Sensing method and system based on millimeter radar waves

Abstract

The invention relates to the technical field of radar sensing, and discloses a sensing method and system based on millimeter radar waves. The method comprises the steps that millimeter wave radar receiving signals are subjected to low-noise amplification, frequency bands are screened through a band-pass filter and then mixed with local oscillation signals to obtain distance and speed information, after analog-to-digital conversion digitization, features are extracted to distinguish a human body target in a closed / open space, and finally multi-scene adaptability analysis is executed to generate human body position state information. According to the invention, through construction of a signal processing link, from signal amplification, filtering, frequency mixing and digital processing to target identification and multi-scene adaptability analysis, comprehensive perception of a human body target is realized, a static human body and a non-human body interference source can be effectively distinguished, a processing strategy is automatically adjusted according to different space environment characteristics, and the processing efficiency is improved. And the accuracy and reliability of human body perception are greatly improved.

Description

Technical Field

[0001] The present application relates to the field of radar sensing technology, and in particular to a sensing method and system based on millimeter radar waves. Background Art

[0002] In the modern intelligent sensing field, accurate detection of human presence and activity is crucial for applications such as smart homes, security monitoring, and business analytics. Traditional human sensing technologies primarily rely on infrared sensing, visual recognition, ultrasound, and LiDAR. While inexpensive, infrared sensors are susceptible to ambient temperature and lighting, making them incapable of distinguishing between stationary individuals and the background. Visual recognition systems are limited by lighting conditions, with performance degrading in low-light environments and posing privacy concerns. Ultrasonic technology has poor penetration and is insensitive to even the slightest movement. While LiDAR offers high accuracy, its bulk and high cost make it unsuitable for homes and small businesses. These traditional solutions often exhibit unstable detection performance under varying environmental conditions, particularly struggling to accurately detect stationary individuals and adapt to diverse spatial scenarios.

[0003] While existing millimeter-wave radar technology offers advantages in penetration, anti-interference, and privacy protection, it mostly uses a unified signal processing model and is unable to adaptively optimize for the characteristics of different spatial environments. The reflection characteristics and multipath effects of radar signals differ significantly in confined and open spaces, requiring different processing strategies. Furthermore, existing radar systems primarily focus on detecting moving targets and have limited ability to recognize stationary humans, especially in complex environments where it is difficult to distinguish between stationary humans and stationary objects. Furthermore, the signal processing chain lacks overall optimization, resulting in poor performance in noise suppression, signal enhancement, and feature extraction. This ultimately impacts the accuracy and stability of human perception and fails to meet the demand for refined perception of human status in intelligent environments. Summary of the Invention

[0004] This application provides a perception method and system based on millimeter radar waves, which is used to achieve comprehensive perception of human targets by building a signal processing chain, from signal amplification, filtering, mixing, digital processing to target recognition and multi-scenario adaptive analysis. It can effectively distinguish between stationary human bodies and non-human interference sources, and automatically adjust the processing strategy according to the characteristics of different spatial environments, greatly improving the accuracy and reliability of human perception.

[0005] In the first aspect, the present application provides a perception method based on millimeter radar waves, which includes: performing signal enhancement on the reflected signal received by the millimeter wave radar array through a low-noise amplifier to obtain an amplified radar echo signal; performing frequency band filtering on the amplified radar echo signal through a bandpass filter to obtain a band-limited echo signal; performing a mixing operation on the band-limited echo signal with a local oscillation signal to obtain an intermediate frequency signal containing distance information and speed information; performing digitization on the intermediate frequency signal through an analog-to-digital converter to obtain an original data stream; performing signal processing and feature extraction on the original data stream to distinguish the characteristics of human targets in confined spaces and open spaces to obtain human target recognition data; performing multi-scenario adaptive analysis based on the human target recognition data to generate human position and status information to obtain human perception results.

[0006] In a second aspect, the present application provides a millimeter radar wave-based perception system, the millimeter radar wave-based perception system comprising: An enhancement module is used to enhance the reflected signal received by the millimeter-wave radar array through a low-noise amplifier to obtain an amplified radar echo signal; a screening module, configured to perform frequency band screening on the amplified radar echo signal through a bandpass filter to obtain a band-limited echo signal; A frequency mixing module, configured to perform a frequency mixing operation on the band-limited echo signal and a local oscillation signal to obtain an intermediate frequency signal containing distance information and speed information; A processing module, configured to digitize the intermediate frequency signal through an analog-to-digital converter to obtain an original data stream; a processing module, configured to perform signal processing and feature extraction on the raw data stream, distinguish the features of human targets in confined spaces and open spaces, and obtain human target recognition data; The analysis module is used to perform multi-scene adaptive analysis based on the human target recognition data, generate human position and status information, and obtain human perception results.

[0007] In a third aspect, a millimeter radar wave-based perception device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the millimeter radar wave-based perception device executes the above-mentioned millimeter radar wave-based perception method.

[0008] In a fourth aspect, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium. When the computer-readable storage medium is run on a computer, the computer executes the above-mentioned millimeter radar wave-based perception method.

[0009] In the technical solution provided in this application, high-precision human body perception function is achieved by constructing a millimeter-wave radar signal processing link. The reflected signals received by the millimeter-wave radar array are amplified through a low-noise amplifier, effectively improving the signal-to-noise ratio and addressing the challenge of detecting weak signals. Bandpass filtering through a bandpass filter enables precise extraction of valid signals, suppresses out-of-band interference sources, and ensures signal purity. Subsequently, by mixing with a local oscillator signal, the high-frequency signal is converted into an intermediate-frequency signal containing distance and velocity information. Analog-to-digital conversion then digitizes the signal, ensuring flexibility and accuracy in signal processing. Furthermore, during signal processing and feature extraction of the raw data stream, the system distinguishes between human targets in confined and open spaces. In particular, through extensive feature parameter optimization training using a support vector machine classifier, the system significantly improves human target recognition accuracy, achieving a breakthrough in distinguishing between stationary humans and non-human interference sources. The algorithm's feature vector construction and classification boundary optimization contribute significantly to the scheme's recognition accuracy. Finally, multi-scenario adaptability analysis, performed based on human target recognition data, enables intelligent adaptation to diverse environmental conditions. In particular, it automatically optimizes parameters in confined and open spaces, addressing the limited adaptability of traditional radar systems due to fixed parameters. The overall solution builds a complete processing system from signal reception to feature recognition to scene adaptation, achieving high-precision and high-reliability human perception in complex environments, overcoming technical defects of traditional technologies such as strong dependence on the environment, weak ability to recognize stationary people, and poor adaptability to multiple scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0011] Figure 1 This is a schematic diagram of an embodiment of a sensing method based on millimeter radar waves in an embodiment of the present application; Figure 2 This is a schematic diagram of an embodiment of a perception system based on millimeter radar waves in an embodiment of the present application; Figure 3 It is a schematic block diagram of the structure of a sensing device based on millimeter radar waves in an embodiment of the present invention. DETAILED DESCRIPTION

[0012] The embodiments of the present application provide a perception method and system based on millimeter radar waves. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0013] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In one embodiment of the present application, a method for sensing based on millimeter radar waves includes: Step S101: Amplify the reflected signal received by the millimeter-wave radar array through a low-noise amplifier to obtain an amplified radar echo signal; Step S102: performing frequency band screening on the amplified radar echo signal through a bandpass filter to obtain a band-limited echo signal; Step S103: performing a mixing operation on the band-limited echo signal and the local oscillation signal to obtain an intermediate frequency signal containing distance information and speed information; Step S104: digitize the intermediate frequency signal through an analog-to-digital converter to obtain an original data stream; Step S105: performing signal processing and feature extraction on the original data stream to distinguish the characteristics of human targets in the confined space and the open space, and obtaining human target recognition data; Step S106: Perform multi-scenario adaptability analysis based on the human target recognition data, generate human position and status information, and obtain human perception results.

[0014] It is understandable that the execution subject of this application can be a perception system based on millimeter radar waves, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0015] Specifically, the reflected signal received by the millimeter-wave radar array is amplified by a low-noise amplifier (LNA), producing an amplified radar echo signal. Multiple receiving antenna elements in the millimeter-wave radar array simultaneously receive target reflections. These receiving antennas are typically arranged in a linear or planar array. The weak signal strength captured by each receiving element typically ranges from -80dBm to -100dBm. These weak signals are pre-amplified by a multi-stage cascaded LNA, whose noise figure is controlled below 2.5dB and whose gain is set to 25dB, ensuring effective signal amplification while minimizing noise introduction. In practical applications of the HLK-HE2012, the amplified 60GHz radar received signal is capable of effectively detecting the presence of people within doorframes, while the XM-HE203's 24GHz signal amplification is suitable for wider-area indoor spatial perception.

[0016] The amplified radar echo signal is filtered through a bandpass filter to obtain a band-limited echo signal. The amplified radar echo signal is then input into a multi-order bandpass filter with a passband range set to ±125MHz of the center frequency and a stopband attenuation greater than 40dB. For example, the XM-HE203's 24GHz radar has a sweep bandwidth of 250MHz. After passing through the bandpass filter, only the valid signal between 23.875GHz and 24.125GHz is retained, effectively suppressing out-of-band noise and interference. The filtered signal is then subjected to in-band ripple compensation to keep the amplitude response fluctuation within the passband within 0.5dB, ensuring frequency response flatness.

[0017] The band-limited echo signal is mixed with a local oscillator signal to generate an intermediate frequency (IF) signal containing distance and velocity information. The sinusoidal reference signal generated by the local oscillator is divided into I and Q quadrature reference signals with a 90-degree phase difference. The band-limited echo signal and these quadrature reference signals are input into two mixers for downconversion, generating baseband signals containing in-phase and quadrature components. For the HLK-HE2012 60GHz signal, the mixed IF signal frequency is proportional to the target distance, generating an IF of approximately 1.67MHz for every 1m of distance, enabling the identification of a person inside or outside the door frame. The IF signal is digitized using an analog-to-digital converter (ADC) to generate a raw data stream. The IF signal is preconditioned by a variable gain amplifier, dynamically adjusting its amplitude to match the ADC's input range. It is then sampled using a 12-bit ADC at a 5MHz sampling rate to generate a discrete digital sequence. These digital sequences are organized according to a frame structure, with each frame length set to 1024 points and an overlap rate of 50% between adjacent frames, forming a framed data structure to ensure the continuity and time resolution of signal processing.

[0018] Signal processing and feature extraction are performed on the raw data stream to distinguish the characteristics of human targets in confined and open spaces, generating human target recognition data. This step performs a fast Fourier transform on the raw data stream for distance domain analysis with 1024 points, generating a range spectrum. Based on the range spectrum, Doppler processing is performed on multiple frames of data to calculate the target's velocity distribution at different range gates, generating a two-dimensional range-velocity map. In the XM-HE203 application, the range resolution is 0.75m, and the two-dimensional range-velocity map can accurately distinguish between stationary and moving human targets within a range of 9m.

[0019] Based on human target recognition data, multi-scenario adaptive analysis is performed to generate human position and status information, resulting in human perception results. This step automatically distinguishes between confined and open spaces, distinguishing between environmental types by analyzing the multipath characteristics and spatial scattering model of the reflected signal. For applications where the HLK-HE2012 is installed on a door frame, the system can accurately identify the presence of humans within an 80×40×150cm area within the door frame, and uses infrared sensing to detect human obstruction at close range. For ceiling-mounted or wall-mounted XM-HE203 installations, the system automatically adjusts detection parameters based on the mounting height, creating a sensing area with a ±75-degree detection angle suitable for ceiling-mounted installation.

[0020] For example, when the HLK-HE2012 is mounted on an office doorframe, the millimeter-wave radar signal is first amplified and filtered, then converted through mixing into an intermediate-frequency signal containing distance and velocity information. After analyzing these signals, the system can accurately distinguish human targets within the doorframe area, preventing false triggering outside the doorframe. When a person enters the doorframe area, the system detects their presence within 0.1 seconds, illuminating the indicator light and simultaneously switching the output voltage from 5V to 0V, triggering the lighting or access control system. Similarly, when the XM-HE203 is mounted on the ceiling of a conference room, it can dynamically adjust its sensing sensitivity based on occupant distribution, accurately detecting human presence within a 9-meter range, supporting intelligent lighting control and energy-saving equipment management.

[0021] In the embodiment of the present application, a high-precision human body perception function is achieved by constructing a millimeter-wave radar signal processing link. The reflected signals received by the millimeter-wave radar array are amplified through a low-noise amplifier, effectively improving the signal-to-noise ratio and addressing the challenge of detecting weak signals. Bandpass filtering through a bandpass filter enables precise extraction of valid signals, suppresses out-of-band interference sources, and ensures signal purity. Subsequently, by mixing with a local oscillator signal, the high-frequency signal is converted into an intermediate-frequency signal containing distance and velocity information. Analog-to-digital conversion then digitizes the signal, ensuring flexibility and accuracy in signal processing. Furthermore, during signal processing and feature extraction of the raw data stream, the system distinguishes between human targets in confined and open spaces. In particular, through extensive feature parameter optimization training using a support vector machine classifier, the system significantly improves human target recognition accuracy, achieving a breakthrough in distinguishing between stationary humans and non-human interference sources. The algorithm's feature vector construction and classification boundary optimization contribute significantly to the scheme's recognition accuracy. Finally, multi-scenario adaptability analysis, performed based on human target recognition data, enables intelligent adaptation to diverse environmental conditions. In particular, it automatically optimizes parameters in confined and open spaces, addressing the limited adaptability of traditional radar systems due to fixed parameters. The overall solution builds a complete processing system from signal reception to feature recognition to scene adaptation, achieving high-precision and high-reliability human perception in complex environments, overcoming technical defects of traditional technologies such as strong dependence on the environment, weak ability to recognize stationary people, and poor adaptability to multiple scenes.

[0022] In a specific embodiment, the process of executing step S101 may specifically include the following steps: Multiple receiving antenna units in the millimeter wave radar array simultaneously receive the target reflected echo to obtain multiple original received signals; The multi-channel original received signals are pre-amplified by a multi-stage cascade low-noise amplifier to obtain a primary amplified signal; Perform phase calibration on the primary amplified signal to eliminate the phase differences between the antenna array elements and obtain amplified signals with consistent phases; The gain factor of the phase-matched amplified signal is automatically adjusted according to the target distance. The gain is reduced for close targets and increased for distant targets, thus obtaining a signal with dynamic range compensation. The dynamic range compensated signals are converged and synthesized through a multi-channel parallel processing structure, and an amplitude weighted summation method is adopted to obtain a synthetic signal after signal enhancement; Temperature compensation is performed on the composite signal after signal enhancement, and the amplifier operating point is automatically adjusted according to the ambient temperature change to obtain the amplified radar echo signal.

[0023] Specifically, the primary step in signal processing is for multiple receiving antenna elements to simultaneously receive target reflected echoes. Millimeter-wave radar arrays typically consist of multiple antenna elements arranged in a 4×4 or 8×8 configuration, with each element responsible for receiving reflected signals within a specific angular range. For example, the HLK-HE2012 60GHz radar utilizes a 4×4 antenna array, with each element covering an angular range of approximately 20 degrees, resulting in a detection range of ±75 degrees. When a human target enters the radar detection area, the transmitted radar signal scatters from the human surface, with some of the reflected signal captured by the receiving antenna. Since the HLK-HE2012 is primarily mounted on door frames for detecting entry and exit, its antenna array is optimized for the 80×40×150cm doorframe area, ensuring that other areas beneath the doorframe are not triggered, effectively preventing false detections. The received multi-path raw signals are extremely weak, typically in the -80dBm to -100dBm range. Direct processing would result in the signals being drowned out by ambient noise. Therefore, it is crucial to pre-amplify these weak signals using a multi-stage cascaded low-noise amplifier. The XM-HE203 utilizes a low-noise amplifier with a noise figure below 2.5dB and a gain setting of 25-30dB. A three-stage cascade amplification structure boosts signal strength to a sufficient level for processing. This multi-stage amplification structure maintains a low noise figure while providing sufficient gain, boosting the original signal from microvolts to millivolts, resulting in a primary amplified signal.

[0024] Phase calibration of the primary amplified signal is a critical step in ensuring accurate directionality of the antenna array. Due to the physical location differences between the receiving elements in the millimeter-wave radar array and the non-uniform electrical characteristics of the receiving channels, phase differences can occur in the received signals. The phase calibration process begins by transmitting a standard reference signal to a standard reflector at a known position, then measuring the phase response of each receiving channel. For example, during factory calibration of the XM-HE203, the phase deviation values ​​of each channel are recorded to form a calibration coefficient table. These calibration coefficients are then applied to each channel signal in real time during actual use, eliminating phase differences between the antenna array elements and achieving phase-aligned amplified signals.

[0025] Automatically adjusting the gain coefficient of a phase-aligned amplified signal based on target distance is a core technology for addressing radar signal attenuation with distance. Millimeter-wave radar signals propagate in space according to the inverse square law, with signal strength rapidly attenuating with increasing distance. The HLK-HE2012 and XM-HE203 employ a dynamic gain control strategy to automatically adjust the gain coefficient based on distance information. Specifically, the gain coefficient is proportional to the square of the distance. When the target distance doubles, the gain coefficient increases by approximately 4x (12dB), corresponding to compensation for the inverse square of the distance. In practical applications, a lower gain setting is used for the short-range gate (0.75m to 2.25m) to avoid signal saturation, while a higher gain setting is used for the long-range gate (6m to 9m) to ensure sufficient signal strength. This dynamic range compensation mechanism enables the radar to maintain relatively consistent sensitivity across the entire detection range of 1m to 9m. Aggregating and synthesizing the dynamically range-compensated signals through a multi-channel parallel processing architecture is a key approach to improving signal quality. In the XM-HE203's 24GHz millimeter-wave radar, the signal quality of each receiving channel after the aforementioned processing varies. The multi-channel parallel processing architecture evaluates the signal-to-noise ratio and signal stability of each channel, assigns different weights, and synthesizes the signals using an amplitude-weighted summation method. Channels with better signal quality receive higher weights, while those with lower signal quality receive lower weights. This ensures that the synthesized signal maximizes the preservation of useful information while suppressing noise interference. Furthermore, if certain channels experience temporary faults or anomalies, this method automatically reduces their weights or even completely eliminates them, enhancing system robustness and producing an enhanced synthesized signal.

[0026] Temperature compensation is performed on the enhanced composite signal. The performance of millimeter-wave radar's internal electronic components, particularly the low-noise amplifier (LNA), can vary significantly with ambient temperature. The XM-HE203's built-in temperature sensor monitors the module's operating temperature in real time. This temperature data is combined with a pre-calibrated temperature-gain response curve to calculate the gain change caused by temperature drift. At room temperature (25°C), the standard operating point is set. For every 10°C increase in temperature, the gain typically decreases by approximately 1-1.5dB; conversely, for every 10°C decrease in temperature, the gain increases by approximately 1-1.5dB. The temperature compensation algorithm automatically adjusts the amplifier's bias current and voltage based on the real-time measured temperature value, maintaining a stable overall gain in the radar receive chain, ultimately producing a stable and reliable amplified radar echo signal.

[0027] For example, when the XM-HE203 is mounted on the ceiling of an office, its 24GHz millimeter-wave radar array continuously transmits frequency-modulated continuous wave (FMCW) signals and receives reflected waves. When a person enters the office, the reflected signals are simultaneously captured by eight receiving antenna elements, generating eight original received signals. These weak signals are amplified by approximately 25dB using a three-stage cascaded low-noise amplifier, boosting the signal strength from approximately -90dBm to -65dBm. The amplified signals are then corrected using a factory-set phase calibration table to eliminate ±3-degree phase variations. The system then detects a target at approximately 5 meters and automatically adjusts the gain factor to approximately 7 times the close-range value to compensate for distance loss. The eight compensated signals are weighted based on the real-time signal-to-noise ratio (SNR), with the three highest-quality channels receiving 70% of the total weight, resulting in a combined high-quality signal. Because the office temperature was 18°C, below the standard value, the system reduced the amplifier bias current by approximately 5% to achieve temperature compensation. This results in a stable and accurate amplified radar echo signal, enabling accurate identification of occupant location and activity, and intelligent lighting control.

[0028] In a specific embodiment, the process of executing step S102 may specifically include the following steps: performing multi-order bandpass filtering on the amplified radar echo signal to obtain a preliminary filtered signal; The in-band ripple compensation is performed on the preliminary filtered signal, and an equal ripple design is used to make the amplitude response fluctuation within the passband less than 0.5dB, obtaining a flat frequency response signal; The flat frequency response signal is filtered out with an adaptive notch filter to remove fixed frequency interference, directionally eliminate power supply clutter and system self-coupling signals, and obtain an interference-suppressed signal; The interference suppression signal is input into a digital tunable filter, and the filter bandwidth is dynamically adjusted according to the real-time detected signal-to-noise ratio to obtain a signal with optimized bandwidth; Perform group delay equalization on the signal with optimized bandwidth to ensure that the group delay deviation of the filter within the passband is less than 10 nanoseconds, thus obtaining a phase-linear filtered signal. The phase-linear filtered signal is subjected to digital filtering post-processing, and the burst interference is suppressed by applying sliding average and median filtering to obtain a band-limited echo signal.

[0029] Specifically, in millimeter-wave radar systems, multi-order bandpass filtering typically employs a 5- to 7-order Chebyshev filter structure. This structure is characterized by constant ripple within the passband and steep attenuation in the stopband. Taking the XM-HE203 24GHz radar as an example, the passband range of the bandpass filter is set to ±125MHz of the center frequency, from 23.875GHz to 24.125GHz. This bandwidth matches the 250MHz frequency modulation bandwidth of the radar transmit signal. The filter's stopband attenuation is required to be greater than 40dB to effectively suppress out-of-band frequency components. During the filtering process, the amplified radar echo signal passes through the filter circuit, retaining only the significant frequency components within the passband and filtering out out-of-band noise and interference to produce a preliminary filtered signal. In-band ripple compensation is a key technology for addressing the issue of amplitude unevenness within the bandpass filter's passband. Although the Chebyshev filter has a steep transition band, it also exhibits ripple within the passband, compromising the signal's spectral integrity. In-band ripple compensation utilizes an equal-ripple design approach. By precisely calculating the parameters of each filter unit during filter design, the amplitude response fluctuation within the passband is controlled within 0.5dB. Specifically, after preliminary spectrum analysis of the filtered signal, the amplitude response values ​​at different frequencies within the passband are obtained. Based on these response values, an anti-characteristic compensation network is designed. This network provides gain at low-amplitude frequencies and attenuation at high-amplitude frequencies, thereby achieving a flat frequency response signal. This compensation is particularly important in the HLK-HE2012 product because the component parameters in the 60GHz band are highly discrete, and the amplitude fluctuation within the uncompensated filter passband can reach 2-3dB.

[0030] Adaptive notch filters filter out fixed-frequency interference, a delicate process for improving signal quality. In millimeter-wave radar systems, common fixed-frequency interference sources include power supply noise (typically 50Hz or 60Hz and its harmonics), digital circuit clock signals (such as crystal oscillator frequencies), and self-coupled signals between the radar's transmit and receive channels. The adaptive notch filter first uses spectrum analysis to identify fixed-frequency interference within a flat-frequency response signal. It then sets narrowband notches at these frequencies, precisely matching the center frequency of the notch to the interference frequency. The bandwidth is set to 1.2-1.5 times the interference signal's bandwidth, and the attenuation depth is typically 20-30dB. For the XM-HE203, particular attention is paid to system self-coupled signals around 250kHz. A precisely designed notch filter attenuates these signals below the noise level, resulting in a signal that suppresses interference.

[0031] The core of adaptive processing is the digitally adjustable filter, which dynamically adjusts its bandwidth based on the real-time signal-to-noise ratio. After the interference-suppressing signal is input into the digitally adjustable filter, the signal-to-noise ratio of the current signal is first calculated. This is done by taking the logarithm of the ratio of signal power to noise power and multiplying it by 10 to obtain the dB value. Based on the calculated signal-to-noise ratio, the filter bandwidth is dynamically adjusted: when the signal-to-noise ratio is high (e.g., greater than 20dB), the filter bandwidth is appropriately expanded to improve distance and velocity resolution; when the signal-to-noise ratio is low (e.g., less than 10dB), the filter bandwidth is reduced to improve noise immunity. In actual applications, the XM-HE203 has a bandwidth adjustment range of 50MHz to 200MHz. Through real-time calculation and adaptive adjustment, it maintains optimal detection performance under different environmental conditions and obtains a signal with optimized bandwidth.

[0032] Group delay equalization ensures linear phase response within the filter's passband, a fundamental technology for ensuring accurate distance measurement. Group delay refers to the time delay variation experienced by different frequency components as a signal passes through a filter. Ideally, all frequency components within the passband should have the same group delay. However, practical filters often exhibit group delay distortion at the passband edges, leading to phase nonlinearity. Group delay equalization, by adding a compensation network to the signal processing chain, limits group delay deviation within the passband to less than 10 nanoseconds. For the 24 GHz XM-HE203, this translates to a distance measurement error of approximately 0.3 meters. Group delay equalization reduces this error to less than 0.03 meters, significantly improving distance measurement accuracy and producing a phase-linear filtered signal. Digital post-processing uses sliding average and median filtering to suppress burst interference. Even in a phase-linear filtered signal, some burst interference may still exist, such as spikes generated by motor startup and shutdown, switching power supply transients, and so on. Sliding average filtering smoothes the signal by taking the arithmetic mean of multiple consecutive sample points. The processing window is typically set to 5-9 sample points. The median filter replaces the current sample value with the median value from the sorted data sample sequence, effectively removing sudden interference while preserving the signal's edge characteristics. In practice, the median filter is first applied to remove spikes, and then a sliding average filter is applied to smooth the signal overall, ultimately yielding a high-quality, band-limited echo signal.

[0033] For example, when the XM-HE203 is installed above a shopping mall entrance, it detects people entering and exiting. When a customer enters the detection area, the radar receives the echo signal, which first undergoes the aforementioned amplification processing and then enters a multi-stage bandpass filter. Due to electromagnetic interference generated by various electrical devices in the mall environment, the filter retains the signal's significant frequency components (23.875-24.125 GHz) while simultaneously filtering out 40 dB of out-of-band interference. Through in-band ripple compensation, the amplitude response flatness across the entire passband is improved to within ±0.25 dB. Next, an adaptive notch filter identifies and eliminates 156 kHz periodic interference from the nearby elevator control cabinet and 8 MHz interference generated by its own clock. The system monitors the ambient signal-to-noise ratio in real time. During low-traffic hours (such as early morning), when the signal-to-noise ratio is high (approximately 25 dB), the filter bandwidth automatically expands to 180 MHz. During high-traffic hours (such as midday), when the signal-to-noise ratio decreases (approximately 15 dB), the bandwidth automatically decreases to 100 MHz. After group delay equalization, the system maintains distance measurement accuracy within ±2cm for targets within a 40-meter range. A combination of 7-point median filtering and 5-point sliding average filtering effectively removes spike interference generated when the mall lighting system starts and stops, providing high-quality, band-limited echo signals for subsequent human detection and crowd counting.

[0034] In a specific embodiment, the process of executing step S103 may specifically include the following steps: The sinusoidal reference signal generated by the local oscillator is divided into I and Q paths through a frequency dividing circuit with a phase difference of 90 degrees to obtain an orthogonal reference signal; The band-limited echo signal and the orthogonal reference signal are input into two orthogonal mixers respectively, and a down-conversion operation is performed to obtain a baseband signal containing in-phase and quadrature components; The baseband signal is filtered out of high-frequency mixing products through a low-pass filter with a cutoff frequency set to 500kHz to obtain a pure I / Q signal; Perform quadrature error correction on the pure I / Q signal to eliminate the amplitude imbalance and phase non-orthogonality between the I / Q channels and obtain the corrected quadrature signal; Perform DC offset compensation on the corrected orthogonal signal to eliminate the DC component introduced by the mixer and amplifier to obtain a zero-mean baseband signal; The zero-mean baseband signal is converted into an analytical signal through Hilbert transform, and the instantaneous amplitude and phase information are extracted to obtain an intermediate frequency signal containing distance and speed information.

[0035] Specifically, a sinusoidal reference signal generated by a local oscillator is divided into an I-path and a Q-path via a frequency divider circuit. The local oscillator generates a high-frequency sine wave at the same frequency as the radar transmit signal—60 GHz in the HLK-HE2012 and 24 GHz in the XM-HE203. The frequency divider uses a 90-degree hybrid to split this high-frequency sine wave into two paths: one retains its original phase as the I-path (in-phase component), while the other is phase-delayed by 90 degrees as the Q-path (quadrature component), forming a quadrature reference signal pair. In the XM-HE203, the 90-degree hybrid utilizes a microstrip coupling structure to achieve precise phase control, keeping the phase error within ±0.5 degrees.

[0036] The band-limited echo signal and the quadrature reference signal are fed into two quadrature mixers, respectively, to downconvert the high-frequency radar signal to baseband. The quadrature mixers consist of two independent multipliers: one for processing the I-channel signal and the other for processing the Q-channel signal. The echo signal is multiplied by the I-channel reference signal to produce the in-phase channel output; similarly, the echo signal is multiplied by the Q-channel reference signal to produce the quadrature channel output. The mixing result consists of a sum frequency and a difference frequency component. The difference frequency component is the desired baseband signal, consisting of both in-phase and quadrature components.

[0037] Filtering the baseband signal through a low-pass filter to remove high-frequency mixing products is essential for obtaining a pure baseband signal. During the mixing process, the sum frequency component is approximately twice the original frequency (for example, the sum frequency for a 24GHz radar is approximately 48GHz). This component contains no target information and interferes with subsequent processing. The low-pass filter retains the difference frequency component, which contains target information, while filtering out the sum frequency component. Setting the cutoff frequency to 500kHz means that only signal components within the 0-500kHz range are retained. In FMCW radar, the difference frequency is proportional to target range. For example, with a 250MHz frequency modulation bandwidth and a 1ms sweep period for the XM-HE203, a target at a distance of 9 meters generates a difference frequency of approximately 150kHz, which falls within the 500kHz cutoff frequency and is therefore effectively retained.

[0038] Quadrature error correction for pure I / Q signals is crucial for ensuring measurement accuracy. Ideally, I / Q signals should be equal in amplitude and orthogonal (90 degrees out of phase). However, in real circuits, I / Q imbalance often exists due to component errors and temperature drift. Quadrature error correction employs a matrix transformation approach. First, an error model is established, and the I / Q channel gain difference and inter-channel crosstalk coefficient are measured. For example, in the XM-HE203, the Q channel gain may be 5% lower than the I channel, and inter-channel crosstalk may be approximately 1-2%. Based on these parameters, a correction matrix is ​​constructed. Matrix operations are then used to correct the original I / Q signals, eliminating gain imbalance and phase non-orthogonality, resulting in a corrected quadrature signal.

[0039] Performing DC offset compensation on the corrected quadrature signals is a key step in eliminating inherent system errors. In practical circuits, zero-point offsets in mixers and amplifiers cause DC components in the I / Q signals. These DC components are irrelevant to target information but can introduce errors in distance measurement. DC offset compensation uses a long-term averaging method to estimate the offset, typically taking 1024 or more sampling points to calculate the average value to ensure accurate estimation. In the HLK-HE2012 door frame installation application, analyzing unmanned signal data over a period of one second can calculate an I channel offset of approximately 25mV and a Q channel offset of approximately -18mV. Subtracting these offset values ​​from the original signals yields a zero-mean baseband signal, effectively eliminating false target detection caused by DC offset.

[0040] Constructing an analytic signal from a zero-mean baseband signal through a Hilbert transform is the mathematical key to extracting range and velocity information. For radar signals that already have both I / Q components, the analytic signal can be directly represented as a complex number consisting of the I channel and the Q channel. The amplitude of the analytic signal represents the echo signal strength and is obtained by taking the square root of the sum of the squares of the I and Q components. The phase is calculated by dividing Q by the inverse tangent of I and contains target range and velocity information. In FMCW radar, range information is expressed as the absolute value of the phase, while velocity information is expressed as the rate of change of the phase over time. The target's radial velocity can be determined by calculating the rate of change of the phase, and the absolute value of the phase is proportional to the target's range. By processing the phase information, an intermediate frequency signal containing both range and velocity information is generated.

[0041] For example, the XM-HE203 is installed above a shopping mall aisle for human presence sensing. A 24GHz radar signal is transmitted and reflected by a customer. A local oscillator generates a 24GHz reference signal, which is split into an I-channel signal with a 0-degree phase and a Q-channel signal with a 90-degree phase using a 90-degree mixer. When a customer walks at 0.5 m / s, 4 meters below the radar, the Doppler effect shifts the frequency of the reflected signal to 24GHz plus or minus 40Hz (the frequency shift is proportional to speed). This reflected signal is mixed with the I / Q reference signals to generate I / Q baseband signals with a difference frequency of approximately 66.7kHz (according to FMCW radar principles, this is the difference frequency corresponding to a 4-meter distance under given frequency modulation parameters). These baseband signals are filtered through a low-pass filter with a cutoff frequency of 500kHz to remove high-frequency mixing products and retain the difference frequency signal containing human presence information. Measurements show an I-channel amplitude of 105mV, a Q-channel amplitude of 98mV, and a phase difference of 88.5 degrees. These errors are corrected using an orthogonal error correction matrix. The corrected signal contains DC offsets of +22mV for the I channel and -15mV for the Q channel. By calculating the average of 1024 sampling points and subtracting these offsets, a zero-mean baseband signal is generated. Finally, a complex analytical signal is constructed, and the calculated human reflection intensity is approximately 100mV, with a phase change rate of approximately 24 degrees per second (corresponding to a radial velocity of 0.5 m / s). Incorporating FMCW radar principles, the human distance is calculated to be 4.02 meters from the 66.7kHz difference frequency. This precise human position and movement information enables accurate crowd monitoring and counting.

[0042] In a specific embodiment, the process of executing step S104 may specifically include the following steps: The intermediate frequency signal is pre-conditioned through a variable gain amplifier, and the amplitude range is dynamically adjusted to adapt to the input range of the analog-to-digital converter to obtain a matching signal; The matching signal is input into a 12-bit analog-to-digital converter for sampling, and the sampling rate is set to 5MHz to obtain a discrete digital sequence; Performing interleaved sampling structure reconstruction on the discrete digital sequence, rearranging multiple channel data in time sequence to obtain multi-channel digital data; Analyze and correct quantization errors of multi-channel digital data, apply dithering technology and resampling methods to reduce quantization noise, and obtain digital signals with high signal-to-noise ratio; The high signal-to-noise ratio digital signal is organized into a frame structure through a buffer memory, with each frame length being 1024 points and the overlap rate of adjacent frames being 50%, thereby obtaining framed data; Perform data integrity check and outlier detection on the framed data, mark and process sampling anomalies and data breakpoints, and obtain the original data stream.

[0043] Specifically, the amplitude range of intermediate frequency (IF) signals typically varies widely, from a few millivolts to hundreds of millivolts, while the ADC has a fixed input range, such as ±1V or ±2V. A variable-gain amplifier (VGA) dynamically adjusts the gain factor based on the signal amplitude, ensuring that the output signal amplitude remains within the ADC's optimal operating range. The HLK-HE2012 and XM-HE203 products utilize a digitally controlled VGA with an adjustable gain range from 0dB to 40dB, with 1dB step accuracy. The gain control logic is based on real-time signal peak level measurement. When the signal peak falls below 25% of the ADC's full-scale range, the amplifier gain is increased; when the signal peak exceeds 75% of full-scale, the amplifier gain is decreased. This dynamic adjustment ensures that the signal is neither too small to lose detail nor too large to cause clipping and distortion, ultimately resulting in a signal with a matching amplitude range that is tailored to the ADC.

[0044] The quantization level of a 12-bit analog-to-digital converter is = 4096 levels, meaning the analog signal's amplitude range is evenly divided into 4096 discrete levels, each corresponding to a digital code. The sampling rate is set to 5MHz, meaning the signal is sampled 5 million times per second, with a sampling interval of 0.2 microseconds. According to the Nyquist sampling theorem, a 5MHz sampling rate is sufficient to reconstruct signals with frequencies up to 2.5MHz without distortion. In the XM-HE203, the highest IF signal frequency typically does not exceed 500kHz, so a 5MHz sampling rate provides ample sampling margin. The sampling process converts the continuous analog signal into a digital sequence that is discrete in both time and amplitude, facilitating subsequent digital signal processing.

[0045] Reconstructing the interleaved sampling structure of discrete digital sequences is a key technique for processing multichannel data. In multichannel radar systems, each receiving channel uses an independent analog-to-digital converter for sampling, but the sampling times of each channel may vary slightly. Interleaved sampling structure reconstruction first precisely marks the sampling times of each channel. Then, based on these time stamps, the multichannel data is reordered in strict chronological order. Taking the 8-channel receiving array in the XM-HE203 as an example, the eight channels may be sampled sequentially at intervals of 1 / 8 of a clock cycle. The reconstruction process reorganizes this data according to the actual sampling time sequence, forming a single-channel data stream with an equivalent sampling rate of 40 MHz (5 MHz × 8). Furthermore, data points with coincident or very close sampling times are combined using a weighted average, leveraging multichannel information to improve signal quality and ultimately produce time-aligned multichannel digital data.

[0046] Quantization error analysis and correction for multi-channel digital data is a meticulous task for improving digital signal quality. Quantization error arises from the approximation error caused by mapping the analog signal to a finite number of quantization levels, manifesting as quantization noise superimposed on the useful signal. Quantization error analysis begins by calculating the theoretical quantization noise level of the signal. For a 12-bit converter, quantization noise is theoretically 1 / 4096 of the signal's full-scale range, or approximately 0.024% of full-scale. Dithering is an effective method for reducing quantization noise. Its core concept is to superimpose a small-amplitude, high-frequency random signal (dither signal) on the original signal, causing the signal to cross the quantization threshold multiple times near the critical point. The effects of quantization noise are then reduced by averaging the multiple sampling results. In HLK-HE2012, the dither signal amplitude is set to 1 / 2 of the minimum quantization level, and the frequency is 1 / 4 of the sampling rate, or 1.25 MHz. The resampling method is to interpolate and reconstruct the quantized data, calculate the estimated value of the signal between the original sampling points, increase the number of sampling points through interpolation, and apply low-pass filtering to eliminate high-frequency artifacts introduced by interpolation, further increase the effective number of bits of the signal, reduce the impact of quantization noise, and obtain a high signal-to-noise ratio digital signal.

[0047] Organizing high-SNR digital signals into a frame structure within the buffer memory is a crucial step in data processing. Frames divide a continuous data stream into fixed-length blocks, facilitating subsequent spectrum analysis and feature extraction. In the XM-HE203, each frame is set to 1024 points, equivalent to 0.2048 milliseconds of signal acquisition time (1024 ÷ 5 MHz). A 50% overlap is set between adjacent frames, meaning the last 512 points of one frame are also the first 512 points of the next. This overlapping structure ensures that short-term signal changes at frame boundaries are not missed. The buffer memory utilizes a double-buffering mechanism. While one buffer is receiving new sampled data, the other buffer is being read and analyzed by the processing unit. This enables parallel data acquisition and processing, improving system real-time performance. During the frame data organization process, each frame's timestamp and sequence number are recorded, facilitating subsequent timing analysis and inter-frame correlation, ultimately resulting in structured, framed data.

[0048] Performing data integrity checks and outlier detection on framed data is the final line of defense for ensuring data quality. Data integrity checks first verify that the number of points in each frame meets the expected 1024 points and that the frame header and footer markers are intact, identifying possible data loss or frame boundary misalignment. Outlier detection uses a statistical thresholding method to identify sampling points where the signal amplitude significantly deviates from the normal range. This method calculates the signal's mean and standard deviation, and marks sampling points that deviate from the mean by more than three times the standard deviation as potential outliers. Detected data discontinuities or outliers are processed using interpolation or median filtering. Interpolation estimates a plausible value for the outlier based on the normal data before and after the outlier, while median filtering replaces the outlier with the median value within a window around the outlier. In the HLK-HE2012 door frame installation application, common anomalies include electromagnetic interference pulses generated by elevator startup and shutdown and interference from external wireless devices. The outlier processing algorithm effectively eliminates the impact of these interferences, resulting in a reliable raw data stream.

[0049] For example, the XM-HE203 is installed above an office building corridor to detect human activity. When an employee passes through the corridor, the radar receives an intermediate frequency (IF) signal with an amplitude of approximately 120mV. The optimal input range of the analog-to-digital converter is ±1V. The variable-gain amplifier calculates that a gain of 8.4x (or 18.5dB) is required to amplify the signal to within approximately 1V. The nearest integer gain value, 18dB, is selected, amplifying the signal to approximately 1008mV. The amplified, matched signal is input to a 12-bit analog-to-digital converter (ADC). The sampling rate is set to 5MHz, and the data is stored in two's complement format. The original signal amplitude of 1008mV corresponds to a digital value of approximately 2067 (50.5% of the full-scale range of 4095). For the 8-channel receive array, the sampling times of each channel are staggered by 25ns (one-eighth of the 5MHz sampling rate). The interleaved sampling structure reorders the data in chronological order, forming a data stream with an equivalent sampling rate of 40MHz. Dithering and 4x resampling reduce quantization noise by approximately 5.5dB, improving the signal-to-noise ratio. The reconstructed data is organized into a 1024-point / frame structure using double buffering, with a 50% frame overlap rate to ensure smooth transitions between frames. A data integrity check revealed that three points in a particular frame deviated from the normal value by more than three standard deviations. These outliers were processed using a five-point median filter, replacing the outlier with the median of five points, two before and two after the outlier. This fully processed raw data stream accurately captures the employee's presence and movement characteristics.

[0050] In a specific embodiment, the process of executing step S105 may specifically include the following steps: The original data stream is analyzed in the distance domain by fast Fourier transform, with the number of transformation points being 1024, to obtain the distance spectrum; Doppler processing is performed on multiple frames of data based on the range spectrum to calculate the velocity distribution of the target at different range gates and obtain a two-dimensional range-velocity map. The distance-speed two-dimensional map is input into the adaptive threshold detection algorithm, and the detection threshold is dynamically set according to the background noise level to obtain the target candidate point set; Apply the density clustering algorithm to the target candidate point set, set the cluster distance parameter to 0.5 meters, and the minimum number of points to 3, and obtain the target clustering result; Based on the target clustering results, human body characteristic parameters are extracted, including micro-Doppler characteristics, distance change characteristics, and reflection intensity characteristics, and feature vectors are constructed to obtain a feature database; The feature database is input into the support vector machine classifier to distinguish human targets from non-human interference sources, and simultaneously identify static and moving human bodies to obtain human target recognition data.

[0051] Specifically, the Fast Fourier Transform (FFT) is an algorithm for efficiently computing discrete Fourier transforms, converting a time series into a collection of frequency components. In FMCW radar, the frequency after FFT processing directly corresponds to the target's range information, with higher frequencies representing farther targets. When performing a 1024-point FFT on the raw data stream, a Hanning window function is applied to the data to reduce spectral leakage. The FFT is then performed to produce a 1024-point complex spectrum. Due to the conjugate symmetry of the FFT result, only the first 512 points are used as the effective range spectrum. In the XM-HE203, a 1024-point FFT at a 5MHz sampling rate provides a range resolution of approximately 0.75 meters, resulting in a maximum detection range of 768 meters (512 × 0.75 × 2). However, considering signal attenuation and processing gain, the actual effective detection range is approximately 9 meters. The range spectrum is expressed as a power spectral density, which is the square of the modulus of the FFT result to form a range spectrum reflecting the signal strength distribution at each range gate.

[0052] Doppler processing extracts velocity information by exploiting the frequency shift caused by target motion. Specifically, 128 consecutive frames of range spectrum data are aligned along the range dimension to form a two-dimensional time-range data matrix. A second FFT transform, called the Doppler FFT, is then performed on each of the 128 consecutive frames of data for each range gate (each frequency point in the range spectrum), this time along the time dimension. The frequency of the Doppler FFT result corresponds to the radial velocity of the target, with positive frequencies indicating approach and negative frequencies indicating retreat. By performing the Doppler FFT on all range gates, a two-dimensional data structure containing both range and velocity information is obtained, known as a two-dimensional range-velocity spectrum. In the XM-HE203, 128 frames of data correspond to approximately 26 milliseconds of observation time (128 × 0.2048 milliseconds), resulting in a Doppler resolution of approximately ±1.9 m / s (±38 Hz), capable of effectively distinguishing between stationary and moving subjects, as well as moving targets of varying speeds.

[0053] Inputting the range-velocity 2D map into the adaptive threshold detection algorithm is a key step in initial target identification. Adaptive threshold detection is based on the constant false alarm rate (CFAR) principle, dynamically setting the detection threshold based on the local background noise level rather than using a fixed threshold. In the XM-HE203 implementation, the Cell Average Constant False Alarm Rate (CA-CFAR) algorithm is used. For each cell in the 2D map, the average noise level of surrounding reference cells (typically excluding several immediately adjacent cells as protection cells) is calculated. This average is then multiplied by a multiplication factor to serve as the detection threshold. The multiplication factor is set based on the desired false alarm rate and is typically set between 13-16 dB (4.5-6.3 times). When a cell value exceeds the calculated threshold, the cell is marked as a potential target point. This process iterates through all cells in the entire 2D map, generating a set of potential target points with position and intensity information, known as the target candidate set. The adaptive threshold effectively addresses differences in background noise levels across different regions, improving detection robustness.

[0054] Applying a density clustering algorithm to the set of target candidate points is an important means to integrate scattered target points into meaningful targets. The core idea of ​​density clustering is to group data based on the density of the points and identify high-density areas separated by low-density areas. In this scheme, a density-based spatial clustering algorithm (DBSCAN) is adopted. It is a density-based clustering algorithm that can discover clusters of arbitrary shapes and process noise points. The DBSCAN algorithm requires two key parameters: the cluster distance parameter (Eps) is defined as 0.5 meters, which indicates the maximum distance between two points considered as neighbors; the minimum point number parameter (MinPts) is set to 3, which indicates the minimum number of neighbors required to form a core point. The algorithm process includes: starting from an unprocessed point, finding all its neighboring points within the Eps distance; if the number of neighbors is greater than or equal to MinPts, the point is marked as a core point and forms a cluster with all its neighbors; recursively add the neighbors of the newly added point to the cluster until it cannot be expanded; then select the next unprocessed point and repeat the above process. In this way, the noise points and real target points in the target candidate point set are effectively separated, and a point group representing the actual physical target is obtained, that is, the target clustering result.

[0055] Extracting human characteristic parameters based on target clustering results is a feature engineering step in target recognition. Human targets differ from other static or moving objects primarily in their unique micro-motion and reflection characteristics. Micro-Doppler characteristics refer to the additional frequency modulation generated by the relative motion of various human body parts (such as arms and legs), which manifests as sidebands around the main frequency on the Doppler spectrum. Extracting micro-Doppler characteristics requires performing a short-time Fourier transform (STFT) on the Doppler spectrum of each clustered target to obtain the time-frequency distribution and calculate characteristic parameters such as spectral width, sideband intensity, and modulation period. Range variation characteristics are obtained by tracking the change in distance between target cluster centers between consecutive frames, calculating the rate of change, acceleration, periodicity, and regularity of change. Reflection intensity characteristics include the target's radar cross-section (RCS) estimate, scattering characteristics at different angles of incidence, and polarization characteristics. These characteristic parameters are combined into a multidimensional feature vector, with each dimension corresponding to a characteristic parameter, forming a complete "fingerprint" describing the target's characteristics. In the implementation of XM-HE203, the feature vector contains 15-20 dimensions, forming a feature database that provides a basis for subsequent classification and recognition.

[0056] Inputting the feature database into a support vector machine classifier is the final step in achieving accurate human object recognition. A support vector machine (SVM) is a supervised learning algorithm that achieves classification by finding a hyperplane that separates samples of different classes by the largest margin. In human recognition applications, the SVM classifier is pre-trained using a large amount of labeled data. It can distinguish between human targets and non-human interference sources (such as fans and swaying plants), as well as between stationary and moving objects. The SVM classification process involves feature normalization, kernel function mapping, and classification decision. Feature normalization scales all features to a uniform range to prevent certain features from dominating the classification results due to their large values. The XM-HE203 uses a radial basis function (RBF) as the kernel function to map the original feature space to a higher-dimensional space, enhancing class separability. The classification decision is based on the distance between the sample and the hyperplane. It outputs not only the classification result (human / non-human, stationary / moving) but also the classification confidence level, facilitating decision-making in practical applications. After SVM classification, complete human object recognition data is obtained, including target category, location, velocity, and confidence level.

[0057] For example, the XM-HE203 is installed on the bedroom ceiling, detecting occupant status and controlling smart home devices. When one person sits in bed reading and another stands by the window, the raw radar data stream is first converted into a range spectrum using a 1024-point FFT. The range spectrum shows two prominent reflection peaks at approximately 2 meters and 3.5 meters, corresponding to the positions of the two people. Doppler processing is performed on 128 consecutive frames of the range spectrum, generating a two-dimensional range-velocity map. The target at 2 meters exhibits a very low Doppler shift (approximately ±0.1 m / s), indicating near-stationary motion. The target at 3.5 meters exhibits a significant Doppler shift (approximately ±0.5 m / s), indicating minimal motion. The two-dimensional map is fed into an adaptive threshold detection algorithm, which sets a detection threshold (approximately -64 dBm) based on the calculated background noise level (approximately -80 dBm). This yields approximately 15 target candidate points. These candidate points were processed using the DBSCAN density clustering algorithm, with a cluster distance parameter set to 0.5 meters and a minimum number of points of 3. This ultimately resulted in two distinct clusters, corresponding to the two human targets. Feature parameters were extracted for each cluster, including micro-Doppler characteristics (the micro-Doppler spectrum width for a seated person is approximately 5Hz, and for a standing person approximately 15Hz), distance variation characteristics, and reflection intensity characteristics, to construct a 15-dimensional feature vector. These feature vectors were input into a trained SVM classifier, and the classification results showed that the target at 2 meters was a "stationary person" with a confidence level of 98%, and the target at 3.5 meters was a "moving person" with a confidence level of 95%. Based on this human target recognition data, the smart home system keeps the room lights on and dynamically adjusts the air conditioning temperature and airflow direction based on occupant activity, achieving precise human perception control.

[0058] In a specific embodiment, the process of executing step S106 may specifically include the following steps: Automatically identify human target recognition data in closed spaces and open spaces, distinguish environment types based on the multipath characteristics of the reflected signal and the spatial scattering model, and obtain scene classification results; According to the scene classification results, the corresponding target tracking parameters are selected. A small amount of measurement noise setting and a high correlation threshold are used for confined spaces, and a large amount of measurement noise setting and a low correlation threshold are used for open spaces to obtain scene adaptive parameters. The human target recognition data and scene adaptive parameters are input into the multi-target tracking algorithm to establish target trajectory association and state estimation to obtain target trajectory data; Perform human behavior analysis on the target trajectory data, extract indicators such as residence time, movement speed, and activity area, construct a human behavior feature set, and obtain behavior analysis results; Match the behavior analysis results with the preset scenario behavior pattern library to identify normal and abnormal behaviors and obtain behavior explanation data; The behavior interpretation data is structured and encapsulated to generate a data packet containing the human body position coordinates, motion state, existence duration and behavior type to obtain the human body perception results.

[0059] Specifically, radar echo characteristics differ significantly between confined spaces (such as the doorframe installation scenario of the HLK-HE2012) and open spaces (such as the ceiling installation scenario of the XM-HE203). These characteristics necessitate automatic environment type identification. The identification process first analyzes the multipath characteristics of the reflected signal. Multipath refers to the phenomenon in which a signal reaches the receiver via both a direct path and multiple reflections. In confined spaces, walls, ceilings, and floors strongly reflect radar signals, resulting in the received signal containing numerous secondary and tertiary reflection components. These multipath components appear as secondary peaks near the primary peak on the range spectrum. The multipath feature vector is generated by calculating the intensity ratio of the primary to secondary peaks on the range spectrum (multipath ratio) and the temporal distribution of the secondary peaks (multipath distribution). The spatial scattering model of the signal is then analyzed, and the statistical characteristics of the angular scattering function (ASF) and the delay scattering function (DSF) are extracted to construct the scattering feature vector. The multipath feature vector and the scattering feature vector are combined and input into the pre-trained scene classifier to calculate the probability scores of confined spaces and open spaces. The type with the highest score is selected as the recognition result to obtain the scene classification result.

[0060] Selecting appropriate target tracking parameters based on scene classification results is a key technology for adapting to different environments. The parameter requirements for target tracking algorithms in confined spaces and open spaces differ significantly. In confined spaces, due to the abundance of multipath reflections and the fact that walls restrict the target's range of motion, tracking algorithms are suited to low measurement noise settings. Measurement noise here refers to the random error in the target's position and velocity measurements. In XM-HE203 applications, position measurement noise in confined spaces is typically set to 0.1-0.2 meters, and velocity measurement noise is set to 0.05-0.1 meters per second. Furthermore, confined spaces are also suitable for high correlation thresholds. The correlation threshold is used to determine whether a new measurement point belongs to an existing track. A high correlation threshold requires a higher match confidence level. Typically, the correlation threshold for confined spaces is set to 0.85-0.95, meaning that only measurement points with a matching probability exceeding 85% are associated with an existing track. Conversely, in open space, targets have greater freedom of movement, and distance and angle measurements are more susceptible to environmental influences. Therefore, a large noise setting (position noise 0.3-0.5 meters, velocity noise 0.15-0.25 meters per second) and a low correlation threshold (0.6-0.8) are suitable to accommodate greater measurement uncertainty and more complex motion patterns. Automatic tracking parameter selection is based on scene classification results, smoothly transitioning between closed and open spaces to obtain scene-adaptive parameters.

[0061] The multi-target tracking algorithm employs the Interactive Multi-Model Kalman Filter (IMM-KF) framework, which combines multiple motion models to handle complex human motion. In HLK-HE2012 and XM-HE203, a combination of a uniform velocity (CV) model and a uniform acceleration (CA) model is typically used to handle uniform linear motion and acceleration / deceleration, respectively. The Kalman filter estimates the target state through two cyclic steps: prediction and update. The prediction step predicts the current state based on the previous state and the motion model; the update step corrects the predicted state based on new measurement data. During this process, the measurement noise matrix in the scene adaptation parameters is used to set the filter's observation noise, influencing the weighting of the measurement data. Target trajectory association utilizes a combination of the nearest neighbor algorithm and the probabilistic data association algorithm (PDA), using the association threshold in the scene adaptation parameters to determine the matching criteria between trajectory and measurement. For newly appearing targets, trajectory initialization is performed; for targets that have lost observation, trajectory prediction and termination are performed. Through this series of processes, discrete target detection points are connected to form a continuous motion trajectory, consisting of a time series of position, velocity, and acceleration, resulting in target trajectory data.

[0062] Human behavior analysis of target trajectory data is a key step in uncovering human activity patterns. First, basic temporal features are extracted, such as dwell time (the cumulative duration of a target's stay within a specific area), activity period (the interval between repeated activities), and activity duration. Then, spatial motion features are extracted, including movement speed (the displacement between trajectory points divided by the interval), trajectory curvature (the degree of trajectory curvature), and activity area (the spatial distribution of trajectory points). Specifically for the XM-HE203 application, micro-motion features are specifically extracted to analyze subtle movements of the human body, even when "at rest." This is crucial for distinguishing stationary humans from stationary objects. Micro-motion features are derived by analyzing minute changes in the target's position and subtle Doppler frequency shifts while at rest. Typically, a stationary human body exhibits micro-motion speeds of 0.05-0.2 m / s. In the HLK-HE2012 door frame detection application, crossing characteristics, such as crossing speed, crossing direction, and crossing frequency, are analyzed to help distinguish regular entry and exit from unusual lingering behavior. These extracted features collectively form a human behavior feature set, describing the temporal and spatial distribution patterns of human activity and generating behavioral analysis results.

[0063] Matching behavioral analysis results with a pre-defined library of scene-based behavioral patterns is a crucial step in understanding the semantics of human behavior. The scene-based behavioral pattern library is a collection of predefined behavioral patterns, each containing feature descriptions and semantic labels. In the XM-HE203 application, the library includes common behaviors such as "straight-line walking," "wandering around," "brief stops," "long periods of sitting still," and "sudden falls." The matching process employs a fuzzy logic inference system to calculate the similarity score between the behavioral analysis results and each pattern. For example, the "wandering around" pattern has characteristics such as "medium-low speed (0.3-0.8 m / s)," "high trajectory repetition rate (>60%)," "small activity area (<3 m diameter)," and "frequent changes of direction (>5 times / minute)." When the behavioral analysis results strongly match these characteristics, the behavior is identified as "wandering around." The matching process also considers the context of the behavior, distinguishing between different semantics for similar behaviors in different scenarios. For example, "wandering" in a public area has different semantics than "wandering" in a private area. The matching results include behavior type identification and abnormality assessment, marking abnormal behaviors that do not match scenario expectations and obtaining behavior explanation data.

[0064] Structured encapsulation of behavioral interpretation data is the final step in generating standardized human perception results. Structured encapsulation converts unstructured behavioral interpretation data into a data packet with a unified format and standardized fields. The encapsulation process first determines the data packet format, typically using a JSON or XML structure with a fixed set of fields. Core fields include body position coordinates (three-dimensional spatial coordinates in meters), motion state (velocity vector and acceleration vector), presence duration (accumulated time from first detection to the present), and behavior type (semantic label derived from behavior matching). The XM-HE203 also includes a confidence field (the confidence score for target and behavior recognition) and attribute fields (such as target number estimate and activity intensity). For targets tracked over long periods of time, a compressed summary of historical trajectory and behavior records is generated simultaneously, facilitating analysis of long-term behavioral patterns by higher-level applications. Finally, metadata such as timestamp, device ID, and data version are added to the data packet to ensure data timeliness and traceability, resulting in complete human perception results.

[0065] The above describes the sensing method based on millimeter radar waves in the embodiment of the present application. The following describes the sensing system based on millimeter radar waves in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of a perception system based on millimeter radar waves includes: The enhancement module 201 is used to enhance the reflected signal received by the millimeter wave radar array through a low noise amplifier to obtain an amplified radar echo signal; A screening module 202 is configured to perform frequency band screening on the amplified radar echo signal through a bandpass filter to obtain a band-limited echo signal; A mixing module 203 is configured to perform a mixing operation on the band-limited echo signal and a local oscillation signal to obtain an intermediate frequency signal containing distance information and speed information; The processing module 204 is configured to digitize the intermediate frequency signal through an analog-to-digital converter to obtain an original data stream; The processing module 205 is used to perform signal processing and feature extraction on the raw data stream, distinguish the characteristics of human targets in confined spaces and open spaces, and obtain human target recognition data; The analysis module 206 is used to perform multi-scenario adaptive analysis based on the human target recognition data, generate human position and status information, and obtain human perception results.

[0066] Through the collaborative cooperation of the above components and the construction of a complete millimeter-wave radar signal processing chain, high-precision human body perception function is achieved. The reflected signals received by the millimeter-wave radar array are amplified through a low-noise amplifier, effectively improving the signal-to-noise ratio and addressing the challenge of detecting weak signals. Bandpass filtering through a bandpass filter enables precise extraction of valid signals, suppresses out-of-band interference sources, and ensures signal purity. Subsequently, by mixing with a local oscillator signal, the high-frequency signal is converted into an intermediate-frequency signal containing distance and velocity information. Analog-to-digital conversion then digitizes the signal, ensuring flexibility and accuracy in signal processing. Furthermore, during signal processing and feature extraction of the raw data stream, the system distinguishes between human targets in confined and open spaces. In particular, through extensive feature parameter optimization training using a support vector machine classifier, the system significantly improves human target recognition accuracy, achieving a breakthrough in distinguishing between stationary humans and non-human interference sources. The algorithm's feature vector construction and classification boundary optimization contribute significantly to the scheme's recognition accuracy. Finally, multi-scenario adaptability analysis, performed based on human target recognition data, enables intelligent adaptation to diverse environmental conditions. In particular, it automatically optimizes parameters in confined and open spaces, addressing the limited adaptability of traditional radar systems due to fixed parameters. The overall solution builds a complete processing system from signal reception to feature recognition to scene adaptation, achieving high-precision and high-reliability human perception in complex environments, overcoming technical defects of traditional technologies such as strong dependence on the environment, weak ability to recognize stationary people, and poor adaptability to multiple scenes.

[0067] above Figure 2 The millimeter radar wave-based perception system in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The millimeter radar wave-based perception device in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0068] Figure 3 This is a schematic diagram of the structure of a millimeter radar wave-based perception device provided by an embodiment of the present invention. This millimeter radar wave-based perception device 300 may vary significantly depending on configuration or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors), memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) storing application programs 333 or data 332. The memory 320 and storage medium 330 may be either transient or persistent storage. The program stored in the storage medium 330 may include one or more modules (not shown), each of which may include a series of instructions for operating on the millimeter radar wave-based perception device 300. Furthermore, the processor 310 may be configured to communicate with the storage medium 330, executing the series of instructions stored in the storage medium 330 on the millimeter radar wave-based perception device 300 to implement the steps of the millimeter radar wave-based perception method described above.

[0069] The millimeter radar wave-based sensing device 300 may further include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be understood by those skilled in the art that Figure 3 The structure of the millimeter radar wave-based sensing device shown does not constitute a limitation on the millimeter radar wave-based sensing device provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0070] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the steps of the millimeter radar wave-based perception method.

[0071] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0072] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the existing technology, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a millimeter radar wave-based sensing device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program code.

[0073] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A sensing method based on millimeter radar waves, characterized in that: The method comprises: The reflected signal received by the millimeter-wave radar array is enhanced by a low-noise amplifier to obtain an amplified radar echo signal; Performing frequency band screening on the amplified radar echo signal through a bandpass filter to obtain a band-limited echo signal; Performing a frequency mixing operation on the band-limited echo signal and the local oscillation signal to obtain an intermediate frequency signal containing distance information and speed information; Digitally processing the intermediate frequency signal through an analog-to-digital converter to obtain an original data stream; Performing signal processing and feature extraction on the raw data stream to distinguish the characteristics of human targets in confined spaces and open spaces, and obtaining human target recognition data; Based on the human target recognition data, multi-scene adaptive analysis is performed to generate human position and status information to obtain human perception results.

2. The millimeter radar wave sensing method according to claim 1, characterized in that: The method of performing signal enhancement on the reflected signal received by the millimeter wave radar array through a low noise amplifier to obtain an amplified radar echo signal includes: Multiple receiving antenna units in the millimeter wave radar array simultaneously receive the target reflected echo to obtain multiple original received signals; Pre-amplifying the multi-channel original received signals through a multi-stage cascade low-noise amplifier to obtain a primary amplified signal; Performing phase calibration on the primary amplified signal to eliminate phase differences between the antenna array elements and obtain amplified signals with consistent phases; Automatically adjusting the gain coefficient of the phase-matched amplified signal according to the target distance, reducing the gain for close targets and increasing the gain for distant targets, thereby obtaining a signal with dynamic range compensation; The dynamic range compensated signals are converged and synthesized through a multi-channel parallel processing structure, and an amplitude weighted summation method is adopted to obtain a synthesized signal after signal enhancement; Temperature compensation processing is performed on the synthesized signal after the signal enhancement, and the amplifier operating point is automatically adjusted according to the ambient temperature change to obtain the amplified radar echo signal.

3. The millimeter radar wave sensing method according to claim 1, characterized in that: The step of performing frequency band screening on the amplified radar echo signal through a bandpass filter to obtain a band-limited echo signal comprises: performing multi-order bandpass filtering on the amplified radar echo signal to obtain a preliminary filtered signal; Performing in-band ripple compensation on the preliminary filtered signal, using an equal ripple design to make the amplitude response fluctuation within the passband less than 0.5 dB, thereby obtaining a flat frequency response signal; The flat frequency response signal is subjected to an adaptive notch filter to filter out fixed frequency interference, directionally eliminate power supply clutter and system self-coupling signals, and obtain an interference-suppressed signal; Inputting the interference-suppressed signal into a digitally adjustable filter, dynamically adjusting the filter bandwidth according to the real-time detected signal-to-noise ratio to obtain a signal with optimized bandwidth; Performing group delay equalization on the signal of the optimized bandwidth to ensure that the group delay deviation of the filter within the passband is less than 10 nanoseconds, thereby obtaining a phase-linear filtered signal; Digital filtering post-processing is performed on the phase-linear filtered signal, and sliding average and median filtering are applied to suppress burst pulse interference to obtain the band-limited echo signal.

4. The millimeter radar wave sensing method according to claim 1, characterized in that: The mixing operation of the band-limited echo signal with the local oscillation signal to obtain an intermediate frequency signal containing distance information and speed information includes: The sinusoidal reference signal generated by the local oscillator is divided into I and Q paths through a frequency dividing circuit with a phase difference of 90 degrees to obtain an orthogonal reference signal; Inputting the band-limited echo signal and the orthogonal reference signal into two orthogonal mixers respectively, performing down-conversion operation, and obtaining a baseband signal containing in-phase and quadrature components; The baseband signal is filtered out of high-frequency mixing products through a low-pass filter with a cutoff frequency set to 500 kHz to obtain a pure I / Q signal; Performing quadrature error correction on the pure I / Q signals to eliminate amplitude imbalance and phase non-orthogonality between I / Q channels, thereby obtaining a corrected quadrature signal; Performing DC offset compensation on the corrected quadrature signal to eliminate the DC component introduced by the mixer and the amplifier to obtain a zero-mean baseband signal; The zero-mean baseband signal is converted into an analytical signal through Hilbert transform, and instantaneous amplitude and phase information are extracted to obtain the intermediate frequency signal containing distance information and speed information.

5. The millimeter radar wave sensing method according to claim 1, characterized in that: The digitizing the intermediate frequency signal through an analog-to-digital converter to obtain an original data stream includes: Preconditioning the intermediate frequency signal through a variable gain amplifier, dynamically adjusting the amplitude range to match the input range of the analog-to-digital converter, and obtaining a matching signal; The matching signal is input into a 12-bit analog-to-digital converter for sampling, with a sampling rate set to 5 MHz to obtain a discrete digital sequence; Performing interleaved sampling structure reconstruction on the discrete digital sequence, rearranging the multiple channel data in time sequence, and obtaining multi-channel digital data; Performing quantization error analysis and correction on the multi-channel digital data, applying dithering technology and resampling methods to reduce quantization noise, and obtaining a digital signal with a high signal-to-noise ratio; The high signal-to-noise ratio digital signal is organized in a frame structure through a buffer memory, with each frame length being 1024 points and an overlap rate of adjacent frames being 50%, to obtain framed data; Data integrity check and outlier detection are performed on the framed data, sampling anomalies and data breakpoints are marked and processed, and the original data stream is obtained.

6. The millimeter radar wave sensing method according to claim 1, characterized in that: The signal processing and feature extraction are performed on the original data stream to distinguish the characteristics of human targets in the confined space and the open space to obtain human target recognition data, including: Performing distance domain analysis on the original data stream by fast Fourier transform with 1024 transformation points to obtain a distance spectrum; Performing Doppler processing on multiple frames of data based on the range spectrum, calculating the velocity distribution of the target at different range gates, and obtaining a two-dimensional range-velocity map; Inputting the distance-speed two-dimensional map into an adaptive threshold detection algorithm, dynamically setting the detection threshold according to the background noise level, and obtaining a target candidate point set; Applying a density clustering algorithm to the target candidate point set, setting the clustering distance parameter to 0.5 meters and the minimum number of points to 3, to obtain the target clustering result; Extracting human body characteristic parameters based on the target clustering results, including micro-Doppler characteristics, distance change characteristics, and reflection intensity characteristics, constructing feature vectors, and obtaining a feature database; The feature database is input into a support vector machine classifier to distinguish human targets from non-human interference sources, and simultaneously identify static human bodies and moving human bodies to obtain the human target recognition data.

7. The millimeter radar wave sensing method according to claim 1, characterized in that: The performing multi-scenario adaptive analysis based on the human target recognition data, generating human position and state information, and obtaining human perception results, includes: Automatically identifying the human target recognition data in closed spaces and open spaces, distinguishing the environment types based on the multipath characteristics of the reflected signal and the spatial scattering model, and obtaining a scene classification result; Selecting corresponding target tracking parameters according to the scene classification results, using a small measurement noise setting and a high correlation threshold for a confined space, and a large measurement noise setting and a low correlation threshold for an open space, to obtain scene adaptive parameters; Inputting the human target recognition data and the scene adaptive parameters into a multi-target tracking algorithm, establishing target trajectory association and state estimation, and obtaining target trajectory data; Performing human behavior analysis on the target trajectory data, extracting indicators such as dwell time, movement speed, and activity area, constructing a human behavior feature set, and obtaining behavior analysis results; Matching the behavior analysis results with a preset scenario behavior pattern library to identify normal behaviors and abnormal behaviors to obtain behavior interpretation data; The behavior interpretation data is structured and encapsulated to generate a data packet containing the human body position coordinates, motion state, existence duration and behavior type, and the human body perception result is obtained.

8. A perception system based on millimeter radar waves, characterized in that: For implementing the millimeter radar wave-based perception method according to any one of claims 1 to 7, the millimeter radar wave-based perception system comprises: An enhancement module is used to enhance the reflected signal received by the millimeter-wave radar array through a low-noise amplifier to obtain an amplified radar echo signal; a screening module, configured to perform frequency band screening on the amplified radar echo signal through a bandpass filter to obtain a band-limited echo signal; A frequency mixing module, configured to perform a frequency mixing operation on the band-limited echo signal and a local oscillation signal to obtain an intermediate frequency signal containing distance information and speed information; A processing module, configured to digitize the intermediate frequency signal through an analog-to-digital converter to obtain an original data stream; a processing module, configured to perform signal processing and feature extraction on the raw data stream, distinguish the features of human targets in confined spaces and open spaces, and obtain human target recognition data; The analysis module is used to perform multi-scene adaptive analysis based on the human target recognition data, generate human position and status information, and obtain human perception results.

9. A sensing device based on millimeter radar waves, characterized in that: It includes a memory and a processor, the memory stores a computer program that can be run on the processor, and the processor implements the millimeter radar wave based perception method described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the processor is caused to perform the millimeter radar wave based perception method according to any one of claims 1 to 7.

Images