Micro-motion sensing system and method integrated on single chip, electronic equipment, storage medium and program product

Abstract

The invention provides a micro-motion sensing system and method integrated on a single chip, electronic equipment, a storage medium and a program product, and relates to the technical field of data processing. According to the scheme, the millimeter wave noise radar front end, the digital signal processor and the hardware AI accelerator are integrated on one chip, so that the hardware volume, the power consumption, the delay and the cost can be reduced, and the requirements of edge equipment are met. According to the system, a broadband random modulation signal with the instantaneous bandwidth not smaller than 1GHz is transmitted by the front end of the millimeter wave noise radar, the sidelobe level of a range profile generated by the system can be remarkably lower than-40dB through simulation / experiment verification, and the industrial problem that a weak target is shielded by a strong target due to high sidelobe of the FMCW radar is fundamentally solved. On the basis, the system calibrates and preprocesses the range profile through the on-chip digital processor, so that the compatibility of the AI accelerator and the radar can be ensured, and necessary pretreatment support is provided for the on-chip hardware AI accelerator. And a plurality of range gates are divided and processed in parallel through an on-chip hardware AI accelerator, so that self-adaptive separation and feature extraction of micro-motion signals in each range gate are realized, and high-precision and high-reliability parallel sensing of a plurality of static or micro-motion targets in a complex scene is realized. The method is particularly suitable for in-vehicle multi-passenger vital sign monitoring, industrial equipment predictive maintenance, human body posture tracking and recognition, gesture recognition and medical monitoring.

Description

Cross-references to related applications

[0001] This application claims priority to Chinese Patent Application No. 2025116617141, filed on November 13, 2025, entitled “Micro-motion sensing system, method, electronic device, storage medium and program product integrated on a single chip”, the entire contents of which are incorporated herein by reference. Technical Field

[0002] This application relates to the field of data processing technology, and more specifically, to a micro-motion sensing system, method, electronic device, storage medium, and program product integrated on a single chip. Background Technology

[0003] With the development of radio sensing technology, the demand for micro-motion target sensing is increasing in fields such as non-contact vital sign monitoring, industrial equipment status monitoring, intelligent security, and gesture interaction recognition. Micro-motion targets typically refer to targets with periodic, minute displacements on their surfaces, such as the millimeter-level or even sub-millimeter-level fluctuations in the human chest cavity caused by breathing and heartbeat, or the micrometer-level vibrations generated by the operation of industrial equipment. In such applications, the system needs to be able to detect and distinguish multiple stationary or micro-motion targets in complex environments and extract their micro-motion characteristics (such as frequency and amplitude) with high precision.

[0004] Currently, in both civilian and industrial fields, frequency-modulated continuous-wave (FMCW) millimeter-wave radar is the mainstream technology for achieving non-contact micro-motion sensing. Its basic principle is to transmit a linearly varying frequency continuous-wave signal, mix it with the received echo signal, and then analyze the phase changes of the echo signal to extract the micro-motion information of the target.

[0005] However, FMCW radar has many inherent drawbacks in practical applications, severely limiting its performance and application scenarios. Firstly, during range dimension processing, FMCW radar generates high range sidelobes due to the use of Fast Fourier Transform (FFT) for signal processing. This means that when strong reflective targets are present, their sidelobes may mask weak signals from nearby targets. For example, in a vehicle interior environment, the sidelobes of highly reflective objects such as seats and walls may obscure weak signals like heartbeats and breathing, causing the system to be unable to effectively distinguish between multiple spatially close stationary targets, significantly impacting its ability to detect weak signals.

[0006] Secondly, the multi-target resolution capability of FMCW radar is also limited. In complex scenarios with multiple micro-motion sources, such as multiple occupants in a vehicle or multiple pieces of equipment operating simultaneously in a factory, the micro-Doppler signals of the target are prone to aliasing in the range-velocity dimension. Traditional signal processing methods are less than satisfactory in separating these aliased signals, making it difficult to accurately distinguish and identify individual micro-motion sources, and thus failing to meet the needs of precise monitoring and analysis of multiple targets.

[0007] Furthermore, FMCW radar has poor resistance to co-channel interference. When multiple FMCW radars operating at the same frequency band work simultaneously, their linear frequency modulated signals interfere with each other, causing a sharp decline in radar performance and severely affecting its applicability in multi-sensor coexistence environments. For example, in the workshop communication scenario of multi-radar vehicles, this interference problem will greatly limit the reliability and stability of the radar system.

[0008] In summary, the current FMCW radar solution cannot effectively distinguish between multiple stationary targets that are close in space, and cannot meet the needs for accurate monitoring and analysis of multiple targets. Summary of the Invention

[0009] The purpose of this application is to provide a micro-motion sensing system, method, electronic device, storage medium, and program product integrated on a single chip, in order to improve the existing schemes using FMCW radar, which cannot effectively distinguish multiple stationary targets that are close in space, and cannot meet the needs of accurate monitoring and analysis of multiple targets.

[0010] In a first aspect, embodiments of this application provide a micro-motion sensing system integrated on a single chip, the micro-motion sensing system integrated on a single chip comprising: The front end of the millimeter-wave noise radar is used to perform cross-correlation calculations between the echo signal and the transmitted broadband random modulated signal to generate a range profile of the target. An on-chip digital signal processor is used to calibrate and preprocess the range image to obtain a processed range image; An on-chip hardware AI accelerator is used to detect the micro-motion features of the target within each distance gate in the distance image, wherein each distance gate corresponds to a spatial detection area.

[0011] In the above implementation process, the millimeter-wave noise radar front-end used in this scheme transmits a broadband random modulation signal and performs cross-correlation processing. Since the performance of the broadband random modulation signal does not depend on a single waveform but on statistical averaging characteristics, the range image generated by its cross-correlation operation has a sharp peak at the target location, while the statistical average in other locations (sidelobes) tends to zero. Therefore, a high-resolution, low-sidelobe range image can be obtained, which fundamentally solves the problem of strong targets masking weak targets caused by high sidelobes in FMCW radar. Furthermore, by calibrating and preprocessing the range image through an on-chip digital processor, the compatibility between the AI ​​accelerator and the radar can be ensured, providing the necessary preprocessing support for the on-chip hardware AI accelerator. And because the range image itself has extremely low sidelobes, based on this high-quality range image, the on-chip hardware AI accelerator can extract the micro-motion features of the detected target in each range gate in parallel, thereby achieving high-precision and high-reliability perception of multiple stationary or micro-moving targets in complex scenarios, and thus meeting the needs for accurate monitoring and analysis of multiple targets.

[0012] Optionally, the millimeter-wave noise radar front end includes: The signal transceiver module is used to generate and transmit broadband random modulation signals and receive echo signals, wherein the center frequency of the broadband random modulation signals is located in the millimeter wave band. The signal correlation processing module, connected to the signal transceiver module, is used to perform cross-correlation calculations between the echo signal and a local copy of the broadband random modulation signal to generate a range profile of the detected target.

[0013] In the above implementation process, the millimeter-wave band broadband random modulation signal transmitted by the signal transceiver module, combined with the cross-correlation operation of the signal correlation processing module, can generate a range image with extremely low range sidelobes and high resolution. This architecture fully utilizes the low sidelobe characteristics of noisy radar, effectively reducing sidelobe interference and improving the detection capability of weak signals, making it particularly suitable for parallel monitoring of multiple targets in complex environments. Simultaneously, the use of broadband random modulation signals enhances the system's anti-interference capability, ensuring stability and reliability in multi-sensor coexistence scenarios.

[0014] Optionally, the signal correlation processing module includes multiple cross-correlators, each of which stores a local copy of a broadband random modulated signal corresponding to a range gate. Each cross-correlator is used to perform cross-correlation operation between the echo signal and the stored local copy to generate a range profile of the target within the range gate.

[0015] In the above implementation, multiple cross-correlators are set up in the signal correlation processing module. Each cross-correlator stores a local copy of the corresponding broadband random modulation signal for a specific range gate, and performs cross-correlation operations on the echo signal and the local copy, thereby generating an independent range image for each range gate. This design enables parallel processing of different spatial regions, significantly improving the system's multi-target detection capability and processing efficiency. Since each range gate has an independent local copy for correlation operations, it can more accurately extract the micro-motion features of the target within the corresponding region, effectively avoiding mutual interference between multi-target signals and improving detection accuracy and reliability.

[0016] Optionally, the millimeter-wave noise radar front-end further includes: an ADC sampling module; When the cross-correlator is a digital correlator, the ADC sampling module is located between the signal transceiver module and the signal correlation processing module. The ADC sampling module is used to sample the echo signal and input it to the signal correlation processing module. When the cross-correlator is an analog correlator, the ADC sampling module is located after the signal correlation processing module and is used to sample the range image.

[0017] In the above implementation process, two optimized signal processing paths are provided through flexible ADC sampling module layout: placing the ADC before the correlation processing enables the flexibility of fully digital processing and facilitates the implementation of complex algorithms; while placing the ADC after the analog correlator significantly reduces the system's requirements for ADC performance and power consumption by performing high-precision sampling on the slowed-down baseband signal, providing feasibility for low-cost, low-power edge intelligent micro-motion sensing applications.

[0018] Optionally, the signal transceiver module includes: A signal generator is used to generate baseband signals that produce random or pseudo-random sequences. A modulation module is used to modulate the baseband signal onto a millimeter-wave carrier wave to form a broadband random modulation signal, wherein the broadband random modulation signal has broadband signal characteristics similar to noise statistics; A transmitting antenna is used to transmit the broadband random modulated signal; A receiving antenna, used to receive echo signals; The transmitting antenna and the receiving antenna are MIMO antennas.

[0019] In the above implementation process, the random / pseudo-random sequence generated by the signal generator conforming to a specific mathematical distribution, combined with millimeter wave modulation and MIMO antenna architecture, not only ensures that the transmitted signal has low probability of noise interception and strong anti-interference characteristics, but also forms extremely high angular resolution through the virtual aperture of multiple antennas. Thus, in complex scenarios, high-precision ranging, multi-target resolution and robust micro-motion feature extraction can be achieved simultaneously.

[0020] Optionally, the broadband random modulation signal is a randomized OFDM signal; the randomized OFDM signal is generated by randomizing the subcarriers of the OFDM symbol.

[0021] In the above implementation process, randomized OFDM signals are used as broadband random modulation signals. By randomizing the subcarriers, the high spectral efficiency and strong anti-multipath capability of OFDM technology are effectively combined with the extremely low probability of interception of noisy radar, which further improves the anti-interference performance and signal concealment of the system.

[0022] Optionally, the signal correlation processing module is specifically used to: perform cross-correlation operation on the echo signal and a local copy of the broadband random modulation signal to recover the random sequence used to generate the randomized OFDM signal; use the random sequence to derandomize the echo signal to obtain a standard OFDM signal in the time domain; perform a fast Fourier transform on the standard OFDM signal in the time domain to obtain a frequency domain symbol; and generate a range profile of the detected target based on the phase and amplitude information of the frequency domain symbol.

[0023] In the above implementation process, the random sequence is recovered through cross-correlation operations and the echo is derandomized to convert the signal into the standard OFDM format. This allows for the acquisition of a high-resolution, low-sidelobe range image using a mature and efficient FFT processing link. This approach fully preserves the low interception and anti-interference characteristics of randomly modulated signals while successfully incorporating the high spectral efficiency, strong resistance to multipath fading, and flexible digital signal processing capabilities of OFDM technology, thus enhancing the system's robustness in complex channel environments.

[0024] Optionally, the on-chip hardware AI accelerator includes multiple processing units, each of which is used to detect the micro-motion features of a target within a range gate. This enables efficient feature extraction.

[0025] Secondly, embodiments of this application provide a micro-motion sensing method, which is applied to the on-chip hardware AI accelerator integrated into the aforementioned micro-motion sensing system on a single chip. The method includes: Receives the processed range image sent by the on-chip digital signal processor; The micro-motion features of the target within each range gate in the range image are detected, wherein each range gate corresponds to a spatial detection area.

[0026] In the above implementation process, since the obtained range image itself has extremely low sidelobes, based on this high-quality range image, the on-chip hardware AI accelerator can extract the micro-motion features of the detected target in each range gate in parallel, thereby realizing high-precision and high-reliability perception of multiple stationary or micro-moving targets in complex scenes, and thus meeting the needs of accurate monitoring and analysis of multiple targets.

[0027] Optionally, detecting the micro-motion features of the target within each range gate in the range image includes: For each range gate, determine the data points representing the detected target within each range image frame; Calculate the instantaneous phase of the data points; Arrange the instantaneous phases of the data points corresponding to the range gate in each frame of range image in chronological order to form a phase sequence; The micro-motion characteristics of the target are obtained based on the phase sequence.

[0028] In the above implementation process, by accurately extracting the instantaneous phase of the target data points from the range image and constructing a phase sequence, the physical micro-movements of the target (such as chest cavity fluctuations and mechanical vibrations) are directly converted into phase time-domain signals with high signal-to-noise ratios. This enables sensitive capture and accurate analysis of weak micro-movement features such as breathing and heartbeat, effectively improving the accuracy and reliability of micro-movement feature monitoring.

[0029] Optionally, determining the data points representing the detected target within the range gate from each frame range image includes: Extract all data points within the distance gate from each frame of the distance image; The peak clustering algorithm or the weighted average algorithm of signals from multiple strong reflection points is used to determine the data points representing the detection target from all the data points.

[0030] In the above implementation process, by selecting the data point with the strongest signal strength within the gate as the analysis object, the target subject with the strongest reflection in the spatial region can be effectively locked, eliminating the influence of background clutter and weak interference, and ensuring that subsequent phase extraction and micro-motion analysis are always aimed at the main detection target, thereby significantly improving the accuracy and reliability of feature detection.

[0031] Optionally, obtaining the micro-motion characteristics of the target based on the phase sequence includes: The number of detection targets is determined based on the phase sequence; If the number of detected targets is one, the phase sequence is input into the neural network model, and the micro-motion features of the detected target are extracted through the neural network model; If the number of detected targets is at least two, the phase sequence is divided into at least two independent phase sequences using variational mode decomposition, blind source separation algorithm or empirical mode decomposition, and each independent phase sequence is input into the corresponding neural network model, and the micro-motion features of the detected targets are extracted through the neural network model.

[0032] In the above implementation process, by intelligently judging the number of targets in the phase sequence and adaptively selecting the processing path, it can not only use neural networks to efficiently process single-target scenes, but also solve the problem of multi-target signal aliasing through advanced signal separation technology. Finally, feature extraction is completed through parallel neural networks, realizing the ability to maintain high accuracy in both single-target and multi-target scenes with adaptive micro-motion feature detection.

[0033] Optionally, the micro-motion features include the user's vital sign data, and after detecting the micro-motion features of the target within each distance gate, the method further includes: The system monitors for abnormalities in the user based on the vital signs data. By monitoring the extracted vital signs data in real time, it achieves a completely non-contact, continuous health status assessment. It can promptly detect dangerous conditions such as respiratory arrest and abnormal heart rate without the user noticing and trigger an alert. It can also be applied to the detection of abnormalities in scenarios such as human posture and behavior perception and gesture recognition, significantly improving user safety.

[0034] Optionally, the micro-motion feature includes the vibration frequency of industrial equipment, and after detecting the micro-motion feature of the target within each distance gate, the method further includes: The industrial equipment is monitored for abnormal conditions based on the vibration frequency. By accurately extracting the vibration frequency characteristics of the industrial equipment in a non-contact manner and performing real-time abnormal condition monitoring based on this, early potential faults such as bearing wear and blade cracks can be detected in a timely manner, thereby enabling predictive maintenance, effectively avoiding sudden equipment downtime, extending equipment service life, and significantly improving production safety and operational efficiency.

[0035] Thirdly, embodiments of this application provide an electronic device, including a processor and a memory, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the steps of the method provided in the second aspect above are performed.

[0036] Fourthly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the method provided in the second aspect above.

[0037] Fifthly, embodiments of this application provide a computer program product, including computer program instructions, which are read and executed by a processor to perform the steps of the method provided in the second aspect above.

[0038] Other features and advantages of this application will be set forth in the following description and will be apparent in part from the description or may be learned by practicing embodiments of this application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 A structural block diagram of a micro-motion sensing system integrated on a single chip, provided for an embodiment of this application; Figure 2 A structural block diagram of a millimeter-wave noise radar front-end provided in an embodiment of this application; Figure 3 This is a schematic diagram of the signal link structure of a micro-motion sensing system provided in an embodiment of this application; Figure 4 A schematic diagram of an on-chip hardware AI accelerator provided in this application embodiment; Figure 5 A flowchart of a micro-motion sensing method provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of an electronic device for performing a micro-motion sensing method, provided as an embodiment of this application. Detailed Implementation

[0041] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0042] It should be noted that the terms "system" and "network" in the embodiments of this invention can be used interchangeably. "Multiple" refers to two or more; therefore, in the embodiments of this invention, "multiple" can also be understood as "at least two". "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / ", unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship.

[0043] It should also be noted that all actions involving the acquisition of signals, information, or data in this application are carried out in compliance with the relevant data protection laws and policies of the country where the application is located, and with the authorization granted by the owner of the relevant device.

[0044] This application provides a micro-motion sensing system integrated on a single chip. The system includes a millimeter-wave noise radar front-end, an on-chip digital signal processor (DSP), and an on-chip hardware AI accelerator. In this solution, the millimeter-wave noise radar front-end, DSP, and AI accelerator are integrated onto a single chip, reducing hardware size, power consumption, latency, and cost to meet the needs of edge devices. The millimeter-wave noise radar front-end used in this solution transmits a broadband random modulation signal and performs cross-correlation processing. Since the performance of the broadband random modulation signal does not depend on a single waveform but on statistical averaging characteristics, the range image generated by its cross-correlation operation shows a sharp peak at the target location, while the statistical average in other locations (sidelobes) tends to zero. Therefore, a high-resolution, low-sidelobe range image can be obtained, fundamentally solving the problem of strong targets masking weak targets caused by high sidelobes in FMCW radar. Calibration and preprocessing of the range image by the on-chip DSP ensures the compatibility of the AI ​​accelerator with the radar and provides necessary preprocessing support for the on-chip hardware AI accelerator. On-chip hardware AI accelerators offer superior power consumption and speed, while also meeting current sensor privacy requirements. Furthermore, running lightweight models using on-chip hardware AI accelerators enables fast and privacy-preserving performance. In addition, on-chip hardware AI accelerators can extract the micro-motion features of the detected targets in parallel from each distance gate, thereby achieving high-precision and high-reliability perception of multiple stationary or micro-moving targets in complex scenarios, thus meeting the needs for accurate monitoring and analysis of multiple targets.

[0045] Please refer to Figure 1 , Figure 1 This is a schematic diagram of a micro-motion sensing system 100 integrated on a single chip, which is provided as an embodiment of this application. The micro-motion sensing system 100 integrated on a single chip includes a millimeter-wave noise radar front-end 110, an on-chip digital signal processor 120, and an on-chip hardware AI accelerator 130.

[0046] The millimeter-wave noise radar front-end 110 is used to perform cross-correlation calculations between the echo signal and the transmitted broadband random modulation signal to generate a range profile of the target.

[0047] The digital signal processor 120 is used to calibrate and preprocess the range image to obtain the processed range image.

[0048] The on-chip hardware AI accelerator 130 is used to detect the micro-motion features of the target within each range gate in the range image, where each range gate corresponds to a spatial detection area.

[0049] The millimeter-wave noise radar front-end 110 in this solution boasts the advantages of high resolution and low sidelobes. High resolution is ensured by the broadband signal emitted by the front-end 110, while low sidelobes are guaranteed by the randomness of the signal, eliminating the need for windowing and thus avoiding main lobe expansion and signal-to-noise ratio loss. Low sidelobes ensure that weak targets can be clearly distinguished even near strong targets, such as a weak reflective target (e.g., a pedestrian or bicycle) near a strong reflector (e.g., a metal wall or large vehicle). The millimeter-wave noise radar front-end 110 used in this solution is crucial for detecting pedestrians on the roadside in complex urban scenarios.

[0050] The millimeter-wave noise radar front-end 110 can transmit a broadband signal without a fixed pattern and similar to noise. This signal is usually generated by a pseudo-random sequence or chaotic source, and its frequency is in the millimeter-wave band.

[0051] The millimeter-wave noise radar front-end 110 can perform cross-correlation calculations between the received echo signal and a locally stored copy of the transmitted signal to obtain the range profile of the detected target. Specifically, the receiving antenna of the millimeter-wave noise radar front-end 110 can capture the weak echo reflected by the detected target, and then perform real-time cross-correlation calculations between the echo and the locally stored copy of the transmitted signal. When a certain time delay component in the echo matches the copy signal, the correlation calculation will output a peak value. By scanning all possible time delays, a curve is obtained, with the horizontal axis representing the time delay (corresponding to the distance) and the vertical axis representing the amplitude of the correlation peak. This curve is the range profile. Due to the characteristics of the noise signal, this range profile has extremely low range sidelobes, meaning that strong reflective targets will not produce wide sidelobes to overwhelm weak targets nearby. For example, in a vehicle interior scenario, the generated range profile can clearly show obvious peak values ​​at 0.8 meters, 1.1 meters, and 1.4 meters, corresponding to the chest positions of the driver's seat, front passenger seat, and rear passenger, respectively, without obscuring the weak reflection of the front passenger seat due to the strong reflection from the driver's seat.

[0052] After acquiring the range image, the millimeter-wave noise radar front-end 110 can send the range image to the on-chip digital signal processor 120 via the high-speed bus on the chip. The on-chip digital signal processor 120 is connected to the millimeter-wave noise radar front-end 110.

[0053] The on-chip digital signal processor 120 (DSP) can be used for front-end preprocessing of the range image. Specifically, this may include performing integrity and validity checks on the range image output from the millimeter-wave noise radar front-end 110, converting the range image into a fixed-point or floating-point format suitable for processing, performing I / Q imbalance correction, DC offset extraction, phase noise compensation, and other RF corrections. It also calculates the noise floor to provide normalized parameters for the on-chip hardware AI accelerator 130.

[0054] In some implementations, the on-chip digital signal processor 120 can also be used to post-process the output of the on-chip hardware AI accelerator 130, such as performing rationality checks and filtering smoothing on the output results of the on-chip hardware AI accelerator 130, performing multi-target data association and state tracking, and managing system output interfaces and external communications. The on-chip digital signal processor 120 and the on-chip hardware AI accelerator 130 form a coordinated processing pipeline. This has the advantage of utilizing the mature reliability of traditional digital signal processors in system calibration and target tracking, while fully leveraging the performance advantages of AI accelerators in complex feature extraction, thus achieving optimal system-level energy efficiency.

[0055] The on-chip hardware AI accelerator 130 is used to extract the micro-motion features of the detected target. In some implementations, multiple distance gates can be first divided from the distance image. The distance gates can be one or more fixed intervals manually divided on the distance image (these fixed intervals can be determined by detecting the distance of targets within the range scanned by the millimeter-wave noise radar front-end 110 during initialization, thus adapting to changes in target position). Each interval corresponds to a spatial detection area. For example, the distance gates can be set according to the application scenario. In an in-vehicle monitoring scenario, the distance range between each seat and the millimeter-wave noise radar front-end 110 can be pre-configured, such as distance gate 1: 0.7-0.9 meters, corresponding to the driver's area; distance gate 2: 0.9-1.1 meters, corresponding to the passenger's area; and distance gate 3: 1.2-1.4 meters, corresponding to the rear left passenger area.

[0056] Each distance gate corresponds to a spatial detection area, so the on-chip hardware AI accelerator 130 can extract the micro-motion features of the detected target within each distance gate. For example, it can extract the micro-motion features such as the heart rate and respiratory rate of the user inside the vehicle. These features can be used by upper-layer applications for health monitoring, safety warnings, etc.

[0057] In this solution, the millimeter-wave noise radar front-end 110, the on-chip digital signal processor 120, and the on-chip hardware AI accelerator 130 are integrated onto a single chip, such as a SoC chip. This reduces hardware size, power consumption, latency, and cost to meet the needs of edge devices.

[0058] In the above implementation process, the millimeter-wave noise radar front-end 110 used in this scheme transmits a broadband random modulation signal and performs cross-correlation processing. Since the performance of the broadband random modulation signal does not depend on a single waveform but on statistical averaging characteristics, the range image generated by its cross-correlation operation has a sharp peak at the target location, while the statistical average in other locations (sidelobes) tends to zero. Therefore, a high-resolution, low-sidelobe range image can be obtained, which fundamentally solves the problem of strong targets masking weak targets caused by high sidelobes in FMCW radar. Furthermore, by calibrating and preprocessing the range image through an on-chip digital processor, the compatibility between the AI ​​accelerator and the radar can be ensured, providing necessary preprocessing support for the on-chip hardware AI accelerator. And because the range image itself has extremely low sidelobes, based on this high-quality range image, the on-chip hardware AI accelerator can extract the micro-motion features of the detected target in each range gate in parallel, thereby achieving high-precision and high-reliability perception of multiple stationary or micro-moving targets in complex scenarios, thus meeting the needs for accurate monitoring and analysis of multiple targets.

[0059] Based on the above embodiments, the millimeter-wave noise radar front-end 110 may include a signal transceiver module 112 and a signal correlation processing module 114. The signal transceiver module 112 can be used to generate and transmit a broadband random modulation signal and receive echo signals, the center frequency of which is located in the millimeter-wave band. The signal correlation processing module 114 is connected to the signal transceiver module 112 and is used to perform cross-correlation calculations between the echo signal and a local copy of the broadband random modulation signal to generate a range profile of the detected target.

[0060] Wideband random modulation signals can be understood as electrical signals without specific patterns, and their statistical characteristics are similar to true random noise. The bandwidth of wideband random modulation signals can be no less than 1 GHz, generally in the range of 1 GHz to 5 GHz. A bandwidth of 4 GHz can provide a resolution of less than 5 cm in the distance dimension, which can distinguish multiple stationary targets that are close in space.

[0061] The millimeter wave band typically ranges from 30 GHz to 300 GHz. The center frequency of the broadband random modulation signal in this scheme can be 60 GHz or 77 GHz. Within this band, it is easy to obtain a large absolute bandwidth, which can more clearly distinguish two targets that are very close in radial distance, thereby achieving accurate segmentation of multiple targets.

[0062] The signal transceiver module 112 can capture the weak signal reflected back from the detection target, called the echo signal. The echo signal can be cross-correlated with a local copy of the broadband random modulation signal. The cross-correlation operation can be used to measure the similarity between two signals at different relative time delays. If the two signals are highly similar at a certain time delay, the operation result will produce a peak.

[0063] The signal correlation processing module 114 can be implemented using an FPGA (Field-Programmable Gate Array), an analog correlator, or a digital correlator to perform cross-correlation operations. The results of the cross-correlation operation over all time delays are then plotted to obtain a range image. This range image is a one-dimensional image where the horizontal axis represents distance (which can be calculated from time delays and the speed of light), and the vertical axis represents the intensity of the reflected signal from the target at that distance (i.e., the amplitude of the cross-correlation peak). Because the autocorrelation function of a random signal has a sharp, approximately "thumbtack"-shaped peak, the generated range image has a sharp main lobe (high resolution) and extremely low side lobes.

[0064] In the above implementation process, the millimeter-wave band broadband random modulation signal transmitted by the signal transceiver module, combined with the cross-correlation operation of the signal correlation processing module, can generate a range image with extremely low range sidelobes and high resolution. This architecture fully utilizes the low sidelobe characteristics of noisy radar, effectively reducing sidelobe interference and improving the detection capability of weak signals, making it particularly suitable for parallel monitoring of multiple targets in complex environments. Simultaneously, the use of broadband random modulation signals enhances the system's anti-interference capability, ensuring stability and reliability in multi-sensor coexistence scenarios.

[0065] Based on the above embodiments, the signal correlation processing module may include multiple cross-correlators. Each cross-correlator stores a local copy of a broadband random modulated signal corresponding to a range gate. Each cross-correlator is used to perform cross-correlation operation between the echo signal and the stored local copy to generate a range profile of the detected target within the range gate.

[0066] During system design or initialization, N distance gates to be monitored can be predefined according to the application scenario, such as distance gate 1 (driver's seat): 0.8 meters, distance gate 2 (passenger's seat): 1.1 meters, and distance gate 3 (rear left seat): 1.4 meters. These N distance gates can be dynamically set by the millimeter-wave noise radar front-end through target scanning. Since the millimeter-wave noise radar front-end is a MIMO antenna, its angle can be used to assist in target localization. Therefore, by locating the target, the distance gate where the target is located can be determined, and the initially set distance gates can be calibrated.

[0067] According to the delay calculation formula, the delay _k = (2*d_k) / c, where d_k represents the distance to the k-th distance gate and c represents the speed of light. The time delay value corresponding to the center distance of each distance gate can be calculated. Based on this time delay value, the local copy of the transmitted signal can be divided, for example, each distance gate corresponds to a time delay value. In terms of hardware configuration, the signal correlation processing module can also include a multi-channel delay network, which can be composed of a transmit signal distributor and a delay linear array.

[0068] A local copy s(t) of the transmitted signal (i.e., a broadband random modulated signal) is fed into this network. The network has N parallel paths, each with a programmable analog delay line. Each delay line is precisely configured to correspond to the delay of a distance gate. _k. Thus, the network output consists of N signals s(t- ),s(t- ), ...,s(t- ).

[0069] During cross-correlation processing, the first input of the k-th cross-correlator is connected to the undelayed, complete echo signal, and the second input of the k-th cross-correlator is connected to the k-th signal s(t-) output from the delay line network. Thus, the kth cross-correlator performs cross-correlation operation on the echo signal and the kth signal output by its delay line network to obtain the range image corresponding to the range gate.

[0070] In the above implementation, multiple cross-correlators are set up in the signal correlation processing module. Each cross-correlator stores a local copy of the corresponding broadband random modulation signal for a specific range gate, and performs cross-correlation operations on the echo signal and the local copy, thereby generating an independent range image for each range gate. This design enables parallel processing of different spatial regions, significantly improving the system's multi-target detection capability and processing efficiency. Since each range gate has an independent local copy for correlation operations, it can more accurately extract the micro-motion features of the target within the corresponding region, effectively avoiding mutual interference between multi-target signals and improving detection accuracy and reliability.

[0071] Based on the above embodiments, the millimeter-wave noise radar front-end 110 may also include an ADC (Analog-to-Digital Converter) sampling module. When the cross-correlator is a digital correlator, the ADC sampling module may be located between the signal transceiver module 112 and the signal correlation processing module 114, and is used to sample the echo signal and input it into the signal correlation processing module 114.

[0072] In this implementation, the signal is sampled first and then correlated. After the receiving antenna receives the echo signal, it can be down-converted to intermediate frequency in the analog domain. The ADC sampling module converts the continuous analog signal into a discrete digital signal. The echo signal is digitized before entering the digital correlator.

[0073] According to the Nyquist sampling theorem, the sampling frequency of an ADC sampling module must be at least twice the signal bandwidth. For example, if the bandwidth of a random signal emitted by a radar is 2 GHz, then the ADC sampling rate needs to reach at least 4 GS / s. This is equivalent to collecting 4 billion data points per second, placing extremely high demands on the ADC's performance.

[0074] The analog correlator can be composed of an integrator consisting of a Gilbert unit multiplier and an operational amplifier. Its output is a baseband analog voltage signal, which can be digitized by an ADC with a sampling rate of only 100 kS / s but a precision of 16 bits. This path has extremely low power consumption and is suitable for automotive and other applications.

[0075] Alternatively, in some other implementations, when the cross-correlator is an analog correlator, the ADC sampling module is located after the signal correlation processing module 114 and is used to sample the range image. In this case, the cross-correlation operation is performed in the analog domain, and the signal correlation processing module 114 can be an analog correlator rather than a digital processor.

[0076] The echo signal received by the receiving antenna is directly input into the signal correlation processing module 114 for processing. The output of the signal correlation processing module 114 is no longer a bandwidth signal, but a slowly changing baseband analog voltage signal. Its voltage value represents the target reflection intensity at a certain distance, that is, the range image.

[0077] The ADC sampling module samples the analog voltage signal in the range image. Since the frequency of this signal change is consistent with the frequency of micro-motion (0.1-0.5 Hz for breathing, 1-2 Hz for heartbeat), the ADC sampling rate only needs to be between 1 kS / s and 100 kS / s. In this scheme, the ADC sampling rate is even lower. A 16-bit or even 24-bit high-precision ADC can be selected to capture the minute voltage changes caused by micro-motion.

[0078] Digital correlators can be implemented in parallel using a large number of multiply-accumulate units in an FPGA to calculate the cross-correlation function in real time. This approach offers high flexibility and higher performance limits.

[0079] In the above implementation process, a flexible ADC sampling module layout provides two optimized signal processing paths: In the digital correlator architecture, the ADC sampling module is placed between the signal transceiver module and the signal correlation processing module, enabling high-speed sampling of the echo signal directly. This provides high-precision digital data for subsequent digital signal processing, ensuring the accuracy and reliability of signal processing, and is suitable for scenarios with high signal processing accuracy requirements. In the analog correlator architecture, the ADC sampling module is placed after the signal correlation processing module, sampling the range image that has already undergone analog correlation processing. This approach reduces the requirement for the ADC sampling rate, decreases system power consumption and cost, simplifies the signal processing flow, and improves system real-time performance and efficiency, making it particularly suitable for power- and cost-sensitive applications. This flexible ADC sampling module configuration allows the millimeter-wave noise radar front-end system to achieve an optimal performance-cost balance under different application scenarios and performance requirements, enhancing the system's adaptability and competitiveness.

[0080] Based on the above embodiments, such as Figure 2 As shown, the signal transceiver module 112 may include a signal generator 1122, a modulation module 1124, a transmitting antenna 1126, and a receiving antenna 1128. The signal generator 1122 is used to generate a baseband signal with a random or pseudo-random sequence. The modulation module 1124 is used to modulate the baseband signal onto a millimeter-wave carrier wave to form a broadband random modulation signal. The broadband random modulation signal has broadband signal with noise-like statistical characteristics. The transmitting antenna 1126 is used to transmit the broadband random modulation signal, and the receiving antenna 1128 is used to receive the echo signal. The transmitting antenna 1126 and the receiving antenna 1128 are MIMO (Multiple-Input Multiple-Output) antennas.

[0081] Signal generator 1122 can be a pseudo-random binary sequence generator, such as a PRBS31 generator or a chaotic sequence generator based on the Lorentz equation. A pseudo-random binary sequence generator can typically be constructed using a linear feedback shift register, generating a long-period binary sequence with good statistical properties according to a preset polynomial. For example, this generator can generate an m-sequence at a rate of 2 Gbps; this digital sequence is the baseband signal, and its values ​​are uniformly distributed between [-1, +1].

[0082] The baseband signal can be considered a random noise signal. The noise can be generated by utilizing the inherent physical noise of electronic components, such as shot noise or thermal noise in diodes or transistors. These are real, unpredictable physical random processes. Then, through amplifiers and filter circuits, this weak noise is shaped into a usable signal with specific power and bandwidth, resulting in a truly aperiodic random signal.

[0083] Wideband random modulated signals are wideband signals with noise-like statistical characteristics. This means that the amplitude, phase, and other parameters of the sequence conform to specific statistical laws, such as uniform distribution or Gaussian distribution, which can ensure that the signal has ideal correlation characteristics.

[0084] The modulation module 1124 can be implemented using a modulator such as a mixer, like a BPSK or QPSK modulator. Modulation refers to the process of mounting a low-frequency baseband signal onto a high-frequency radio wave for transmission via an antenna. The millimeter-wave carrier can be a relatively high-frequency sine wave, typically located in the millimeter-wave band, serving as the signal carrier. Specifically, the baseband signal can be fed into a modulator, and a millimeter-wave voltage-controlled oscillator can generate a frequency-stable high-frequency carrier. In the modulator, the baseband signal is multiplied by the carrier, shifting its spectrum to the millimeter-wave band. For example, a baseband m-sequence modulated onto a 77GHz carrier generates a broadband random modulated signal with a center frequency of 77GHz and a bandwidth of up to 4GHz.

[0085] The transmitting antenna 1126 and the receiving antenna 1128 are MIMO antennas, which are multiple transmitting antennas and multiple receiving antennas. The principle of MIMO technology is that by using different combinations of N transmitting antennas and M receiving antennas, N*M independent virtual transceiver channels can be created, which can greatly improve the angular resolution of the system without actually deploying N*M physical antennas.

[0086] The generated broadband random modulated signal can be fed to the transmit antenna array and can be multiplexed by time division or code multiplexing. Then, multiple transmit antennas transmit mutually orthogonal random signals in an orderly or simultaneous manner.

[0087] Each antenna in the receiving antenna array independently receives echo signals from the scene. The signal received by each receiving channel is a mixture of all the transmitting antenna signals reflected by the target. For example, four receiving antennas may receive signals simultaneously. Since the specific signal transmitted by each transmitting antenna is known, the back-end processor can "demix" the mixed signal, separating the independent channel signals between each pair of "transmit-receive" antennas.

[0088] In the above implementation process, the random / pseudo-random sequence generated by the signal generator conforming to a specific mathematical distribution, combined with millimeter wave modulation and MIMO antenna architecture, not only ensures that the transmitted signal has low probability of noise interception and strong anti-interference characteristics, but also forms extremely high angular resolution through the virtual aperture of multiple antennas. Thus, in complex scenarios, high-precision ranging, multi-target resolution and robust micro-motion feature extraction can be achieved simultaneously.

[0089] In some other implementations, the broadband randomized modulation signal described above can be implemented using a randomized OFDM (Orthogonal Frequency Division Multiplexing) signal, which is generated by randomizing the subcarriers of the OFDM symbol.

[0090] OFDM itself is a multi-carrier modulation technique that divides a high-speed data stream into multiple mutually orthogonal subcarriers with different frequencies for parallel transmission, and has the advantages of high spectral efficiency and strong resistance to multipath fading.

[0091] A random sequence can be generated by a signal generator and applied to the data symbols of each subcarrier within each OFDM symbol period. In practice, the subcarriers of the OFDM signal are preprocessed with randomization to give them noise-like statistical characteristics; for example, phase randomization can be used alone, or a combination of amplitude and phase randomization can be used. At the receiver, this randomization is decoupled by cross-correlation processing with a known reference signal, and then a high-resolution range profile is obtained using the OFDM processing link.

[0092] Specifically, random phase modulation refers to multiplying the mapped data symbols on each subcarrier (e.g., QPSK (Quadrature Phase Shift Keying) symbols) by a random phase rotation factor (e.g., ...). , where θ is randomly and uniformly distributed in [0, 2π); random amplitude modulation refers to weighting the amplitude of data symbols with a random number.

[0093] After randomized modulation, the complex symbols on all subcarriers form a frequency domain vector. This frequency domain vector is then converted into a discrete time domain signal by an inverse fast Fourier transform, and a cyclic prefix is ​​added to counteract time delay spread, ultimately generating a randomized OFDM baseband signal in the time domain. This baseband signal is then up-converted to the millimeter-wave band (e.g., 77 GHz) and radiated by a transmitting antenna.

[0094] In the above implementation process, randomized OFDM signals are used as broadband random modulation signals. By randomizing the subcarriers, the high spectral efficiency and strong anti-multipath capability of OFDM technology are effectively combined with the extremely low probability of interception of noisy radar, which further improves the anti-interference performance and signal concealment of the system.

[0095] In some implementations, to adapt to the back-end processing of randomized OFDM signals, the signal correlation processing module can be used to perform cross-correlation operations on the echo signal and a local copy of the broadband randomized modulated signal to recover the random sequence used to generate the randomized OFDM signal. The random sequence is then used to derandomize the echo signal to obtain a standard OFDM signal in the time domain. A fast Fourier transform is then performed on the standard OFDM signal in the time domain to obtain a frequency domain symbol. Based on the phase and amplitude information of the frequency domain symbol, a range profile of the detected target is generated.

[0096] The echo captured by the receiving antenna is down-converted to obtain a mixed signal of baseband randomized OFDM signal, noise, and interference. The signal correlation processing module first performs the core cross-correlation operation: it cross-correlates the received echo signal with a locally stored copy of the transmitted signal. This copy has the same random sequence and structure as the transmitted randomized OFDM signal, thus enabling the recovery or extraction of the random modulation factors applied to the subcarriers. From the output of the correlation operation, the random sequence (such as the random phase rotation factor and / or amplitude weighting factor on each subcarrier) initially applied at the transmitting end can be extracted or recovered.

[0097] Subsequently, the extracted random sequence is used to perform an inverse operation on the received echo signal. Specifically, for each received symbol on a subcarrier, it is divided by the corresponding random phase factor (and amplitude compensation is performed if amplitude randomization is included). This step is equivalent to stripping away the random "mask" added during transmission, restoring the signal to a pure time-domain standard OFDM signal.

[0098] Finally, the recovered time-domain standard OFDM signal undergoes conventional OFDM reception processing, primarily including Fast Fourier Transform (FFT) to convert the time-domain signal back to the frequency domain. At this point, the signal amplitude and phase on each subcarrier contain the target's reflection coefficient information. Since each OFDM subcarrier essentially corresponds to a specific frequency point, the frequency offset or phase change of its reflected signal directly maps to the target range. Therefore, a high-resolution range profile can be synthesized by analyzing all subcarriers.

[0099] In the above implementation process, the random sequence is recovered through cross-correlation operations, and the echo is derandomized to convert the signal into standard OFDM format. This allows for the acquisition of a low-sidelobe range profile using a mature and efficient FFT processing link. This approach fully preserves the low intercept and anti-interference characteristics of randomly modulated signals while successfully incorporating the high spectral efficiency, strong resistance to multipath fading, and flexible digital signal processing capabilities of OFDM technology, thus improving the system's robustness in complex channel environments. Furthermore, the randomization of subcarrier modulation effectively controls the peak-to-average power ratio of the system's transmitted signal, reducing the requirements for the power amplifier.

[0100] Please refer to Figure 3 , Figure 3 The diagram shows the structure of the transmit link and receive processing link of the millimeter-wave noise radar front end. Figure 3 In the process, the transmitted signal first passes through the signal generator, modulation module and transmitting antenna in the transmission link. The echo signal is processed by the receiving processing link and then input into the on-chip digital signal processor 120, and then into the on-chip hardware AI accelerator 130.

[0101] In the hardware implementation, the on-chip AI accelerator can be implemented using hardware chips such as FPGA, ASIC (Application-Specific Integrated Circuit), or NPU (Neural Processing Unit). In this approach, the range image obtained by the millimeter-wave noise radar front-end 110 can be directly input to the on-chip digital signal processor 120 via a dedicated high-speed data channel for preprocessing before being input into the AI ​​accelerator. In the hardware, the address range of each range gate can be configured in an independent register. When a frame of range image data flows in, a set of data distributors simultaneously copy and route the data to the corresponding processing channel based on these addresses. Thus, each range gate's data has an independent, dedicated computing unit, and these computing units can work simultaneously and in parallel, thereby achieving efficient processing of micro-motion features.

[0102] In the above implementation process, the on-chip hardware AI accelerator can perform high-speed parallel processing of signals from multiple distance gates and directly complete complex micro-motion feature extraction algorithms at the sensor edge, thereby achieving low latency, low power consumption and high efficiency, and providing reliable edge intelligent processing capabilities for multi-target real-time monitoring scenarios.

[0103] In some implementations, the on-chip hardware AI accelerator includes multiple processing units, each of which is used to detect the micro-motion features of a target within a range gate.

[0104] In some implementations, the on-chip hardware AI accelerator includes a range gate partitioning module, a signal separation unit, and a model inference engine. In this implementation, the processing unit is the model inference engine, and multiple model inference engines can be set up to detect the micro-motion features of the target within multiple range gates.

[0105] The range gate partitioning module can be used to divide the range image into multiple range gates, and then input the range image corresponding to each range gate into a processing unit for processing. Of course, if the radar uses multiple cross-correlators as mentioned above, the range gate partitioning module can be omitted, and the range image of each range gate can be directly input into the signal separation unit. The range gate partitioning module can be implemented through digital filtering (such as FFT). The signal received by the radar is a time-domain signal, which can be converted into a frequency-domain signal through Fast Fourier Transform. The frequency component directly corresponds to the target range, and each frequency unit (or a group of frequency units) constitutes a range gate.

[0106] The signal separation unit can refer to the VMD (Variational Mode Decomposition) signal separation unit. VMD is a signal decomposition algorithm that can automatically decompose a complex mixed signal into several relatively simple sub-signal components of different frequencies.

[0107] The model inference engine can refer to a one-dimensional convolutional neural network. The clean, pre-processed micro-motion signals (which may be a combination of multiple VMD components) are fed into this model inference engine. The engine is pre-trained and can automatically learn and recognize the micro-motion features in these signals.

[0108] Understandably, the model inference engine in this implementation can also be used to extract micro-motion features and perform anomaly recognition based on these features.

[0109] In some implementations, such as Figure 4 As shown, the on-chip hardware AI accelerator may include a distance gate partitioning module, multiple processing units (PE arrays), a signal separation unit, and a model inference engine. In this embodiment, the processing unit is a PE array.

[0110] The implementation methods of the distance gate partitioning module and the signal separation unit are similar to those described above, and will not be repeated here.

[0111] Multiple processing units can be understood as multiple independent micro-motion sensing engines, or PE arrays. Each engine corresponds to a range gate and contains dedicated phase sequence buffers and target quantity judgment logic. The PE array is a parallel computing network composed of a large number of small, dedicated computing units. This module is specifically used to extract signals generated by the target's minute movements (such as heartbeat, breathing, and limb swings during walking) from the data of a single range gate. These engines work together through a shared weight buffer and a unified scheduler to achieve synchronous processing of all range gate signals.

[0112] In this embodiment, after the micro-motion features are extracted by the processing unit, the mixed micro-motion features are separated by the signal separation unit, and then input into the model inference engine for anomaly recognition. That is, the model inference engine has been pre-trained and can automatically learn and recognize the behavior or target category or anomaly pattern corresponding to these signal patterns, such as walking, falling, or abnormality.

[0113] In some other implementations, if the signal correlation processing module described above is implemented using multiple cross-correlators, then the range image output by each cross-correlator is input into the on-chip digital signal processor for preprocessing, and the range image of the subsequent range gate is input into a processing unit for further processing. Since each cross-correlator generates a range image of a range gate, and this range image can be output to a processing unit, one processing unit can be used to perform further processing on the range image of a range gate. Multiple processing units can achieve parallel processing, thereby greatly improving processing efficiency.

[0114] The following section provides a detailed explanation of the process of micro-motion feature extraction in on-chip hardware AI accelerators.

[0115] Please refer to Figure 5 , Figure 5 A flowchart of a micro-motion sensing method provided in this application embodiment, which is applied to the above-mentioned on-chip hardware AI accelerator, includes the following steps: Step S210: Receive the processed range image sent by the on-chip digital signal processor.

[0116] Step S220: Detect the micro-motion features of the target at each range gate in the range image.

[0117] The process of generating a range image at the front end of the millimeter-wave noise radar can be referred to the relevant description in the above embodiments. For the sake of brevity, it will not be elaborated further here.

[0118] After acquiring the range image, the millimeter-wave noise radar front-end can directly input the range image into an on-chip digital signal processor for preprocessing. The processed range image is then input into an on-chip hardware AI accelerator, which performs subsequent micro-motion feature extraction after acquiring the range image. In some embodiments, the on-chip hardware AI accelerator first divides the range image into multiple range gates, and then detects the micro-motion features of the target within each range gate, with each range gate corresponding to a spatial detection area. In other embodiments, the on-chip hardware AI accelerator can directly acquire the range images of each range gate (as in the architecture of multiple cross-correlators and multiple processing units mentioned above), thus enabling direct micro-motion feature extraction from the range image of each range gate.

[0119] Each acquired range image frame can be input into an on-chip hardware AI accelerator. The on-chip hardware AI accelerator performs range gating on each range image frame, or it can directly acquire each range image frame for each range gate. The range gates can be some pre-configured range ranges. That is, the range image can be divided according to the range gates, thus obtaining a range image segment for each range gate. Then, the micro-motion features of the detected target can be extracted from the range image segment of that range gate.

[0120] The detection target mentioned in this solution can be any target that needs to be monitored in certain scenarios. For example, in a vehicle interior scenario, the detection target could be a user. In this case, the millimeter-wave noise radar front-end can be installed on the vehicle's ceiling, pointing towards the passenger compartment. The distance threshold can be pre-configured between the millimeter-wave noise radar front-end and the driver's seat, front passenger seat, and rear seats; these distance ranges constitute the distance threshold. Micro-motion characteristics could include the user's heart rate, respiratory rate, etc. In an industrial equipment monitoring scenario, the detection target could be industrial equipment. In this case, the millimeter-wave noise radar front-end can be installed pointing towards critical equipment in the factory, such as a high-speed motor spindle. The distance threshold is the distance range from the millimeter-wave noise radar front-end to the industrial equipment, and micro-motion characteristics could include the vibration frequency of the industrial equipment. In an elderly care monitoring scenario, the detection target could be a user requiring monitoring. In this case, the millimeter-wave noise radar front-end can be installed on the ceiling of the living room or bedroom. The distance threshold is the distance range from the millimeter-wave noise radar front-end to areas such as sofas, beds, and dining tables, and micro-motion characteristics could include the user's heart rate, respiratory rate, etc.

[0121] In the above implementation process, since the obtained range image itself has extremely low sidelobes, based on this high-quality range image, the on-chip hardware AI accelerator can extract the micro-motion features of the detected target in each range gate in parallel, thereby realizing high-precision and high-reliability perception of multiple stationary or micro-moving targets in complex scenes, and thus meeting the needs of accurate monitoring and analysis of multiple targets.

[0122] Based on the above embodiments, in the method of detecting the micro-motion features of the target within each range gate, the on-chip hardware AI accelerator can first determine the data points representing the target within each range gate from each frame of range image, then calculate the instantaneous phase of the data points, arrange the instantaneous phases of the data points corresponding to the range gate in each frame of range image in chronological order to form a phase sequence, and then obtain the micro-motion features of the target based on the phase sequence.

[0123] Each frame of range image is input into the on-chip hardware AI accelerator. The on-chip hardware AI accelerator divides each frame of range image into range gates according to the pre-configured range gates. In this way, each frame of range image can be divided into range images with multiple range gates. For example, each frame of range image can be divided into range images with 4 range gates. Then, the data points representing the detected target are determined from the range images of each range gate.

[0124] Data points refer to sampling points selected within the range profile of each range gate to represent the primary target (such as a user) within that range gate. These data points can be used to calculate the instantaneous phase. Instantaneous phase refers to the phase value of a complex signal at a given moment. Radar signals are typically in complex form, containing in-phase and quadrature components, i.e., I / Q signals, where I represents the in-phase component and Q represents the quadrature component. Together, they form a complex number Z = I + jQ.

[0125] For a given data point, its I and Q values ​​can be obtained, and its instantaneous phase can be calculated using the four-quadrant arctangent function. .

[0126] For each range gate, multiple instantaneous phases are calculated across multiple range images, and these instantaneous phases can then be arranged in chronological order. That is, the instantaneous phases of representative data points for the same range gate across multiple consecutive range images are arranged in chronological order to form a time-series signal. For example, the radar generates range images at a fixed frame rate (e.g., 30 frames / second). For the passenger-side range gate, the phase calculation steps are repeated for each frame to obtain a phase value. These values ​​are then stored sequentially in an array. For instance, within 5 seconds, the system collects 150 phase values ​​(30 frames / second * 5 seconds). This array contains 150 elements. It's a phase sequence. The waveform of this sequence is actually the microscopic trajectory of the front passenger's chest cavity surface moving back and forth due to breathing and heartbeat.

[0127] After obtaining the phase sequence corresponding to each range gate, the phase sequence can be analyzed to extract the micro-motion features of the target. For example, the phase sequence can be input into a neural network model or machine learning model, and the micro-motion features can be extracted through the pre-trained neural network model or machine learning model.

[0128] In the above implementation process, by accurately extracting the instantaneous phase of the target data points from the range image and constructing a phase sequence, the physical micro-movements of the target (such as chest cavity fluctuations and mechanical vibrations) are directly converted into phase time-domain signals with high signal-to-noise ratios. This enables sensitive capture and accurate analysis of weak micro-movement features such as breathing and heartbeat, effectively improving the accuracy and reliability of micro-movement feature monitoring.

[0129] Based on the above embodiments, in the above method of determining the data points representing the target within each range gate from each frame range image, all data points within the range gate can be extracted from each frame range image, and then a peak clustering algorithm or a weighted average algorithm for the signals of multiple strong reflection points can be used to determine the data points representing the target from all data points.

[0130] In the method of using peak clustering algorithm to determine the data points of the detected target, all data points are first extracted within the range image of the range gate. Then, these data points are clustered according to the signal amplitude value, thereby finding multiple strong reflection points generated by the same target. These strong reflection data points usually correspond to the parts with the strongest reflection. For example, within the passenger seat range gate, if N data points are clustered into one class, then these data points are selected as the representative data points of the range gate in this frame's range image.

[0131] Specifically, each data point can be plotted in a two-dimensional space of "distance index - signal amplitude". The algorithm searches for points with amplitude values ​​significantly higher than the noise floor and closely adjacent in distance, grouping them into a cluster. A cluster represents a potential physical sub-target or part of a target. For example, within the passenger seat's distance from the door, the algorithm might identify two clusters: cluster A (possibly corresponding to the chest cavity, with high and concentrated amplitude) and cluster B (possibly corresponding to a raised arm, with slightly lower amplitude and a slightly different position).

[0132] Then, the total energy (sum of squared amplitudes) of all points within each cluster is calculated, and the cluster with the largest total energy is selected as the main cluster, representing the primary detection target. Data points within the main cluster can be used as data points for the detection target, or the data point with the largest signal amplitude can be selected as the final representative point.

[0133] In determining the data points for the target detection using a weighted averaging algorithm based on signals from multiple strong reflection points, we can first use the square of the signal amplitude (i.e., power) of each data point as its weight, as power better reflects the energy contribution of the signal. Then, we can separately perform a weighted averaging of the I and Q values ​​of all data points to obtain a virtual, more representative data point. This improves stability and avoids introducing noise due to minor fluctuations at individual points.

[0134] In the above implementation process, by accurately extracting the instantaneous phase of the target data points from the range image and constructing a phase sequence, the physical micro-movements of the target (such as chest cavity fluctuations and mechanical vibrations) are directly converted into phase time-domain signals with high signal-to-noise ratios. This enables sensitive capture and accurate analysis of weak micro-movement features such as breathing and heartbeat, effectively improving the accuracy and reliability of micro-movement feature monitoring.

[0135] Based on the above embodiments, if some detection targets are relatively close, in order to achieve accurate differentiation, the number of detection targets is first determined according to the phase sequence. If the number of detection targets is one, the phase sequence is input into the neural network model, and the micro-motion features of the detection target are extracted through the neural network model. If the number of detection targets is at least two, the phase sequence is divided into at least two independent phase sequences using variational mode decomposition, blind source decomposition algorithm or empirical mode decomposition method, and each independent phase sequence is input into the corresponding neural network model, and the micro-motion features of the detection target are extracted through the neural network model.

[0136] After obtaining the phase sequence corresponding to each range gate, the number of detected targets within that range gate can be determined for each phase sequence. For example, by analyzing the statistical or spectral characteristics of the phase sequence, it can be determined that it is a mixture of one or more detected targets.

[0137] In some implementations, the entropy of the power spectrum of the phase sequence can be calculated. The spectrum of a single periodic signal is simple and has a low entropy, while the spectrum of a signal mixed from multiple different frequency sources is complex and has a high entropy.

[0138] In some implementations, principal component analysis (PCA) can be performed on the signal. If the first principal component explains the majority of the variance, it is likely a single source; if more than two principal components are needed to explain most of the variance, it indicates the presence of multiple sources. For example, in a car's rear seat distance gate, if the system calculates that the spectral entropy of its phase sequence is high and indicates that two principal components are needed, it can be determined that there are two targets detected within the distance gate.

[0139] In some implementations, an FFT operation can be performed on the signal to count the number of peak values ​​that significantly exceed the noise floor.

[0140] The number of detection targets can be determined based on the phase sequence using the methods described above. If there is only one detection target, the original phase sequence can be directly input into a pre-trained neural network model, such as a convolutional neural network model (e.g., a one-dimensional convolutional neural network 1D-CNN or a temporal convolutional network TCN). This neural network model is a lightweight model; its input is the pre-processed phase sequence, and its output is the analyzed micro-motion features (e.g., respiratory rate, heart rate). This model is trained on a dedicated dataset and deployed on the on-chip hardware AI accelerator. The neural network model can directly predict the micro-motion features of the detection target from the phase sequence.

[0141] If there are at least two targets within a range gate, the phase sequence can be separated using variational mode decomposition or blind source separation algorithm. Variational mode decomposition can decompose a complex signal into K quasi-orthogonal intrinsic mode functions (IMFs) with specific center frequencies. Each IMF can represent a target. Blind source separation algorithm can recover the unknown source signal from the mixed signal.

[0142] Specifically, the on-chip hardware AI accelerator can run variational mode decomposition (VMD) to decompose the mixed phase sequence into K IMF components, where the value of K is determined by the number of known detection targets. The iterative computation process of VMD is implemented as highly parallel matrix operations on the AI ​​accelerator to ensure real-time performance.

[0143] Furthermore, blind source decomposition algorithms refer to a class of mathematical methods that recover individual source signals from the observed mixed signal when both the source signals and the mixing method are unknown. Their core assumption is that the multiple source signals are statistically independent, and an optimization algorithm is used to find a separation matrix that maximizes the independence between the output signals.

[0144] Empirical Mode Decomposition (EMD) is a fully data-driven adaptive signal decomposition method specifically designed to handle nonlinear and non-stationary time-domain signals. Its core idea is to adaptively decompose any complex signal into a finite number of intrinsic mode functions (EMFs) with different time scales and a residual trend term (representing the long-term trend or DC component of the signal).

[0145] For details on the implementation process of signal separation using variational mode decomposition, blind source decomposition, and empirical mode decomposition, please refer to the implementation process in related technologies; detailed examples will not be provided here.

[0146] Each separated phase sequence is simultaneously fed into an independent neural network model (these neural network models can be the same or optimized for different frequency ranges). Each neural network model works independently and outputs the micro-motion characteristics of the corresponding detection target.

[0147] In the above implementation process, by intelligently judging the number of targets in the phase sequence and adaptively selecting the processing path, it can not only use neural networks to efficiently process single-target scenes, but also solve the problem of multi-target signal aliasing through advanced signal separation technology. Finally, feature extraction is completed through parallel neural networks, realizing the ability to maintain high accuracy in both single-target and multi-target scenes with adaptive micro-motion feature detection.

[0148] Based on the above embodiments, after acquiring the micro-motion characteristics of the target, the on-chip hardware AI accelerator can send the micro-motion characteristics to the control module, which then performs operations such as anomaly monitoring based on the micro-motion characteristics. Micro-motion characteristics have different applications in different scenarios; some typical application scenarios are listed below.

[0149] Micro-motion characteristics include the user's vital signs data, such as heart rate and respiratory rate, which can be used to detect abnormalities in the user. By monitoring the extracted vital signs data in real time, a completely non-contact, continuous health status assessment is achieved. It can promptly detect dangerous conditions such as respiratory arrest and abnormal heart rate without the user noticing, and trigger warnings, significantly improving user safety.

[0150] Scenario 1: Medical and health monitoring field.

[0151] A millimeter-wave noise radar front-end can replace some electrode patches, enabling monitoring of a patient's heart rate, respiratory rate, etc., in the aforementioned manner. Users do not need to wear any devices, allowing for 24 / 7 physiological parameter monitoring under normal living conditions. Micro-motion characteristics can be analyzed to analyze the user's health status and assess vital signs. For example, if the patient's heart rate or respiratory rate exceeds the normal threshold range, it can be considered abnormal, triggering an alarm. This method achieves environmental perception capabilities comparable to visual sensors while completely protecting privacy, and is unaffected by light conditions.

[0152] Scenario 2: Monitoring the vital signs of occupants inside the vehicle.

[0153] The above method can acquire the micro-motion characteristics of each occupant in the vehicle, such as heart rate and respiratory rate, and then perform anomaly detection. For example, if an abnormality in the driver's heart rate (such as a sudden increase or decrease) is detected, an alarm message can be sent to the vehicle's main unit through the output module. This method can monitor all occupants simultaneously, independently, and imperceptibly, solving the problem of signal interference and indistinguishability in multi-person scenarios for FMCW radar.

[0154] Micro-motion characteristics can include the vibration frequency of industrial equipment, which can be used to monitor the abnormal condition of industrial equipment. By accurately extracting the vibration frequency characteristics of industrial equipment in a non-contact manner and using this as a basis for real-time abnormal condition monitoring, early potential failures such as bearing wear and blade cracks can be detected in a timely manner, thereby enabling predictive maintenance, effectively avoiding sudden equipment downtime, extending equipment lifespan, and significantly improving production safety and operational efficiency.

[0155] Scenario 3: Predictive maintenance of industrial equipment.

[0156] The above method allows for the acquisition of micro-motion characteristics such as vibration frequencies from various industrial equipment. The control module can then identify characteristic frequencies representing faults such as bearing wear and blade cracks. When the fault characteristic energy exceeds a set threshold, a warning signal is output, recommending shutdown for maintenance. This non-contact measurement method avoids the hassle of installing sensors, and its extremely high distance resolution allows for the precise extraction of minute vibrations on component surfaces from strong equipment body echoes.

[0157] In addition, it can be applied to the field of biosensing, such as monitoring the physiological activities (heartbeat, respiration) of small laboratory animals (e.g., mice) without interference in the laboratory. It can also be applied to border / perimeter security, such as detecting and distinguishing intruders from small animals, reducing false alarms. Furthermore, it can be applied to drone landing guidance, human posture recognition, and intelligent gesture recognition.

[0158] Please refer to Figure 6 , Figure 6 This is a schematic diagram of an electronic device for performing a micro-motion sensing method, provided as an embodiment of this application. The electronic device may include: at least one processor 310, such as a CPU; at least one communication interface 320; at least one memory 330; and at least one communication bus 340. The communication bus 340 is used to establish communication between these components. In this embodiment, the communication interface 320 is used for signaling or data communication with other node devices. The memory 330 may be a high-speed RAM or a non-volatile memory, such as at least one disk storage device. Optionally, the memory 330 may also be at least one storage device located remotely from the aforementioned processor. The memory 330 stores computer-readable instructions, which, when executed by the processor 310, cause the electronic device to perform the aforementioned method process.

[0159] Understandable. Figure 6 The structure shown is for illustrative purposes only; the electronic device may also include components that are more advanced than those shown. Figure 6 The more or fewer components shown, or having the same Figure 6 The different configurations shown. Figure 6 The components shown can be implemented using hardware, software, or a combination thereof.

[0160] This application provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it performs the method process executed by the electronic device in the above method embodiments.

[0161] This embodiment discloses a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, and when the program instructions are executed by a computer, the computer can perform the methods provided in the above-described method embodiments, such as including: Receives the processed range image sent by the on-chip digital signal processor; The micro-motion features of the target within each range gate in the range image are detected, wherein each range gate corresponds to a spatial detection area.

[0162] In summary, this application provides a micro-motion sensing system, method, electronic device, storage medium, and program product integrated on a single chip. The millimeter-wave noise radar front-end used in this solution transmits a broadband random modulation signal and performs cross-correlation processing. Since the performance of the broadband random modulation signal does not depend on a single waveform but on statistical averaging characteristics, the range image generated by its cross-correlation operation shows a sharp peak at the target location, while the statistical average in other locations (sidelobes) approaches zero. Therefore, a high-resolution, low-sidelobe range image can be obtained, fundamentally solving the problem of strong targets masking weak targets caused by high sidelobes in FMCW radar. Furthermore, by calibrating and preprocessing the range image through an on-chip digital processor, the compatibility between the AI ​​accelerator and the radar can be ensured, providing necessary preprocessing support for the on-chip hardware AI accelerator. And because the range image itself has extremely low sidelobes, based on this high-quality range image, the on-chip hardware AI accelerator can extract the micro-motion features of the detected target in parallel from each range gate, thereby achieving high-precision and high-reliability sensing of multiple stationary or micro-moving targets in complex scenarios, thus meeting the needs for accurate monitoring and analysis of multiple targets.

[0163] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0164] Furthermore, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0165] Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0166] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0167] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A micro-motion sensing system integrated on a single chip, characterized in that, The micro-motion sensing system includes: The front end of the millimeter-wave noise radar is used to perform cross-correlation calculations between the echo signal and the transmitted broadband random modulated signal to generate a range profile of the target. An on-chip digital signal processor is used to calibrate and preprocess the range image to obtain a processed range image; An on-chip hardware AI accelerator is used to detect the micro-motion features of the target within each distance gate in the distance image, wherein each distance gate corresponds to a spatial detection area.

2. The micro-motion sensing system integrated on a single chip according to claim 1, characterized in that, The millimeter-wave noise radar front-end includes: The signal transceiver module is used to generate and transmit broadband random modulation signals and receive echo signals, wherein the center frequency of the broadband random modulation signals is located in the millimeter wave band. The signal correlation processing module, connected to the signal transceiver module, is used to perform cross-correlation calculations between the echo signal and a local copy of the broadband random modulation signal to generate a range profile of the detected target.

3. The micro-motion sensing system integrated on a single chip according to claim 2, characterized in that, The signal correlation processing module includes multiple cross-correlators. Each cross-correlator stores a local copy of a broadband random modulated signal corresponding to a range gate. Each cross-correlator is used to perform cross-correlation operation between the echo signal and the stored local copy to generate a range profile of the target within the range gate.

4. The micro-motion sensing system integrated on a single chip according to claim 3, characterized in that, The millimeter-wave noise radar front-end also includes: an ADC sampling module; When the cross-correlator is a digital correlator, the ADC sampling module is located between the signal transceiver module and the signal correlation processing module. The ADC sampling module is used to sample the echo signal and input it to the signal correlation processing module. When the cross-correlator is an analog correlator, the ADC sampling module is located after the signal correlation processing module and is used to sample the range image.

5. The micro-motion sensing system integrated on a single chip according to claim 2, characterized in that, The signal transceiver module includes: A signal generator is used to generate baseband signals that produce random or pseudo-random sequences. A modulation module is used to modulate the baseband signal onto a millimeter-wave carrier wave to form a broadband random modulation signal, wherein the broadband random modulation signal has broadband signal characteristics similar to noise statistics; A transmitting antenna is used to transmit the broadband random modulated signal; A receiving antenna, used to receive echo signals; The transmitting antenna and the receiving antenna are MIMO antennas.

6. The micro-motion sensing system integrated on a single chip according to claim 2, characterized in that, The broadband random modulation signal is a randomized OFDM signal; the randomized OFDM signal is generated by randomizing the subcarriers of the OFDM symbol.

7. The micro-motion sensing system integrated on a single chip according to claim 6, characterized in that, The signal correlation processing module is specifically used for: performing cross-correlation operations on the echo signal and a local copy of the broadband random modulation signal to recover the random sequence used to generate the randomized OFDM signal; using the random sequence to derandomize the echo signal to obtain a standard OFDM signal in the time domain; performing a fast Fourier transform on the standard OFDM signal in the time domain to obtain a frequency domain symbol; and generating a range profile of the detected target based on the phase and amplitude information of the frequency domain symbol.

8. The micro-motion sensing system integrated on a single chip according to claim 1, characterized in that, The on-chip hardware AI accelerator includes multiple processing units, each of which is used to detect the micro-motion features of a target within a range gate.

9. A micro-motion sensing method, characterized in that, The method is applied to the on-chip hardware AI accelerator integrated into a single chip in the micro-motion sensing system as described in any one of claims 1-8, and the method includes: Receives the processed range image sent by the on-chip digital signal processor; The micro-motion features of the target within each range gate in the range image are detected, wherein each range gate corresponds to a spatial detection area.

10. The method according to claim 9, characterized in that, The detection of the micro-motion features of the target within each range gate in the range image includes: For each range gate, determine the data points representing the detected target within each range image frame; Calculate the instantaneous phase of the data points; Arrange the instantaneous phases of the data points corresponding to the range gate in each frame of range image in chronological order to form a phase sequence; The micro-motion characteristics of the target are obtained based on the phase sequence.

11. The method according to claim 10, characterized in that, The step of determining the data points representing the detected target within the range gate from each frame of range image includes: Extract all data points within the distance gate from each frame of the distance image; The peak clustering algorithm or the weighted average algorithm of signals from multiple strong reflection points is used to determine the data points representing the detection target from all the data points.

12. The method according to claim 10, characterized in that, The step of obtaining the micro-motion characteristics of the target based on the phase sequence includes: The number of detection targets is determined based on the phase sequence; If the number of detected targets is one, the phase sequence is input into the neural network model, and the micro-motion features of the detected target are extracted through the neural network model; If the number of detected targets is at least two, the phase sequence is divided into at least two independent phase sequences using variational mode decomposition, blind source separation algorithm or empirical mode decomposition, and each independent phase sequence is input into the corresponding neural network model, and the micro-motion features of the detected targets are extracted through the neural network model.

13. The method according to claim 9, characterized in that, The micro-motion features include the user's vital sign data. After detecting the micro-motion features of the target within each distance gate, the method further includes: The user is monitored for abnormalities based on the vital signs data.

14. The method according to claim 9, characterized in that, The micro-motion characteristics include the vibration frequency of industrial equipment. After detecting the micro-motion characteristics of the target within each distance gate, the method further includes: The industrial equipment is monitored for abnormal conditions based on the vibration frequency.

15. An electronic device, characterized in that, It includes a processor and a memory, the memory storing computer-readable instructions that, when executed by the processor, perform the method as described in any one of claims 9-14.

16. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it performs the method as described in any one of claims 9-14.

17. A computer program product, characterized in that, It includes computer program instructions, which, when read and executed by a processor, perform the method as described in any one of claims 9-14.

Images