Low-height millimeter wave radar tumble detection device and method

Abstract

The invention relates to the technical field of radar detection, and discloses a low-height millimeter-wave radar fall detection device, which comprises a multi-antenna array module, a narrow-beam low-elevation radar unit, an anti-interference component and a signal processing module. According to the low-height millimeter-wave radar falling detection method, a signal processing module comprises a dynamic background modeling unit which constructs a Gaussian mixture model through static scanning, carries out pixel-level background differencing in real time and judges a foreground target; and the micro-Doppler feature extraction unit is used for performing short-time Fourier transform on the human body echo signals, generating a time-frequency diagram and extracting micro-Doppler frequency shift features. Compared with the prior art, the method has the advantages that the problem of low detection precision caused by environmental interference when an existing millimeter-wave radar is installed at a low height is solved, accurate recognition of human body tumble at the installation height smaller than or equal to 80 cm is achieved, and meanwhile the method is compatible with a complex indoor environment.

Description

Technical Field

[0001] This invention relates to the field of radar detection technology, specifically to a low-altitude millimeter-wave radar fall detection device and method. Background Technology

[0002] Millimeter-wave radar is widely used in intelligent security and health monitoring due to its strong penetration, excellent anti-interference ability, and immunity to light conditions. Traditional fall detection solutions mainly rely on visual sensors or millimeter-wave radar installed at high altitudes (≥1.8 meters). Visual solutions pose a risk of privacy leakage, while traditional millimeter-wave radar, although able to acquire human posture data from a top-down perspective when installed at high altitudes, suffers from the following problems in low-altitude (≤80 cm) installation scenarios:

[0003] Ground reflections cause strong clutter, which drowns out the human body echo signal;

[0004] Furniture obstructs monitoring, creating blind spots.

[0005] It is difficult to distinguish between low-position movements (such as bending over) and falling postures.

[0006] In existing technologies, millimeter-wave radar installed at low heights suffers from low signal-to-noise ratio, insufficient detection accuracy, and high false detection rate, failing to meet the needs of scenarios such as small apartments and nursing bedside locations. Therefore, there is an urgent need for a device and method capable of achieving high-precision fall detection in low-height installation environments. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to overcome the above-mentioned technical difficulties and provide a low-height millimeter-wave radar fall detection device and method, which solves the problem of low detection accuracy caused by environmental interference when existing millimeter-wave radar is installed at low heights, and achieves accurate identification of human falls at an installation height of ≤80cm, while being compatible with complex indoor environments (such as dense furniture and diverse floor materials).

[0008] To solve the above-mentioned technical problems, the technical solution provided by the present invention is as follows:

[0009] A low-altitude millimeter-wave radar fall detection device includes:

[0010] The multi-antenna array module adopts a 4Tx8Rx MIMO architecture with an antenna spacing of λ / 2 (λ is the millimeter wave wavelength). It is integrated into a rectangular metal cavity with a height of 3cm and dynamically generates a focused beam pointing to the waist height of the human body through a beamforming algorithm.

[0011] The narrow-beam, low-elevation radar unit has a center frequency of 77 GHz, a beam angle of 12°, and a physical installation angle tilted upwards by 15° so that the beam center axis is aligned with the waist of the human body. The detection range is limited to a cylindrical area 0.3-3 meters away from the radar and 0.5-1.8 meters in vertical height.

[0012] The anti-interference component includes a metal shield and a radio frequency filter circuit. The metal shield covers the bottom and two sides of the radar module, leaving only a forward 120° detection window. The radio frequency filter circuit 2 integrates a 50-90GHz bandpass filter.

[0013] The signal processing module is used to perform dynamic background modeling, micro-Doppler feature extraction, 3D pose calculation, and fall detection.

[0014] A low-altitude millimeter-wave radar fall detection method, wherein the signal processing module includes:

[0015] The dynamic background modeling unit constructs a Gaussian mixture model (GMM) through static scanning, performs pixel-level background subtraction in real time, and determines the foreground target;

[0016] The micro-Doppler feature extraction unit performs short-time Fourier transform (STFT) on the human body echo signal to generate a time-frequency diagram and extract micro-Doppler frequency shift features;

[0017] The three-dimensional attitude calculation unit generates three-dimensional point clouds based on virtual aperture synthesis and back projection algorithms of multi-antenna arrays. It extracts joint points through adaptive radius filtering and prior human structure and calculates pitch angle, roll angle and joint angular velocity.

[0018] The fall detection unit uses an SVM classifier to fuse pitch angle, joint angular velocity, and micro-Doppler entropy features for fall detection.

[0019] The advantages of this invention compared to the prior art are:

[0020] 1. The present invention improves the clutter suppression ratio by 25dB through multi-antenna beamforming and metal shield 1;

[0021] 2. The narrow beam and low elevation angle design of this invention reduces ground reflection signal attenuation by ≥40%;

[0022] 3. The 3D point cloud and micro-Doppler fusion algorithm of this invention improves the fall detection accuracy to 98.7%;

[0023] 4. The adaptive background modeling of this invention reduces the false detection rate to less than 2 times / day. Attached Figure Description

[0024] Figure 1 This is a flowchart of the present invention.

[0025] Figure 2 This is a schematic diagram of the structure of the multi-antenna array module of the present invention.

[0026] Figure 3 This is a schematic diagram of the narrow-beam, low-elevation radar unit of the present invention.

[0027] As shown in the figure: 1. Metal shielding cover; 2. Radio frequency filter circuit; 3. Radio frequency filter circuit. Detailed Implementation

[0028] In the description of this invention, it should be understood that the terms "center," "lateral," "upper," "lower," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more. Additionally, the term "comprising" and any variations thereof are intended to cover non-exclusive inclusion.

[0029] The present invention will now be described in further detail with reference to the embodiments and accompanying drawings.

[0030] A low-altitude millimeter-wave radar fall detection device includes:

[0031] Multi-antenna array module: such as Figure 1 As shown, a 4Tx8Rx MIMO architecture is adopted, with an antenna spacing of λ / 2 (λ being the millimeter-wave wavelength), integrated into a rectangular metal cavity 3 with a height of 3cm (dimensions: 3cm×3cm×3cm). A focused beam pointing towards waist height (approximately 90cm) is dynamically generated using beamforming algorithms (such as Capon beamforming) to suppress clutter reflections from the ground (reflectivity > 0.8) and furniture (such as metal table legs).

[0032] Narrow-beam, low-elevation radar unit: Utilizes a millimeter-wave radar chip with a center frequency of 77GHz and a beam angle of 12° (such as the TIIWR1843), with the physical installation angle tilted upwards at 15° (see attached image). Figure 2 (Diagram of radar elevation angle) aligns the central axis of the beam with the waist of the human body, limiting the detection range to a cylindrical area 0.3-3 meters from the radar and 0.5-1.8 meters in vertical height, thus avoiding direct ground reflection;

[0033] Anti-interference components:

[0034] Metal shield 1: A 1mm thick aluminum shield is used to cover the bottom and two sides of the radar module, leaving only the forward 120° detection window, which attenuates the reflected signal from the ground by ≥20dB.

[0035] RF filtering circuit 2: Integrates a 50-90GHz bandpass filter (insertion loss <3dB) in the receiving link to filter out narrowband interference such as Wi-Fi (2.4 / 5GHz) and Bluetooth (2.4GHz);

[0036] The signal processing module is used to perform dynamic background modeling, micro-Doppler feature extraction, 3D pose calculation, and fall detection.

[0037] A low-altitude millimeter-wave radar fall detection method, wherein the signal processing module includes:

[0038] The dynamic background modeling unit constructs a Gaussian mixture model (GMM) through static scanning, performs pixel-level background subtraction in real time, and determines foreground targets. Specifically, it includes:

[0039] Initialization phase: A 10-minute static scan is conducted to collect radar echoes from stationary objects such as the ground and furniture, and a Gaussian mixture model (GMM) is constructed. Each background pixel corresponds to three Gaussian components (mean μ, variance Σ, weight ω).

[0040] Real-time processing: Perform pixel-level background subtraction on the current frame signal and calculate D(x)=||f(x)-B(x)|| / σ(x). If D(x)>2.5, it is determined to be a foreground target (human signal); otherwise, it is background clutter.

[0041] The micro-Doppler feature extraction unit performs a short-time Fourier transform (STFT) on the human echo signal, generates a time-frequency diagram, and extracts the micro-Doppler frequency shift features. Specifically, it includes:

[0042] Perform STFT (Short Time Fourier Transform) on the human body echo signal to generate a time-frequency graph (time resolution 100ms, frequency resolution 2Hz).

[0043] A peak detection algorithm is used to identify the micro-Doppler frequency shifts corresponding to limb swings (e.g., arm swings produce a ±3Hz frequency shift, and leg movements produce a ±5Hz frequency shift), and a feature vector V = [f1, f2, ..., fn] is constructed, where fi is the frequency of the i-th micro-motion component.

[0044] The 3D attitude calculation unit generates 3D point clouds based on virtual aperture synthesis and back projection algorithms using a multi-antenna array. It extracts joint points through adaptive radius filtering and prior knowledge of human structure, and calculates pitch angle, roll angle, and joint angular velocity. Specifically, it includes:

[0045] This step addresses the issue of human targets being easily occluded and having blurred pose features in low-altitude scenes. It achieves accurate pose reconstruction through multi-dimensional point cloud processing and skeletal dynamics constraints. The specific execution process is as follows:

[0046] First, 3D point cloud reconstruction and denoising.

[0047] Virtual aperture synthesis based on 4Tx8RxMIMO array: The transmitting antenna transmits linear frequency modulated continuous wave (FMCW) signals in a time-division manner, and the receiving antenna synchronously collects the echoes. The target azimuth angle (θ_az) and elevation angle (θ_el) are calculated using the phase difference of each antenna pair, and a 32×32 virtual array (32 sampling points in the horizontal direction and 32 sampling points in the vertical direction) is synthesized, improving the spatial resolution to 10cm×10cm (better than the 30cm×30cm of traditional wide-beam radar).

[0048] Back projection (BP) algorithm imaging: After performing range-Doppler processing on the virtual array data, the echoes from each channel are focused onto a three-dimensional spatial grid (coordinates (x, y, z), grid accuracy 5 cm) using the BP algorithm to generate an initial point cloud;

[0049] The formula is as follows: P(x,y,z)=∑i=132∑j=132Si,j(r,φ)·exp(-jλ4π(xsinθaz,i+ycosθaz,i+zsinθel,j));

[0050] Among them, S i，j (r, φ) represents the distance-Doppler spectrum of the virtual antenna in the i-th row and j-th column, where r is the 77 GHz millimeter wave wavelength (approximately 3.9 mm).

[0051] Dynamic noise filtering: For clutter point clouds formed by ground reflections in low-altitude scenes (mostly concentrated in the region z≤30cm), adaptive radius filtering is adopted:

[0052] 1) Set a radius threshold R = 20cm (based on the prior of the minimum distance between human limbs), and count the number of neighbors within 20cm around each point cloud;

[0053] 2) If the number of neighbors is less than 5 (lower than the point cloud density of the human torso / limbs), it is judged as clutter and removed, and the effective human point cloud is retained (signal-to-noise ratio improvement ≥15dB);

[0054] Second, human feature point extraction (joint point recognition).

[0055] Region segmentation based on prior knowledge of human structure: Taking advantage of the characteristic that human targets are mainly distributed in the z=0.5-1.8m range in low-altitude scenes, the point cloud is divided into 3 regions according to height:

[0056] 1) Upper body region (z = 1.2-1.8m): includes feature points such as the shoulder joint and neck;

[0057] 2) Trunk region (z = 0.8-1.2m): includes characteristic points such as the hip joint and waist;

[0058] 3) Lower limb region (z = 0.5-0.8m): includes characteristic points such as the knee joint and ankle joint;

[0059] Feature point clustering and filtering:

[0060] Perform Euclidean clustering on the point cloud of each region (clustering distance threshold 15cm) and extract the cluster centers as candidate feature points;

[0061] Based on the skeletal dynamics model (with pre-defined human joint connections: such as a hip-knee joint distance of 40-60cm and a shoulder joint distance of 30-45cm), candidate points that conform to human structural constraints are selected, and the final joint coordinates (e.g., H(x)) are determined. h y h ,z h ), K (x) of the knee joint k ,y k ,z k Shoulder joint S(x) s ,y s ,z s );

[0062] Third, attitude parameter calculation

[0063] Pitch angle (θ) calculation: Using the human body's longitudinal axis (the line connecting the hip joint H to the shoulder joint S) as a reference, calculate its angle with the vertical direction (z-axis) through vector dot product:

[0064] To address the issue of human figures being easily obscured when leaning forward in low-height scenarios, when the shoulder joint point is missing, the line connecting the hip and knee joints is automatically used to replace the vertical axis in the calculation (error correction ≤ 5°).

[0065] Roll angle Calculation: Based on the projection of the hip joint H onto the horizontal plane (H'(x) h ,y h ,0)) and the projection of shoulder joint S (S'(x s y s ,0)), calculate the angle between the connecting line and the x-axis:

[0066] Joint angular velocity (ω) calculation: The joint coordinates of three consecutive frames (100ms time interval) are differentially analyzed, and the angular velocity is calculated using the center difference method.

[0067] Where, θ k+1 θ k-1 The joint angle between adjacent frames is Δt = 0.1s, ensuring a capture accuracy of ±5° / s for rapid posture changes during falls (such as knee flexion angular velocity > 150° / s).

[0068] The fall detection unit employs an SVM classifier, fusing pitch angle, joint angular velocity, and micro-Doppler entropy features for fall detection. Specifically, it includes:

[0069] Set a fall feature threshold:

[0070] Pitch angle θ > 45° and duration > 0.3s;

[0071] Joint angular velocity ω > 150° / s (knee or hip joint);

[0072] Micro-Doppler entropy E > 3.5 (characterizing increased motion disorder);

[0073] An SVM classifier is used, which fuses multiple feature vectors for decision-making. The discriminant function is:

[0074] f(x) = w1θ + w2ω + w3E + b, where w1 = 0.4, w2 = 0.35, w3 = 0.25, and b = -2.8.

[0075] The adaptive calibration mechanism specifically includes:

[0076] Initial calibration: After installation, an environmental scan is automatically performed to record the furniture positions (identifying stationary objects through point cloud clustering algorithms) and generate ROI masks (excluding ground and furniture areas);

[0077] Online update: The background model is updated hourly, and the incremental GMM algorithm (μ_new=μ_old+α(x-μ_old)) is used to adapt to environmental changes (such as furniture movement or changes in reflectivity caused by carpeting).

[0078] In specific implementation of the embodiments of the present invention:

[0079] Example 1

[0080] Family bedroom scene (installation height 50cm)

[0081] Hardware deployment: Install the radar device on the side of the bedside table (50cm high), with the shield facing the bed and the beam tilted at a 15° angle towards the upper body of the bedridden person.

[0082] Testing process:

[0083] When the elderly person sat up from the bed, the radar detected an elevation angle θ = 30°, a joint angular velocity ω = 80° / s, and a micro-Doppler entropy value E = 2.1, which was determined to be a normal movement.

[0084] If an elderly person accidentally leans forward and falls when getting out of bed, with θ = 60°, ω = 200° / s, and E = 4.2, a fall alarm will be triggered, with a response time of <200ms.

[0085] Example 2

[0086] A scene in a nursing home corridor (installation height 40cm)

[0087] Environmental challenges: The corridor has metal handrails on both sides (reflectivity 0.9) and the floor is tiled (reflectivity 0.85).

[0088] Anti-interference measures: The main lobe is directed to the center of the corridor (20° away from the handrail direction) by beamforming, and a ROI mask for the handrail area is set in the algorithm (areas less than 0.5 meters from the radar are regarded as background).

[0089] Test results: In 100 simulated fall tests, 99 were correctly detected and 1 was falsely detected (due to a brief increase in micro-Doppler entropy caused by rapid running), with a false detection rate of 1%.

[0090] The present invention and its embodiments have been described above, and this description is not restrictive. If those skilled in the art are inspired by this description and design similar embodiments without departing from the spirit of the invention, such embodiments should fall within the protection scope of the present invention.

Claims

1. A low-altitude millimeter-wave radar fall detection device, characterized in that, include: The multi-antenna array module adopts a 4Tx8Rx MIMO architecture with an antenna spacing of λ / 2. It is integrated into a rectangular metal cavity (3) with a height of 3cm and dynamically generates a focused beam pointing to the waist height of the human body through a beamforming algorithm. The narrow-beam, low-elevation radar unit has a center frequency of 77 GHz, a beam angle of 12°, and a physical installation angle tilted upwards by 15° so that the beam center axis is aligned with the waist of the human body. The detection range is limited to a cylindrical area 0.3-3 meters away from the radar and 0.5-1.8 meters in vertical height. The anti-interference component includes a metal shield (1) and a radio frequency filter circuit (2). The metal shield (1) covers the bottom and two sides of the radar module, leaving only a forward 120° detection window. The radio frequency filter circuit (2) integrates a 50-90GHz bandpass filter. The signal processing module is used to perform dynamic background modeling, micro-Doppler feature extraction, 3D pose calculation, and fall detection.

2. A low-altitude millimeter-wave radar fall detection method according to claim 1, characterized in that, The signal processing module includes: The dynamic background modeling unit constructs a Gaussian mixture model through static scanning, performs pixel-level background subtraction in real time, and determines the foreground target. The micro-Doppler feature extraction unit performs short-time Fourier transform on the human body echo signal, generates a time-frequency diagram, and extracts micro-Doppler frequency shift features; The three-dimensional attitude calculation unit generates three-dimensional point clouds based on virtual aperture synthesis and back projection algorithms of multi-antenna arrays. It extracts joint points through adaptive radius filtering and prior human structure and calculates pitch angle, roll angle and joint angular velocity. The fall detection unit uses an SVM classifier to fuse pitch angle, joint angular velocity, and micro-Doppler entropy features for fall detection.

3. The low-altitude millimeter-wave radar fall detection method according to claim 2, characterized in that, The dynamic background modeling unit specifically includes: Initialization phase: A 10-minute static scan is conducted to collect radar echoes from stationary objects such as the ground and furniture, and a Gaussian mixture model is constructed, with each background pixel corresponding to 3 Gaussian components; Real-time processing: Perform pixel-level background subtraction on the current frame signal and calculate D(x)=||f(x)-B(x)|| / σ(x). If D(x)>2.5, it is determined to be a foreground target (human signal); otherwise, it is background clutter.

4. The low-altitude millimeter-wave radar fall detection method according to claim 2, characterized in that, The micro-Doppler feature extraction unit specifically includes: STFT was performed on the human body echo signal to generate a time-frequency diagram; A peak detection algorithm is used to identify the micro-Doppler frequency shift corresponding to the limb swing, and a feature vector V=[f1,f2,...,fn] is constructed, where fi is the frequency of the i-th micro-motion component.

5. A low-altitude millimeter-wave radar fall detection method according to claim 2, characterized in that, The three-dimensional attitude calculation unit specifically includes: First, 3D point cloud reconstruction and denoising. Virtual aperture synthesis based on 4Tx8Rx MIMO array: The transmitting antenna transmits linear frequency modulated continuous wave signals in a time-division manner, and the receiving antenna synchronously collects the echoes. The target azimuth and elevation angles are calculated using the phase difference of each antenna pair, and a 32×32 virtual array is synthesized, improving the spatial resolution to 10cm×10cm. Back projection algorithm imaging: After performing range-Doppler processing on the virtual array data, the echoes from each channel are focused onto a three-dimensional spatial grid using the BP algorithm to generate an initial point cloud; The formula is as follows: P(x,y,z)=∑i=132∑j=132Si,j(r,ϕ)⋅exp(−jλ4π(xsinθaz,i+ycosθaz,i+zsinθel,j)); Where Si,j(r,ϕ) is the range-Doppler spectrum of the virtual antenna in the i-th row and j-th column, and r is the 77 GHz millimeter wave wavelength; Dynamic noise filtering: Adaptive radius filtering is used for clutter point clouds formed by ground reflections in low-altitude scenes. 1) Set a radius threshold R=20cm, and count the number of neighbors within 20cm of each point cloud; 2) If the number of neighbors is less than 5, it is identified as clutter and removed, while the valid human point cloud is retained; Second, human feature point extraction; Region segmentation based on prior knowledge of human structure: Taking advantage of the characteristic that human targets are mainly distributed in the z=0.5-1.8m range in low-altitude scenes, the point cloud is divided into 3 regions according to height: 1) Upper body region (z=1.2-1.8m): includes feature points such as the shoulder joint and neck; 2) Trunk region (z=0.8-1.2m): includes characteristic points such as the hip joint and waist; 3) Lower limb region (z=0.5-0.8m): includes feature points such as the knee joint and ankle joint; Feature point clustering and filtering: Perform Euclidean clustering on the point cloud of each region and extract the cluster centers as candidate feature points; By combining skeletal dynamics models, candidate points that conform to human structural constraints are selected, and the final joint coordinates are determined (e.g., Knee joint shoulder joint ; Third, attitude parameter calculation Pitch angle calculation: Using the human body's longitudinal axis as a reference, calculate the angle between it and the vertical direction (z-axis) through vector dot product: ; To address the issue of human figures being easily obscured when leaning forward in low-height scenes, when the shoulder joint is missing, the line connecting the hip and knee joints is automatically used to replace the vertical axis in the calculation. Roll angle calculation: based on the projection of the hip joint H onto the horizontal plane Projection of shoulder joint S Calculate the angle between the connecting line and the x-axis: ; Joint angular velocity calculation: The joint coordinates of three consecutive frames are differentially analyzed, and the angular velocity is calculated using the central difference method. ; in, , The joint angles of adjacent frames. =0.1s, ensuring an accuracy of ±5° / s in capturing rapid posture changes during a fall.

6. The low-altitude millimeter-wave radar fall detection method according to claim 2, characterized in that, The fall detection unit specifically includes: Set a fall feature threshold: Pitch angle θ > 45° and duration > 0.3s; Joint angular velocity ω > 150° / s (knee or hip joint); The micro-Doppler entropy value E > 3.5; An SVM classifier is used, which fuses multiple feature vectors for decision-making. The discriminant function is: f (x) = w1θ + w2ω + w3E + b, where w1 = 0.4, w2 = 0.35, w3 = 0.25, and b = -2.

8.

7. A low-altitude millimeter-wave radar fall detection method according to claim 2, characterized in that, It also includes an adaptive calibration mechanism, specifically including: Initial calibration: After installation, an environmental scan is automatically performed to record furniture positions and generate ROI masks; Online update: The background model is updated once an hour, and the incremental GMM algorithm (μ_new=μ_old+α(x-μ_old)) is used to adapt to environmental changes (such as furniture movement or changes in reflectivity caused by carpeting).

Images