Fall detection method based on Bluetooth AOA

Abstract

The invention provides a tumble detection method based on Bluetooth AOA, relates to the technical field of wearable equipment information processing, and provides a device and a method for realizing indoor tumble detection through a Bluetooth AOA indoor positioning system. The device comprises two Bluetooth AOA base stations with 4 * 4 array antennas and a Bluetooth AOA beacon. A user wears an AOA beacon, whether falling occurs or not can be detected in real time, and meanwhile, the system can judge six indoor common postures of the user. When a fall event is detected, the system may immediately determine a specific location where a fall occurs. The method is crucial to timely rescue.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wearable device information processing, and in particular to a fall detection method based on Bluetooth AOA. BACKGROUND

[0002] Fall is one of the most dangerous risks in the life of the elderly. An automatic monitoring system capable of quickly detecting the occurrence of a fall and recording the location of the fall needs to be designed. The system can notify medical personnel to carry out rescue when the elderly fall, and provide strong protection for the safety of the daily life of the elderly.

[0003] The current commonly used indoor fall detection methods are divided into three types: video monitoring based method, motion sensor based method and radar based method. The video monitoring based method is to shoot a video through a camera, and then use a machine learning method to recognize the fall action. This method extracts feature values from the video images captured, and then sends them into a machine learning model to classify human posture. This method has high accuracy, but video shooting requires a certain lighting environment. At the same time, this method will infringe on the privacy of the monitored person.

[0004] The motion sensor based method uses self-made wearable devices or commercial mobile phones to realize fall detection. These wearable devices usually contain motion sensors, MCUs with certain computing power and wireless communication modules. The data generated by the motion sensor needs to be continuously collected by the MCU, and then the collected data is judged by a certain algorithm to determine whether a fall has occurred. Therefore, the power consumption of such devices is generally high and needs to be charged frequently. When used indoors, the elderly sometimes forget to carry a commercial mobile phone, which can easily cause the failure of this detection method. The radar based method is commonly used in various modes of radar systems such as ultra-wideband radar, millimeter wave radar and frequency modulated continuous wave radar. This method does not require devices to be worn on the human body, so it is less likely to be forgotten by the user than wearable devices. Moreover, it does not infringe on privacy, but the cost of various radar devices is high. It is also a major challenge to implement its complex algorithms to achieve high accuracy.

[0005] The existing fall detection solutions mainly have the following problems:

[0006] The first problem is the privacy protection problem. Video fall detection can infringe on personal privacy. It cannot be used in private scenarios. The second problem is that the power consumption of wearable fall detectors is high and needs to be charged frequently. The cost of radar fall detection solutions is high. The third problem is that the fall detection solution does not have a positioning function and cannot accurately determine the location of the user's fall in the indoor environment. In large indoor environments such as nursing homes, it can affect the time of rescue. SUMMARY

[0007] The application discloses a Bluetooth AOA-based method capable of detecting falls of a wearer by wearing related equipment.

[0008] In order to achieve the above-mentioned purpose, the application adopts the following technical scheme: the required equipment comprises a Bluetooth AOA wearable device and two AOA antenna base stations. The Bluetooth AOA wearable device only needs to contain a Bluetooth AOA beacon. There are usually three wearing modes: in the first mode, the beacon is placed on a hat and worn on the head; in the second mode, the beacon is hung in front of the chest in the form of a necklace; and in the third mode, the beacon is worn on the arm in the form of a bracelet. The Bluetooth AOA antenna base station usually has two deployment modes: one is a mode in which the base station is parallel to the ground, and the other is a mode in which the base station is perpendicular to the ground. The mode in which the base station is parallel to the ground can place the Bluetooth receiving base station on the ceiling, and in this mode, the signal of the beacon is not easily blocked by furniture, and the range of the signal detected by the base station is large. This deployment mode is suitable for nursing homes, halls or living rooms. The mode in which the base station is perpendicular to the ground usually suspends the Bluetooth receiving base station on the wall. In this method, the suspension height of the Bluetooth receiving base station is usually not too high, and is suitable for small-range detection, such as toilets and other areas.

[0009] The specific algorithm comprises the following steps:

[0010] S1, collecting IQ values through the Bluetooth AOA antenna base station;

[0011] S2, incident angle calculation: converting into the direction angle and the elevation angle of the incident signal through the ISSS algorithm;

[0012] S3, feature extraction: extracting features in a window sliding mode, and extracting the mean value and the variance of the incident angle;

[0013] S4, posture classification;

[0014] S5, noise filtering: filtering out distortion data contained in the converted feature values;

[0015] S6, fall judgment: constructing XGBoost tree models and LightGBM tree models to classify and identify static actions and dynamic actions, and realizing fall detection;

[0016] Finally, position tracking is performed: the least square method is used to calculate the coordinates of the user.

[0017] Compared with the prior art, the present application has the following advantages:

[0018] The application can not only detect whether a person has fallen in a room, but also determine the falling position. The application has very high detection accuracy and the following advantages:

[0019] 1. The technical scheme of the present application has high recognition accuracy for the occurrence of falling, and does not compromise user privacy. It avoids privacy leakage caused by devices such as cameras.

[0020] 2. At the same time, the traditional wearable anti-falling device is mainly based on acceleration sensor to collect human action data, and then upload the data to the server through Bluetooth or WIFI. The power consumption of such device is high. In the technical scheme of the present application, the device worn by the user only needs to perform BLE broadcast. Therefore, the system device has lower energy consumption, longer wearing time and lower cost.

[0021] 3. The technical scheme of the present application can quickly locate the falling position, which can significantly reduce the rescue response time. In an emergency, every second may affect the health and safety of the faller. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 (a) is the way of deploying the Bluetooth AOA base station parallel to the ground; Figure 1 (b) is the way of deploying the Bluetooth AOA base station perpendicular to the ground;

[0023] Figure 2 It is a schematic diagram of the double-base station deployment in the range space;

[0024] Figure 3 It is a schematic diagram of a 4x4 rectangular array antenna;

[0025] Figure 4 (a) is the change curve of the incident elevation angle and azimuth angle of the tag in the falling posture; Figure 4 (b) is the change curve of the incident elevation angle and azimuth angle of the tag in the walking posture;

[0026] Figure 5 It is a schematic diagram of window mode segmentation action;

[0027] Figure 6 It is a block diagram of the overall flow of the algorithm;

[0028] Figure 7 It is a block diagram of the posture classification model. DETAILED DESCRIPTION

[0029] The following is a further description of the technical scheme of the present application in conjunction with the drawings, but the present application is not limited to these embodiments.

[0030] Technical route:

[0031] 1. Equipment Deployment: This invention requires one Bluetooth AOA wearable device and two AOA antenna base stations. The Bluetooth AOA wearable device only needs to include one Bluetooth AOA beacon. There are typically three wearing methods: the first is to place the beacon on a hat and wear it on the top of the head; the second is to hang the beacon around the neck like a necklace; and the third is to wear the beacon on the arm like a bracelet. The Bluetooth AOA antenna base stations typically have two deployment methods, such as... Figure 1 (a) Figure 1 As shown in (b), there are two deployment methods: one where the base station is parallel to the ground, and the other where it is perpendicular to the ground. The ground-parallel deployment allows the Bluetooth receiving base station to be placed on the ceiling. This method makes the beacon signal less susceptible to obstruction by furniture, and the base station has a wider signal detection range. This deployment method is suitable for nursing homes, lobbies, or living rooms, such as... Figure 2 As shown, the vertical deployment of base stations typically involves suspending the Bluetooth receiving base station on a wall. This method usually results in a relatively low suspension height for the Bluetooth receiving base station, making it suitable for small-area detection, such as in restrooms.

[0032] 2: S1 Bluetooth AOA Antenna Base Station IQ Value Acquisition: The antenna base station consists of a 4x4 rectangular array antenna, such as... Figure 3 As shown. The Bluetooth 5.1 signal includes a Continuous Tone Extension (CTE) group. This group contains a 4 μs guard period followed by an 8 μs reference period. All signals acquired during the reference period are collected on the first antenna. Therefore, the phase shift per unit time (1 μs or 2 μs) can be calculated using the IQ values ​​acquired during the reference period. The initial phase is calculated using the IQ values ​​acquired during the 8 reference periods:

[0033]

[0034] Then, phase correction is performed, with the initial correction value Correction=0 and the count value count=0.

[0035] Starting from zi=2, the calculation continues until zi=8. If count is positive, then count = count + 1; otherwise, count remains unchanged.

[0036] Then correct the value. Corrected angle:

[0037]

[0038] Then calculate the absolute value of the difference between adjacent angles:

[0039]

[0040] The final phase shift can be expressed as:

[0041] After the reference period ends, each antenna in the antenna array sequentially acquires the IQ value. However, since the Bluetooth chip can only access one antenna channel, the antenna array needs to use electronic switches to switch between antennas, allowing each antenna to acquire signals sequentially. Therefore, at the end of the reference period, the CTE consists of alternating switching slots and sampling slots, each slot being 1 or 2 μs long. Each sampling slot contains an IQ value, which is the IQ value of the incident signal. The received signal x(t) of the antenna array can be calculated using the phase shift and the IQ value of the incident signal.

[0042]

[0043] in M=16 is the number of array antennas, and S is the number of sampling periods.

[0044] 3: S2 Incident Angle Calculation: The mathematical model of the signal x(t) received by the antenna linear array can be expressed as:

[0045]

[0046] in It is a signal transmitted from the air. It is the steering vector of the antenna array. This represents the Gaussian noise of the signal. When there are N snapshots, the data sample covariance matrix can be approximated by the sampling covariance matrix using the following formula:

[0047]

[0048] In indoor spaces, the presence of various reflected signals can lead to multipath propagation, resulting in coherent signals. In such cases, spatial smoothing needs to be incorporated into the algorithm to correlate the signals. When the antenna array is divided into sections along the x and y axes... Individual and When there are subarrays, the data covariance matrix after forward spatial smoothing can be expressed as:

[0049]

[0050] The covariance matrix is ​​decomposed into eigenvalues ​​as follows:

[0051]

[0052] in yes The eigenvector matrix corresponding to the first D largest eigenvalues ​​is: The matrix, yes the back of the matrix of eigenvectors corresponding to the smaller eigenvalues is the matrix of eigenvectors corresponding to the larger eigenvalues. Thus is the signal subspace, and the pseudo-spectrum of the ISSS algorithm is given by

[0053]

[0054] The values of φ and seta when P takes the maximum value are the azimuth and elevation of the true signal label.

[0055] 4: S3 Feature Extraction: There are six actions in the indoor environment, which are sitting, lying, standing, falling, walking, and sitting down. This scheme classifies the six actions of the user to realize the judgment of the user falling. We observe that the angle changes of the tag elevation and azimuth under different postures are different. As shown in (a) and (b), the incident signal angle change curves of falling and walking actions are different. In this step, features are extracted for distinction. In fact, an action can be completed in about 2 seconds. We lengthen the segmented time to 4 seconds to include as many features of an action as possible. We extract features in the way of window sliding, as shown in (c). Each window takes 2 seconds, and the data in 4 seconds can be divided into 3 overlapping windows, each with 1 s overlap. Then the data features of each window are extracted. The incident angle calculation contains the following features: Figure 4 Figure 4 Figure 5

[0056] Given a set X = {X1, …Xn} containing n samples, the mean is the average of the sum of all elements in the set. The formula is as follows:

[0057]

[0058] Variance is a statistical measure of the degree of dispersion of data, which represents the average of the square of the difference between each data point and the mean. If you want to know the dispersion between a group of data, you can use "variance" to represent it. The formula is as follows:

[0059]

[0060] Skewness is a statistical index used to describe the asymmetry of probability distribution, which reflects the degree of data distribution deviating from symmetric distribution. The formula is as follows:

[0061]

[0062] ​​​Kurtosis is a statistical measure used to describe the thickness of the tail of a probability distribution and the degree of sharpness of the peak, reflecting the probability of extreme values (outliers) compared to the normal distribution. It can reflect the degree of change of angle data. The formula is:

[0063]

[0064] 5: S4 dynamic and static action classification:

[0065] When classifying postures, we first determine whether it is a dynamic or static posture. When a static posture occurs, the azimuth and elevation angles remain basically unchanged. Therefore, the variance tends to 0 in the entire action time. Dynamic posture will not have this situation. Through this feature, we first distinguish dynamic and static actions, as shown in the following formula:

[0066]

[0067] where T represents the threshold of variance, Dynamic is dynamic posture, and Static is static posture. Take the mean of the variance of each time window, and when the mean is greater than the threshold T, it is judged as a dynamic posture. When the mean of the variance is less than the threshold T, it is judged as a static posture.

[0068] 6: S5 noise filtering: The collected incident angles contain a lot of noise and interference, and the converted feature values contain a lot of distorted data, which are filtered in the following way. Any variance feature value exceeding the threshold means that the sampling data has noise. Filter out the data point. The threshold values of dynamic and static actions are different. Set the dynamic variance threshold to: . The static action variance threshold is: .

[0069] 7: S6 fall judgment:

[0070] A motion is divided into 3 time windows, and the features of each window are extracted. Then the feature values of the three windows are input into the classification model at the same time. This scheme builds XGBoost tree and LightGBM tree models to classify actions through the following steps. XGBoost tree is used to distinguish static actions, and LightGBM tree is used to classify dynamic actions, including falls.

[0071] 1) First, give the initialization prediction probability-->2) Then calculate the gradient and Hessian matrix-->3) Then build the XGBoost tree or LightGBM tree-->4) Calculate the output value of the leaf node-->5) Update the prediction probability-->6) Calculate the gradient and Hessian matrix using the new prediction probability->7) Repeat the above steps 3-6.-->8) Until the residual error is reduced to a certain extent or the maximum number of iterations is reached. In this way, the six actions of sitting, lying, standing, falling, walking and sitting can be classified with high accuracy, so as to perform fall detection.

[0072] 8: Position tracking:

[0073] The coordinate point of the user in the room is represented by (x, y, z), and the azimuth and elevation angle of the beacon worn by the user to the Bluetooth AOA receiving base station are represented by and , and the coordinate value of the i-th Bluetooth AOA receiving base station in the 3D space is represented by The user coordinates are calculated using the least squares method, and the formula is as follows:

[0074]

[0075] Among them: , , ,

[0076] Using this formula, the real-time coordinate value of the user can be calculated.

[0077] The specific embodiments described herein are merely illustrative of the spirit of the present application. Those skilled in the art of the present application can make various modifications or supplements to the described specific embodiments or use similar ways to replace them, but will not deviate from the spirit of the present application or exceed the scope defined by the appended claims.

Claims

1. A fall detection method based on Bluetooth AOA, characterized in that, Equipment deployment: Includes one Bluetooth AOA wearable device and two AOA antenna base stations; the Bluetooth AOA wearable device has a Bluetooth AOA beacon; when the base stations are deployed parallel to the ground, the Bluetooth receiving base station is placed on the ceiling of the indoor detection space; when the base stations are deployed perpendicular to the ground, the Bluetooth receiving base station is suspended from the wall of the indoor detection space. Fall detection methods include the following steps: S1. Collect IQ values ​​via Bluetooth AOA antenna base station; S2. Incident Angle Calculation: Converted into the direction and elevation angles of the incident signal using the ISSS algorithm; S3. Feature Extraction: Features are extracted using a sliding window method, extracting the mean and variance of the incident angle; S4. Posture classification; S5. Noise Filtering: Filter out distorted data contained in the transformed feature values; S6. Fall detection: Construct an XGBoost tree model and a LightGBM tree model to classify and identify static and dynamic actions, thereby achieving fall detection; Finally, position tracking is performed: the user coordinates are calculated using the least squares method.

2. The fall detection method based on Bluetooth AOA according to claim 1, characterized in that... In step S1, the antenna base station consists of a 4x4 rectangular array antenna. All signals acquired during the reference period are collected on the first antenna. Therefore, the phase shift per unit time can be calculated using the IQ values ​​acquired during the reference period. The initial phase is calculated using the IQ values ​​acquired over 8 reference periods. Then, phase correction is performed, with the correction value Correction=0 and the count value count=0. Starting from zi=2, the calculation continues until zi=8. Then count = count + 1; otherwise, count remains unchanged. Then correct the value. Corrected angle: Then calculate the absolute value of the difference between adjacent angles: ； The final phase shift can be expressed as: ; After the reference period ends, each antenna in the antenna array will sequentially acquire the IQ value. However, since the Bluetooth chip can only access one antenna channel, the antenna array needs to switch each antenna electronically to ensure that each antenna acquires the signal sequentially. Therefore, at the end of the reference period, the CTE consists of alternating switching slots and sampling slots, each slot being 1 or 2 μs long. Each sampling slot contains an IQ value, which is the IQ value of the incident signal. The received signal x(t) of the antenna array can be calculated using the phase shift and the IQ value of the incident signal. ； in M=16 is the number of array antennas, and S is the number of sampling periods.

3. The fall detection method based on Bluetooth AOA according to claim 1, characterized in that, In step S2, the mathematical model of the signal x(t) received by the linear array can be expressed as: ； in It is a signal transmitted from the air. It is the steering vector of the antenna array. The signal represents Gaussian noise; when there are N snapshots, the data sample covariance matrix can be approximated by the sampling covariance matrix and calculated using the following formula: ； In indoor spaces, the presence of various reflected signals can lead to multipath propagation, resulting in coherent signals. In such cases, spatial smoothing needs to be incorporated into the algorithm to correlate the signals. This is especially important when the antenna array is divided into sections along the x and y axes. One and When there are subarrays, the data covariance matrix after forward spatial smoothing can be expressed as: ； The covariance matrix is ​​decomposed into eigenvalues ​​as follows: ； in yes The eigenvector matrix corresponding to the first D largest eigenvalues ​​is: The matrix, yes After The eigenvector matrix corresponding to the smaller eigenvalues ​​is: The matrix; therefore For the signal subspace, the final pseudo-spectrum of the ISSS algorithm is given by the following formula: ； When P reaches its maximum value, φ and seta are the azimuth and elevation angles of the actual signal tag.

4. The fall detection method based on Bluetooth AOA according to claim 1, characterized in that, In step S3, users typically exhibit six indoor behaviors: sitting, lying down, standing, falling, walking, and sitting up. These six behaviors are categorized, and the user's fall behavior is determined by the changes in the tag's elevation and azimuth angles collected by the antenna under different postures. The following features are calculated based on the incident angle: Given a set X = {X1, …Xn} containing n samples, the mean is the average of the sums of all elements in this set, as shown in the following formula: ； Variance is a statistic that measures the dispersion of data. It is represented by the average of the squares of the differences between each data point and the mean. If you want to know how dispersed a set of data is, you can use "variance" to represent it. The formula is: ； Skewness is a statistical indicator used to describe the asymmetry of a probability distribution, reflecting the degree to which the data distribution deviates from a symmetrical distribution. The formula is as follows: ； Kurtosis is a statistical indicator used to describe the thickness of the tails and the sharpness of the peaks in a probability distribution. It reflects the probability of extreme values ​​(outliers) occurring in a data distribution compared to a normal distribution; it can reflect the drasticness of changes in angular data. The formula is as follows: 。 5. The fall detection method based on Bluetooth AOA according to claim 1, characterized in that, In step S4, when a static attitude occurs, the azimuth and elevation angles remain essentially unchanged, indicating a static attitude; otherwise, it is considered dynamic. Therefore, its variance approaches 0 throughout the entire action time, whereas dynamic attitudes do not exhibit this characteristic. Based on this feature, dynamic and static actions are first distinguished, as shown in the following formula: ； Where T represents the variance threshold, Dynamic represents the dynamic pose, and Static represents the static pose; Take the mean variance of each time window. If the mean variance is greater than the threshold T, it is judged as a dynamic attitude; if the mean variance is less than the threshold T, it is judged as a static attitude.

6. The fall detection method based on Bluetooth AOA according to claim 1, characterized in that, In step S5, the collected incident angle contains a lot of noise and interference, and the converted feature values ​​contain a lot of distorted data. Filtering is performed as follows: if any variance feature value exceeds a threshold, the sampled data is considered noisy and the data point is filtered out. The thresholds for dynamic and static actions are different; the dynamic variance threshold is set to... The static action variance threshold is .

7. The fall detection method based on Bluetooth AOA according to claim 1, characterized in that, In step S7, the user's indoor coordinates are represented by (x, y, z), and the azimuth and elevation angles from the beacon worn by the user to the Bluetooth A0A receiving base station are represented by... and Indicates, using Let represent the coordinates of the i-th Bluetooth AOA receiving base station in 3D space; the user coordinates are calculated using the least squares method, as shown in the following formula: in: , , , ; This formula can be used to calculate the user's real-time coordinates.

Images