A fall detection method based on millimeter-wave radar

Abstract

This invention discloses a fall detection method based on millimeter-wave radar, comprising the following steps: S1: The millimeter-wave radar continuously transmits frequency-modulated continuous wave signals and receives echo signals reflected by the human body within the monitoring area; S2: The echo signals are processed to generate human point cloud time-series data containing distance, orientation, velocity, and micro-motion information; S3: Multi-dimensional feature vectors representing human posture and motion state are extracted from the human point cloud time-series data; S4: The multi-dimensional feature vectors are input into a primary fall recognition model to calculate a preliminary fall risk probability value; S5: If the preliminary fall risk probability value exceeds a first set threshold, it is determined to be a suspected fall event, and a secondary precise authentication process is immediately initiated; the secondary precise authentication process includes: initiating non-contact vital sign monitoring and retrieving the user's recent behavioral dynamic baseline model for comparative analysis. This invention enables more accurate fall detection.

Description

Technical Field

[0001] This invention relates to the field of monitoring methods, and more specifically to a fall monitoring method based on millimeter-wave radar. Background Technology

[0002] In home-based and institutional elderly care settings, falls have become the primary accidental risk threatening the safety of the elderly. Their reduced reaction time makes them more susceptible to accidents after a fall. However, current elderly care monitoring faces numerous technical challenges: traditional contact-based monitoring solutions (such as wearable sensors and emergency call devices) rely on the elderly's active cooperation, leading to extremely low compliance due to discomfort or forgetting to wear them, thus failing to achieve 24 / 7 coverage; video surveillance solutions, while providing intuitive monitoring, pose serious privacy risks and are limited by lighting conditions (nighttime, backlighting), obstructions, etc., making it difficult to meet the needs of non-intrusive monitoring; single-sensor (such as infrared or ultrasound) monitoring solutions can only capture simple movement states, lacking multi-dimensional characterization of human posture and physiological state, resulting in a high false alarm rate for fall detection. This not only increases the workload of caregivers but also easily leads to the "boy who cried wolf" effect, causing caregivers to overlook real falls and indirectly increasing care costs.

[0003] In addition, existing monitoring technologies generally lack personalized adaptation capabilities, fail to consider the differences in individual behavioral habits of the elderly (such as some elderly people moving slowly and having small daily activity ranges), and the fixed models cannot be iteratively optimized, which can easily affect the recognition accuracy in long-term use; at the same time, most solutions do not establish a graded response mechanism, and cannot match rescue resources according to the severity of the fall, resulting in untimely rescue or waste of resources. Therefore, a fall monitoring method based on millimeter-wave radar is proposed. Summary of the Invention

[0004] The present invention solves the above-mentioned technical problems through the following technical solution, and the present invention includes the following steps:

[0005] S1: The millimeter-wave radar continuously transmits frequency-modulated continuous wave signals and receives echo signals reflected by the human body in the monitored area;

[0006] S2: Perform signal processing on the echo signal to generate human point cloud time-series data containing distance, orientation, velocity and micro-motion information;

[0007] S3: Extract multi-dimensional feature vectors representing human posture and motion state from human point cloud temporal data;

[0008] S4: Input the multi-dimensional feature vector into the first-level fall recognition model to calculate the preliminary fall risk probability value;

[0009] S5: If the initial fall risk probability value exceeds the first set threshold, it is determined to be a suspected fall event, and the second-level precise authentication process is immediately initiated; the second-level precise authentication process includes: initiating non-contact vital sign monitoring and retrieving the user's recent behavioral dynamic baseline model for comparative analysis;

[0010] S6: Based on the output of the two-level precise authentication process, a comprehensive judgment is made. If it is confirmed as a fall event, an alarm message containing the event type and emergency level is generated and sent through a preset communication link.

[0011] Furthermore, the extraction of multi-dimensional feature vectors in step S3 includes the following calculation process:

[0012] Calculate the dynamic characteristics of human body centroid height: By performing centroid tracking on continuous point cloud frames, obtain the sequence of human body centroid height changes over time, and then calculate the average rate of change of centroid height and the acceleration of the rate of change within a specific time window T.

[0013] Calculating human posture divergence features: Projecting each frame of point cloud data onto a horizontal plane to form a two-dimensional point set, calculating the covariance matrix of the coordinates of this point set and analyzing its eigenvalue distribution, yields the divergence index D, which characterizes the degree of posture extension or contraction. xy ;

[0014] Calculating motion pattern characteristics: Extracting micro-Doppler signals generated by the slight movements of various parts of the human body from radar echoes, analyzing their spectrum within a short time window, and quantifying the regularity of motion by calculating the entropy value of the spectral energy distribution in each sub-band, thus obtaining the micro-Doppler spectral entropy E. md ;

[0015] rate of change of centroid height , change in acceleration a H Attitude divergence index D xy and micro-Doppler spectral entropy E md Together they form a multi-dimensional feature vector ,in, and a H The specific calculation formula is as follows:

[0016] ;

[0017] Where Ht represents the centroid height at time t, and T is the length of the time window.

[0018] Furthermore, the first-level fall detection model in step S4 is a lightweight gradient boosting decision tree model, which is obtained through offline training; the model calculates the initial fall risk probability value P. preThe process includes: inputting a multi-dimensional feature vector X into an ensemble model composed of multiple decision tree weak learners; each weak learner performs layer-by-layer judgment on the input features and outputs a weight value for a leaf node; summing the output weight values ​​of all weak learners, and then mapping the summation result to the [0,1] interval using the Sigmoid function to obtain the preliminary fall risk probability value P. pre Its calculation formula is expressed as:

[0019] ;

[0020] Where N is the total number of weak learners, w i denoted as the weight value output by the i-th weak learner for the current input multi-dimensional feature vector X.

[0021] Furthermore, the non-contact vital sign monitoring initiated in step S5 includes:

[0022] Separate the phase modulation component, which is mainly caused by chest cavity fluctuations, from the radar echo signal;

[0023] Time-frequency analysis was performed on the phase modulation component to extract estimated respiratory rate and heart rate values ​​within a specific time window before and after the suspected fall event.

[0024] The physiological sign disorder index is calculated by comparing the currently estimated respiratory rate and heart rate with the user's baseline values ​​at rest, calculating the absolute value of the relative deviation, and then weighting and summing the two deviations to quantify the degree of abnormality in the physiological state. The formula is as follows:

[0025] ;

[0026] Among them, f breathe and f heart This is the current estimate. and α is the individual baseline value, and β are preset weighting coefficients.

[0027] Furthermore, the dynamic baseline model of recent user behavior retrieved in step S5 is constructed through continuous learning of users' daily activity patterns. Its comparative analysis used in the secondary precision authentication process specifically includes:

[0028] Extract the feature sequence of point cloud motion trajectory in the period of time before the suspected fall event occurred;

[0029] Calculate the behavioral deviation: Align and compare the feature sequences to be compared with the feature sequences of typical activity patterns in the same time period stored in the dynamic baseline model. Calculate the minimum cumulative distance between the two using the dynamic time warping algorithm. This distance value is the behavioral deviation S. dev ;

[0030] Constructing a comprehensive and accurate certification score: incorporating the physiological sign disorder index I phy Deviation from behavior S dev A comprehensive verification score is obtained by linearly combining the results according to preset weights. The calculation formula is as follows:

[0031] ;

[0032] in, and This is for adjusting the coefficient.

[0033] Furthermore, the comprehensive decision in step S6 is specifically as follows:

[0034] If accurate certification comprehensive score Exceeding the second set threshold Th high If so, it is directly identified as a high-urgency fall event;

[0035] If accurate certification comprehensive score Given the third set threshold Th low With the second set threshold Th high In between, local audio-visual alerts will be activated, and a proactive response assessment period will begin;

[0036] During the proactive response assessment period, assess the user's self-recovery capabilities: continuously monitor and remind users for a set duration. Human point cloud data is used to calculate a recovery index R. recover ;

[0037] This index is calculated by comprehensively evaluating the net rate of increase in center of mass height over the time period and the stability of body posture at the end of the period. Specifically, the net rate of increase in center of mass height is characterized by the ratio of the net increase in height to the evaluation duration, and posture stability is characterized by the posture divergence index D at the end of the period. xy The two are represented by a weighted sum to obtain the recovery index R. recover The calculation formula is:

[0038] ;

[0039] in, This represents the net increase in centroid height during the assessment period. and These are the weighting coefficients;

[0040] If you get up, the recovery index If the value is below the fourth set threshold, it is determined that the user failed to get up effectively on their own and is confirmed as a fall event; otherwise, it is determined as a false alarm and the alarm is canceled.

[0041] Furthermore, the emergency level of the alarm information generated in step S6 is based on the preliminary fall risk probability value P. pre Comprehensive score for accurate certification The emergency level is divided into at least three levels: high, medium, and low, as determined by fuzzy logic rules. Different levels correspond to different alarm notification strategies and contact notification priorities.

[0042] Furthermore, the method also includes a model optimization step:

[0043] Data Collection: The system collects the final feedback results for each alarm event and outputs the corresponding multi-dimensional feature vector X and physiological sign disorder index I. phy Behavioral Deviation S dev and recovery index (If applicable) Store as a labeled training sample;

[0044] Online optimization of judgment parameters: For samples marked as false alarms or voluntarily cancelled, their corresponding accurate authentication comprehensive scores are used. and recovery index To optimize the objective, an online sequential extreme learning machine algorithm is employed, which updates the network output weights using a recursive least squares approach, thereby dynamically adjusting the computation. Weighting coefficients and calculation Weighting coefficients , This makes the system tend to make conservative judgments in similar scenarios;

[0045] Incremental update of the recognition model: The first-level fall recognition model is updated periodically using newly added, correctly labeled confirmed fall samples in an incremental learning manner to improve its recognition performance.

[0046] Compared with existing technologies, this invention has the following advantages: This fall detection method based on millimeter-wave radar transmits and receives signals from millimeter-wave radar to generate human point cloud time-series data. After extracting multi-dimensional feature vectors, it is initially judged by a first-level fall recognition model. When a fall is suspected, a second-level precise authentication process is initiated, which includes non-contact vital sign monitoring and behavioral dynamic baseline model comparison. The system makes a precise judgment by combining a comprehensive score and a recovery index. It can also dynamically adjust parameters and update the model through model optimization steps. It has the beneficial effects of accurate monitoring, effective reduction of false alarm rate, targeted alarm strategy according to emergency level, more convenient and private non-contact monitoring, and continuous model optimization to improve recognition performance, making the system more worthy of promotion and use. Attached Figure Description

[0047] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation

[0048] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0049] like Figure 1 As shown, this embodiment provides a technical solution: a fall detection method based on millimeter-wave radar, comprising the following steps:

[0050] S1: The millimeter-wave radar continuously transmits frequency-modulated continuous wave signals and receives echo signals reflected by the human body in the monitored area;

[0051] S2: Perform signal processing on the echo signal to generate human point cloud time-series data containing distance, orientation, velocity and micro-motion information;

[0052] S3: Extract multi-dimensional feature vectors representing human posture and motion state from human point cloud temporal data;

[0053] S4: Input the multi-dimensional feature vector into the first-level fall recognition model to calculate the preliminary fall risk probability value;

[0054] S5: If the initial fall risk probability value exceeds the first set threshold, it is determined to be a suspected fall event, and the second-level precise authentication process is immediately initiated; the second-level precise authentication process includes: initiating non-contact vital sign monitoring and retrieving the user's recent behavioral dynamic baseline model for comparative analysis;

[0055] S6: Based on the output of the two-level precise authentication process, a comprehensive judgment is made. If it is confirmed as a fall event, an alarm message containing the event type and emergency level is generated and sent through a preset communication link.

[0056] Step S3, extracting multi-dimensional feature vectors, includes the following calculation process:

[0057] Calculate the dynamic characteristics of human body centroid height: By performing centroid tracking on continuous point cloud frames, obtain the sequence of human body centroid height changes over time, and then calculate the average rate of change of centroid height and the acceleration of the rate of change within a specific time window T.

[0058] Calculating human posture divergence features: Projecting each frame of point cloud data onto a horizontal plane to form a two-dimensional point set, calculating the covariance matrix of the coordinates of this point set and analyzing its eigenvalue distribution, yields the divergence index D, which characterizes the degree of posture extension or contraction. xy ;

[0059] Calculating motion pattern characteristics: Extracting micro-Doppler signals generated by the slight movements of various parts of the human body from radar echoes, analyzing their spectrum within a short time window, and quantifying the regularity of motion by calculating the entropy value of the spectral energy distribution in each sub-band, thus obtaining the micro-Doppler spectral entropy E. md ;

[0060] rate of change of centroid height , change in acceleration a H Attitude divergence index D xy and micro-Doppler spectral entropy E md Together they form a multi-dimensional feature vector ,in, and a H The specific calculation formula is as follows:

[0061] ;

[0062] Where Ht represents the centroid height at time t, and T is the length of the time window;

[0063] By extracting the rate of change of centroid height ΔH and the acceleration a H Attitude divergence index D xy and micro-Doppler spectral entropy E md Four core features construct a multi-dimensional feature vector, comprehensively depicting the human body state from three key dimensions: vertical motion dynamics, human posture and morphology, and regularity of movement patterns. This avoids the limitation that a single feature cannot distinguish similar behaviors (such as bending over to pick up an object versus falling), providing highly recognizable and strongly correlated input for subsequent fall recognition models. It significantly improves the accuracy of initial fall judgments, and the feature calculation logic is rigorous, highly operable, and adaptable to engineering implementation needs.

[0064] The rate of change of the height of the center of mass ΔH and the acceleration a H calculate:

[0065] Set the time window length T=1.5s, and obtain the data through point cloud frame centroid tracking:

[0066] The height of the center of mass H at time t t =0.4m;

[0067] At time tT, the height of the center of mass H t-T =1.5m;

[0068] The height of the center of mass H at time t-2T t-2T =1.6m;

[0069] Calculated according to the formula:

[0070] Rate of change of centroid height: ;

[0071] First, calculate the rate of change at time tT: ;

[0072] Change in acceleration: ;

[0073] Negative ΔH and a H This indicates that the height of the human body's center of gravity drops rapidly and accelerates, which is consistent with the characteristics of vertical motion when falling (ΔH is approximately -0.2 to -0.3 m / s when bending over normally, with no obvious negative acceleration).

[0074] Attitude divergence index D xy calculate:

[0075] Select a two-dimensional point set projected onto a horizontal plane from a given frame of point cloud: {(0.8,1.2),(0.9,1.3),(0.7,1.1),(1.0,1.4),(0.6,1.0)}

[0076] Calculation steps:

[0077] Mean calculation: ;

[0078] Covariance matrix calculation: ,in:

[0079] ;

[0080] ;

[0081] ;

[0082] Right now ;

[0083] Eigenvalue solving: through the characteristic equation ;

[0084] Attitude divergence index: (ε is to avoid the minimum value where the denominator is 0)

[0085] The largest D xy The value indicates that the human body is in an extremely curled-up posture, consistent with the curled-up state of the body after a fall (D when standing normally). xy Approximately 3-5, and approximately 6-8 when walking.

[0086] Micro-Doppler spectral entropy E md calculate:

[0087] The spectrum of the micro-Doppler signal within a short time window is divided into 10 sub-bands, and the energy distribution (after normalization) is: [0.02, 0.03, 0.05, 0.08, 0.4, 0.3, 0.06, 0.03, 0.02, 0.01];

[0088] Calculated using the entropy formula: ( (For the child to bring normalized energy)

[0089] Calculation process:

[0090] ;

[0091] The results indicate that during a fall, human movement becomes chaotic, and the spectral energy distribution is more uniform, E md It will rise to above 2.5; during normal activities (such as walking), the movement pattern shows that energy is concentrated in a few subbands, E md Typically between 1.0 and 1.8, this example matches the movement pattern characteristics of a suspected fall.

[0092] The first-level fall detection model in step S4 is a lightweight gradient boosting decision tree model, which is obtained through offline training; the model calculates the initial fall risk probability value P. pre The process includes: inputting a multi-dimensional feature vector X into an ensemble model composed of multiple decision tree weak learners, with each weak learner making layer-by-layer judgments on the input features and outputting the weight value of a leaf node;

[0093] The output weights of all weak learners are summed, and the sum is then mapped to the [0,1] interval using the Sigmoid function to obtain the initial fall risk probability value P. pre Its calculation formula is expressed as:

[0094] ;

[0095] Where N is the total number of weak learners, w i Let be the weight value output by the i-th weak learner for the current input multi-dimensional feature vector X;

[0096] A lightweight gradient boosting decision tree model is adopted as the first-level fall detection model, which balances recognition accuracy and computational efficiency and is suitable for real-time monitoring scenarios. By integrating multiple decision tree weak learners and mapping them with the Sigmoid function, the feature vector is transformed into an intuitive fall risk probability value in the [0,1] interval, avoiding the one-sidedness of a single model. Moreover, the model has a fast inference speed after offline training and can quickly output the preliminary judgment result, providing an accurate and efficient basis for the subsequent second-level accurate authentication process, effectively balancing real-time performance and reliability.

[0097] Continuing with the previous scenario (human body suspected of falling, multi-dimensional feature vectors) The total number of weak learners in the first-level fall recognition model is set to N=5. After offline training, the output weight values ​​of each weak learner for the input feature vector X are as follows: w1=2.1, w2=1.8, w3=2.3, w4=1.9, w5=2.2 (the weights are positive and relatively large, because the input features meet the typical features of a fall, the weak learners tend to judge the risk of a fall).

[0098] Sum of weak learner weights: ;

[0099] Substitute the values ​​into the Sigmoid function to calculate the initial probability of falling: ;

[0100] Verification of the calculation process: ,therefore

[0101] A value close to 1 indicates an extremely high risk of falling. If the first threshold is set to 0.8, this value exceeds the threshold and is judged as a suspected fall event, initiating the second-level precise authentication process. If it is a normal activity scenario (such as bending over to pick up an object), the feature vector... The output weights of the weak learner are w1=0.3, w2=0.2, w3=0.4, w4=0.1, and w5=0.2, respectively. Summing them up gives... Calculated If the value is below the threshold of 0.8, it will not be judged as a suspected fall, thus avoiding false alarms and demonstrating the model's ability to effectively distinguish between falls and normal activities.

[0102] The non-contact vital sign monitoring initiated in step S5 includes:

[0103] Separate the phase modulation component, which is mainly caused by chest cavity fluctuations, from the radar echo signal;

[0104] Time-frequency analysis was performed on the phase modulation component to extract estimated respiratory rate and heart rate values ​​within a specific time window before and after the suspected fall event.

[0105] The physiological sign disorder index is calculated by comparing the currently estimated respiratory rate and heart rate with the user's baseline values ​​at rest, calculating the absolute value of the relative deviation, and then weighting and summing the two deviations to quantify the degree of abnormality in the physiological state. The formula is as follows:

[0106] ;

[0107] Among them, f breathe and f heart This is the current estimate. and For each individual, α and β are preset weighting coefficients;

[0108] The system extracts vital signs information such as respiratory rate and heart rate from radar echoes in a non-contact manner, and calculates the physiological sign disorder index I by combining it with the user's personal resting baseline. phy It can quantify the degree of abnormality in the human body's physiological state after a suspected fall without contact, avoiding misjudgment of falls due to external behavior, providing objective and accurate physiological dimensions for secondary accurate authentication, while protecting user privacy, adapting to private monitoring scenarios such as home, and improving the reliability and applicability of secondary verification.

[0109] By separating the phase modulation component related to chest cavity fluctuations from radar echoes in a non-contact manner, respiratory rate and heart rate are extracted and combined with personal baseline values ​​to calculate the physiological sign disorder index. This method can quantify the degree of abnormality in human physiological state without contact, avoid misjudgment caused by individual differences, and provide objective physiological dimension basis for accurate secondary authentication. It protects user privacy and is suitable for private monitoring scenarios such as home, effectively improving the reliability and pertinence of secondary verification.

[0110] Continuing from the previous suspected fall scenario (multi-dimensional feature vector) Preliminary probability of falling The parameters are set as follows:

[0111] User's personal resting state baseline: respiratory rate Heart rate ;

[0112] The preset weighting coefficients are α=0.5 (respiratory rate deviation weight) and β=0.5 (heart rate deviation weight).

[0113] Calculation of the physiological sign disorder index in suspected fall scenarios:

[0114] Based on time-frequency analysis of the phase modulation components of the radar echo, the estimated values ​​of current vital signs are extracted:

[0115] Current respiratory rate ;

[0116] Current heart rate ;

[0117] Calculated according to the original formula:

[0118] ;

[0119] Calculation process:

[0120] ;

[0121] ;

[0122] ;

[0123] Comparison of normal activity scenarios (such as bending over to pick up an item), feature vectors );

[0124] Extract current vital sign estimates:

[0125] Current respiratory rate ;

[0126] Current heart rate ;

[0127] Calculation process:

[0128] ;

[0129] ;

[0130] ;

[0131] I in suspected fall scenario phy =0.4 (a relatively large value, indicating that the fall caused significant physiological disruption), I under normal activity scenarios phy ≈0.0456 (extremely small value, indicating stable physiological state). This index can accurately distinguish between physiological abnormalities caused by actual falls and physiological fluctuations during normal activities, providing key physiological dimension support for accurate secondary certification and avoiding misjudgment based on a single behavioral feature.

[0132] The dynamic baseline model of recent user behavior retrieved in step S5 is constructed by continuously learning the user's daily activity patterns. Its comparative analysis used in the secondary precision authentication process specifically includes:

[0133] Extract the feature sequence of point cloud motion trajectory in the period of time before the suspected fall event occurred;

[0134] Calculate the behavioral deviation: Align and compare the feature sequences to be compared with the feature sequences of typical activity patterns in the same time period stored in the dynamic baseline model. Calculate the minimum cumulative distance between the two using the dynamic time warping algorithm. This distance value is the behavioral deviation S. dev ;

[0135] Constructing a comprehensive and accurate certification score: incorporating the physiological sign disorder index I phy Deviation from behavior S dev A comprehensive verification score is obtained by linearly combining the results according to preset weights. The calculation formula is as follows:

[0136] ;

[0137] in, and For adjustment coefficients;

[0138] The behavioral dynamic deviation is obtained by calculating the similarity between the current behavioral feature time series and the individual's proprietary behavioral dynamic baseline sequence using a dynamic time warping algorithm. Then, it is integrated with the degree of physiological sign disorder according to preset weights to form a comprehensive verification score, realizing in-depth verification of "behavioral temporal features + physiological state" in two dimensions. The DTW algorithm adapts to the time scale differences of behavioral actions, and the personalized baseline model matches the user's behavioral habits to avoid misjudgment by general standards. The comprehensive verification score quantifies the degree of abnormality to make up for the limitations of a single dimension, accurately distinguishes between real falls and similar behaviors, greatly reduces the false alarm rate, and the baseline model can be dynamically updated to ensure the adaptability and recognition stability of long-term monitoring.

[0139] Continuing with the previous unified scenario (suspected fall scenario: initial fall risk probability P) pre ≈0.99995, physiological sign disorder Iphy=0.4;

[0140] Normal activity scenario: bending over to pick up an object, Iphy≈0.0456, with the following parameters set:

[0141] Behavioral characteristics time series: Selecting the rate of change of centroid height Attitude divergence D xy This constitutes a two-dimensional temporal feature with a sequence length N=5 (time interval 0.5s).

[0142] Personal behavioral dynamic baseline sequence (normal activity):

[0143] ;

[0144] Preset weights: (Physiological dimension weight) (Behavioral dimension weights);

[0145] Euclidean distance (feature point similarity): ;

[0146] DTW distance (time series similarity):

[0147] ;

[0148] Behavioral dynamic deviation (normalized): , The preset maximum DTW distance;

[0149] Overall verification score: ;

[0150] Comprehensive verification score calculation for suspected fall scenarios:

[0151] Current behavioral feature time series:

[0152] ;

[0153] DTW distance and behavioral dynamic deviation:

[0154] Final value of the cumulative distance matrix: ;

[0155] Behavioral dynamic deviation: ;

[0156] Overall verification score:

[0157] ;

[0158] Calculation of comprehensive verification score under normal activity scenario (bending over to pick up an object):

[0159] Current behavioral feature time series:

[0160] ;

[0161] DTW distance and behavioral dynamic deviation:

[0162] Final value of the cumulative distance matrix: ;

[0163] Behavioral dynamic deviation: ;

[0164] Overall verification score:

[0165] ;

[0166] Results Explanation: A comprehensive verification score threshold of 0.3 was set for suspected fall scenarios. It was determined to be a genuine fall;

[0167] Normal activity scenarios By eliminating false positives, the two-dimensional scoring system effectively improves the accuracy of fall detection.

[0168] The comprehensive decision in step S6 is as follows:

[0169] If accurate certification comprehensive score Exceeding the second set threshold Th high If so, it is directly identified as a high-urgency fall event;

[0170] If accurate certification comprehensive score Given the third set threshold Th low With the second set threshold Th high In between, local audio-visual alerts will be activated, and a proactive response assessment period will begin;

[0171] During the proactive response assessment period, assess the user's self-recovery capabilities: continuously monitor and remind users for a set duration. Human point cloud data is used to calculate a recovery index R. recover ;

[0172] This index is calculated by comprehensively evaluating the net rate of increase in center of mass height over the time period and the stability of body posture at the end of the period. Specifically, the net rate of increase in center of mass height is characterized by the ratio of the net increase in height to the evaluation duration, and posture stability is characterized by the posture divergence index D at the end of the period. xy The two are represented by a weighted sum to obtain the recovery index R. recover The calculation formula is:

[0173] ;

[0174] in, This represents the net increase in centroid height during the assessment period. and These are the weighting coefficients;

[0175] If you get up, the recovery index If the value is below the fourth preset threshold, it is determined that the user failed to get up effectively on their own, and it is confirmed as a fall event; otherwise, it is determined as a false alarm and the alarm is canceled.

[0176] By setting multiple threshold levels, a differentiated comprehensive judgment is made on the accurate authentication comprehensive score. Falls exceeding the second threshold are directly identified as high-urgency falls, enabling rapid response. Falls between the third and second thresholds trigger audio-visual alerts and an active response assessment period, combined with the recovery index R. recover Taking into account both the net rise rate of the center of gravity and posture stability, it accurately distinguishes between "real falls where the person cannot get up on their own" and "false alarm scenarios where the person can recover on their own." This avoids missed reports of high-emergency falls and significantly reduces unnecessary and invalid alarms, improving the flexibility and accuracy of the judgment. At the same time, it adapts to different fall severity and user recovery capabilities, ensuring the practicality and reliability of the monitoring.

[0177] Continuing with the previous unified scene parameters (accurate authentication comprehensive score for suspected fall scenarios) Intermediate threshold scenario 1: Intermediate threshold scenario 2: Normal activity scenario: bending down to pick up an item. Added judgment parameters:

[0178] Set threshold: Second set threshold The third threshold is set. The fourth threshold is set (the recovery index threshold). ;

[0179] The active response assessment period is τ=3s, with weighting coefficients θ=0.5 (weight of net rise rate of center of mass height) and η=0.5 (weight of attitude stability).

[0180] Core Formula (Recovery Index): ;

[0181] in, The net increase in centroid height during the assessment period. (After normalization, the closer to 1, the more curled up the posture; the closer to 0, the more extended and stable the posture.)

[0182] High-urgency fall scenarios ;

[0183] because If the comprehensive score exceeds the second set threshold, it is directly identified as a high-urgency fall event, and an alarm message is immediately sent through the preset communication link (such as being pushed to the family's mobile phone and the community emergency center simultaneously) to ensure a rapid response in high-risk scenarios.

[0184] Intermediate threshold scenario 1 (unable to get up on their own) );

[0185] because Activate local audio-visual alerts (such as a buzzer and flashing indicator lights) and simultaneously enter the proactive response evaluation period:

[0186] Monitoring data during the assessment period: After a fall, the user had difficulty getting up on their own, and the center of gravity height increased slightly from 0.4m to 0.5m, a net increase. The final posture remains curled up, and the normalized posture divergence D xy =0.9;

[0187] Recovery Index Calculation:

[0188] ;

[0189] result: If the system determines that the user failed to get up on their own, it is confirmed as a medium-to-emergency fall event, and an alarm message is sent.

[0190] Intermediate threshold scenario 2 (able to stand up independently, );

[0191] because Initiate local audio-visual alerts and enter the proactive response assessment period:

[0192] Monitoring data during the assessment period: A user's suspected fall was triggered by "quickly bending over to pick up an item." After being alerted, the user immediately stood up on their own, and the center of gravity height increased from 0.6m to 1.7m, a net increase of [value missing]. ;

[0193] At the end of the term, the student resumed standing posture, and the normalized postural divergence D was calculated. xy =0.2;

[0194] Recovery Index Calculation:

[0195] ;

[0196] result: Approaching but still below Further extend the evaluation period by 1 second (this can be flexibly set in actual scenarios). After the extension... Recalculated, we get: It was ultimately determined to be a false alarm, the alarm was canceled, and the scene was recorded for model optimization.

[0197] Low-scoring non-fall scenarios ;

[0198] because If the conditions for activating the sound and light alert are not met, it will be directly judged as a non-fall event and no alarm action will be taken to avoid interference with normal activities.

[0199] By correcting The accurate value, which is related to The threshold determination logic is more rigorous; combined with the dynamic assessment of the recovery index, it ensures that high-risk falls (such as...) are effectively managed. It ensures no scenario is missed, and through secondary evaluation, it filters out suspected scenarios that can be autonomously recovered, effectively balancing the timeliness and accuracy of alarms.

[0200] By combining multi-level thresholds with the recovery index, a rapid response to high-emergency falls is achieved, while accurately filtering out false alarm scenarios that can recover on their own. This effectively balances the timeliness and accuracy of alarms and avoids issues of missed or false alarms.

[0201] The emergency level of the alarm information generated in step S6 is based on the preliminary fall risk probability value P. pre Comprehensive score for accurate certification The emergency level is divided into at least three levels: high, medium, and low, as determined by fuzzy logic rule mapping. Different levels correspond to different alarm notification strategies and contact notification priorities.

[0202] By combining preliminary fall risk probability and precise authentication comprehensive score as dual core parameters, the system achieves accurate mapping of emergency levels through fuzzy logic rules, classifying events into high, medium, and low-risk levels for differentiated responses. This ensures rapid linkage of emergency resources (family members, emergency centers) for high-risk falls, while avoiding excessive use of rescue resources for medium and low-risk events. It allows rescuers to anticipate the severity of events in advance, improving the targeting and efficiency of rescue efforts. At the same time, it adapts to the response needs of different fall scenarios (serious falls, minor falls, accidental triggers), enhancing the system's practicality and user-friendliness, and reducing the interference of invalid alarms on users and rescuers.

[0203] Continuing with the previous scenario parameters, set the following rules and parameters:

[0204] Fuzzy logic mapping rules (based on common fuzzy inference strategies in the industry, quantifying the association between two parameters and urgency level):

[0205] High urgency: and (Both parameters indicate high risk);

[0206] Medium urgency: and (Both parameters are of medium risk);

[0207] Low urgency: or (Single or dual parameters indicate low risk, excluding non-fall events);

[0208] Alarm strategy and contact priority:

[0209] High urgency: Priority 1 (emergency center) + Priority 2 (immediate family members) + Priority 3 (community health service center), simultaneously pushing event location and physiological signs data;

[0210] Medium urgency: Priority 2 (immediate family members) + Priority 3 (community health service center), push event information + automatic escalation if not confirmed within 10 minutes;

[0211] Low urgency: Only priority 2 (immediate family members) receive an app pop-up notification; no real-time response is required.

[0212] High-urgency fall scenarios:

[0213] Core parameters: , ;

[0214] Fuzzy logic judgment: and This satisfies the high urgency rule;

[0215] Alarm Execution: Immediately and simultaneously dial emergency numbers (push user location, I)phy Physiological data such as 0.4) are used to send a voice alarm with location information to the family's mobile phone. The community health service center receives the incident work order at the same time, realizing rapid multi-channel linkage rescue.

[0216] Moderate to severe fall scenarios:

[0217] Key parameters: Simulates a scenario of "minor fall but inability to get up on one's own". (Level 1 model indicates high risk) ;

[0218] Fuzzy logic judgment: and This meets the medium urgency rule;

[0219] Alarm execution: A vibration and voice alarm is sent to the family member's mobile phone ("The user is suspected of falling and cannot get up on their own. Please check on them immediately"). The community health service center receives the event notification. If the family member does not click "Handled" within 10 minutes, the system will automatically escalate the rescue request to the community, balancing rescue efficiency and resource consumption.

[0220] Low-urgency fall scenarios:

[0221] Key parameters: Simulates the scenario of "getting up on their own after a minor bump". (Level 1 model determines medium risk) (Between low and medium thresholds);

[0222] Fuzzy logic judgment: but ,and This satisfies the low urgency rule;

[0223] Alarm execution: Only a pop-up notification is sent to the family's APP ("The user exhibits signs of a possible fall, has gotten up on their own, and is in stable condition"), without any additional rescue coordination, to avoid excessive disturbance, while letting the family know the user's status.

[0224] Non-fall low-risk scenarios:

[0225] Key parameters: Normal activity (bending over to pick up an object). , ;

[0226] Fuzzy logic judgment: This meets the low urgency (non-fall) rule;

[0227] Alarm execution: There is no active alarm action; the event is only recorded in the system background and can be viewed in the "History" section of the family's APP, completely avoiding interference from invalid alarms.

[0228] Through dual-parameter fuzzy logic mapping, different risk scenarios are accurately matched with corresponding rescue resources. High-urgency events achieve rapid response within the "golden rescue time," medium- and low-urgency events are rationally allocated resources, and non-fall scenarios are free from interference. This not only ensures user safety but also improves the efficiency of rescue resource utilization, demonstrating the core advantages of tiered response.

[0229] The method also includes a model optimization step:

[0230] Data Collection: The system collects the final feedback results for each alarm event and outputs the corresponding multi-dimensional feature vector X and physiological sign disorder index I. phy Behavioral Deviation S dev and recovery index (If applicable) Store as a labeled training sample;

[0231] Online optimization of judgment parameters: For samples marked as false alarms or voluntarily cancelled, their corresponding accurate authentication comprehensive scores are used. and recovery index To optimize the objective, an online sequential extreme learning machine algorithm is employed, which updates the network output weights using a recursive least squares approach, thereby dynamically adjusting the computation. Weighting coefficients and calculation Weighting coefficients , This makes the system tend to make conservative judgments in similar scenarios;

[0232] Incremental update of the recognition model: The first-level fall recognition model is updated periodically using newly added, correctly labeled confirmed fall samples in an incremental learning manner to improve its recognition performance;

[0233] By collecting labeled training samples and combining them with an online sequential extreme learning machine algorithm to dynamically optimize the weight coefficients (γ, δ, θ, η) of the accurate authentication comprehensive score and the recovery index, the system makes more conservative judgments on false alarm scenarios. At the same time, the first-level fall recognition model is updated in an incremental learning manner to achieve closed-loop iterative optimization of system performance. This reduces the accumulation of false alarms over long-term use and adapts to changes in user behavior and new fall scenarios, ensuring the long-term reliability, accuracy, and personalized adaptability of the monitoring system.

[0234] By collecting monitoring event samples with real labels, and combining them with an online sequential extreme learning machine algorithm to dynamically iteratively optimize the weight coefficients of the accurate authentication comprehensive score and the recovery index, the system makes more conservative judgments on false alarm scenarios caused by the daily activities of the elderly. At the same time, the first-level fall recognition model is continuously updated through incremental learning, realizing a closed-loop upgrade of system performance. This not only adapts to the long-term changes in the individual behavioral habits of the elderly, but also continuously improves the accuracy of fall recognition, reduces the false alarm rate, reduces the ineffective workload of caregivers, and ensures the long-term reliability of non-contact, non-intrusive, and all-weather monitoring. This further reduces the care costs of home-based and institutional elderly care, and the optimization process does not affect the continuity of monitoring or the protection of user privacy.

[0235] Continuing with the previous unified scene parameters, set and optimize relevant parameters:

[0236] Initial weighting coefficient: γ = 0.5 (I phy (weight) (S) dev (Weight) (Weight of net increase rate of centroid height) (Attitude stability weights);

[0237] Online Sequential Extreme Learning Machine (OS-ELM) parameters: number of input layer nodes = 2, number of hidden layer nodes = 10, activation function is Sigmoid, initial weight matrix W(0) is a random small value matrix (value range [-0.1, 0.1]), initial covariance matrix P(0) = I (2×2 identity matrix);

[0238] Incremental learning trigger condition: A first-level model update is initiated every 8 valid labeled samples collected.

[0239] Core optimization formula:

[0240] Recursive Least Squares (RLS) weight update formula:

[0241] ;

[0242] ;

[0243] Where W is the weight coefficient vector to be optimized, X is the sample input feature vector, y is the expected output value, and k is the current optimization sample index;

[0244] Incremental update of loss function for Level 1 model (adapted to lightweight gradient boosting decision tree):

[0245] ;

[0246] Where M is the number of newly added valid samples, yi is the sample label (fall = 1, normal / false alarm = 0), and f(xi) is the model's output score for sample xi.

[0247] The system collected 8 valid samples (labeled) from scenarios involving the elderly, and selected 3 core samples for example calculations:

[0248] Sample 1 (real fall, label=1): Multi-dimensional feature vector Physiological signs disorder level I phy =0.4, behavioral deviation S dev =0.626, Accurate Authentication Overall Score Recovery index ;

[0249] Sample 2 (routine bending over to pick up an item, false alarm, label=0): Multi-dimensional feature vector Physiological signs disorder level I phy =0.0456, behavioral deviation S dev =0.0625, Accurate Authentication Overall Score (No need to get up to recover your index);

[0250] Sample 3 (slowly standing up after prolonged sitting, false alarm, label=0): Multi-dimensional feature vector Physiological signs disorder level I phy =0.06, behavioral deviation S dev =0.08, Accurate Authentication Overall Score Recovery index (Evaluation period) End of term ).

[0251] Online optimization of judgment parameters (for false alarm sample 3, optimization) ):

[0252] optimization (Adaptation to accurate authentication comprehensive score calculation):

[0253] Optimization goal: Reduce the overall score of false alarm scenarios and decrease the probability of falsely triggering the secondary response;

[0254] Sample input feature vector Current overall score output The expected output y(k) = 0.05 (which corresponds to the low score range of normal activities for the elderly).

[0255] Initial state: weight vector covariance matrix ;

[0256] Calculation process:

[0257] Calculate intermediate variables :

[0258] ;

[0259] Update the covariance matrix P(1):

[0260] ;

[0261] ;

[0262] Calculate output deviation :

[0263] ;

[0264] Update weight vector :

[0265] ;

[0266] ;

[0267] Verification of the optimized overall score:

[0268] ;

[0269] The score is slightly lower than the original score of 0.07, which is closer to the score range of normal activity scenarios.

[0270] optimization (Adapted to recovery index calculation);

[0271] Optimization goal: Improve the recovery index in false alarm scenarios, making it easier to determine autonomous recovery and reduce false alarms;

[0272] Sample input feature vector Current recovery index output The expected output y(k) = 0.65 (higher than the fourth set threshold of 0.6, clearly indicating autonomous recovery);

[0273] Initial state: weight vector covariance matrix ;

[0274] Calculation process:

[0275] Calculate intermediate variables :

[0276] ;

[0277] Update the covariance matrix P(1):

[0278] ;

[0279] ;

[0280] Calculate output deviation :

[0281] ;

[0282] Update weight vector :

[0283] ;

[0284] ;

[0285] Optimized recovery index verification:

[0286] ;

[0287] If the value exceeds the fourth set threshold of 0.6, it can be directly determined as an autonomous recovery, thus avoiding false alarms.

[0288] Incremental update of the identification model (based on 8 new samples):

[0289] The newly added sample set contains 3 real fall samples and 5 samples, including 3 false alarm samples and 2 normal samples, which are added to the original model training set for incremental updates;

[0290] The original model outputs a score for sample 1 (a real fall). loss value ;

[0291] The updated model outputs a score for sample 1. loss value The losses were significantly reduced;

[0292] Model output probability validation: Preliminary fall risk probability of sample 1 after update , than the original More closely resembles the high-risk characteristics of real falls; Sample 2 This is a further reduction from the original 0.768, thus reducing the probability of false alarms.

[0293] Dynamic optimization using the OS-ELM algorithm The equal weighting coefficient reduces the overall score of false alarm scenarios for the elderly and increases the recovery index, making the system's judgment more conservative and effectively reducing false alarms;

[0294] After incremental learning updates the first-level fall recognition model, the confidence in recognizing real falls is improved, the probability of misjudging normal activities is reduced, and it adapts to the long-term changes in the behavioral habits of the elderly.

[0295] The optimization of the entire process does not affect the continuity of non-contact monitoring or privacy protection, continuously improves monitoring accuracy, and further reduces the care costs of home-based and institutional elderly care.

[0296] 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 technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0297] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0298] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A fall detection method based on millimeter-wave radar, characterized in that, Includes the following steps: S1: The millimeter-wave radar continuously transmits frequency-modulated continuous wave signals and receives echo signals reflected by the human body in the monitored area. S2: Perform signal processing on the echo signal to generate human point cloud time-series data containing distance, orientation, velocity and micro-motion information; S3: Extract multi-dimensional feature vectors representing human posture and motion state from human point cloud temporal data; S4: Input the multi-dimensional feature vector into the first-level fall recognition model to calculate the preliminary fall risk probability value; S5: If the initial fall risk probability value exceeds the first set threshold, it is determined to be a suspected fall event, and the second-level precise authentication process is immediately initiated; The Level 2 precision authentication process includes: initiating non-contact vital sign monitoring and retrieving the user's recent behavioral dynamic baseline model for comparative analysis; S6: Based on the output of the two-level precise authentication process, a comprehensive judgment is made. If it is confirmed as a fall event, an alarm message containing the event type and emergency level is generated and sent through a preset communication link. The non-contact vital sign monitoring initiated in step S5 includes: Separate the phase modulation component, which is mainly caused by chest cavity fluctuations, from the radar echo signal; Time-frequency analysis of the phase modulation components was performed to extract estimated respiratory rate and heart rate values ​​within a specific time window before and after a suspected fall event. Calculate the physiological sign disorder index: By comparing the currently estimated respiratory rate and heart rate with the user's personal resting state baseline values, the absolute value of their relative deviations is calculated, and the two deviations are weighted and summed to obtain the physiological sign disorder index, thereby quantifying the degree of abnormality of the physiological state. The dynamic baseline model of recent user behavior retrieved in step S5 is constructed by continuously learning the user's daily activity patterns. Its comparative analysis used in the secondary precision authentication process specifically includes: Extract the feature sequence of point cloud motion trajectories in the period of time before a suspected fall event occurred; Calculating behavioral deviation: Align and compare the feature sequences of point cloud motion trajectories over a previous period with the feature sequences of typical activity patterns for the same time period stored in the dynamic baseline model. Calculate the minimum cumulative distance between the two using a dynamic time warping algorithm; this minimum cumulative distance is the behavioral deviation S. dev ; Constructing a comprehensive and accurate certification score: incorporating the physiological sign disorder index I phy Deviation from behavior S dev A comprehensive verification score is obtained by linearly combining the coefficients according to preset weights. The calculation formula is as follows: ； in, and For adjustment coefficients; The comprehensive decision in step S6 is as follows: If accurate certification comprehensive score Exceeding the second set threshold Th high If so, it is directly identified as a high-urgency fall event; If accurate certification comprehensive score Given the third set threshold Th low With the second set threshold Th high In between, local audio-visual alerts will be activated, and a proactive response assessment period will begin; During the proactive response assessment period, assess the user's self-recovery capabilities: continuously monitor and remind users for a set duration. Human point cloud data is used to calculate the recovery index R. recover ; The recovery index sets the duration based on a comprehensive assessment. The net rate of increase in center of mass height and the stability of body posture at the end of the period are used for calculation. Specifically, the net rate of increase in center of mass height is characterized by the ratio of the net height increase value to the evaluation time, and posture stability is measured by the posture divergence index D at the end of the period. xy The two are represented by a weighted sum to obtain the recovery index R. recover The calculation formula is: ； in, This represents the net increase in centroid height during the assessment period. and These are the weighting coefficients; If you get up, the recovery index If the value is below the fourth preset threshold, it is determined that the user failed to get up effectively on their own and is confirmed as a fall event; otherwise, it is determined as a false alarm and the alarm is canceled.

2. A fall detection method based on millimeter-wave radar according to claim 1, characterized in that: Step S3, extracting multi-dimensional feature vectors, includes the following calculation process: Calculate the dynamic characteristics of human body centroid height: By performing centroid tracking on continuous point cloud frames, obtain the sequence of human body centroid height changes over time, and then calculate the average rate of change of centroid height and the acceleration of the rate of change within a specific time window T. Calculating human posture divergence features: Projecting each frame of point cloud data onto a horizontal plane to form a two-dimensional point set, calculating the covariance matrix of the coordinates of the two-dimensional point set and analyzing its eigenvalue distribution, yields the divergence index D, which characterizes the degree of posture extension or contraction. xy ; Calculating motion pattern characteristics: Extracting micro-Doppler signals generated by the slight movements of various parts of the human body from radar echoes, analyzing their spectrum within a short time window, and quantifying the regularity of motion by calculating the entropy value of the spectral energy distribution in each sub-band, thus obtaining the micro-Doppler spectral entropy E. md ; rate of change of centroid height , change in acceleration a H Attitude divergence index D xy and micro-Doppler spectral entropy E md Together they form a multi-dimensional feature vector .

3. A fall detection method based on millimeter-wave radar according to claim 2, characterized in that: The first-level fall detection model in step S4 is a lightweight gradient boosting decision tree model, which is obtained through offline training. The model calculates the initial fall risk probability value P. pre The process includes: inputting a multi-dimensional feature vector X into an ensemble model composed of multiple decision tree weak learners, with each weak learner making layer-by-layer judgments on the input features and outputting the weight value of a leaf node; The output weights of all weak learners are summed, and the sum is then mapped to the [0,1] interval using the Sigmoid function to obtain the initial fall risk probability value P. pre .

4. A fall detection method based on millimeter-wave radar according to claim 3, characterized in that: The emergency level of the alarm information generated in step S6 is based on the preliminary fall risk probability value P. pre Comprehensive score for accurate certification The emergency level is divided into at least three levels: high, medium, and low, as determined by fuzzy logic rules. Different levels correspond to different alarm notification strategies and contact notification priorities.

5. A fall detection method based on millimeter-wave radar according to claim 4, characterized in that: The method also includes a model optimization step: Data Collection: The system collects the final feedback results for each alarm event and outputs the multi-dimensional feature vector X and the physiological sign disorder index I corresponding to the alarm event. phy Behavioral Deviation S dev and recovery index The associated storage is a labeled training sample; Online optimization of judgment parameters: For samples marked as false alarms or voluntarily cancelled, their corresponding accurate authentication comprehensive scores are used. and recovery index To optimize the objective, an online sequential extreme learning machine algorithm is employed, which updates the network output weights using a recursive least squares approach, thereby dynamically adjusting the computation. Weighting coefficients and calculation Weighting coefficients , This makes the system tend to make conservative judgments in similar scenarios; Incremental update of the recognition model: The first-level fall recognition model is updated periodically using newly added, correctly labeled confirmed fall samples in an incremental learning manner to improve its recognition performance.

Images