Low bandwidth multipath anchor-gated millimeter wave radar heart rate estimation method and system

Abstract

The application discloses a low-bandwidth multi-path anchor point gating millimeter wave radar heart rate estimation method and system, and belongs to the field of millimeter wave radar vital sign detection. The method is as follows: a target distance unit is determined by a millimeter wave radar, a multi-receiving channel beam forming anchor point path under the same transmission time slot is constructed at the radar end, a single channel phase path and / or a multi-transmission and multi-reception compensation path formed by time sequence phase compensation of a time-sharing transmission virtual array, and a low-bandwidth observation stream not containing complete original ADC data and complete radar cubes is formed with a distinguishable path level field; a time-frequency feature of each path is constructed at the rear end, a main path output is generated from the anchor point path, an auxiliary output and a gating value are generated from the characteristics of the auxiliary path itself, the corresponding auxiliary path contribution is adjusted by using the gating value, and a heart rate estimation value is obtained by fusion. The application can reduce transmission and storage overhead, and improve the stability of non-contact heart rate estimation in a complex scene.

Description

Technical Field

[0001] This invention relates to the field of millimeter-wave radar vital sign detection technology, specifically to a low-bandwidth multipath anchor-gated millimeter-wave radar heart rate estimation method and system. Background Technology

[0002] Heart rate is one of the most important vital signs reflecting the physiological state of the human body. Traditional heart rate detection typically relies on electrocardiogram electrodes, photoplethysmography (PPG) sensors, or other contact sensors. While these methods offer high accuracy in typical scenarios, they require contact with the skin and are easily limited by factors such as wearing comfort, skin condition, activity level, and the convenience of long-term monitoring. Millimeter-wave radar, on the other hand, can detect minute displacements of the chest wall caused by breathing and heartbeats without contact, and is therefore widely used in vital sign detection, health monitoring, smart cockpits, and home care applications.

[0003] In non-contact heart rate estimation using millimeter-wave radar, the chest wall displacement caused by heartbeats is relatively small, and the corresponding phase changes are easily affected by various factors. For example, changes in human posture can alter the chest wall reflection area; range cells may jump between adjacent range gates; indoor multipath reflections can introduce non-target phase components; and time-division multiple-input multiple-output radars may also be affected by phase differences caused by transmission timing. In addition, respiratory components, body motion noise, and radar hardware noise can also make it difficult to stably extract the heartbeat frequency component.

[0004] Existing non-contact heart rate estimation methods for millimeter-wave radar can be broadly categorized as follows: phase demodulation and correction of phase signals from a single channel or selected range cell, combined with bandpass filtering, respiratory and harmonic suppression, variational mode decomposition or other mode decomposition, adaptive filtering, and spectral estimation to extract heartbeat components; improving observation quality through adaptive range cell selection, multi-range cell combination, multi-receiver channel processing, or multi-input multi-output array processing; and utilizing machine learning or deep learning models to classify or regress heart rate based on time-domain, frequency-domain, or time-frequency features. The above methods can improve estimation accuracy under certain conditions, but they still have the following shortcomings: First, in some research-based implementations using offline experimental links for millimeter-wave radar raw data, it is still necessary to use external high-speed raw data acquisition boards like the DCA1000 to transmit multi-channel ADC sampling data or relatively complete radar data to a PC or server for processing, which increases additional acquisition hardware, interface bandwidth, data storage overhead, and system deployment complexity; Second, some schemes mainly focus on processing single paths, single range cells, or unified fusion weights, lacking a multi-physical path observation organization method suitable for real-time implementation at the radar end; Third, when directly splicing or fusing multi-channel, multi-range cell, or multi-path features with fixed weights, auxiliary observations affected by motion, multipath, range cell jumps, or time-division transmission compensation errors may negatively contribute to the final results; Fourth, some data-driven methods focus on back-end network structure design and do not fully utilize the differences in stability and failure modes of different radar physical observation paths.

[0005] Therefore, a new non-contact heart rate estimation scheme for millimeter-wave radar is needed. This scheme would eliminate the need for an external high-speed raw data acquisition board like the DCA1000 to continuously transfer multi-channel raw ADC sampling data or the complete radar cube to the back-end processing equipment. Instead, the radar-side processor would extract multiple phase observation paths with different physical meanings during real-time onboard processing and output them to the back-end processing equipment as a low-bandwidth observation stream. The back-end would then use a stable path as an anchor point to perform gating adjustment on the auxiliary paths before fusing them together. This approach balances system integration, low-bandwidth transmission, radar-side deployability, and estimation stability in complex scenarios. Summary of the Invention

[0006] The purpose of this invention is to provide a low-bandwidth multipath anchor-gated millimeter-wave radar heart rate estimation method to solve problems such as the reliance on external high-speed raw data acquisition links during online operation of some research-based schemes, the insufficient stability of single-phase observation paths in attitude change and multipath environments, and the easy introduction of unreliable auxiliary information by direct splicing of multiple paths or fixed-weight fusion. This method reduces the data transmission and storage burden between the radar end and the back end while improving the stability of heart rate estimation in complex scenarios.

[0007] The technical solution adopted in this invention is as follows: A heart rate estimation method for low-bandwidth multipath anchor-gated millimeter-wave radar, comprising the following steps:

[0008] S1, Target range cell determination: The echo data of the human target area is collected by millimeter-wave radar, and the target range cell where the human target is located is determined based on the echo data;

[0009] S2, Multi-path Phase Observation Construction: Based on the target range cell, multiple phase observation paths are extracted at the radar end. These multiple phase observation paths include a multi-receiver channel beamforming anchor path and at least one auxiliary path. The multi-receiver channel beamforming anchor path is obtained by coherent synthesis of complex observations from multiple receive channels at the target range cell within the same transmission time slot, without crossing different transmission time slots. The auxiliary path includes a single-channel phase path and / or a multi-transmission / multi-receiver compensation path. The multi-transmission / multi-receiver compensation path is based on a virtual array channel formed by time-division multiplexing, and the phase relationship between virtual channels corresponding to different transmission time slots is corrected according to the phase compensation amount related to the time-division multiplexing timing.

[0010] S3, Low-bandwidth observation stream generation and output: The phase information or equivalent complex observation information corresponding to multiple phase observation paths are packaged into a low-bandwidth observation stream with path-level fields that can distinguish each observation path, and output from the radar end to the back-end processing device. The low-bandwidth observation stream does not include the complete original ADC data and the complete radar cube, so as to reduce the amount of data transmitted and the bandwidth requirements.

[0011] S4, Multipath Preprocessing and Time-Frequency Characterization: The back-end processing equipment performs preprocessing and time-frequency characterization on each path observation in the low-bandwidth observation stream to obtain the multipath time-frequency characteristics corresponding to each observation path.

[0012] S5, Main path output, auxiliary path output and gate value generation: The main path output is generated using the time-frequency characteristics corresponding to the anchor point path of the multi-receive channel beamforming as the anchor point characteristics; For each auxiliary path, the auxiliary path output and the corresponding gate value are generated according to the time-frequency characteristics of the auxiliary path itself.

[0013] S6, Gated Fusion and Heart Rate Estimation: Each gate value is used to adaptively adjust the contribution of its corresponding auxiliary path output to the heart rate estimation result, and the gated auxiliary path output is fused with the main path output to obtain the heart rate estimate of the human target.

[0014] Furthermore, in step S1, the target range unit determination specifically includes: performing range processing and target detection on the echo data. The target detection is constant false alarm rate (CFAR) detection. Within a preset range, candidate range units that meet the target reflection conditions are selected based on the signal-to-noise ratio of the detection points, and the candidate range units are locked in conjunction with continuous frame stability. When a candidate range unit jumps relative to the previously locked range unit, it is first used as a jump candidate. The target range unit is updated only after the jump candidate meets the stability conditions in multiple consecutive frames. When the number of consecutive frames in which no valid target is detected reaches a preset loss threshold, the phase processing state is reset and the target range unit is marked as invalid.

[0015] Furthermore, in step S2, the multi-receiver channel beamforming anchor path is composed of multiple receive channels under the same transmission time slot, without crossing different transmission time slots; based on the complex observations of the multiple receive channels at the target range cell, a covariance matrix is ​​constructed and diagonally loaded; Capon power scanning is performed within a preset candidate angle set, the candidate angle with the highest power is selected as the target angle, and the multi-receiver channel observations are coherently synthesized based on the minimum variance distortionless response beamforming weight vector corresponding to the target angle to obtain the complex output of the anchor path and the anchor path phase sequence.

[0016] Furthermore, in step S2, the single-channel phase path is obtained from the complex echoes of a set of transmitting and receiving antennas at the target range cell, and a single-channel phase sequence is generated based on the in-phase and quadrature components of the complex echoes.

[0017] Furthermore, in step S2, the multiple-transmission-multiple-receiver compensation path is composed of virtual array channels formed by time-division multiplexing; the radar estimates the target radial velocity based on the phase change of consecutive frames, determines the time-division multiplexing phase compensation amount based on the target radial velocity and the time interval between different transmission time slots, and uses the time-division multiplexing phase compensation amount to correct the phase relationship between virtual channels corresponding to different transmission time slots before constructing the virtual array spatial statistics; the beamforming weight vector is determined based on the virtual array spatial statistics corrected by the phase relationship between channels, and beamforming synthesis is performed on the virtual array observation using the beamforming weight vector, the phase of the synthesized complex output is calculated, and the phase continuity is maintained through one or more of the following processes: phase unfolding, differential phase accumulation, and range cell jump stitching, to generate a multiple-transmission-multiple-receiver compensation phase sequence.

[0018] Furthermore, in step S3, the low-bandwidth observation stream encapsulates the phase information or complex output information of the multi-receiver channel beamforming anchor path and each auxiliary path into mutually distinguishable path-level fields to maintain the path correspondence between each observation path and the corresponding time-frequency features of the back end. The low-bandwidth observation stream also includes one or more state variables among target range unit, target angle, beamforming power index, radar end processing time, and target loss flag. The radar end includes a millimeter-wave radar front end with an integrated digital signal processor. The radar end completes target range unit determination, multi-path phase observation extraction, and low-bandwidth observation stream output. The back-end processing equipment receives the low-bandwidth observation stream and performs preprocessing, time-frequency characterization, and anchor point gating fusion based on the low-bandwidth observation stream.

[0019] Furthermore, in step S4, the preprocessing and time-frequency characterization include performing time alignment, resampling, phase expansion, detrending, bandpass filtering, sliding window segmentation, and short-time Fourier transform on the multipath observations to obtain multipath time-frequency characteristics corresponding to the heartbeat frequency range; and performing peak preservation processing on the time spectrum obtained by the short-time Fourier transform to preserve the amplitude of the spectral peak and its neighboring frequency positions in each time slice, and attenuating the frequency positions far from the spectral peak.

[0020] Furthermore, in steps S5 and S6, the main path output is an anchor path classification output, a regression output, or a combination of both, and the auxiliary path output is an auxiliary path classification output, a regression output, or a combination of both. For each auxiliary path, the pooling feature formed by the time-frequency features of the auxiliary path itself is used to generate a gate value corresponding to the auxiliary path through a gate mapping and a Sigmoid function. The gate value is applied to the corresponding auxiliary path output to adjust the contribution of the auxiliary path output to the final heart rate estimation result.

[0021] Furthermore, when the main path output and auxiliary path output adopt classification output, the anchor path generates the anchor path classification output, and each auxiliary path generates the auxiliary path classification output and corresponding gating value. First, the corresponding auxiliary path classification output is adjusted using each gating value. Then, the adjusted auxiliary path classification output is fused with the anchor path classification output. Softmax normalization is performed on the fused classification output to obtain the probability of each heart rate category. Based on the category probability, maximum probability decoding or expectation decoding is performed to obtain the heart rate estimate.

[0022] Another objective of this invention is achieved through a millimeter-wave radar non-contact heart rate estimation system. This system is capable of performing the aforementioned low-bandwidth multipath anchor-gated millimeter-wave radar heart rate estimation method. The system includes a millimeter-wave radar front-end module, a radar-end signal processing module, a low-bandwidth output module, a back-end processing module, an anchor-gated fusion module, and a heart rate output module.

[0023] The millimeter-wave radar front-end module is used to collect echo data of the human target area;

[0024] The radar-side signal processing module, connected to the millimeter-wave radar front-end module, is used to lock the target range unit corresponding to the human target based on the echo data, and extract multiple phase observation paths based on the target range unit. The multiple phase observation paths include a multi-receiver channel beamforming anchor point path formed under the same transmission time slot and not spanning transmission time slots, and at least one auxiliary path. The auxiliary path includes a single-channel phase path and / or a multi-transmission multi-receiver compensation path formed through time-division transmission phase compensation.

[0025] The low-bandwidth output module is connected to the radar-end signal processing module and is used to encapsulate the phase information or equivalent complex observation information of each phase observation path into low-bandwidth observation streams with mutually distinguishable path-level fields and transmit them outward. The low-bandwidth observation streams do not contain complete original ADC data and complete radar cubes.

[0026] The back-end processing module is communicatively connected to the low-bandwidth output module, and is used to receive the low-bandwidth observation stream, and perform preprocessing and time-frequency characterization on the observation data of each path to generate corresponding multi-path time-frequency features;

[0027] The anchor point gating fusion module is configured on the back end and is used to generate the main path output based on the anchor point path time-frequency characteristics, generate the corresponding auxiliary path output and gating value based on the time-frequency characteristics of each auxiliary path, adjust the contribution of the corresponding auxiliary path output using the gating value, and fuse the main path and the adjusted auxiliary path output to obtain the heart rate estimate.

[0028] The heart rate output module is connected to the anchor point gating fusion module and is used to output the final target heart rate estimate of the human body.

[0029] The beneficial effects of this invention are:

[0030] Compared with the prior art, the present invention has at least the following beneficial effects.

[0031] First, this invention processes high-volume echo data at the radar end into path-level phase information or equivalent complex observation information that can distinguish the physical source, and outputs it as a low-bandwidth observation stream. While retaining the information required for path feature construction and gating fusion at the back end, it avoids continuously outputting complete original ADC data or complete radar cube, which helps to reduce transmission bandwidth, storage overhead and back-end interface pressure.

[0032] Second, this invention sets the beamforming path formed by multiple receiving channels under the same transmission time slot as the anchor path. This path utilizes the spatial coherence information of multiple receiving channels and does not cross different transmission time slots. Therefore, it is physically different from the single-channel path and the virtual array path that needs to handle the additional phase difference of time-division transmission. This provides a physical basis for setting the asymmetric fusion relationship between the main path and the auxiliary path in the backend. The anchor path does not necessarily mean that it has the lowest estimation error in all scenarios.

[0033] Third, this invention uses single-channel phase paths and / or multiple transmit and multiple receive compensation paths as auxiliary paths, and makes each auxiliary path generate a corresponding gate value based on its own time-frequency characteristics, thereby adjusting the contribution of the auxiliary path output to the final estimation result, rather than using fixed weights for all paths or pre-synthesizing them into a single observation. This is beneficial in limiting the adverse contributions of auxiliary paths when they are affected by factors such as attitude, multipath, time-division transmission compensation errors, or range cell continuity.

[0034] Fourth, this invention preserves the path identity of each path at the radar end and transmits it with distinguishable path-level fields. The back end then constructs time-frequency features and auxiliary gating corresponding to each observation path. Therefore, there is a clear path correspondence and processing dependency between the radar end multi-path construction, low-bandwidth observation stream and back-end anchor point gating, rather than simply listing independent functional modules side by side.

[0035] Fifth, this invention clearly defines the functional boundaries between the radar end and the back-end processing equipment: the radar end completes target range unit determination, multi-physical path observation extraction, necessary time-division transmission phase compensation and low-bandwidth output, and the back-end processing equipment completes path preprocessing, time-frequency characterization and anchor point gating fusion, thereby forming an online processing chain suitable for the collaborative deployment of integrated DSP millimeter-wave radar with PCs, edge computing devices or embedded hosts. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the overall structure of a millimeter-wave radar non-contact heart rate estimation system provided in an embodiment of the present invention, including a radar end and a back-end processing device.

[0037] Figure 2 This is a schematic diagram of the radar-side multipath phase observation and extraction process provided in an embodiment of the present invention, including a single-channel phase path, a multi-receiver channel beamforming anchor point path, and a multi-transmitter multi-receiver compensation path.

[0038] Figure 3 This is a schematic diagram of the low-bandwidth observation stream output and back-end processing flow provided in an embodiment of the present invention.

[0039] Figure 4This is a schematic diagram of the anchor point gating fusion network structure provided in an embodiment of the present invention, wherein the anchor point path generates the main path output, and the auxiliary path generates the auxiliary path output and the gating value.

[0040] Figure 5 The flowchart illustrates the millimeter-wave radar non-contact heart rate estimation method provided in this embodiment of the invention. Detailed Implementation

[0041] The invention will be further described below with reference to embodiments. This invention collaboratively integrates radar-side physical observation construction, path-level low-bandwidth transmission, and back-end anchor-point gating. Instead of pre-combining multi-channel echoes into a single phase or single heart rate result, or continuously outputting complete raw ADC data or a complete radar cube, the radar retains path-level observations with different physical formation mechanisms on the same target range cell: a single-channel phase path is formed by a single transmit-receive channel; multiple receive channels under the same transmit time slot form a multi-receive channel beamforming anchor path, which does not cross different transmit time slots; and a multi-transmit multi-receive compensated path is formed by virtual array channels formed by time-division transmission and phase compensation related to the transmission timing. Each path forms a low-bandwidth observation stream with mutually distinguishable path-level fields, enabling the back-end processing equipment to maintain the physical source and path correspondence of each path, construct corresponding time-frequency characteristics, and use the multi-receive channel beamforming path as the anchor path to perform gating adjustment on each auxiliary path. This integrates radar-side multi-physical path observation, low-bandwidth online transmission, and back-end reliability into a complete technology chain with front-end and back-end dependencies. The following embodiments are used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Where there is no conflict, technical features in different embodiments can be combined with each other. To avoid unnecessarily limiting the scope of protection to a specific platform, a millimeter-wave radar front-end with an integrated digital signal processor is used as an example for illustration. The present invention can also be applied to other millimeter-wave radar platforms with equivalent radar front-end processing capabilities.

[0042] Example 1: Overall System Architecture

[0043] like Figure 1 and Figure 5 As shown, in one embodiment, the millimeter-wave radar front-end is an FMCW millimeter-wave radar, operating in the millimeter-wave frequency band, and has multiple receiving channels and at least one transmitting channel. The radar end includes a radio frequency transceiver unit, an analog-to-digital conversion unit, a range processing unit, a target detection unit, a phase path extraction unit, and a low-bandwidth output unit. The back-end processing device can be a PC, an edge computing device, an embedded host, or other devices with computing capabilities.

[0044] The radar determines target range cells, extracts multipath observations, and outputs low-bandwidth observation streams frame by frame or several frames. The back-end processing equipment constructs a sliding time window based on the continuously received low-bandwidth observation streams to perform multipath preprocessing, time-frequency characterization, anchor-gated fusion, and heart rate output. Instead of continuously outputting complete raw ADC data or a complete radar cube, the radar outputs a low-bandwidth observation stream consisting of a small amount of path-level phase information, complex observation information, and state variables.

[0045] For ease of description, in the following embodiments, the single-channel phase path formed by complex observations of a single transmit-receive channel is referred to as path A; the multi-receive channel beamforming anchor path formed by multiple receive channels in the same transmit time slot is referred to as path B; and the multi-transmit multi-receive compensation path formed by a time-division multiplexed virtual array combined with phase compensation and continuous processing is referred to as path D. A, B, and D are merely designations used to distinguish different observation paths and do not constitute a limitation on the number of paths, the number of channels, or the implementation method.

[0046] The aforementioned paths A, B, and D are not duplicates of the same observation result, but rather preserve observations with different physical formation mechanisms within the same target range cell. Path A retains the direct local phase information of a single transmit-receive channel; path B utilizes multiple receive channels for spatial coherent synthesis within the same transmit time slot, without crossing different transmit time slots; path D utilizes a virtual array formed across transmit time slots and corrects for additional phase relationships introduced by time-division multiplexing. Because the physical sources and potential error sources of the three paths differ, the radar does not pre-compress them into a single path result. Instead, it maintains the path identity to form a low-bandwidth observation stream, which is then used by the backend to construct time-frequency features and perform anchor-gated fusion.

[0047] In one specific embodiment, the millimeter-wave radar employs a time-division multiplexing configuration with a 77 GHz frequency band, two transmit channels, and four receive channels, forming eight virtual array channels. The frame period is approximately 33 ms, and the frame rate is approximately 30.3 Hz. This configuration is merely an example of an implementable method, and the present invention is not limited to the aforementioned operating frequency, number of antennas, frame period, or specific chip model.

[0048] Example 2: Target Distance Cell Determination

[0049] like Figure 2 As shown, in one embodiment, the radar performs a range-oriented Fast Fourier Transform on the echo data to obtain range domain data, and performs constant false alarm rate (CFAR) detection within a preset range to obtain a list of detection points containing range cell indices and signal-to-noise ratio (SNR) information. The radar then selects the range cell corresponding to the detection point with the highest SNR within the preset range as the current candidate target range cell.

[0050] To reduce the impact of short-term range cell jumps on phase continuity, a range cell locking state machine is set up at the radar end. When the difference between the current candidate target range cell and the previous locked range cell does not exceed a preset adjacent threshold, the locked range cell is directly updated or maintained; when the difference exceeds the adjacent threshold, the current candidate target range cell is recorded as a jump candidate, and the locked range cell is only updated after the jump candidate meets the stability condition for several consecutive frames.

[0051] In one specific embodiment, the range of the distance cell search can be set to the 6th to the 40th distance cell; when a candidate distance cell changes by no more than one distance cell relative to the previous locked distance cell, it is updated directly; when a large jump occurs, it is updated after the same jump candidate has been stable for 10 consecutive frames; when no valid target is detected for 30 consecutive frames, the phase unrolling, moving target indication, differential phase accumulation, velocity estimation, and distance cell stitching related states are reset, and the target distance cell is marked as invalid. All of the above values ​​are adjustable implementation parameters.

[0052] When a target range cell is invalid, the radar can also generate a target loss flag and output it along with the low-bandwidth observation stream. This allows the back-end processing equipment to identify whether the target is lost or the range cell is invalid, and to perform window filtering, status marking, or reduce the participation of the corresponding observation in subsequent processing as needed.

[0053] Example 3: Single-channel phase path

[0054] In one embodiment, path A is obtained from the complex echoes of the corresponding channels of a set of transmit and receive antennas at the target range cell. The radar reads the in-phase component of this channel at the target range cell. and orthogonal components Both are then output as part of a low-bandwidth observation stream to the back-end processing device.

[0055] After receiving the in-phase and quadrature components, the back-end processing device calculates and expands the A-path phase:

[0056]

[0057] in, Indicates that path A is at time... The expanded phase. Path A has a simple structure and low computational cost, and can provide direct single-channel phase observations. Its observation quality may vary with human posture, single-channel signal-to-noise ratio, and multipath conditions. Therefore, in this embodiment, it participates in back-end gating fusion as an auxiliary path, rather than as the sole source of heart rate estimation.

[0058] It should be noted that path A is calculated from single-channel I / Q observations and is not confused with other debugging phase fields output by the radar.

[0059] Example 4: Anchor Path for Multi-Receiver Channel Beamforming

[0060] like Figure 2 As shown, in one embodiment, path B consists of multiple receiving channels within the same transmission time slot, without spanning different transmission time slots. In a specific embodiment with four receiving channels, the radar obtains a vector at the target range cell composed of complex observations from the four receiving channels. and utilize Construct the covariance matrix using snapshots:

[0061]

[0062] Here, the superscript H denotes the conjugate transpose. To improve the numerical stability of the matrix solution, the covariance matrix is ​​diagonally loaded:

[0063]

[0064] in, Let I be the diagonal loading coefficient, and I be the identity matrix. In one specific embodiment, the loading amount can be dynamically determined based on the trace of the covariance matrix and preset positive terms, rather than limiting the fixed coefficients to a single solution.

[0065] Perform a Capon spatial spectrum scan within a predefined set of candidate angles. For candidate angles... Its guiding vector is The Capon spatial spectral power is:

[0066]

[0067] Select to make The candidate angle with the larger value is used as the target angle for path B. And calculate the MVDR weight vector:

[0068]

[0069] The weight vector is used to coherently combine multiple receiving channels to obtain the complex output of the B path:

[0070]

[0071] The phase of path B can be obtained from the in-phase and quadrature components of the complex output through arctangent calculation. The radar can output the complex output of path B, the phase of path B, or both. Path B is formed by multiple receiving channels within the same transmission time slot, without crossing different transmission time slots. It achieves physical characteristics distinct from single-channel direct observation of path A and cross-transmission time slot virtual array observation of path D through spatial coherent synthesis and direction selection. Since path B utilizes the spatial information of multiple receiving channels and does not require the cross-transmission time slot phase compensation process corresponding to path D, it is set as the anchor path in this scheme, providing a fixed structural reference for the output of the main path at the back end. "Anchor path" indicates its structural role in this scheme and does not imply that the path necessarily has the lowest error in all scenarios.

[0072] Example 5: Compensation Path for Multiple Sending and Receiving

[0073] like Figure 2 As shown, in one embodiment, the D path consists of a virtual array channel formed by time-division multiplexing. In a specific configuration of 2 transmit channels and 4 receive channels, 8 virtual array channels can be formed. Different transmit antennas transmit at different times, and the displacement of the human chest wall or target reflection point within the transmission interval introduces an additional phase difference between different transmission time slots.

[0074] The radar first estimates the target's radial velocity based on the raw phase changes of a single channel across consecutive frames. Let the... The phase change of the frame after wrapping correction relative to the previous frame is The radar wavelength is The frame period is Then the instantaneous radial velocity can be expressed as:

[0075]

[0076] To reduce instantaneous estimation fluctuations, the velocity estimation results can be recursively smoothed. In one specific embodiment, the weight of historical velocity estimates can be 0.7, and the weight of the current instantaneous velocity estimate can be 0.3; both weights can be adjusted according to actual conditions.

[0077] Let the time interval between different transmission channels be . The compensation amount for the additional phase in time-division transmission can be expressed as:

[0078]

[0079] When constructing the 8-channel spatial covariance matrix, complex rotation compensation is applied to the virtual channel observations belonging to the later transmission time slot:

[0080]

[0081] in, This represents the virtual channel index that requires time-division multiplexing compensation. The compensated channel observations are grouped into a vector, and a D-path space covariance matrix is ​​constructed using multiple snapshots. Subsequently, the covariance matrix is ​​diagonally loaded, subjected to Capon spatial spectrum scanning, and MVDR weights are solved; its basic form can be found in Example 4.

[0082] In this embodiment, time-division multiplexing phase compensation is used to correct the inter-channel phase relationship in the spatial covariance matrix of the virtual array. Based on the corrected spatial statistics, Capon spatial spectrum scanning and MVDR weighting are performed. Then, the obtained weights are used to coherently synthesize the current virtual array observations to obtain a result containing in-phase components. and orthogonal components The complex output of the D path.

[0083] In another embodiment, the time-division transmission phase compensation can be directly applied to the complex observations of the virtual channels in the corresponding transmission time slots, and the spatial covariance matrix can be constructed using the compensated virtual array observations for beamforming synthesis. Both of these methods reduce the impact of the additional phase difference from time-division transmission on the D-path by correcting the phase relationship between the virtual channels corresponding to different transmission time slots.

[0084] After phase expansion and moving target indication processing of the synthesized complex output of the D path, the differential arctangent phase accumulation method can be used to maintain phase continuity. When there are no range cell jumps between adjacent time steps, the differential phase can be expressed as:

[0085]

[0086] in, and These represent the changes in the in-phase and quadrature components at adjacent time points, respectively. To prevent positive numbers with a denominator of zero, the differential phases at each time step by step are accumulated to obtain the continuous phase of the D path.

[0087] When the target range cell changes between adjacent range cells, the complex output of the current range cell and the adjacent range cells can be calculated using the same beamforming weight vector to estimate the phase difference before and after the jump, and update the range cell splicing offset, thereby reducing the phase discontinuity caused by range cell switching. The radar output finally performs the D-path phase after the above continuous processing.

[0088] The D-path provides virtual array observation information across transmission time slots, and its spatial observation range differs from the B-path's multi-receiver channel observation within the same transmission time slot. To form a D-path usable for backend fusion, the radar first determines the time-division transmission phase compensation based on the target radial velocity and the time interval between different transmission time slots. Before constructing the virtual array spatial statistics, it corrects the phase relationship between corresponding virtual channels in different transmission time slots, and then performs spatial statistics, beamforming, and phase continuity processing. Since the D-path may still be affected by velocity estimation errors, TDM compensation errors, and range cell continuity, this embodiment does not perform equal-weighted fixed fusion with the B-path. Instead, it is retained as an auxiliary path, and its output contribution is adjusted by the backend based on the gating value generated by the D-path's own characteristics.

[0089] Example 6: Low-bandwidth observation stream

[0090] like Figure 3 As shown, in one specific embodiment, the radar end encapsulates the low-bandwidth observation stream into a TLV data packet, outputs it via a UART interface, and transmits it to the back-end processing device through an onboard UART-to-USB bridge circuit. In other embodiments, when using a radar platform or processor with equivalent radar end processing capabilities, the low-bandwidth observation stream can also be transmitted via USB, Ethernet, wireless communication interface, on-chip bus, board-level bus, or other equivalent communication methods.

[0091] In a specific data frame embodiment used for debugging and algorithm verification, each frame outputs 12 32-bit floating-point numbers, with the following fields in order: A path single-channel I component, A path single-channel Q component, debugging phase obtained by conventional phase expansion of the eight-channel beamforming complex output, D path continuous phase, B path I component, B path Q component, B path phase, target range cell, target angle obtained by 8-channel spatial spectrum scanning, 8-channel Capon spatial spectrum power index, radar-side correlation processing time, and target loss flag.

[0092] In this embodiment, the third and fourth fields both originate from the complex output of the eight-channel beamforming, but employ different phase continuity processing. The third field is the phase obtained by conventional phase expansion of the eight-channel complex output, primarily used for debugging or comparison. The fourth field is the D-path continuous phase obtained by differential phase accumulation and range cell stitching based on the eight-channel complex output, and serves as the input for D-path backend processing. Neither of these fields indicates the addition of an independent physical observation path. The A-path phase is calculated at the backend based on the single-channel I / Q components in the first and second fields. The target angle and Capon spatial spectrum power index are obtained from eight-channel spatial spectrum scanning. The B-path phase can be calculated from the B-path I and Q components; this specific data frame retains both the B-path I / Q components and the B-path phase, primarily for debugging, algorithm verification, and result cross-validation, and does not imply that the aforementioned mutually derivable fields must be output simultaneously during actual deployment.

[0093] The pure payload of 12 32-bit floating-point numbers is 48 bytes per frame. At a frame rate of approximately 30.3 Hz, the pure payload rate is approximately 1454.4 B / s, or about 1.45 kB / s, excluding packet headers, checksums, and other communication encapsulation overhead. The aforementioned 12-field data frame retains some mutually derivable observations for ease of debugging and algorithm verification. In another embodiment, only necessary fields from the phase or complex observations of each path can be retained as needed by the backend processing; for example, when only the phase scalars of paths A, B, and D are output, the pure payload rate is approximately 363.6 B / s, or about 0.36 kB / s. The above values ​​are used to illustrate the reduction in data volume relative to the complete raw ADC data or the complete radar cube; "low bandwidth" is not limited to a fixed absolute rate threshold.

[0094] The number of fields, field order, data type, and communication protocol of the low-bandwidth observation stream can all be adjusted according to the needs of backend processing. For example, it can output only the phase of each path, output the complex observations of each path, or output the path observations and necessary state variables simultaneously. Regardless of the specific encapsulation method used, the observation information of the anchor path and each auxiliary path is retained separately as path-level fields that can be distinguished from each other. This allows the backend to determine the physical observation path to which each field belongs and maintain the correspondence between that path and subsequent preprocessing, time-frequency features, auxiliary path outputs, and gate values. Therefore, the low-bandwidth processing in this embodiment does not pre-synthesize multiple paths into a single path result, but rather retains the path identity information required for the backend to perform sub-path reliability adjustment while significantly reducing the amount of data.

[0095] Example 7: Back-end preprocessing and time-frequency characterization

[0096] like Figure 3As shown, in one embodiment, after receiving the low-bandwidth observation stream, the backend processing device first performs time alignment and resampling on the multipath observations. For path A, the backend... and Calculate and expand the phase; for the B path, the back-end uses the B path phase output from the radar, or recalculates the phase based on the B path I / Q; for the D path, the back-end uses the D path phase output from the radar after continuous processing.

[0097] In one specific embodiment, the original frame rate at the radar end is approximately 30.3Hz, and the backend resamples each path to 20Hz. Detrending and fourth-order bandpass filtering are performed on each path, with the bandpass range set to 0.75Hz to 2.5Hz. Then, a 30-second sliding window is used to segment each path sequence, and the window is updated according to a preset step size; during online inference, a new inference window can be formed every 1 second.

[0098] For path Preprocessed signal Its short-time Fourier transform can be expressed as:

[0099]

[0100] in, Indicated by The center analysis window, Frequency is represented. In one specific embodiment, a 5-second Hamming analysis window, a 1-second step size, and a 2400-point Fast Fourier Transform are used. A 30-second window forms 26 time slices and retains 241 frequency positions. The time-frequency feature size corresponding to the three paths is 3×26×241.

[0101] To highlight the candidate spectral peak structure within each time slice, a peak-preserving soft mask can be applied to the time spectrum. Within each time slice, the spectral peak position is determined, the amplitude of the peak and its neighboring frequency positions is preserved, and the amplitude of positions far from the peak is reduced according to a preset attenuation coefficient. In one specific embodiment, three frequency positions to the left and right of the peak can be preserved, and the remaining positions are multiplied by 0.2. All of the above parameters can be adjusted according to the frequency resolution and actual deployment conditions.

[0102] The backend can also use the global mean and standard deviation calculated from the training data to apply consistent Z-score normalization to both training and inference inputs. Reference heart rate and reference photoelectric waveforms from the training data can be used for label generation and training window quality filtering, but only the low-bandwidth observation stream from the millimeter-wave radar is required during the online inference phase; no reference photoelectric sensor input is needed.

[0103] Example 8: Anchor Point Gated Fusion

[0104] like Figure 4As shown, in one embodiment, the back-end processing device inputs the multi-path time-frequency features corresponding one-to-one with each observation path into the anchor-gated fusion model. The model includes an anchor branch and at least one auxiliary branch. The anchor branch receives the time-frequency features of path B and generates the main path output; each auxiliary branch receives the time-frequency features of its corresponding path A and / or path D, and generates an auxiliary path output and a gating value corresponding to that auxiliary path. The path correspondence maintained at the back-end by the path-level low-bandwidth observation streams allows each auxiliary gating to act on an auxiliary output consistent with its physical source.

[0105] In one specific embodiment, the B path branch employs a two-level 2D convolution, normalization, non-linear activation, temporal attention pooling, and classification output structure, and can be configured with an auxiliary regression output; the A and D path branches employ a lighter 2D convolution and pooling structure. For each auxiliary path, the gating value is generated by gating the pooling features of that auxiliary path itself, rather than being given by fixed weights, and is not required to be obtained through explicit similarity calculation between the anchor path and the auxiliary path.

[0106] For the There are 1 auxiliary path, and its pooling characteristic is 1. Gating mapping is Then the corresponding gate value It can be represented as:

[0107]

[0108] in, Since it is a sigmoid function, therefore:

[0109]

[0110] Each auxiliary path's gate value is generated independently. Normalization of these gate values ​​is not required, nor is it required that their sum be 1. The gate values ​​are used to adjust the contribution of the corresponding auxiliary path's output; they are not required to be interpreted as strictly probabilistic or independent physical reliability indicators.

[0111] In a specific embodiment employing three paths (A, B, and D) for categorical heart rate estimation, the classification output of path B serves as the anchor classification output, while paths A and D generate their respective auxiliary classification outputs and gating values. Let the classification output of path B be... The outputs for path A and path D are respectively and The corresponding gate values ​​are respectively and The fusion classification output can then be represented as:

[0112]

[0113] First, the classification outputs of paths A and D are adjusted separately using their respective gating values. Then, the adjusted auxiliary classification outputs are fused with the anchor point classification output of path B. This fusion method maintains the main output of path B as the structural anchor point, and the anchor point classification output of path B is directly used as the fusion reference term without participating in gating scaling. Paths A and D participate in the final classification output with independent gating methods. This scheme first completes the gating adjustment of the auxiliary path output and its fusion with the anchor point path output, and then performs Softmax normalization on the fused classification output to obtain the probability of each heart rate category.

[0114] Perform Softmax normalization on the fused classification output. For the... Each heart rate category, its probability It can be represented as:

[0115]

[0116] in, This represents the total number of heart rate categories. In one specific embodiment, 40 BPM to 160 BPM is discretized into 241 categories at 0.5 BPM intervals, i.e. = 241.

[0117] Heart rate can be estimated using either maximum probability class decoding or, preferably, expectation decoding. When using maximum probability class decoding, the estimated result can be obtained based on the heart rate value corresponding to the class with the highest probability. When using expectation decoding, let the first... The heart rate value corresponding to each category is Then the heart rate estimate can be expressed as:

[0118]

[0119] in, For the first The probability of each category, For the first The heart rate values ​​corresponding to each category.

[0120] During model training, reference heart rates corresponding to the radar window time are used to generate supervisory labels. As a concrete example, Gaussian soft labels can be generated based on the distance between the reference heart rate and each discrete heart rate category, and training is performed using a joint objective function consisting of fusion classification loss, B-path assisted classification loss, and B-path assisted regression loss. The weights of each loss term can be adjusted based on training performance and are not limited to a single value. In a specific training implementation, cross-entropy loss and Huber regression loss can be used, and random channel deactivation can be applied to either the A-path or D-path to enable end-to-end joint updates of the feature extraction branch, auxiliary output branch, and gating network.

[0121] After training, the model parameters and normalized statistics are saved. During online inference, only the millimeter-wave radar multipath time-frequency features that have undergone the same preprocessing and normalization are input into the model; a reference heart rate sensor is not required. Through joint training, the gating network can learn to adjust its output contribution based on the auxiliary path features; this statement indicates the model's design function and does not imply that the gating value will necessarily change in a preset direction under arbitrary disturbance conditions.

[0122] Example 9: Systems, Electronic Device Systems, and Storage Media

[0123] like Figure 1 As shown, in one embodiment, the millimeter-wave radar non-contact heart rate estimation system includes a millimeter-wave radar front-end module, a radar-end signal processing module, a low-bandwidth output module, a back-end processing module, and a heart rate output module. The millimeter-wave radar front-end module transmits millimeter-wave signals and receives reflected echoes from human targets; the radar-end signal processing module determines target range cells, extracts multipath phase observations, and generates path state variables; the low-bandwidth output module forms a low-bandwidth observation stream from path phase information, complex observation information, and state variables; the back-end processing module preprocesses, performs time-frequency characterization, and anchor-point gating fusion on the low-bandwidth observation stream; and the heart rate output module outputs heart rate estimates and can send these estimates to a display interface, storage unit, or host computer application.

[0124] In another embodiment, the electronic device system includes a radar end processor, a back-end processor, and memory. The radar end processor can be an on-chip digital signal processor, microcontroller, or equivalent processor for millimeter-wave radar; the back-end processor can be one or more of a CPU, GPU, NPU, embedded processor, or edge computing processor. The radar end processor and the back-end processor can reside in two communicating physical devices or be integrated into different processing units within the same device, and can transmit low-bandwidth observation streams via UART, USB, Ethernet, wireless network, on-chip bus, or board-level bus.

[0125] When the computer program stored in the memory is executed by the radar-side processor, it achieves target range cell determination, multipath observation extraction, and low-bandwidth output; when executed by the back-end processor, it achieves preprocessing, time-frequency characterization, anchor point gating fusion, and heart rate output.

[0126] In yet another embodiment, a computer-readable storage medium stores a computer program. When executed by a corresponding processor, the computer program implements one or more steps in the above-described millimeter-wave radar non-contact heart rate estimation method.

[0127] It should be noted that the specific radar models, operating frequencies, number of transmit and receive channels, frame periods, resampling frequencies, window lengths, short-time Fourier transform parameters, number of network layers, number of categories, number of low-bandwidth observation stream fields, and loss weights in the above embodiments are all illustrative and implementable examples, and are not intended to limit the scope of protection of this invention. Those skilled in the art can make equivalent substitutions or modifications based on specific hardware platforms, monitoring distances, sampling rates, and deployment conditions without departing from the technical concept of this invention. The scope of protection of this invention is defined by the claims.

Claims

1. A heart rate estimation method for low-bandwidth multipath anchor-gated millimeter-wave radar, characterized in that, Includes the following steps: S1, Target range cell determination: Echo data of the human target area is collected by millimeter-wave radar, and the target range cell where the human target is located is determined based on the echo data; S2, Multi-path Phase Observation Construction: Based on the target range cell, multiple phase observation paths are extracted at the radar end. These multiple phase observation paths include a multi-receiver channel beamforming anchor path and at least one auxiliary path. The multi-receiver channel beamforming anchor path is obtained by coherent synthesis of complex observations from multiple receive channels at the target range cell within the same transmission time slot, without crossing different transmission time slots. The auxiliary path includes a single-channel phase path and / or a multi-transmission / multi-receiver compensation path. The multi-transmission / multi-receiver compensation path is based on a virtual array channel formed by time-division multiplexing, and the phase relationship between virtual channels corresponding to different transmission time slots is corrected according to the phase compensation amount related to the time-division multiplexing timing. S3, Low-bandwidth observation stream generation and output: The phase information or equivalent complex observation information corresponding to multiple phase observation paths are packaged into a low-bandwidth observation stream with path-level fields that can distinguish each observation path, and output from the radar end to the back-end processing device. The low-bandwidth observation stream does not include the complete original ADC data and the complete radar cube. S4, Multipath Preprocessing and Time-Frequency Characterization: The back-end processing equipment performs preprocessing and time-frequency characterization on each path observation in the low-bandwidth observation stream to obtain the multipath time-frequency characteristics corresponding to each observation path. S5, Main path output, auxiliary path output and gate value generation: The main path output is generated using the time-frequency characteristics corresponding to the anchor point path of the multi-receive channel beamforming as the anchor point characteristics; For each auxiliary path, the auxiliary path output and the corresponding gate value are generated according to the time-frequency characteristics of the auxiliary path itself. S6, Gated Fusion and Heart Rate Estimation: Using the main path output as the anchor point, each gate value is used to adaptively adjust the contribution of its corresponding auxiliary path output to the heart rate estimation result, and the gated auxiliary path output is fused with the main path output to obtain the heart rate estimate of the human target.

2. The heart rate estimation method for low-bandwidth multipath anchor-gated millimeter-wave radar according to claim 1, characterized in that, In step S1, the target range unit determination specifically includes: performing range processing and target detection on the echo data. The target detection is constant false alarm rate (CFAR) detection. Within a preset range, candidate range units that meet the target reflection conditions are selected based on the signal-to-noise ratio of the detection points, and the candidate range units are locked in conjunction with the stability of consecutive frames. When a candidate range unit jumps relative to the previously locked range unit, it is first used as a jump candidate. The target range unit is updated after the jump candidate meets the stability conditions in multiple consecutive frames. When the number of consecutive frames in which no valid target is detected reaches a preset loss threshold, the phase processing state is reset and the target range unit is marked as invalid.

3. The heart rate estimation method for low-bandwidth multipath anchor-gated millimeter-wave radar according to claim 1, characterized in that, In step S2, the multi-receiver channel beamforming anchor path is composed of multiple receive channels under the same transmission time slot, without crossing different transmission time slots; based on the complex observations of the multiple receive channels at the target range cell, a covariance matrix is ​​constructed and diagonally loaded; Capon power scanning is performed within a preset candidate angle set, the candidate angle with the highest power is selected as the target angle, and the multi-receiver channel observations are coherently synthesized based on the minimum variance distortionless response beamforming weight vector corresponding to the target angle to obtain the complex output of the anchor path and the anchor path phase sequence.

4. The heart rate estimation method for low-bandwidth multipath anchor-gated millimeter-wave radar according to claim 1, characterized in that, In step S2, the single-channel phase path is obtained from the complex echoes of a set of transmitting and receiving antennas at the target range cell, and a single-channel phase sequence is generated based on the in-phase and quadrature components of the complex echoes.

5. The heart rate estimation method for low-bandwidth multipath anchor-gated millimeter-wave radar according to claim 1, characterized in that, In step S2, the multiple-transmission-multiple-receiver compensation path is composed of virtual array channels formed by time-division multiplexing. The radar estimates the target radial velocity based on the phase change of consecutive frames, and determines the time-division multiplexing phase compensation amount based on the target radial velocity and the time interval between different transmission time slots. Before constructing the virtual array spatial statistics, the phase relationship between virtual channels corresponding to different transmission time slots is corrected using the time-division multiplexing phase compensation amount. The beamforming weight vector is determined based on the virtual array spatial statistics corrected by the phase relationship between channels. The beamforming weight vector is used to perform beamforming synthesis on the virtual array observations, calculate the phase of the synthesized complex output, and maintain phase continuity through one or more of the following processes: phase unfolding, differential phase accumulation, and range cell jump stitching, to generate a multiple-transmission-multiple-receiver compensation phase sequence.

6. The heart rate estimation method for low-bandwidth multipath anchor-gated millimeter-wave radar according to claim 1, characterized in that, In step S3, the low-bandwidth observation stream encapsulates the phase information or complex output information of the multi-receiver channel beamforming anchor path and each auxiliary path into mutually distinguishable path-level fields to maintain the path correspondence between each observation path and the corresponding time-frequency features of the back end. The low-bandwidth observation stream also includes one or more state variables among target range unit, target angle, beamforming power index, radar end processing time, and target loss flag. The radar end includes a millimeter-wave radar front end with an integrated digital signal processor. The radar end completes target range unit determination, multi-path phase observation extraction, and low-bandwidth observation stream output. The back-end processing equipment receives the low-bandwidth observation stream and performs preprocessing, time-frequency characterization, and anchor point gating fusion based on the low-bandwidth observation stream.

7. The heart rate estimation method for low-bandwidth multipath anchor-gated millimeter-wave radar according to claim 1, characterized in that, In step S4, the preprocessing and time-frequency characterization include time alignment, resampling, phase expansion, detrending, bandpass filtering, sliding window segmentation, and short-time Fourier transform of the multipath observations to obtain multipath time-frequency features corresponding to the heartbeat frequency range. The time spectrum obtained by the short-time Fourier transform is subjected to peak preservation processing, which preserves the amplitude of the spectral peak and its neighboring frequency positions in each time slice, and attenuates the frequency positions far away from the spectral peak.

8. The heart rate estimation method for low-bandwidth multipath anchor-gated millimeter-wave radar according to claim 1, characterized in that, In steps S5 and S6, the main path output is the anchor path classification output, regression output, or a combination of both, and the auxiliary path output is the auxiliary path classification output, regression output, or a combination of both. For each auxiliary path, the pooling feature formed by the time-frequency features of the auxiliary path itself is gated and mapped using a sigmoid function to generate a gate value corresponding to the auxiliary path. The gate value is applied to the corresponding auxiliary path output to adjust the contribution of the auxiliary path output to the final heart rate estimation result.

9. The heart rate estimation method for low-bandwidth multipath anchor-gated millimeter-wave radar according to claim 8, characterized in that, When the main path output and auxiliary path output adopt classification output, the anchor path generates the anchor path classification output, and each auxiliary path generates the auxiliary path classification output and corresponding gating value. First, the corresponding auxiliary path classification output is adjusted using each gating value. Then, the adjusted auxiliary path classification output is fused with the anchor path classification output. Softmax normalization is performed on the fused classification output to obtain the probability of each heart rate category. Based on the category probability, maximum probability decoding or expectation decoding is performed to obtain the heart rate estimate.

10. A millimeter-wave radar non-contact heart rate estimation system, characterized in that, The system is used to perform the low-bandwidth multipath anchor-gated millimeter-wave radar heart rate estimation method as described in any one of claims 1-9. The system includes a millimeter-wave radar front-end module, a radar-end signal processing module, a low-bandwidth output module, a back-end processing module, an anchor-gated fusion module, and a heart rate output module. The millimeter-wave radar front-end module is used to collect echo data of the human target area; The radar-side signal processing module, connected to the millimeter-wave radar front-end module, is used to lock the target range unit corresponding to the human target based on the echo data, and extract multiple phase observation paths based on the target range unit. The multiple phase observation paths include a multi-receiver channel beamforming anchor point path formed under the same transmission time slot and not spanning transmission time slots, and at least one auxiliary path. The auxiliary path includes a single-channel phase path and / or a multi-transmission multi-receiver compensation path formed through time-division transmission phase compensation. The low-bandwidth output module is connected to the radar-end signal processing module and is used to encapsulate the phase information or equivalent complex observation information of each phase observation path into low-bandwidth observation streams with mutually distinguishable path-level fields and transmit them outward. The low-bandwidth observation streams do not contain complete original ADC data and complete radar cubes. The back-end processing module is communicatively connected to the low-bandwidth output module, and is used to receive the low-bandwidth observation stream, and perform preprocessing and time-frequency characterization on the observation data of each path to generate corresponding multi-path time-frequency features; The anchor point gating fusion module is configured on the back end and is used to generate the main path output based on the anchor point path time-frequency characteristics, generate the corresponding auxiliary path output and gating value based on the time-frequency characteristics of each auxiliary path, adjust the contribution of the corresponding auxiliary path output using the gating value, and fuse the main path and the adjusted auxiliary path output to obtain the heart rate estimate. The heart rate output module is connected to the anchor point gating fusion module and is used to output the final target heart rate estimate of the human body.

Images