological modeling is performed on the projected injection features through multi-scale deep convolution to extract the peak and valley morphology and local change trend of the short-term breathing waveform, and local features are obtained. S85. Global state modeling of the projected injection features is performed through a selective state-space mechanism. The respiratory dynamics within a preset time range are modeled, and the ratio of the state maintenance amount to the observation injection amount is adaptively adjusted according to the current observation reliability to obtain global features. S86. Fuse local features with global features and output the fused features to obtain the restored breathing waveform.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention provides a non-line-of-sight (LOS) method for recovering respiratory waveforms of mobile human beings based on millimeter-wave radar. Using only a single FMCW millimeter-wave radar, it achieves contactless respiratory waveform recovery in non-line-of-sight scenarios caused by walls, corridor corners, metal cabinets, furniture, or the orientation of the human body, avoiding reliance on wearable devices and visual equipment. Through MVDR enhancement, multi-peak preservation, and Hungarian cross-frame correlation, weak and fragmented non-line-of-sight multipath components are organized into time-continuous observations, improving respiratory observability in mobile scenarios. Phase sampling from the original virtual antenna channel balances multipath trajectory stability and phase accuracy. The BreathDiff-SSM model combines a conditional diffusion model with a selective state-space module, learning clean respiratory waveform priors and adaptively recovering them based on observation reliability. This suppresses interference from macroscopic body movement and path switching, achieving waveform-level recovery rather than just respiratory rate-level recovery. This invention requires no wearable sensors, no cameras, and no multi-radar collaboration. It can recover waveform-level respiratory information from weak, multipath, time-varying, and motion-contaminated non-line-of-sight millimeter-wave observations, offering advantages such as simple deployment, privacy protection, and applicability to real indoor environments. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 This is a schematic flowchart of a non-line-of-sight moving human breathing waveform recovery method based on millimeter-wave radar, provided by an embodiment of the present invention. Figure 2 This is a schematic diagram of the flowchart of a non-line-of-sight moving human respiratory waveform recovery method based on millimeter-wave radar provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the principle of non-line-of-sight human breathing perception provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of the two-dimensional angle-range spectrum of the non-line-of-sight angle-range response after MVDR beamforming provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the three-dimensional amplitude response surface of the non-line-of-sight angle-distance response after MVDR beamforming provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of cross-frame multipath peak correlation provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the associated non-line-of-sight multipath trajectory provided in an embodiment of the present invention; Figure 8 A schematic diagram of the selective state space module structure provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of a corridor corner occlusion scenario provided in an embodiment of the present invention for a non-line-of-sight mobile human breathing perception experiment; Figure 10 This is a schematic diagram of a metal cabinet obstructing a power distribution room in a non-line-of-sight mobile human breathing perception experiment, provided in an embodiment of the present invention. Figure 11 This is a schematic diagram of a daily indoor posture occlusion scenario for a non-line-of-sight mobile human breathing perception experiment provided in an embodiment of the present invention; Figure 12 This is a schematic diagram of the breathing waveform recovery result in a corridor corner occlusion scenario provided in an embodiment of the present invention; Figure 13 This is a schematic diagram of the breathing waveform recovery result in a power distribution room with metal cabinet obstruction, provided in an embodiment of the present invention. Figure 14 This is a schematic diagram of the breathing waveform recovery result in a daily indoor posture occlusion scenario provided by an embodiment of the present invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0021] Reference Figures 1 to 3 This embodiment provides a method for recovering the respiratory waveform of a non-line-of-sight moving human body based on millimeter-wave radar. The method includes: S11. The millimeter-wave radar transmits signals to the target object, obtains the echo after reflection or diffraction, processes the echo, and generates a virtual channel range-time response.

[0022] S12. Perform multipath sensing processing on the virtual channel distance-time response to obtain multi-channel radio frequency observations.

[0023] S13. The multi-channel radio frequency observations were processed using the BreathDiff-SSM model to restore the respiratory waveform.

[0024] In this embodiment, a single commercial FMCW millimeter-wave radar first acquires non-line-of-sight echoes generated by a moving human body and surrounding structures. The original ADC data is reconstructed into multi-dimensional data organized by frame, chirp, virtual antenna channel, and fast-time sampling, forming a virtual channel range-time complex response. Then, weak indirect reflections are enhanced by minimum variance distortionless response beamforming. Multiple candidate peaks are extracted in the angle-range spectrum, and Hungarian matching is used to globally correlate candidate peaks of adjacent frames one-to-one, obtaining a time-continuous non-line-of-sight multipath trajectory.

[0025] Subsequently, complex values ​​are sampled from the original virtual antenna distance-time response based on the associated trajectory. Phase extraction and time unfolding are then performed, and the phase sequences corresponding to multiple virtual antennas and multiple multipath trajectories are stacked to form a multi-channel radio frequency (RF) observation. Finally, the multi-channel RF observation is input into the BreathDiff-SSM respiratory waveform recovery model. Conditional diffusion denoising is used to learn a clean respiratory waveform prior. Through a time-domain conditional pyramid, a frequency cue pyramid, and a selective state-space module, the model adaptively controls the retention amount of historical respiratory states and the injection amount of current observations based on the reliability of non-line-of-sight observations. This suppresses interference from macroscopic human movement, path switching, and multipath mismatch, and outputs a moving human respiratory waveform.

[0026] As a preferred embodiment, forming a virtual channel distance-time response specifically includes: S21. Deploy a single commercial FMCW millimeter-wave radar on one side of the target scene, so that the human body being tested is in a non-line-of-sight area where the direct path of the radar wave is blocked.

[0027] S22. Control the millimeter-wave radar to transmit linear frequency modulated continuous wave signals and receive radio frequency echo signals after reflection or diffraction by human body and environmental structures.

[0028] S23. Mix the radio frequency echo signal with the local oscillator signal of the millimeter-wave radar to obtain the intermediate frequency analog signal, and then perform analog-to-digital conversion sampling to generate the original ADC data.

[0029] S24. Using MIMO virtual array technology, the multi-channel echo components in the original ADC data are separated and rearranged to obtain spatial dimension information. Then, the data is reassembled according to the dimensional order of frame, chirp, virtual antenna channel, and fast time sampling to construct multi-dimensional radar data.

[0030] S25. The fast time dimension of the multidimensional radar data is processed by range FFT, while the slow time dimension is processed by static clutter suppression and the slow time axis is phase-sorted to finally generate the virtual channel range-time complex response.

[0031] In this embodiment, a single commercial FMCW millimeter-wave radar is deployed on one side of an indoor scene. The human subject being measured can be located at a corridor corner, between metal cabinets in a power distribution room, or in a typical indoor obstructed location. When the direct path of the radar to the chest cavity is unavailable, the periodic micro-displacements of the chest cavity caused by human respiration will still modulate the phase of several indirect reflection paths. The system utilizes these weak indirect paths to recover the respiratory waveform. During radar data acquisition, the radar operates in the 77GHz to 81GHz frequency band, employing a 2-transmit, 4-receive MIMO configuration to form 8 virtual antenna channels. The frame rate can be set to 100Hz, and the range resolution can be set to 7.82cm. The original radar ADC data is rearranged to form a four-dimensional structure of frames, chirps, virtual channels, and fast-time sampling. The range-time complex response of each virtual channel is obtained through range-to-FFT.

[0032] Specifically, the linear frequency modulated continuous wave transmitted by the FMCW millimeter-wave radar can be represented as: (1) in, The imaginary unit, For phase conversion constant, For carrier frequency, For time variables, This represents the frequency modulation slope.

[0033] After target path delay, deskewing, and low-pass filtering, the intermediate frequency signal can be approximated as: (2) in, It is an intermediate frequency signal. Echo amplitude, For beat frequency, For transmission distance, For wavelength, The phase conversion coefficients for the round-trip path are used; the beat frequency is used to determine the distance, and the phase term is used to characterize the micro-displacement of thoracic respiration. The radar range resolution satisfies: (3) in, For radar range resolution, At the speed of light, This represents the FMCW bandwidth. Under non-line-of-sight multipath conditions, the received slow-time signal can be represented as the superposition of multiple effective paths: (4) in, For the received signal at the nth slow time sampling point, For slow-time sampling indexes. For moving human bodies, the first... The effective propagation length of a path is composed of a fixed geometric length, macroscopic human movement, and microscopic respiratory displacement. (5) (6) in, Indicates the valid path number. For path complex coefficients, For effective propagation length, Let p be the fixed geometric propagation length of the p-th path. This refers to path changes caused by macroscopic bodily movements. This is the projection of the respiratory micro-displacement onto this path. For noise and unmodeled interference, Let be the phase term corresponding to the nth slow-time sampling point of the p-th path. This model shows that breathing only manifests as weak phase modulation within a strong motion trend. The above formula illustrates that in non-line-of-sight movement scenarios, breathing exhibits weak phase modulation superimposed on a strong macroscopic motion trend; therefore, it is necessary to first enhance and organize available multipaths before performing waveform restoration.

[0034] As a preferred embodiment, multipath sensing processing is performed on the virtual channel distance-time response, specifically including: S31. Perform minimum variance distortionless response beamforming on each antenna domain snapshot in the virtual channel range-time response to obtain the angle-range domain response spectrum.

[0035] Generate the angle-range domain response spectrum, specifically including: S41. Determine the set of scanning angles based on the detection field of view of the millimeter-wave radar and the preset angle step size.

[0036] S42. Estimate the sample covariance matrix by taking a snapshot of the antenna domain at each range cell in the t-th frame of the virtual channel range-time response and introducing a diagonal loading term to obtain the loaded covariance matrix.

[0037] S43. Based on each scanning angle in the set of scanning angles, construct the array steering vector corresponding to the current angle, combine the array steering vector with the covariance matrix after loading, and calculate the beam weight with minimum variance distortion-free response.

[0038] S44. The antenna domain snapshot is weighted and summed using beam weights to obtain the array response in the current scanning angle direction.

[0039] S45. By traversing the set of scanning angles, obtain the array response of the corresponding distance unit under multiple scanning angles, and form the angular response spectrum of the corresponding distance unit.

[0040] S46. Traverse all distance cells in frame t, arrange the angular response spectra of all distance cells according to the distance dimension, and obtain the angular-distance domain response spectrum of frame t.

[0041] S32. After removing the DC component from the angle-distance domain response spectrum along the time dimension, extract multiple candidate peaks from the angle-distance domain response spectrum of each frame to form a candidate multipath peak set.

[0042] Multi-peak extraction is performed on the angle-distance spectrum of each frame, specifically including: S51. For any angle-distance unit, if the response amplitude is greater than the response amplitude of the adjacent preceding and following distance units in the distance dimension, and greater than the response amplitude of the adjacent preceding and following angles in the angle dimension, then the angle-distance unit is determined to be a local maximum.

[0043] S52. At the same time, the amplitude of the candidate peak must also be greater than the product of the median amplitude of all units in the current frame angle-distance spectrum and the first preset coefficient.

[0044] S53. Sort all candidate peaks that simultaneously satisfy the local maximum condition and the adaptive amplitude threshold condition in descending order of amplitude, and retain only the first preset number of candidate peaks.

[0045] S33. Calculate the weighted distance based on the angle difference and distance difference of each candidate peak, and construct the cross-frame association cost matrix.

[0046] S34. The Hungarian matching algorithm is used to solve the global minimum cost of the cross-frame association cost matrix, obtain the one-to-one optimal association between candidate peaks, and form a time-continuous non-line-of-sight multipath trajectory.

[0047] Forming time-continuous non-line-of-sight multipath trajectories, specifically including: S61. Obtain the set of candidate multipath peaks extracted in frame t. Select several candidate peaks with the highest amplitude in the first frame to initialize the active trajectory. The number of active trajectories shall not exceed the preset maximum number of maintained trajectories.

[0048] S62. For each activity trajectory at the current moment, calculate the reference center using the effective observation position of the most recent frame of the current trajectory. The reference center includes a reference angle and a reference distance.

[0049] S63. Based on the angle difference and distance difference between the current candidate peak and the reference center of each activity trajectory, calculate the matching cost between all candidate peaks and all activity trajectories, and construct the cost matrix.

[0050] S64. The Hungarian algorithm is used to solve for the global minimum cost of the cost matrix, and the one-to-one optimal match between the candidate peak and the activity trajectory is obtained.

[0051] S65. For a successfully matched candidate peak, update the location information of the corresponding activity trajectory. For an activity trajectory that does not match a candidate peak, maintain its state for a preset time period and wait for subsequent matching.

[0052] S66. When the number of active trajectories is less than the preset maximum number of maintained trajectories, one or more new trajectories are initialized using candidate peaks that have not been matched with active trajectories.

[0053] S67. Traverse all frames and return the associated time-continuous non-line-of-sight multipath trajectory.

[0054] S35. Based on the distance position corresponding to the non-line-of-sight multipath trajectory, extract and expand the phase sequence from the virtual channel distance-time response.

[0055] S36. Stack all virtual channel distance-time responses and phase sequences of all non-line-of-sight multipath trajectories to generate multi-channel radio frequency observations.

[0056] Each trajectory in the non-line-of-sight multipath trajectory contains time-varying range cell position information, generating multi-channel radio frequency observations, specifically including: S71. For each non-line-of-sight multipath trajectory, based on the distance position of the target trajectory in each frame, sample in the virtual channel distance-time response to obtain the complex value of each virtual channel at the corresponding distance position.

[0057] S72. Extract the complex-valued phase principal values ​​and perform phase expansion operation along the time dimension to obtain the continuous phase sequence under the combination of each virtual channel and each non-line-of-sight multipath trajectory.

[0058] S73. Stack all virtual antenna channels and phase sequences under all multipath trajectory combinations according to the virtual channel dimension to form a multi-channel radio frequency observation.

[0059] In this embodiment, refer to Figure 4 and Figure 5 During the MVDR beamforming process, for the first Frame number Antenna domain snapshot of a distance cell First, define the array steering vector corresponding to the scanning angle θ: (7) in, Indicates angle Array steering vector in the direction, Indicates the spacing between adjacent virtual array elements. Indicates the wavelength of millimeter-wave radar signals. Indicates the number of virtual antenna channels. Indicates the imaginary unit, superscript This indicates a transpose operation. It is performed by changing the scan angle. This allows us to obtain array responses in different directions. Then, based on the... Frame, First Antenna domain snapshot matrix at each distance cell Estimate the sample covariance matrix A diagonal loading term is introduced to improve the numerical stability of the matrix inversion process, as shown in equation (8).

[0060] (8) in, Indicates the first Frame, First Antenna domain snapshot matrix corresponding to each range cell This indicates the number of chirps used for covariance estimation within the corresponding frame. Indicates by The estimated sample covariance matrix; This represents the covariance matrix after diagonal loading, with superscript indicating the superscript. This indicates the conjugate transpose. Represents the matrix trace operation. Represents the identity matrix. This represents the diagonal loading coefficients. The diagonal loading term can alleviate the problem of ill-conditioned covariance matrix under conditions of insufficient sample size or strong multipath correlation, thereby improving the stability of MVDR beamforming.

[0061] This yields the MVDR beam weights and angle-range response: (9) in, This indicates that the distance unit in the t-th frame is at the scanning angle. MVDR beamforming weights This represents the inverse of the covariance matrix after loading. The MVDR beamforming weights make... The signal in the directional direction maintains a distortion-free response, while suppressing interference reflections from other directions as much as possible. The average antenna snapshot of the t-th frame and the r-th range cell is denoted as... The MVDR weighting algorithm is used to weight the data, and the time mean is removed from each angle-distance cell to obtain the angle-distance response. As shown in equation (10).

[0062] (10) The above processing can suppress competing reflections and enhance candidate directional responses, thereby improving the observability of weak non-line-of-sight breathing load paths, such as... Figure 4 and Figure 5 As shown. After beamforming is completed, the system removes the time mean for each angle-range cell and performs multi-peak extraction on the angle-range spectrum of each frame. If an angle-range cell satisfies the following local maximum constraint, it can be used as a candidate peak: (11) in, The distance dimension is the neighborhood offset. The offset is the angular neighborhood. For the i-th scanning angle, At the same time, the amplitude of the candidate peak must also meet an adaptive threshold: (12) in, The adaptive amplitude threshold for frame t. and All parameters are heuristically set thresholds. All candidate peaks meeting the criteria are sorted by amplitude, and only the top-ranked peaks are retained. Preset number of candidate peaks: (13) in, It is the set of candidate peaks in frame t. For the preset quantity, The peak is a candidate peak. This design avoids relying solely on the strongest peak and preserves multiple potential indirect pathways for respiratory load.

[0063] like Figure 6 As shown, the system models cross-frame multipath peak correlation as a linear allocation problem. Let the total number of peaks extracted in frame t be... There are 1 m-th candidate multipath peaks, where the position of the m-th candidate peak is denoted as ... These correspond to its angular coordinates and distance coordinates, respectively. Assuming K activity trajectories were maintained at the previous time step, for the k-th activity trajectory, the system calculates the reference center using the effective observation position of its nearest L frames, denoted as [reference center]. .

[0064] in, This represents the reference angle of the k-th trajectory before frame t. This represents the reference distance. The matching cost between the current candidate peak m and the activity trajectory k is defined as: (14) in, Here is the cost matrix. For angle weighting, The distance weights are then used. The Hungarian matching algorithm is subsequently employed to solve the following globally minimum cost one-to-one matching problem: (15) in, , Indicates the first The candidate peak was assigned to the first... There are 10 activity trajectories, with a value of 0 indicating a mismatch. The constraint requires that each activity trajectory matches at most one candidate peak, and each candidate peak is assigned to at most 10 activity trajectories, thus forming a one-to-one matching relationship. By solving this globally minimum cost allocation problem, we can avoid the identity swapping and misassociation problems caused by point-by-point greedy matching, and improve the temporal continuity of non-line-of-sight multipath trajectories in motion.

[0065] Reference Figure 7 After obtaining the time-continuous non-line-of-sight multipath trajectories, instead of directly using the MVDR output phase as the breathing observation, the phase sequence of each virtual antenna-multipath combination is obtained by sampling the distance-time complex response of the original virtual antenna based on the distance position of each trajectory, taking the phase of the sampled complex values, and expanding them along time. (16) All phase sequences are stacked by channel to form a multi-channel radio frequency observation: (17) in, Indicates the first The original range-time complex response of each virtual antenna channel at the r-th range cell in the t-th frame. For virtual antenna channel index, , For the index of the associated multipath trajectory, ; This indicates the distance cell position of the k-th multipath trajectory in the t-th frame; ( ) indicates taking the complex phase; unwrap( ) indicates the phase expansion operation; Indicates the first The virtual antenna channel and the first The phase sequence is obtained by combining multiple path trajectories.

[0066] all After being stacked according to channel dimension, a multi-channel radio frequency observation matrix is ​​formed. ,in , Indicates the total number of channels. The number of time frames represents the number of rows in the multi-channel RF observation, which is equal to the product of the total number of virtual channels and the total number of multipath trajectories. The number of columns is equal to the total number of time frames. Each row in the multi-channel RF observation corresponds to a phase-time sequence of a virtual channel and a non-line-of-sight multipath trajectory. Through multipath sensing preprocessing and a conditional diffusion state-space recovery model, the human breathing waveform is recovered from the dynamically changing non-line-of-sight indirect reflections. This construction method uses angle-distance trajectories to ensure multipath continuity while preserving the true phase modulation in the original virtual antenna channels as much as possible.

[0067] As a preferred embodiment, the BreathDiff-SSM model includes: The conditional diffusion model is used to guide the conditional diffusion model to gradually recover a clean breathing waveform from a noisy breathing waveform by using multi-channel radio frequency observations as conditional features, thereby generating a prior distribution.

[0068] The conditional diffusion model includes: The temporal conditional pyramid is used to map multi-channel radio frequency observations to different resolution levels of U-Net, providing aligned temporal conditional features for features at each scale.

[0069] The frequency cue pyramid is used to decompose the conditional features of multi-channel RF observations into frequency cue features through filtering.

[0070] The selective state-space module is used to model the long-term dynamic changes of respiratory signals and changes in observation reliability based on prior distributions, and to recover respiratory waveforms.

[0071] The selective state space module consists of conditional frequency injection, local morphological modeling, global state modeling, and feature fusion to recover the respiratory waveform, specifically including: S81. Obtain the diffusion characteristics of the current layer of the backbone network of the conditional diffusion model, as well as the temporal conditional characteristics and frequency cue characteristics at the same scale as the current layer.

[0072] S82. Normalize, nonlinearly activate, and perform one-dimensional convolutional projection on the features obtained from the backbone network in sequence to obtain the projected features.

[0073] S83. After injecting the time-domain conditional features and frequency cue features into the projected features, the projected injected features are obtained.

[0074] S84. Local morphological modeling is performed on the projected injection features through multi-scale deep convolution to extract the peak and valley morphology and local change trend of the short-term breathing waveform, thereby obtaining local features.

[0075] S85. Global state modeling of the projected injection features is performed through a selective state-space mechanism. The modeling is performed on the respiratory dynamics within a preset time range, and the ratio of the state maintenance quantity to the observation injection quantity is adaptively adjusted according to the current observation reliability to obtain global features.

[0076] S86. Fuse local features with global features and output the fused features to obtain the restored breathing waveform.

[0077] In this embodiment, BreathDiff-SSM models the non-line-of-sight movement breathing recovery task as a conditional diffusion waveform recovery problem. For non-line-of-sight radio frequency observations after motion interference, at any time t, the non-line-of-sight radio frequency observation can be simplified to equation (18): (18) in, This represents the radio frequency observation characteristics at time t. This indicates a potential respiratory state that needs to be restored. This represents the observed coefficients that change over time. This represents interference terms caused by human movement, multipath switching, and path mismatch. Residual noise is represented. The BreathDiff-SSM model shows that in non-line-of-sight movement scenarios, the current radio frequency observation is not equivalent to a clean breathing signal, but is formed by the superposition of respiratory micro-movements, overall human motion, multipath propagation changes, and noise. Therefore, respiratory recovery should not be regarded as a direct filtering of a single-point observation, but as a temporal recovery process of the potential respiratory state. Ideally, the estimation of the current respiratory state can be determined by the historical respiratory state and the current radio frequency observation, and the abstract form of the expression is shown in (19): (19) in, This represents the estimated respiratory state obtained at time t. This represents the retention coefficient for historical states. This represents the injection coefficient for the current observation. When the current radio frequency observation is strongly affected by motion interference or multipath instability, the recovery process should rely more on the historical respiratory state; when the current observation contains effective respiratory information, the contribution of the current observation should be enhanced. To achieve the above adaptive state preservation and observation injection mechanism, this invention introduces a selective state space module in BreathDiff-SSM, the basic form of which is shown in equation (20): (20) in, Indicates the first The hidden state at any given moment. Indicates the input-related state transition coefficients. This represents the relevant observation injection coefficients. These represent the mapping coefficients from state to output. This represents the output feature derived from the hidden state. The coefficients in the basic form of the selective state-space module dynamically change with the input features, thus adaptively adjusting the relative contributions of historical states and current observations based on the reliability of the current observation. In other words, the selective state-space module differs from the non-hidden state module. The correspondence between visual distance breathing recovery tasks is shown in Equation (21): (twenty one) Among them, hidden state This corresponds to the potential respiratory state to be recovered. State transition coefficients Corresponding to the historical state preservation coefficient Observation injection coefficient Corresponding to the current observed injection coefficient Therefore, the selective state-space module can preserve stable respiratory dynamics during time-series modeling while adaptively injecting current radio frequency evidence based on changes in the quality of non-line-of-sight observations.

[0078] Based on the above mechanism, BreathDiff-SSM utilizes a conditional diffusion model to learn the priors for generating clean respiratory waveforms, and employs a selective state-space module to model long-term respiratory dynamics and changes in observation reliability. This enables stable respiratory waveform recovery under conditions of motion disturbance, multipath switching, and weak non-line-of-sight observation. The conditional diffusion model internally constructs a temporal conditional pyramid and a frequency cue pyramid. The temporal conditional pyramid maps multi-channel RF observations to different resolution levels of the U-Net, providing aligned multipath temporal evidence for features at each scale. The frequency cue pyramid decomposes conditional features into low-frequency rhythmic components and high-frequency variation components through lightweight filtering, highlighting respiratory cycle-related information and helping to suppress unstable components caused by human motion, multipath switching, and transient disturbances.

[0079] like Figure 8 As shown, the selective state-space module consists of four parts: conditional-frequency injection, local morphological modeling, global state modeling, and feature fusion. First, the module normalizes, nonlinearly activates, and performs one-dimensional convolutional projection on the backbone input features. Then, it injects temporal conditional features and frequency cue features of the same scale into the backbone features, enabling subsequent branches to recover waveforms under radio frequency observation constraints. The local branches use multi-scale deep convolution to extract the peak-valley morphology and local variation trends of short-term respiratory waveforms; the global branches use a selective state-space mechanism to model respiratory dynamics over a longer time range and adaptively adjust the state preservation and observation injection ratio based on the current observation reliability.

[0080] In this embodiment, the global branch generates a content sequence based on the input features. State maintenance control sequence and update control sequence And perform the following input-related status updates: (twenty two) (twenty three) in, ( ) represents the Sigmoid function. This represents element-wise multiplication; This is a state preservation factor used to control historical states. The degree of preservation at the current moment; This is the observation injection term, used to represent the contribution of the current input features to the state update; Let represent the hidden state at time t. To maintain control over the state, To update the control quantity, This refers to the content sequence. Specifically, when the current non-line-of-sight radio frequency observation is significantly affected by motion interference or multipath mismatch, Increased size allows modules to retain more historical respiratory states; when current observations contain relatively reliable respiratory evidence, Enhancement: The module injects more current observation information into the state.

[0081] The selective state-space module performs forward and backward state scans within the recovery window, and then concatenates and fuses the bidirectional scan results using one-dimensional convolution, thereby utilizing both the preceding and following temporal contexts within the window. Finally, the local and global branch outputs are further refined through concatenation, gating fusion, and a feedforward network, and the recovered features are output via residual connections. Through this design, the selective state-space module can take into account both short-term waveform details and long-term respiratory dynamics, improving the stability of respiratory waveform recovery under conditions of weak non-line-of-sight observation, motion contamination, and multipath instability.

[0082] During the training phase, this embodiment uses the reference respiratory waveform acquired by the synchronized breathing chest strap sensor as the monitoring signal. Let the clean reference respiratory waveform be... Multi-channel radio frequency condition observation is ,in Let T represent the actual respiratory sequence of length T, and X represent the conditional input consisting of the aforementioned multipath phase sequence. For the diffusion step... The forward diffusion process involves... Gaussian noise is gradually added to obtain a noisy breathing waveform. As shown in equation (24).

[0083] (twenty four) in, Indicates the first Noisy breathing waveform under one diffusion step; Indicates the diffusion time step; This represents noise sampled from a standard Gaussian distribution; This represents a standard Gaussian distribution with a mean of 0 and a covariance of 1. This represents the cumulative coefficient in the diffusion noise scheduling, used to control the ratio of the original waveform to the noise component in the current diffusion step.

[0084] The definition of the propagation noise prediction loss is shown in Equation (25).

[0085] (25) Model With noisy waveforms diffusion step and radio frequency condition observation As input, predict the noise added during the forward diffusion process. .in, ( ) represents the noise prediction network in BreathDiff-SSM. Indicates the learnable parameters of the network; This represents the predicted loss for diffuse noise; This represents the squared L2 norm, used to constrain the error between the predicted noise and the actual noise. The clean breathing waveform estimate can be derived from the noise predicted by the model. To enhance the consistency between the recovered waveform and the reference waveform in terms of amplitude and overall shape, this embodiment introduces a waveform consistency loss, as shown in equation (26).

[0086] (26) in, This represents the estimated respiratory waveform obtained from model recovery; This represents the reference respiratory waveform captured by the chest strap. This represents the point-by-point absolute error, used to constrain the amplitude deviation of the recovered waveform; The Pearson correlation coefficient represents the relationship between the recovered waveform and the reference waveform, and is used to measure the overall morphological similarity between the two. The definition of the Pearson correlation coefficient is shown in equation (27).

[0087] (27) in, Indicates the recovered waveform at the 1st The values ​​at each time point Indicates the reference waveform at the 1st The values ​​at each time point; and These represent the mean values ​​of the recovered waveform and the reference waveform over the time dimension, respectively; the summation sign... This indicates that the entire time series is accumulated. The closer the value of PCC is to 1, the more consistent the recovered waveform is with the reference waveform in terms of periodic changes, peak and valley positions, and overall trend. The final training objective is shown in equation (28).

[0088] (28) in, Indicates total training loss The weighting coefficients representing waveform consistency loss are used to balance the relationship between diffusion noise prediction and waveform restoration constraints. Through the above loss design, the model not only learns the noise removal capability during the diffusion process, but also constrains the point value error and morphological correlation of the restored waveform, thereby improving the accuracy of restoring the peak and valley positions, periodic rhythm, and local trends of the respiratory waveform.

[0089] In the experimental embodiment, the system used an IWR1443BOOST millimeter-wave radar and a DCA1000EVM to acquire raw ADC data, collecting over 20 hours of data from five subjects in various non-line-of-sight scenarios and under various motion conditions. Experimental results show that the present invention can recover respiratory waveforms highly consistent with reference signals in scenarios such as corridor corners, metal cabinet obstructions in power distribution rooms, and everyday indoor posture occlusion, with an average waveform Pearson correlation coefficient of approximately 0.84, while maintaining low respiratory rate error and respiratory peak timing error.

[0090] like Figures 9 to 14 As shown, in three non-line-of-sight scenarios—corridor corner occlusion, electrical room metal cabinet occlusion, and everyday indoor posture occlusion—the respiratory waveform recovered by this invention maintains good consistency with the chest strap reference waveform in terms of peak and trough positions, periodic rhythm, and overall morphology. Compared to traditional methods that rely solely on a single path or motion cancellation, this invention, through multipath trajectory association and the BreathDiff-SSM recovery model, can suppress macroscopic motion interference and recover waveform-level respiratory information under conditions of human movement and non-line-of-sight occlusion.

[0091] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for recovering the respiratory waveform of a moving human body without line of sight based on millimeter-wave radar, characterized in that, The method includes: S11. The millimeter-wave radar transmits signals to the target object, obtains the echo after reflection or diffraction, processes the echo, and generates a virtual channel range-time response. S12. Perform multipath sensing processing on the virtual channel distance-time response to obtain multi-channel radio frequency observations; S13. The multi-channel radio frequency observations were processed using the BreathDiff-SSM model to restore the respiratory waveform.

2. The method for restoring non-line-of-sight moving human respiratory waveforms based on millimeter-wave radar according to claim 1, characterized in that, The virtual channel distance-time response is formed, specifically including: S21. Deploy a single commercial FMCW millimeter-wave radar on one side of the target scene, so that the human body being tested is in a non-line-of-sight area where the direct path of the radar wave is blocked. S22. Control the millimeter-wave radar to transmit linear frequency modulated continuous wave signals and receive radio frequency echo signals after reflection or diffraction by human body and environmental structures. S23. Mix the radio frequency echo signal with the local oscillator signal of the millimeter-wave radar to obtain the intermediate frequency analog signal, and then perform analog-to-digital conversion sampling to generate the original ADC data. S24. Using MIMO virtual array technology, the multi-channel echo components in the original ADC data are separated and rearranged to obtain spatial dimension information. Then, the data is reassembled according to the dimensional order of frame, chirp, virtual antenna channel and fast time sampling to construct multi-dimensional radar data. S25. The fast time dimension of the multidimensional radar data is processed by range FFT, while the slow time dimension is processed by static clutter suppression and the slow time axis is phase-sorted to finally generate the virtual channel range-time complex response.

3. The method for restoring non-line-of-sight moving human respiratory waveforms based on millimeter-wave radar according to claim 1, characterized in that, Multipath sensing processing is performed on the distance-time response of the virtual channel, specifically including: S31. Perform minimum variance distortionless response beamforming on each antenna domain snapshot in the virtual channel range-time response to obtain the angle-range domain response spectrum; S32. After removing the DC component from the angle-distance domain response spectrum along the time dimension, extract multiple candidate peaks from the angle-distance domain response spectrum of each frame to form a candidate multipath peak set. S33. Calculate the weighted distance based on the angle difference and distance difference of each candidate peak, and construct the cross-frame association cost matrix; S34. The Hungarian matching algorithm is used to solve the global minimum cost of the cross-frame association cost matrix, obtain the one-to-one optimal association between candidate peaks, and form a time-continuous non-line-of-sight multipath trajectory. S35. Based on the distance position corresponding to the non-line-of-sight multipath trajectory, extract and expand the phase sequence from the virtual channel distance-time response; S36. Stack all virtual channel distance-time responses and phase sequences of all non-line-of-sight multipath trajectories to generate multi-channel radio frequency observations.

4. The method for restoring non-line-of-sight moving human respiratory waveforms based on millimeter-wave radar according to claim 3, characterized in that, Generate the angle-range domain response spectrum, specifically including: S41. Determine the set of scanning angles based on the detection field of view of the millimeter-wave radar and the preset angle step size; S42. Estimate the sample covariance matrix by taking a snapshot of the antenna domain at each range cell in the t-th frame of the virtual channel range-time response and introducing a diagonal loading term to obtain the loaded covariance matrix. S43. Based on each scanning angle in the set of scanning angles, construct the array steering vector corresponding to the current angle, combine the array steering vector with the covariance matrix after loading, and calculate the beam weight with minimum variance distortion-free response. S44. The antenna domain snapshots are weighted and summed using beam weights to obtain the array response in the current scanning angle direction; S45. By traversing the set of scanning angles, the array response of the corresponding distance unit under multiple scanning angles is obtained, forming the angle-dimensional response spectrum of the corresponding distance unit; S46. Traverse all distance cells in frame t, arrange the angular response spectra of all distance cells according to the distance dimension, and obtain the angular-distance domain response spectrum of frame t.

5. The method for recovering non-line-of-sight moving human respiratory waveforms based on millimeter-wave radar according to claim 3, characterized in that, Multi-peak extraction is performed on the angle-distance spectrum of each frame, specifically including: S51. For any angle-distance unit, if the response amplitude is greater than the response amplitude of the adjacent preceding and following distance units in the distance dimension, and greater than the response amplitude of the adjacent preceding and following angles in the angle dimension, then the angle-distance unit is determined to be a local maximum. S52. At the same time, the amplitude of the candidate peak must also be greater than the product of the median amplitude of all units in the current frame angle-distance spectrum and the first preset coefficient. S53. Sort all candidate peaks that simultaneously satisfy the local maximum condition and the adaptive amplitude threshold condition in descending order of amplitude, and retain only the first preset number of candidate peaks.

6. The method for recovering non-line-of-sight moving human respiratory waveforms based on millimeter-wave radar according to claim 3, characterized in that, Forming time-continuous non-line-of-sight multipath trajectories, specifically including: S61. Obtain the set of candidate multipath peaks extracted in the t-th frame. Select several candidate peaks with the highest amplitude in the first frame to initialize the active trajectory. The number of active trajectories shall not exceed the preset maximum number of maintained trajectories. S62. For each activity trajectory at the current moment, calculate the reference center using the effective observation position of the most recent frame of the current trajectory. The reference center includes a reference angle and a reference distance. S63. Based on the angle difference and distance difference between the current candidate peak and the reference center of each activity trajectory, calculate the matching cost of all candidate peaks and all activity trajectories, and construct the cost matrix. S64. The Hungarian algorithm is used to solve for the global minimum cost of the cost matrix, and the one-to-one optimal match between the candidate peak and the activity trajectory is obtained. S65. For a successfully matched candidate peak, update the location information of the corresponding activity trajectory. For an activity trajectory that does not match a candidate peak, maintain its state for a preset time length and wait for subsequent matching. S66. When the number of active trajectories is less than the preset maximum number of maintained trajectories, one or more new trajectories are initialized using candidate peaks that have not been matched with active trajectories. S67. Traverse all frames and return the associated time-continuous non-line-of-sight multipath trajectory.

7. The method for restoring non-line-of-sight moving human respiratory waveforms based on millimeter-wave radar according to claim 3, characterized in that, Each trajectory in the non-line-of-sight multipath trajectory contains time-varying range cell position information, generating multi-channel radio frequency observations, specifically including: S71. For each non-line-of-sight multipath trajectory, based on the distance position of the target trajectory in each frame, sample in the virtual channel distance-time response to obtain the complex value of each virtual channel at the corresponding distance position. S72. Extract the complex-valued phase principal value and perform phase expansion operation along the time dimension to obtain the continuous phase sequence under the combination of each virtual channel and each non-line-of-sight multipath trajectory. S73. Stack all virtual antenna channels and phase sequences under all multipath trajectory combinations according to the virtual channel dimension to form a multi-channel radio frequency observation.

8. The method for recovering non-line-of-sight moving human respiratory waveforms based on millimeter-wave radar according to claim 1, characterized in that, The BreathDiff-SSM model includes: The conditional diffusion model is used to use multi-channel radio frequency observations as conditional features to guide the conditional diffusion model to gradually recover clean breathing waveforms from noisy breathing waveforms and generate a prior distribution. The selective state-space module is used to model the long-term dynamic changes of respiratory signals and changes in observation reliability based on prior distributions, and to recover respiratory waveforms.

9. The method for recovering non-line-of-sight moving human respiratory waveforms based on millimeter-wave radar according to claim 1, characterized in that, The conditional diffusion model includes: Temporal conditional pyramids are used to map multi-channel radio frequency observations to different resolution levels of U-Net, providing aligned temporal conditional features for features at each scale. The frequency cue pyramid is used to decompose the conditional features of multi-channel RF observations into frequency cue features through filtering.

10. The method for recovering non-line-of-sight moving human respiratory waveforms based on millimeter-wave radar according to claim 8, characterized in that, The selective state space module consists of conditional frequency injection, local morphological modeling, global state modeling, and feature fusion, which restores... Respiratory waveforms, specifically including: S81. Obtain the diffusion characteristics of the current layer of the backbone network of the conditional diffusion model, as well as the temporal conditional characteristics and frequency cue characteristics at the same scale as the current layer. S82. Normalize, nonlinear activation and one-dimensional convolutional projection are performed on the features obtained from the backbone network in sequence to obtain the projected features. S83. After injecting the time-domain conditional features and frequency cue features into the projected features, the projected injected features are obtained. S84. Local morphological modeling is performed on the projected injection features through multi-scale deep convolution to extract the peak and valley morphology and local change trend of the short-term breathing waveform, and local features are obtained. S85. Global state modeling of the projected injection features is performed through a selective state-space mechanism. The respiratory dynamics within a preset time range are modeled, and the ratio of the state maintenance amount to the observation injection amount is adaptively adjusted according to the current observation reliability to obtain global features. S86. Fuse local features with global features and output the fused features to obtain the restored breathing waveform.

Images