Fall detection method of millimeter wave radar fusing HRV nonlinear features and 3D point cloud

Abstract

This invention relates to the field of fall detection technology, specifically disclosing a millimeter-wave radar fall detection method that integrates HRV nonlinear features and 3D point clouds. The method includes the following steps: acquiring the inter-heart rate interval (IBI) sequence and 3D point cloud of the target in parallel based on millimeter-wave radar signals; extracting HRV nonlinear physiological features based on the IBI sequence and biomechanical features based on the 3D point cloud; aligning the HRV nonlinear physiological features and biomechanical features to a unified time axis and performing feature coupling calculations to obtain four types of coupling indices; the four types of coupling indices are: cardiac-gait phase coupling, exercise-induced heart rate response, cardiac-triggered motion variability, and joint entropy; and performing fall detection on the target based on a three-level progressive detection mechanism. This invention effectively captures abnormal physiological signals before a fall by establishing an intrinsic correlation between physiological and motion signals, effectively solving the problem of warning lag caused by traditional methods relying solely on motion features.

Description

Technical Field

[0001] This invention relates to the field of fall detection technology, and more specifically to a millimeter-wave radar fall detection method that integrates HRV nonlinear features and 3D point clouds. Background Technology

[0002] As the population ages and the number of elderly people continues to expand, the issue of safety monitoring for the elderly at home and in nursing homes is becoming increasingly prominent. Among these, fall risk monitoring has become a core need concerning the life, health, and quality of life of the elderly. Therefore, developing efficient and reliable fall detection technology has become an urgent need for the development of an aging society.

[0003] Existing fall detection technologies are mainly divided into two categories: one is a detection solution based on wearable devices. This type of solution requires users to wear the device continuously and collect motion data through built-in sensors to determine if a fall has occurred. However, it suffers from poor wearing comfort, which makes most elderly people reluctant to wear it, and it is also prone to false alarms due to motion interference. The other type is a detection solution based on visual cameras. This solution poses a risk of privacy leakage, violates the needs of the elderly and their families for private space, and its detection accuracy drops significantly in dimly lit or obstructed environments, with prominent issues of missed and false alarms.

[0004] Against this backdrop, millimeter-wave radar technology, with its core advantages of non-contact detection, no privacy risks, and immunity to lighting conditions and obstructions, accurately captures radar echoes related to human movement and physiological signals without contacting the human body or collecting visual information. This aligns perfectly with the core needs of home monitoring for the elderly and offers strong compatibility across various application scenarios. However, current fall detection solutions based on millimeter-wave radar still suffer from key technological shortcomings. Most existing solutions are limited to the extraction and judgment of single motion dynamic features, relying primarily on concrete motion indicators such as changes in the body's center of gravity height, limb movement trajectory, and movement speed. This neglects the physiological and motor coupling relationships that precede a fall, leading to the core problems of delayed detection and warning, and high false alarm and missed detection rates. From the perspective of human physiology and motor coordination mechanisms, falls are not sudden, single events of motor instability, but rather involve a clear progressive causal chain: when the human body faces the risk of falling, the autonomic nervous system will first make a stress response, manifested as abnormal fluctuations in physiological indicators such as heart rate variability (HRV). These abnormal physiological signals appear before limb motor instability and are early warning characteristics of fall risk; subsequently, motor characteristics such as decreased limb coordination and instability of movement posture will gradually appear, eventually leading to a fall.

[0005] Secondly, while existing multi-dimensional sensor feature detection solutions take into account heart rate changes before a fall, they suffer from high system integration complexity, high cost, poor data synchronization and compatibility, and poor wearing experience. Furthermore, feature coupling analysis is lacking; multiple sensors merely perform simple data superposition without establishing an intrinsic correlation mechanism between physiological and motion signals. This makes it impossible to capture the progressive pattern of physiological abnormalities preceding motor instability before a fall, and the solutions remain limited to single-dimensional feature judgment.

[0006] Therefore, how to capture the coupling relationship between physiology and movement, realize early warning of falls, reduce false alarm rates, and meet the urgent needs of an aging society for efficient, reliable, and privacy-friendly fall monitoring is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of the above problems, the present invention proposes a millimeter-wave radar fall detection method that integrates HRV nonlinear features and 3D point clouds, so as to overcome the above problems or at least partially solve them.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A millimeter-wave radar fall detection method integrating HRV nonlinear features and 3D point clouds includes the following steps: Based on millimeter-wave radar signals, the inter-heartbeat IBI sequence and 3D point cloud of the target to be detected are acquired in parallel. HRV nonlinear physiological features were extracted based on the IBI sequence, and biomechanical features were extracted based on the 3D point cloud. The nonlinear physiological features of HRV and the biomechanical features were aligned to a unified time axis, and feature coupling calculations were performed to obtain four types of coupling indices. The four types of coupling indices are: cardiac-gait phase coupling, exercise-induced heart rate response, cardiac-triggered motion variability, and joint entropy. Fall detection is performed on the target object based on a three-level progressive detection mechanism; The first level of detection is to calculate whether the mean of the HRV nonlinear physiological characteristics is within a preset multiple of the standard deviation. If it is not, the first level alarm is triggered. The second level of detection is as follows: after the first level alarm, each type of coupling index is compared with its respective preset threshold. A voting mechanism is used. If more than half of the coupling indexes are abnormal, the second level alarm is triggered. The third level of detection is as follows: after the second level alarm, it is determined whether the biomechanical characteristics match the preset biomechanical rules and fall pattern library. If they match, the third level alarm is triggered.

[0009] Furthermore, the extraction process of the HRV nonlinear physiological features includes: After preprocessing the collected millimeter-wave radar data, it is input into a pre-trained neural network model to output the IBI sequence; Based on the IBI sequence, Poincaré plot features and multi-scale entropy are calculated, and detrended fluctuation analysis and recursive quantitative analysis are performed on the IBI sequence to obtain detrended fluctuation features and recursion matrix; the Poincaré plot features, the multi-scale entropy, the detrended fluctuation features, and the proportion of recursive points on the diagonal of the recursion matrix to all recursive points are used as the nonlinear physiological features of HRV.

[0010] Furthermore, when training the neural network model, the sample data includes millimeter-wave radar signal samples of different genders, body types, clothing materials, postures, and environments, and the data is labeled according to time points, including activity type, physiological state, movement state, and body posture.

[0011] Furthermore, the neural network model includes, in turn, an input and batch normalization layer, a first convolutional activation layer, a max pooling layer, a depthwise separable convolutional layer, an average pooling layer, a second convolutional activation layer, a global average pooling and fully connected layer, and a dropout and output layer. The input and batch normalization layer is used to standardize the input signal, providing input data with a unified format and fixed dimensions for all subsequent layers; The first convolutional activation layer is used to extract local phase change features related to the IBI sequence, expanding the single-channel input into multiple feature channels to capture different types of heartbeat-related patterns; The max pooling layer is used to take the maximum value of each channel output by the first convolutional activation layer within a window of 5 sampling points, thus preserving the peak characteristics of the heartbeat signal; The depthwise separable convolutional layer is used to independently convolve each feature channel, capture the temporal correlation features of the signal, and strengthen the feature mapping relationship between the heartbeat cycle and the IBI sequence. The average pooling layer is used to downsample the features output by the depth-separable convolutional layer to smooth feature map fluctuations. The second convolutional activation layer is used to capture heartbeat cycle features over a longer time span; The global average pooling layer and the fully connected layer are used to average the entire time series of each feature channel and perform a non-linear combination. The dropout and output layer is used to randomly drop out some neurons and output multiple consecutive predicted IBI values.

[0012] Furthermore, the extraction process of the biomechanical features includes: Fast Fourier transforms of the original millimeter-wave radar signal in the range, Doppler, and angle dimensions are performed to obtain a three-dimensional spectrum. The constant false alarm rate (CFAR) detection algorithm is used to extract the target point from the three-dimensional spectrogram. The DBSCAN algorithm is used to cluster point clouds belonging to the same target, distinguishing human body parts such as head, torso, upper limbs and lower limbs, and outputting a structured set of point cloud clusters as the initial observation set. Each point cloud cluster represents a potential body part. Inter-frame correlation is performed on each point cloud cluster to generate continuous motion trajectories for each human body part; Based on the continuous motion trajectory, the centroid dynamics features, attitude angle features, segmental coordination features, stability margin features, and motion complexity features are extracted as the biomechanical features.

[0013] Furthermore, the centroid dynamics features include the three-dimensional coordinates of the centroid, vertical velocity, vertical acceleration, and the rate of change of centroid height; wherein, the three-dimensional coordinates of the centroid are calculated by: allocating weight coefficients to each cluster of point clouds through a human biomechanical model, and weighting the three-dimensional spatial coordinate vectors of each human body part point cloud cluster with their respective weight coefficients to obtain the overall three-dimensional coordinates of the centroid. The attitude angle feature is the change in torso angular velocity, which is calculated as follows: principal component analysis is performed on the torso point cloud cluster, the first principal component vector is extracted as the longitudinal axis direction of the torso, the angle between the longitudinal axis of the torso and the horizontal plane is taken as the pitch angle, and the angle between the projection of the longitudinal axis of the torso on the horizontal plane and the forward direction is taken as the roll angle; the torso angular velocity is calculated based on the rate of change of the pitch angle and the rate of change of the roll angle over time. The segmental coordination feature is calculated as follows: extract the point cloud trajectories of the head and pelvis, calculate the vertical displacement signals respectively, perform cross-correlation analysis to obtain the phase difference between the head and pelvis; calculate the coupling degree between the limbs and trunk based on the Pearson correlation coefficient of the limb and trunk velocities. The stability margin feature is the shortest distance between the projection point of the centroid on the ground and the boundary of the supporting polygon. The motion complexity feature is the Shannon entropy of the center-of-mass velocity.

[0014] Furthermore, the cardiac-gait phase coupling is used to quantify the phase-locked relationship between the heartbeat cycle and the gait cycle. The calculation process is as follows: extract the vertical displacement signal from the foot point cloud trajectory, and use Hilbert transform to obtain the analytical signal to obtain the first phase value at the current moment; based on the IBI sequence, linear interpolation is used to obtain the second phase value at the current moment, and the synchronization index R is obtained by calculating the phase difference between the first phase value and the second phase value; if the synchronization index R is less than a preset value, it indicates that the coordination between cardiac and gait is disrupted. The exercise-induced heart rate response is used to determine the effect of exercise intensity on heart rate and to assess baroreflex sensitivity. The calculation process includes: selecting the vertical acceleration of the center of mass to represent exercise intensity and the interbeat period to represent heart rate response, calculating the coherence and transfer function of the two in the low-frequency band, obtaining the average gain, and deriving the baroreflex sensitivity BRS. The cardiac-triggered motion variability is used to detect the stability of trunk micro-movements after each heartbeat, reflecting the strength of myocardial coupling. The calculation process includes: detecting the heartbeat time in the IBI sequence, taking a time window after each heartbeat time, extracting the trunk acceleration signal, calculating the standard deviation of acceleration within each window, and calculating the Z-score of the standard deviation relative to the baseline. The Z-score is used to quantify the degree to which the standard deviation of the current cardiac-triggered micro-movements deviates from the individual's normal baseline. The joint entropy is used to measure the degree of interdependence between the physiological and motor subsystems from an information theory perspective.

[0015] Furthermore, the second-level detection is used to determine whether neuromuscular control is disordered, and the judgment criteria for various coupling indicators are as follows:

[0016] Where CPC represents the cardiac-gait phase coupling index, R represents the synchronization index between cardiac and gait; MIHR represents the exercise-induced heart rate response, BRS represents baroreflex sensitivity; and Z represents the Z-score of the cardiac-triggered motion variability index. This represents the joint entropy.

[0017] Furthermore, the biomechanical rules include multiple rules, each rule corresponding to a feature threshold and a weight. If the biomechanical feature satisfies the feature threshold of one or more rules, the weights corresponding to the currently satisfied one or more rules are added together to obtain a rule score of 0-1. The fall pattern library contains multiple patterns. The biomechanical features are matched with each pattern, and the score with the highest degree of matching is taken as the pattern matching score. The rule score and pattern matching score are weighted and summed to obtain the overall confidence level of the third-level detection. If the overall confidence level is greater than or equal to a preset confidence threshold, a third-level alarm is triggered.

[0018] Furthermore, the biomechanical rules include: rapid descent rule, rapid rotation rule, support instability rule, loss of coordination rule, and motion simplification rule; wherein, the rapid descent rule is: vertical velocity of the center of mass > characteristic threshold one; the rapid rotation rule is: trunk angular velocity > characteristic threshold two; the support instability rule is: stability margin < characteristic threshold three; the loss of coordination rule is: coupling degree between limbs and trunk < characteristic threshold four; and the motion simplification rule is: motion complexity < characteristic threshold five. The fall pattern library includes the following patterns: forward fall, backward fall, side fall, slip, and slow fainting.

[0019] As can be seen from the above technical solution, compared with the prior art, the present invention has the following beneficial effects: 1. This invention establishes an intrinsic correlation between physiological and motor signals by coupling analysis of HRV nonlinear physiological and biomechanical characteristics and combining four coupling indicators: cardiac-gait phase coupling, exercise-induced heart rate response, cardiac-triggered motor variability, and joint entropy. This can effectively capture abnormal physiological signals before a fall, fundamentally improve the predictive ability of detection, and effectively solve the problem of early warning lag caused by traditional methods that rely solely on motor characteristics.

[0020] 2. This invention employs a three-level progressive decision logic to filter out interfering events step by step, significantly reducing the false alarm rate and the missed alarm rate.

[0021] 3. This invention performs parallel processing of the intercardiac IBI sequence and 3D point cloud through dual channels, simultaneously extracting physiological and motor features. It eliminates the need for multi-sensor collaboration, significantly improving operating speed. Compared to multi-sensor solutions that require deploying multiple devices in the body and environment, this invention fundamentally solves problems such as high hardware costs and complex power supply wiring, ensuring device stability.

[0022] 4. This invention achieves non-contact, privacy-free monitoring based on millimeter-wave radar, solving the problems of poor comfort, low user compliance, and susceptibility to motion interference in wearable devices, while avoiding detection interruptions caused by device detachment. Compared to visual camera solutions, this invention does not collect any visual images, but extracts features only through radar I / Q data, eliminating the risk of privacy leakage at the source. It can be adapted to various private scenarios and can also simultaneously solve the problem of sharp drop in accuracy of visual solutions in dim or occluded scenarios. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0024] Figure 1 This is a flowchart of a millimeter-wave radar fall detection method that integrates HRV nonlinear features and 3D point clouds, provided in an embodiment of the present invention. Figure 2 This is an architecture diagram of the neural network model provided in this embodiment of the invention; Figure 3 This is a trend diagram of the changes in Poincaré plot features before and after a fall, provided in an embodiment of the present invention. Figure 4 This is a trend diagram of multi-scale entropy changes before and after a fall, provided in an embodiment of the present invention. Figure 5 This is a graph showing the trend of scaling index changes before and after a fall, provided in an embodiment of the present invention. Figure 6 This is a graph showing the trend of the center of gravity height change before and after a fall, provided in an embodiment of the present invention. Figure 7 This is a graph showing the trend of vertical velocity change of the center of gravity before and after a fall, provided in an embodiment of the present invention. Figure 8 This is a graph showing the trend of trunk angular velocity changes before and after a fall, provided in an embodiment of the present invention. Figure 9 This is a three-dimensional coordinate map of the point cloud of the target to be detected provided in this embodiment of the invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] like Figure 1 As shown in the figure, this invention discloses a millimeter-wave radar fall detection method that integrates HRV nonlinear features and 3D point clouds, including the following steps: S1. Based on the millimeter-wave radar signal, acquire the inter-heartbeat IBI sequence and 3D point cloud of the target to be detected in parallel; S2. Extracting nonlinear physiological features of HRV based on IBI sequences and extracting biomechanical features based on 3D point clouds; S3. Align the nonlinear physiological and biomechanical characteristics of HRV to a unified time axis and perform feature coupling calculations to obtain four types of coupling indices: cardiac-gait phase coupling, exercise-induced heart rate response, cardiac-triggered motion variability, and joint entropy. S4. Fall detection is performed on the target object based on a three-level progressive detection mechanism; The first level of detection is to calculate whether the mean of the nonlinear physiological characteristics of HRV is within a preset multiple of the standard deviation. If it is not, the first level alarm is triggered. The second level of detection is as follows: after the first level alarm, each type of coupling indicator is compared with its own preset threshold. A voting mechanism is used. If more than half of the coupling indicators are abnormal, the second level alarm is triggered. The third level of detection is as follows: after the second level alarm, it is determined whether the biomechanical characteristics match the preset biomechanical rules and fall pattern library. If they match, the third level alarm is triggered.

[0027] The following is a further explanation of each of the above steps.

[0028] S1, based on millimeter-wave radar signals, acquires the inter-heartbeat IBI sequence and 3D point cloud of the target to be detected in parallel.

[0029] The process of obtaining the intercardiac IBI sequence includes: 1) Acquire millimeter-wave radar signals and perform preprocessing. The specific preprocessing process includes: For processing random length sequence signals, a sliding window segmentation algorithm is used. Let the original signal length be L, the sampling rate be f_s, and the target be a 10s signal segment (approximately 12 complete heartbeat cycles). The target window length is N_target = 10000 points, and the overlap rate is r_overlap = 0.5. The window step size is S_step = N_target × (1 - r_overlap) = 5000 points. The number of windows is N_windows = max(1, floor((L - N_target) / S_step) + 1), where floor represents the floor function. The starting position of the k-th window is start_k = max(0, min(L - N_target, k × S_step)).

[0030] The spatial filtering mechanism of digital beamforming (DBF) is used to enhance the signal in the direction of the chest cavity and suppress interference. Specifically, the 12 virtual channels correspond to different spatial positions of the radar receiving array. The weighting vector adopts the Hanning window to reduce sidelobe interference, as shown in formula (1). The beam pointing is set to straight ahead (0° azimuth angle, 0° elevation angle), which is consistent with the typical spatial orientation of the human chest cavity in the monitoring scenario.

[0031] (1) in, For the Hanning window coefficient, , The beam pointing angle is 0 degrees, so the exponent is 1. This represents the original signal collected by the nth virtual receiving channel at time t. This is used for phase compensation of signals from different channels. It ensures that the echo signals in the chest cavity direction (θ=0°) are superimposed in phase during weighted summation, resulting in a significant amplitude enhancement. Meanwhile, interference signals from other directions cancel each other out or are significantly attenuated after summation due to phase mismatch, thereby achieving directional enhancement of signals in the chest cavity region and effective suppression of environmental clutter and interference from other directions.

[0032] For the beam-shaped signal x beam (t) Perform a Fast Fourier Transform (FFT) on the range dimension to obtain the range-dimensional spectrum. Specifically, the continuous-time signal is divided into discrete frames according to the radar frame structure, with each frame containing Nr range sampling points (Nr=256 in this invention). Perform an Nr-point FFT on each frame signal in the range dimension to convert the time-domain signal into a range-domain spectrum X. beam (r,t), where r=0,1,...,Nr 1. Corresponding to different distance gates, it completes the conversion from the time domain to the distance domain, and achieves accurate positioning of the target distance.

[0033] The distance gate is a discretized representation of the distance-dimensional spectrum, where the distance-dimensional spectrum X... beam Each index r in (r,t) corresponds to a distance gate. Calculate the signal variance of each distance gate and select the distance gate with the largest variance as the thoracic distance gate. ,Right now Essentially, it locks the discrete distance interval where the chest cavity is located from the distance dimension spectrum, thereby achieving distance domain focusing of the chest cavity signal. The specific calculation process is shown in formula (2).

[0034] (2) In the above formula, This represents the variance operator, used to calculate the dispersion of the signal amplitude within a range gate r, reflecting the energy concentration of the target signal within that range gate. This represents the signal at time t at distance r after beamforming.

[0035] Determining the distance r from the thoracic cavity chest Then, the signal X corresponding to the distance gate is extracted from the distance-dimensional spectrum. beam (r chest ,t), through complex phase operation (t)=arg(Xbeam (r chest The original phase sequence is obtained by ,t)); this operation only retains the phase information of the thoracic distance gate and filters out the interference of other distance gates, providing a clean phase input for subsequent phase unwrapping and heartbeat feature extraction.

[0036] Phase sequence extracted from the chest cavity distance gate (t), phase unwrapping is performed to avoid the jump phenomenon, see formula (3), and then a fourth-order Butterworth bandpass filter is used with a passband of about 0.8Hz-4Hz to extract the effective heartbeat signal segment.

[0037] (3) In the above formula, This represents the thoracic phase sequence after phase unwrapping at time t. This represents the rounding operator, used to calculate integer multiples of phase transitions, achieving integer period compensation of the phase. Represents time t 1. The thoracic phase value after phase unwrapping. This represents the original phase sequence extracted directly from the thoracic distance gate.

[0038] 2) Input the filtered phase sequence into the pre-trained neural network model and output the IBI sequence; In this embodiment of the invention, the input to the neural network model is a 10-second phase signal (10,000 points), and the output is 7 consecutive IBI values. IBI stands for Inter-Beat Interval, which refers to the time interval between two consecutive heartbeats (R wave peaks). The 7 consecutive IBI values ​​output in this invention represent the predicted time interval of the next 7 heartbeat cycles and are the core input data for calculating the nonlinear characteristics of HRV.

[0039] like Figure 2 As shown, the neural network model includes, in sequence, an input and batch normalization layer L1, a first convolutional activation layer L2, a max pooling layer L3, a depthwise separable convolutional layer L4, an average pooling layer L5, a second convolutional activation layer L6, a global average pooling and fully connected layer L7, and a dropout and output layer L8.

[0040] The input and batch normalization layer L1 is used to standardize the input signal, providing input data with a uniform format and fixed dimensions for all subsequent layers. Specifically, this layer calculates the mean and variance for each mini-batch of data, normalizes the output, and sets parameters. The value is 1e-5, which improves the stability of network training and the model's adaptability to different signal amplitudes.

[0041] The first convolutional activation layer L2 is used to extract local phase change features associated with the IBI sequence, expanding the single-channel input into multiple feature channels to capture different types of heartbeat-related patterns; specifically, the input size of the first convolutional activation layer is 10000. 1. Using a convolutional kernel of size 25 and stride 5, and swish as the activation function, the single-channel input is expanded to 32 feature channels to capture different types of heartbeat-related patterns. The output size is 1996. 32.

[0042] The max-pooling layer L3 is used to maximize the output of each channel of the first convolutional activation layer within a window of 5 sampling points. This effectively preserves the peak characteristics of the heartbeat signal, accelerates dimensionality reduction, and suppresses noise. The output size is 399. 32.

[0043] The depthwise separable convolutional layer L4 is used to independently convolve each feature channel, capturing the temporal correlation features of the signal and strengthening the feature mapping relationship between the cardiac cycle and the IBI sequence. Specifically, this layer is configured with a depth multiplier of 1, 64 output channels, a kernel size of 32, a stride of 1, no padding, a Swish activation function, and an output size of 368. 64.

[0044] The average pooling layer L5 is used to downsample the features output by the depthwise separable convolutional layer, taking the average value within the window to further reduce dimensionality and smooth feature map fluctuations. The pooling size is set to 3, the stride to 3, with no padding, and the output size is 122. 64.

[0045] The second convolutional activation layer, L6, is used to capture heartbeat cycle features over a longer time span, further extracting high-level features. It uses 128 filters, a kernel size of 50, a stride of 1, no padding, and the Swish activation function, with an output size of 73. 128.

[0046] The global average pooling layer and the fully connected layer L7 are used to average the entire time series of each feature channel, compressing the entire time series of each channel into a scalar, and outputting 128 feature values. Then, the 128 global features are non-linearly combined, with 64 neurons set, and the activation function is Swish, outputting 64 neurons.

[0047] The dropout and output layer L8 is used to randomly drop out some neurons and output multiple consecutive predicted IBI values. Specifically, the dropout layer is set to dropout=30%, randomly dropping out some neurons to force the network to learn more robust features. The output layer uses a linear activation function, has 7 neurons, and outputs 7 consecutive predicted IBI values.

[0048] The entire neural network model is designed specifically for extracting IBI values, capturing patterns across different time spans in heartbeat signals by concatenating convolutional kernels of different scales.

[0049] It should be noted that when training the neural network model, a TI AWR2243 radar board with a configuration of 3 transmitters and 4 receivers was used. Frequency modulated continuous wave (FMCW) was transmitted at a sampling frequency of 2000Hz, and 5000 sets of randomly lengthed sample data were collected. These samples included signal samples from different genders, body types, clothing materials, postures (sitting, standing, lying down, etc.), and different environments to ensure data integrity. Data was labeled according to time points, including activity type, physiological state, motion state, and body posture. The collected data was stored as a CSV file, with rows representing the time axis and columns including location information, transmitted radar frame number, and target ID. Specific phase and amplitude values ​​could be read from each row and column. 80% of the dataset was randomly allocated as the training set, 10% as the validation set, and 10% as the test set. The preprocessing method for the data samples was the same as mentioned in section 1).

[0050] S2. Extract nonlinear physiological features of HRV based on IBI sequence and extract biomechanical features based on 3D point cloud.

[0051] 1) The extraction process of HRV nonlinear physiological features includes: calculating Poincaré plot features and multi-scale entropy based on IBI sequences, and performing detrended fluctuation analysis and recursive quantitative analysis on IBI sequences to obtain detrended fluctuation features and recursion matrix; the Poincaré plot features, multi-scale entropy, detrended fluctuation features, and the proportion of recursion points on the diagonal of the recursion matrix to all recursion points are used as HRV nonlinear physiological features.

[0052] ①The calculation process of the Poincaré diagram features is as follows: Let the IBI sequence be SD1 reflects the short-term fluctuation of adjacent IBIs, see formula (4), SD2 reflects the long-term fluctuation range of the IBI sequence, see formula (5), SD1 / SD2 represents the balance between short-term and long-term heart rate fluctuations.

[0053] (4) (5) like Figure 3 As shown, the Poincaré plot features the SD1 / SD2 ratio, representing the balance between short-term and long-term heart rate fluctuations. During the normal period, the ratio remains stable at around 1.2. During the transition period within the normal range, the ratio increases from 1.2 to 1.8, with SD1 increasing and SD2 decreasing. Before a fall, the ratio further increases to 2.2, which is significantly abnormal and indicates a sharply increased risk of falls.

[0054] ②The calculation process of multiscale entropy (MSE) is as follows: Multiscale entropy is a combination of sample entropy (SampEn) based on different time scales: first, coarse-grained sequences of different scales (i=1,2,3) are constructed using formula (6), and then the sample entropy (SampEn) is calculated for each coarse-grained sequence, finally obtaining a set of entropy values ​​under multiple scales.

[0055] Set scale (1, 2, 3), construct a coarse-grained sequence as shown in formula (6), and set the embedding dimension to 2 and the similarity tolerance r=0.2. The sequence standard deviation is used to construct an m-dimensional vector and count the number of matching vectors. Then, an m+1-dimensional vector is constructed and the number of matching vectors is counted. The sample entropy SampEn(m,r,N) is obtained, as shown in formula (7).

[0056] (6) (7) In the above formula, i represents the scale index of the coarse-grained sequence, and j represents the sample index of the coarse-grained sequence. For a length of N orig The original IBI sequence, N orig This represents the total length of the original IBI sequence. N represents the coarse-grained sequence. Let be the probability of template matching in the (m+1)-dimensional vector space. Let be the probability of template matching in the m-dimensional vector space.

[0057] The core of constructing coarse-grained sequences is "average aggregation over time scales". Taking scale τ as an example: the original IBI sequence [x1,x2,...,x] is... N The sample is divided into continuous, non-overlapping segments of length τ; the average of the τ IBI samples within each segment is taken to obtain a new sample; the averages of all segments are arranged in order to form a coarse-grained sequence y at this scale. j (i) For example, when the scale is 2, the coarse-grained sequence of the original sequence [x1,x2,x3,x4] is: [(x1+x2) / 2,(x3+x4) / 2]).

[0058] The object of the entire sample entropy calculation is the coarse-grained sequence generated by formula (6). The multi-scale entropy (MSE) is the three sample entropy values ​​calculated repeatedly for the three coarse-grained sequences y(1), y(2), and y(3) corresponding to scales τ=1, 2, and 3, respectively. This reflects the complexity of the original IBI sequence at different time scales.

[0059] Specifically, the calculation process for sample entropy is as follows: a. Setting parameters: embedding bit length m=2 and similarity tolerance are both empirical parameters, and recommended values ​​should be set under normal circumstances.

[0060] b. Construct an m-dimensional vector from the sequence In the sequence, a series of vectors of length m are constructed sequentially. These vectors represent the pattern formed by m consecutive data points in the sequence.

[0061] c. Count the number of vector matches. For each vector Xm(i), calculate its distance d with all other vectors, and count the vectors that satisfy d ≤ r, denoted as p. The average probability can be calculated from this.

[0062] Similarly, an m+1 dimensional vector can be constructed.

[0063] like Figure 4 As shown, multiscale entropy (MSE, scales 1-3) measures the complexity of heart rate signals at different time scales. A higher entropy value indicates greater flexibility in autonomic nervous system regulation. The three curves correspond to scale 1 (short-term), scale 2 (medium-term), and scale 3 (long-term), respectively. During the normal period, the entropy values ​​at each scale are stable, the signal complexity is high, and the autonomic nervous system is flexible. During the transition period, the entropy values ​​at the low-to-medium scales (1-2) suddenly increase, and the nervous system activates more muscles to maintain balance. In the early stage of a fall, the entropy values ​​at all three scales decrease in a step-by-step manner, the signal tends to become more regular, and the corresponding autonomic regulatory ability temporarily fails.

[0064] ③ The process of performing detrended fluctuation analysis on the IBI series is as follows: After removing the mean from the IBI sequence ( ), find the cumulative sum ( The equation is then segmented according to a scale *s*, and a linear fit is performed on each segment to obtain the residual, thus yielding the average fluctuation *F(s)*. In a log-log coordinate system, a linear relationship is fitted between *log(F(s))* and *log(s)*, with the slope representing the scaling exponent. ,in Reflecting the short-term correlation of heart rate signals, corresponding to a scale s=4~16. Reflecting the long-term correlation of heart rate signals, corresponding to scales s = 16~64. s represents the scale, dividing the cumulative sum sequence Y(k) into several non-overlapping intervals of length s, where s is the independent variable.

[0065] Within each interval of length s, the local trend is obtained through linear fitting, and then the detrended volatility, i.e., the standard deviation of the residuals, is calculated within that interval. The root mean square of the volatility calculated for all intervals is taken to obtain the overall volatility function value F(s) at that scale s.

[0066] like Figure 5As shown, in the detrended fluctuation analysis, during the normal period, α1=0.9 and α2=1.1, which are within the normal range, and the heart rate signal self-similarity is stable. α1 represents short-term regulatory capacity, and α2 represents long-term regulatory capacity. If α1 further decreases and α2 increases sharply, it will cause long-term regulation disorder, resulting in an abnormality of low α1 and high α2.

[0067] ④ The process of recursive quantitative analysis of IBI sequences is as follows: By selecting an embedding dimension m=5 and a time delay τ=1, the phase space vector is reconstructed. The phase space vector is then represented as (x1,x2,x3,x4,x5), (x2,x3,x4,x5,x6),..., and the recursive matrix is ​​calculated as shown in formula (8). It is a step function. The threshold is usually set to 0.2 × the standard deviation of the sequence.

[0068] (8) Determinism (DET) is calculated as the proportion of recursive points on the diagonal (length ≥ l_min=2) of the recursion matrix out of all recursive points, and the proportion of recursive points on the vertical segment (length ≥ v_min==2) of laminarity (LAM) out of all recursive points. DET measures the predictability or determinism of system dynamics. It calculates the proportion of recursive points on the diagonal segment of the recursion matrix (representing the system's repeated evolution along similar trajectories in phase space) out of the total number of recursive points. The calculation method is as follows: In the recursion matrix R, identify all lengths... l ≥ l _min, l For the diagonals where _min=2, count the number of all black points (points with a value of 1) on these line segments, and then divide by the total number of black points in the entire recursive graph. A high DET value indicates that the dynamic behavior of the system is regular and deterministic.

[0069] In laminar flow (LAM) within a recurrence matrix R, a vertical line segment refers to a continuous sequence of black points along the same column j and direction i. These segments indicate that the system state changes very little over multiple consecutive time intervals, approaching a state of rest. Identify all diagonals with lengths v ≥ v_min and v_min = 2, count the number of all black points (points with a value of 1) on these segments, and then divide by the total number of black points in the entire recurrence graph.

[0070] 2) The process of extracting biomechanical features includes: ① Perform fast Fourier transforms on the original millimeter-wave radar signal in the range dimension, Doppler dimension, and angle dimension to obtain a three-dimensional spectrum.

[0071] Specifically, a Fast Fourier Transform (FFT) is performed on the sampling points of each chirp (a chirp refers to a complete cycle of a linearly modulated transmit signal), an FFT is performed on each range gate on different chirs, and an FFT is performed on each range-velocity unit on different channels, where c = 3 × 10 8 Given m / s, B=4GHz, λ=3.9mm, and d=λ / 2, the 3D coordinates can be obtained: x=R×sinθ×cosφ, y=R×cosθ×cosφ, and z=R×sinφ. The three-dimensional FFT (range, velocity, angle) processing results directly provide parameters in spherical coordinates for each detected target point: radial range R, azimuth θ, and elevation φ. After performing range-dimensional FFT and Doppler-dimensional FFT on the raw radar data, a two-dimensional data grid is formed. Each small cell in this grid is called a range-velocity cell (or range-Doppler cell), which uniquely corresponds to a specific range and velocity value.

[0072] After obtaining the aforementioned two-dimensional grid, for each range-velocity element, its complex signal values ​​across all receiving channels need to be extracted, forming a complex sequence of length C (number of channels). This sequence is then subjected to FFT processing. Due to the minute spatial spacing between different antennas, the echo signals from the same target arriving at different antennas will have a phase difference determined by the target's azimuth angle. Performing an FFT (i.e., angle-dimensional FFT or beamforming) on ​​this phase difference sequence (across channel dimensions) allows the resolution of the target's angular information, i.e., azimuth or elevation angle.

[0073] ② Calculate the power of each three-dimensional element and set 2 2 2 protection units and 4 4 The training unit has 4 units, and the average power of the training unit is used as the background noise power estimate, based on the set false alarm probability P_fa=10e -6 A power threshold T is determined; if the power threshold is greater than T, the point is considered a valid target point in the three-dimensional spectrum. Guard cells are a ring of cells immediately surrounding the detected cell. They are set up to prevent target energy leakage into the training cells, which would increase the background noise estimate and cause the target to be missed. Training cells are located outside the guard cells and are considered pure background noise regions. The average signal power within these training cells is calculated as the background noise power estimate.

[0074] ③ Since a target may consist of multiple point clouds, the DBSCAN algorithm is used to cluster point clouds belonging to the same target, distinguishing between human body parts such as the head, torso, upper limbs, and lower limbs. This outputs a structured set of point cloud clusters as the initial observation set. Each point cloud cluster represents a potential body part, such as... Figure 9 As shown, the neighborhood radius ε can be set to 0.03 meters and the minimum number of points MinPts. The Euclidean distance between points can be calculated to find points in the neighborhood with a number of points ≥ MinPts, and the clustering can be expanded.

[0075] ④ Perform inter-frame association for each point cloud cluster to generate continuous motion trajectories for each human body part; specifically, in order to obtain continuous motion trajectories, a combination of multiple hypothesis tracking (MHT) and probabilistic data association filtering (PDAF) is adopted. First, multiple hypothesis tracking (MHT) is used to determine the multiple clouds of the same target and associate them with existing trajectories. Based on the degree of matching between observation and prediction, high probability hypotheses are retained, similar hypotheses are merged, and multiple cloud clusters of the same target are generated.

[0076] Specifically, the degree of matching between observations and predictions is as follows: Based on the Kalman filter, the possible state of the target at the current moment is predicted according to the state of the target at the previous moment; Secondly, the matching probability between all observed point clouds in the current frame and the existing predicted state is calculated based on the Mahalanobis distance between the predicted and observed positions, and a state update with weighted multiple possible association hypotheses is formed for each target, as shown in formula (9).

[0077] (9) Where, d m This refers to the Mahalanobis distance between each observed point cloud and each predicted state, β. i,j This refers to the association probability, P. D Let λ be the detection probability, λ be the clutter density, and L be the probability density value given by the probability density function of a Gaussian distribution given the presence of a target.

[0078] By iteratively predicting, associating, and updating, and combining pruning strategies to eliminate low-probability hypotheses, the maximum number of hypotheses is set to 20, the pruning threshold is set to 0.1, and the trajectory is terminated if there are 5 consecutive frames without association. Finally, smooth and continuous three-dimensional motion trajectories are output for body parts such as the head, torso, and limbs.

[0079] ⑤ Based on the continuous motion trajectory, extract the centroid dynamics features, attitude angle features, segmental coordination features, stability margin features, and motion complexity features as biomechanical features.

[0080] The dynamic characteristics of the center of mass include the three-dimensional coordinates of the center of mass, the vertical velocity, the vertical acceleration, and the rate of change of the center of mass height. The three-dimensional coordinates of the center of mass are calculated by assigning weight coefficients to each cluster of point clouds through a human biomechanical model, and then weighting the three-dimensional spatial coordinate vectors of each human body part point cloud cluster with their respective weight coefficients to obtain the overall three-dimensional coordinates of the center of mass.

[0081] Specifically, the weights are allocated according to the human biomechanical model, namely, the head accounts for 7% of the body weight, the trunk accounts for 50% of the body weight, the upper limbs each account for 5% of the body weight, and the lower limbs each account for 16.5% of the body weight, so as to calculate the center of mass (COM) and its coordinates, as shown in formula (9). Then, the rate of change of the center of mass height, vertical velocity, and vertical acceleration are calculated.

[0082] (10) Where, m i This refers to the normalized mass weights assigned to the i-th human body part, such as the head, trunk, upper limbs, and lower limbs, based on a human biomechanical model. i It refers to the three-dimensional spatial coordinate vector of the point cloud cluster of the i-th human body part obtained by radar point cloud tracking.

[0083] like Figure 6 As shown, the smoothed signal remains stable at the height of an adult standing center of mass during the normal period, with small fluctuations; when sitting normally, the smoothed signal decreases slowly; when a fall occurs, the smoothed signal decreases rapidly from 1.0 meter with a steep slope.

[0084] like Figure 7 As shown, when the vertical velocity of the center of mass is negative, it indicates downward movement; the larger the absolute value, the faster the descent. Under normal circumstances, the smoothed signal fluctuates around 0 with no obvious downward trend. When the smoothed signal becomes slowly negative, the corresponding center of mass descends slowly; when it becomes rapidly negatively accelerated, it corresponds to a falling state.

[0085] The attitude angle features include pitch angle and roll angle. The calculation method is as follows: perform principal component analysis on the torso point cloud cluster, extract the first principal component vector as the longitudinal axis direction of the torso, take the angle between the longitudinal axis of the torso and the horizontal plane as the pitch angle, and take the angle between the projection of the longitudinal axis of the torso on the horizontal plane and the forward direction as the roll angle. The calculation method of pitch angle is shown in formula (11), and the calculation method of roll angle is shown in formula (12).

[0086] (11) (12) Among them, v x v y v zThese are the components of vector v in the X, Y, and Z directions in a three-dimensional Cartesian coordinate system, where v represents the first principal component vector obtained after principal component analysis of the torso point cloud cluster. This vector points in the direction of the longest point cloud distribution, i.e., the longitudinal axis of the torso.

[0087] Next, the rate of change of the pitch angle over time is calculated, expressed as: , used to measure the rate of tipping in the forward-backward direction of the torso; calculates the rate of change of the roll angle over time, expressed as . This is used to measure the rate of lateral tilting of the torso. Then, based on... and Calculate the trunk angular velocity to quantify the body's rotational speed and tendency to tip over.

[0088] The formula for calculating the torso angular velocity is: .

[0089] like Figure 8 As shown, the change in trunk angular velocity reflects the trunk's tilting and rotation speed. The larger the absolute value, the more violent the tilt. Under normal circumstances, the smooth signal is stable and there is no obvious rotation trend. When the smooth signal gradually increases and becomes a sharp jump, the body is in a obviously tilting posture.

[0090] The segmental coordination feature is calculated as follows: extract the point cloud trajectories of the head and pelvis, calculate the vertical displacement signals respectively, perform cross-correlation analysis, and obtain the phase difference between the head and pelvis, as shown in formula (13); calculate the coupling degree between the limbs and trunk based on the Pearson correlation coefficient of the limb and trunk velocities. Specifically, obtain the centroid velocity sequences of the point cloud clusters of a certain limb part and the point cloud clusters of the trunk part respectively, and calculate the velocity amplitude of each sequence, i.e., the magnitude of the velocity vector. The coupling degree is the Pearson correlation coefficient between these two velocity amplitude sequences, as shown in formula (14). (13) Where s_head(t) is the signal of head displacement in the Z-axis direction as a function of time, s_pelvis(t) is the signal of pelvic displacement in the Z-axis direction as a function of time, and τ is the time delay parameter. The cross-correlation coefficient at a time delay τ measures the similarity between pelvic signals and head signals when the pelvic signals are advanced by τ time.

[0091] (14) in, For the speed range of a certain limb, M represents the velocity amplitude of the torso, and M refers to the total number of common data points between the two velocity sequences used in the calculation. This represents the mean of the velocity sequence of the center of mass of the corresponding torso. This represents the mean of the velocity sequence of the center of mass of the corresponding limb.

[0092] The stability margin characteristic is the shortest distance between the projection point of the centroid on the ground and the boundary of the supporting polygon. The motion complexity feature is the Shannon entropy of the centroid velocity, and then the randomness of the motion is evaluated. First, the centroid velocity sequence is calculated and its velocity amplitude is obtained. The velocity amplitude is divided into 20 intervals, the probability distribution of the velocity value is calculated, and the Shannon entropy is obtained, as shown in formula (15).

[0093] (15).

[0094] In the above formula, This represents the probability that the centroid velocity amplitude falls into the i-th interval, which is the proportion of the number of times the velocity value appears in that interval relative to the total number of samples.

[0095] S3. Align the nonlinear physiological and biomechanical features of HRV to a unified time axis. The original HRV features are updated every 10 seconds. All features are resampled to 100Hz through linear interpolation. The features are synchronized and aligned through timestamps to ensure that the alignment error is <10ms.

[0096] Subsequently, feature coupling calculations were performed to obtain four types of coupling indices: cardiac-gait phase coupling, exercise-induced heart rate response, cardiac-triggered motion variability, and joint entropy.

[0097] Cardiac-gait phase coupling (CPC) is used to quantify the phase-locked relationship between the heartbeat cycle and the gait cycle. The calculation process is as follows: the vertical displacement signal is extracted from the foot point cloud trajectory, and the analytical signal is obtained by Hilbert transform. The first phase value at the current moment is obtained, as shown in formula (16). Based on the IBI sequence, the second phase value at the current moment is obtained by linear interpolation. The synchronization index R is obtained by calculating the phase difference between the first and second phase values, as shown in formula (17). If the synchronization index R is less than the preset value, it indicates that the coordination between cardiac and gait is disrupted. Under normal circumstances, R is about 0.4-0.8. If R < 0.3, it indicates that the coordination between cardiac and gait is disrupted.

[0098] (16) Where A(t) is the amplitude of the radar echo signal at time t. The first phase value at time t, Let be the optimal error signal at time t. The Hilbert transform of the optimal error signal, where j is the imaginary unit, i.e., j 2 =-1.

[0099] (17) in, Let be the phase difference at time t. Let t be the second phase value at time t, R be the cyclic correlation coefficient of the phase difference, and T be the number of sampling points within the statistical period.

[0100] Exercise-induced heart rate response (MIHR) is used to determine the effect of exercise intensity on heart rate and assess baroreflex sensitivity. The calculation process includes: selecting the vertical acceleration of the center of mass to represent exercise intensity and the interbeat interval to represent heart rate response; calculating the coherence and transfer function of the two in the low-frequency band; obtaining the average gain; and deriving the baroreflex sensitivity BRS, as shown in formula (18). If BRS < 3.0, it indicates a decline in cardiovascular regulatory capacity.

[0101] (18) in, It refers to the square of the coherence coefficients for all potential frequencies f. Represents the effective discrete frequency point f k The square of the coherence coefficient at a given point is given by f, where f is a continuous independent variable covering the entire analysis frequency range. It can vary continuously and represents all potential frequencies in the signal frequency domain analysis. k Effective discrete frequency points are a subset of all frequency points selected from all frequency points, taking only specific discrete values, and are limited in number.

[0102] For acceleration-intercardia cross-spectral density, Let H(f) be the self-spectral density of the vertical acceleration of the center of mass, and H(f) be the transfer function. The interval autospectral density is represented by BRS, which represents baroreflex sensitivity. This represents the transfer function with frequency f as the independent variable, describing the frequency domain mapping relationship from the input signal to the output signal. Represents the effective discrete frequency point f k The transfer function value at that location.

[0103] Heart-triggered motion variability (HTMV) is used to detect the stability of trunk micro-movements after each heartbeat, reflecting the strength of myocardial coupling. The calculation process includes: detecting the heartbeat moments in the IBI sequence, taking a time window (e.g., 50ms) after each heartbeat moment, extracting trunk acceleration signals, and calculating the standard deviation of acceleration within each window. The Z-score of the standard deviation relative to the baseline is calculated, and the Z-score is used to quantify the degree to which the standard deviation of the current heartbeat-triggered micro-movements deviates from the individual's normal baseline. If the Z-score increases abnormally, it can be determined that autonomic nervous system disorder has occurred. Specifically, through custom settings, trunk acceleration signals are continuously collected for at least 1-2 minutes in a normal, stable standing state. All heartbeat moments within this period are detected, and the standard deviation of acceleration is calculated in a 50ms window after each heartbeat to obtain a set of baseline standard deviations. The mean and standard deviation of the baseline data are calculated, as well as the real-time Z-score, as shown in formula (19).

[0104] (19) in, The degree of deviation of the micro-fluctuation at time k from the individual's resting normal state. Let be the standard deviation at time k. The standard deviation of the baseline. This represents the baseline mean.

[0105] Joint entropy is used to measure the degree of interdependence between the physiological and motor subsystems from an information theory perspective. Specifically, representative values ​​SD1 / SD2 and stability margin SM are selected to form a two-dimensional variable (X,Y). X and Y are discretized into M and N intervals, respectively, and the joint probability distribution and joint entropy are calculated, followed by normalization of the joint entropy. An abnormally high joint entropy indicates decoupling between the two systems. A threshold is set to determine this; this threshold indicates a significant weakening of the statistical correlation between the physiological (SD1 / SD2) and motor (stability margin) subsystems compared to the user's normal steady state. An abnormally high entropy value indicates that the two systems are decoupling.

[0106] Specifically, the value ranges of X and Y are uniformly divided into M and N intervals respectively using the equal-width binning method, and all data pairs (x, y, y) are traversed. t ,y t Determine which two-dimensional interval (i,j) it falls into, where i is a variable of x. t The X interval number to which it belongs, j is y t The Y-interval number to which it belongs. Count the number n samples falling into each two-dimensional interval (i,j). ij Calculate the joint probability P i,j P i,j =n ij / T 总 T 总 This represents the total number of sample logs.

[0107] Secondly, the joint entropy H(X,Y) is calculated according to the definition in information theory, as shown in formula (20). H under normal conditions is obtained using data from the normal time period. base Calculate its change H is the normalized joint entropy.

[0108] (20) S4. Fall detection is performed on the target object based on a three-level progressive detection mechanism.

[0109] Level 1 physiological detection (L1) focuses on detecting physiological characteristics, specifically identifying autonomic nervous system imbalances early on based on HRV nonlinear characteristics. Inputting the SD1 / SD2 ratio, MSE values ​​at scales 1-3, short-term scaling index α1, and deterministic DET, it calculates the standard deviation of the current HRV characteristic values. If the mean is not within ±2 standard deviations, a Level 1 alarm is triggered. Autonomic nervous system abnormalities can occur tens of seconds to minutes before a fall, and L1 provides the earliest possible warning.

[0110] The second-level neuromuscular synergy test (L2) confirms whether neuromuscular control is disordered. Input the R value of CPC, BRS of MIHR, Z score of HTMV, and joint entropy H. Each coupling value is compared with a preset threshold, as shown in formula (21). A voting mechanism is adopted. If more than half of the coupling indicators are abnormal after L1 warning, it indicates that physiological abnormalities have affected neuromuscular control and the risk of falling has increased.

[0111] (twenty one) Level 3 instability confirmation detection (L3) uses motion dynamics features to ultimately confirm whether a fall has occurred. The preset biomechanical rules include multiple rules, each rule corresponding to a feature threshold and a weight. If the biomechanical features meet the feature thresholds of one or more rules, the weights corresponding to the currently met rules are added together to obtain a rule score of 0-1. The fall pattern library contains multiple patterns. Biomechanical features are matched with each pattern, and the score with the highest degree of matching is used as the pattern matching score. The rule score and pattern matching score are weighted and summed to obtain the overall confidence score of the third-level detection. If the overall confidence score is greater than or equal to the preset confidence threshold, the third-level alarm is triggered.

[0112] Specifically, the absolute value of the vertical velocity of the center of mass reflects the descent speed, the absolute value of the trunk angular velocity reflects the rotational speed, the stability margin reflects the resistance to tipping, the limb-trunk coupling degree reflects coordination, and the motion complexity reflects the randomness of the motion. The rules include: rapid descent rule (vertical velocity of the center of mass > 1.2 m / s, weight 0.30), rapid rotation rule (trunk angular velocity > 2.0 radians / s, weight 0.20), unstable support rule (stability margin < 0.05 m, weight 0.25), loss of coordination rule (limb-trunk coupling degree < 0.6, weight 0.15), and motion simplification rule (motion complexity < 2.0, weight 0.10). The weights of the triggering rules are summed to obtain a rule score of 0-1.

[0113] The system matches the data against a predefined fall pattern library, including forward fall, backward fall, side fall, slip, and slow fainting patterns. Forward fall is a combination of rapid descent, rapid rotation, and instability; backward fall is a combination of rapid descent, rapid rotation, and loss of coordination; side fall is a combination of rapid descent, even faster rotation, and severe instability; slip is a combination of moderate descent, low complexity, and low coordination; and slow fainting is a combination of slow descent, instability, and low complexity. Higher matching scores result in higher scores, ranging from 0 to 1. The rule score is weighted at 60%, and the pattern matching score at 40%. The rule score and pattern matching score are then weighted and summed. The overall confidence score is calculated as 0.6 × rule score + 0.4 × matching score. An alarm is triggered if the overall confidence score is ≥ 0.7 and the alarm is triggered within 1 second after the second-level alarm.

[0114] A fall event is determined only if all of the following conditions are met simultaneously: three alarm levels are triggered sequentially, the time interval between each level is within 1 second, the duration is at least 3 seconds, and the overall confidence level of the third level is ≥0.7. The thresholds for triggering alarms based on these characteristics are updated every 7 days, i.e., ±2 standard deviations of the current mean. If a false alarm occurs, all thresholds are increased by 5%; if a false alarm occurs, all thresholds are decreased by 5%.

[0115] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0116] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A millimeter-wave radar fall detection method integrating HRV nonlinear features and 3D point clouds, characterized in that, Includes the following steps: Based on millimeter-wave radar signals, the inter-heartbeat IBI sequence and 3D point cloud of the target to be detected are acquired in parallel. HRV nonlinear physiological features were extracted based on the IBI sequence, and biomechanical features were extracted based on the 3D point cloud. The nonlinear physiological features of HRV and the biomechanical features were aligned to a unified time axis, and feature coupling calculations were performed to obtain four types of coupling indices. The four types of coupling indices are: cardiac-gait phase coupling, exercise-induced heart rate response, cardiac-triggered motion variability, and joint entropy. The cardiac-gait phase coupling is used to quantify the phase-locked relationship between the heartbeat cycle and the gait cycle; The calculation process is as follows: extract the vertical displacement signal from the foot point cloud trajectory, and use Hilbert transform to obtain the analytical signal to obtain the first phase value at the current moment; based on the IBI sequence, linear interpolation is used to obtain the second phase value at the current moment, and the synchronization index R is obtained by calculating the phase difference between the first phase value and the second phase value; if the synchronization index R is less than the preset value, it indicates that the coordination of heartbeat and gait is disrupted. The exercise-induced heart rate response is used to determine the effect of exercise intensity on heart rate and to assess baroreflex sensitivity. The calculation process includes: selecting the vertical acceleration of the center of mass to represent the intensity of motion and the interbeat period to represent the heart rate response, calculating the coherence and transfer function of the two in the low frequency band, obtaining the average gain, and deriving the pressure reflection sensitivity BRS. The cardiac-triggered motion variability is used to detect the stability of trunk micro-movements after each heartbeat, reflecting the strength of myocardial coupling. The calculation process includes: detecting the heartbeat time in the IBI sequence, taking a time window after each heartbeat time, extracting the trunk acceleration signal, calculating the standard deviation of acceleration within each window, and calculating the Z-score of the standard deviation relative to the baseline. The Z-score is used to quantify the degree to which the standard deviation of the current cardiac-triggered micro-movements deviates from the individual's normal baseline. The joint entropy is used to measure the degree of interdependence between the physiological and motor subsystems from an information theory perspective. Fall detection is performed on the target object based on a three-level progressive detection mechanism; The first level of detection is to calculate whether the mean of the HRV nonlinear physiological characteristics is within a preset multiple of the standard deviation. If it is not, the first level alarm is triggered. The second level of detection is as follows: after the first level alarm, each type of coupling index is compared with its respective preset threshold. A voting mechanism is used. If more than half of the coupling indexes are abnormal, the second level alarm is triggered. The third level of detection is as follows: after the second level alarm, it is determined whether the biomechanical characteristics match the preset biomechanical rules and fall pattern library. If they match, the third level alarm is triggered.

2. The millimeter-wave radar fall detection method integrating HRV nonlinear features and 3D point clouds as described in claim 1, characterized in that, The extraction process of the HRV nonlinear physiological features includes: After preprocessing the collected millimeter-wave radar data, it is input into a pre-trained neural network model to output the IBI sequence; Based on the IBI sequence, Poincaré plot features and multi-scale entropy are calculated, and detrended fluctuation analysis and recursive quantitative analysis are performed on the IBI sequence to obtain detrended fluctuation features and recursion matrix; the Poincaré plot features, the multi-scale entropy, the detrended fluctuation features, and the proportion of recursive points on the diagonal of the recursion matrix to all recursive points are used as the nonlinear physiological features of HRV.

3. The millimeter-wave radar fall detection method integrating HRV nonlinear features and 3D point clouds as described in claim 2, characterized in that, When training the neural network model, the sample data includes millimeter-wave radar signal samples of different genders, body types, clothing materials, postures, and environments, and the data is labeled according to time points, including activity type, physiological state, movement state, and body posture.

4. The millimeter-wave radar fall detection method integrating HRV nonlinear features and 3D point clouds as described in claim 2, characterized in that, The neural network model includes, in turn, an input and batch normalization layer, a first convolutional activation layer, a max pooling layer, a depthwise separable convolutional layer, an average pooling layer, a second convolutional activation layer, a global average pooling and fully connected layer, and a dropout and output layer. The input and batch normalization layer is used to standardize the input signal, providing input data with a unified format and fixed dimensions for all subsequent layers; The first convolutional activation layer is used to extract local phase change features related to the IBI sequence, expanding the single-channel input into multiple feature channels to capture different types of heartbeat-related patterns; The max pooling layer is used to take the maximum value of each channel output by the first convolutional activation layer within a window of 5 sampling points, thus preserving the peak characteristics of the heartbeat signal; The depthwise separable convolutional layer is used to independently convolve each feature channel, capture the temporal correlation features of the signal, and strengthen the feature mapping relationship between the heartbeat cycle and the IBI sequence. The average pooling layer is used to downsample the features output by the depth-separable convolutional layer to smooth feature map fluctuations. The second convolutional activation layer is used to capture heartbeat cycle features over a longer time span; The global average pooling and fully connected layers are used to average the entire time series of each feature channel and perform non-linear combinations. The dropout and output layer is used to randomly drop out some neurons and output multiple consecutive predicted IBI values.

5. The millimeter-wave radar fall detection method integrating HRV nonlinear features and 3D point clouds as described in claim 1, characterized in that, The extraction process of the biomechanical features includes: Fast Fourier transforms of the original millimeter-wave radar signal in the range, Doppler, and angle dimensions are performed to obtain a three-dimensional spectrum. The constant false alarm rate (CFAR) detection algorithm is used to extract the target point from the three-dimensional spectrogram. The DBSCAN algorithm is used to cluster point clouds belonging to the same target, distinguishing human body parts such as head, torso, upper limbs and lower limbs, and outputting a structured set of point cloud clusters as the initial observation set. Each point cloud cluster represents a potential body part. Inter-frame correlation is performed on each point cloud cluster to generate continuous motion trajectories for each human body part; Based on the continuous motion trajectory, the centroid dynamics features, attitude angle features, segmental coordination features, stability margin features, and motion complexity features are extracted as the biomechanical features.

6. The millimeter-wave radar fall detection method fusing HRV nonlinear features and 3D point clouds as described in claim 5, characterized in that, The centroid dynamics features include the three-dimensional coordinates of the centroid, the vertical velocity, the vertical acceleration, and the rate of change of the centroid height; wherein, the three-dimensional coordinates of the centroid are calculated by assigning weight coefficients to each cluster of point clouds through a human biomechanical model, and then weighting the three-dimensional spatial coordinate vectors of each human body part point cloud cluster with their respective weight coefficients to obtain the three-dimensional coordinates of the centroid. The attitude angle feature is the torso angular velocity, which is calculated as follows: perform principal component analysis on the torso point cloud cluster, extract the first principal component vector as the torso longitudinal axis direction, take the angle between the torso longitudinal axis and the horizontal plane as the pitch angle, and take the angle between the projection of the torso longitudinal axis on the horizontal plane and the forward direction as the roll angle; calculate the torso angular velocity based on the rate of change of the pitch angle and the rate of change of the roll angle over time. The segmental coordination feature is calculated as follows: extract the point cloud trajectories of the head and pelvis, calculate the vertical displacement signals respectively, perform cross-correlation analysis to obtain the phase difference between the head and pelvis; calculate the coupling degree between the limbs and trunk based on the Pearson correlation coefficient of the limb and trunk velocities. The stability margin feature is the shortest distance between the projection point of the centroid on the ground and the boundary of the supporting polygon. The motion complexity feature is the Shannon entropy of the center-of-mass velocity.

7. The millimeter-wave radar fall detection method integrating HRV nonlinear features and 3D point clouds as described in claim 1, characterized in that, The second-level detection is used to determine whether neuromuscular control is disordered. The criteria for judging various coupling indicators are as follows: Where CPC represents the cardiac-gait phase coupling index, R represents the synchronization index between cardiac and gait; MIHR represents the exercise-induced heart rate response, BRS represents baroreflex sensitivity; and Z represents the Z-score of the cardiac-triggered motion variability index. This represents the joint entropy.

8. The millimeter-wave radar fall detection method integrating HRV nonlinear features and 3D point clouds as described in claim 1, characterized in that, The biomechanical rules include multiple rules, each rule corresponding to a feature threshold and a weight. If the biomechanical feature satisfies the feature threshold of one or more rules, the weights corresponding to the currently satisfied one or more rules are added together to obtain a rule score of 0-1. The fall pattern library contains multiple patterns. The biomechanical features are matched with each pattern, and the score with the highest degree of matching is taken as the pattern matching score. The rule score and pattern matching score are weighted and summed to obtain the overall confidence level of the third-level detection. If the overall confidence level is greater than or equal to a preset confidence threshold, a third-level alarm is triggered.

9. The millimeter-wave radar fall detection method integrating HRV nonlinear features and 3D point clouds as described in claim 8, characterized in that, The biomechanical rules include: rapid descent rule, rapid rotation rule, instability rule, loss of coordination rule, and motion simplification rule; wherein, the rapid descent rule is: vertical velocity of the center of mass > characteristic threshold one; the rapid rotation rule is: trunk angular velocity > characteristic threshold two; the instability rule is: stability margin < characteristic threshold three; the loss of coordination rule is: coupling degree between limbs and trunk < characteristic threshold four; and the motion simplification rule is: motion complexity < characteristic threshold five. The fall pattern library includes the following patterns: forward fall, backward fall, side fall, slip, and slow fainting.

Images