WO2020147098A1

WO2020147098A1 - Ground vibration signal-based human body fall detection system

Info

Publication number: WO2020147098A1
Application number: PCT/CN2019/072282
Authority: WO
Inventors: 伍楷舜; 陈文强; 王璐; 陈鸿杰; 黄彦道
Original assignee: 深圳大学
Priority date: 2019-01-18
Filing date: 2019-01-18
Publication date: 2020-07-23

Abstract

A ground vibration signal-based human body fall detection method and human body fall detection system, the method comprising: three geophones collecting ground vibration signals (S1); performing filtering noise reduction and endpoint segmentation processing on the collected vibration signals (S2); performing an alignment processing on the vibration signals after endpoint segmentation (S3); performing signal feature extraction on the aligned vibration signals (S4); grouping the extracted feature signals into a training set and training to obtain a hidden Markov model (S5); using ground vibration signals during actual use as test data to detect whether the test data are valid vibration signals, and then processing the test data by using a method of processing training data (S6); on the basis of the hidden Markov model, calculating the probability that the test data are fall signals, and if the probability is greater than a threshold, entering a second confirmation stage, otherwise directly determining that the test data are not related to a fall (S7); by means of an EoA positioning algorithm, calculating the displacement of a target within a short period of time after a suspected fall, and if the movement distance is less than a threshold, a fall is confirmed for a second time, otherwise determining that a fall has not occurred (S8). The described detection method uses the hidden Markov model to calculate the probability of vibration signals being related to a fall, and then uses an EoA positioning algorithm to perform second confirmation, thus fall recognition is accurate and the rate of false alarms is low. In addition, a worn device is not required, privacy is protected and user experience is improved.

Description

Human fall detection system based on ground vibration signal

Technical field

The invention relates to the field of information processing technology and intelligent perception, and in particular to a method and system for detecting a human body fall based on ground vibration signals.

Background technique

The global population aging trend is irreversible. According to United Nations statistics, as of 2017, there are more than 960 million people over the age of 60 in the world, and it is expected that this number will double to 2.1 billion in 2050. For the elderly, falling is a health-threatening event they face every day. According to statistics from the World Health Organization, one-third of the elderly over the age of 65 falls at least once a year. And 60% of falls occur at home. Therefore, we urgently need to invent a method and system that is convenient, effective and can automatically detect the fall of the elderly. The existing fall detection methods are mainly vision-based, wearable device-based and wireless signal-based. Vision-based detection methods can effectively detect fall events, but there are major privacy issues. It is also difficult to deploy for sensitive environments such as wet floors and bathrooms where fall incidents are high. In addition, the camera also has visual blind spots and is in a dark environment. The fall detection system based on wearable devices uses the inertial sensors in the device to detect whether the user has fallen. However, it is not user-friendly to carry the device, especially the elderly in the home environment. Keep the device worn on the body, once the device is taken off, the test will fail. Although the fall detection method based on radio communication technology does not require the user to wear the device, the high false alarm rate has been criticized for a long time, and the multipath effect of the signal also makes it difficult to work normally in the complex and dynamic home environment.

Summary of the invention

In order to solve the above technical problems, the present invention proposes to use geophones to collect ground vibration signals and process them through information technology to achieve a more accurate, low false alarm rate, no privacy risk, no user training, and no wearable equipment for human fall detection. The method and system specifically adopt the following technical solutions:

A human fall detection method based on ground vibration signals includes the following steps:

S1: Three geophones collect ground vibration signals;

S2: Perform filtering and noise reduction and end-point segmentation processing on the collected vibration signal to determine the starting point and ending point of the effective vibration signal;

S3, aligning the effective vibration signal after the end point is cut;

S4, superimpose the effective vibration signals of the three geophones, and then perform signal feature extraction;

S5, forming a training set of the extracted feature signals to train a hidden Markov model;

S6: Use ground vibration signals during actual use as test data, and process the test data by processing training data;

S7: Calculate the probability of falling based on the Hidden Markov Model, and if the above is greater than the threshold, perform a second confirmation, otherwise it is not recognized as a fall;

S8: Use the EoA positioning algorithm to track the coordinates in a short period of time. If the target moving range is less than the threshold, it is confirmed as a fall for the second time, otherwise it is not a fall.

As a preference, in the step S3, the effective vibration signal after the end point segmentation is aligned by the overall cross-correlation method, and the specific operation of the alignment processing is to calculate the offset between the two effective vibration signals, Then move the current effective vibration signal, and only take the integral part common between the two vibration signals after the movement.

As a preference, in the step S3, the formula

And O(A,B)=P(A,B)-n to calculate the offset O(A,B) between two effective vibration signals, where a and b represent the effective vibration of two signal lengths n Signal, a(i) represents the amplitude of the i-th point of the effective vibration signal a, b(i) represents the amplitude of the i-th point of the effective vibration signal b, and C(a,b) represents the effective vibration signal a and The correlation degree of the effective vibration signal b; A means that the length of the part of the effective vibration signal a on both sides is filled with zero, and then a first signal with a length of 3n is obtained; B is the effective vibration signal b of length n; P(A ,B) represents the signal position of length n with the highest correlation between the first signal A and the second signal B; O(A,B) is the calculated offset between the first signal A and the second signal B .

As a preference, in the step S4, after the effective vibration signals from the three geophones are superimposed, they are processed by a pre-emphasis filter, and then the signals are decomposed into multiple layers of different frequency components through discrete wavelet transform, respectively. The corresponding energy is calculated at a certain time interval, and finally the energy characteristics in the time domain and the frequency domain are obtained by the formula Feature=10log ₁₀ E, where E is the energy of each time interval at different frequency levels obtained initially, and Feature is The corresponding feature after logarithmic transformation.

As a preference, the number of states of the hidden Markov model in step S5 is 7, the initial state is 1, and the output observation probability function of each state is composed of multiple Gaussian mixture functions, where the number of Gaussian mixture functions It is equal to the number of layers decomposed by discrete wavelet transform in step S4, and each Gaussian mixture function has 3 Gaussian components.

As a preference, it is characterized in that, in the step S7, the probability that the test data is a fall signal is calculated based on the hidden Markov model obtained in step S5, and if it is greater than the threshold, the second confirmation is performed through step S8, otherwise directly It is not considered a fall.

As a preference, it is characterized in that, in the step S8, the EoA positioning algorithm (full name Energy-of-Arrival) proposed by the present invention is used for positioning, which is based on the signal attenuation model Amp(d)=Amp ₀ e ^{− α×d} , where Amp ₀ represents the initial amplitude of the vibration signal, d represents the distance the vibration propagates, and α is the attenuation coefficient based on the propagation medium, then the following formula:

Among them, d ₁ , d ₂ , and d ₃ are the distances from the vibration source to the three detectors A, B, C, and ln E _AB = -2α (d ₁ -d ₂ ) is the ratio of signal energy received by the A and B detectors The logarithmic value of, the coordinate position of the vibration source can be solved by the above formula.

As a preference, in the step S8, the EoA positioning algorithm calculates the coordinate position of the suspected fall after the test data is recognized as a suspicious fall by the hidden Markov model, and then continues to calculate the coordinates of the target within a short period of time , If the displacement exceeds the set threshold, it indicates that the user can move normally and is not considered a fall; otherwise, the second confirmation is passed and it is confirmed as a fall.

As a preference, in the step S2, a Butterworth filter is used to perform filtering and noise reduction processing on the collected vibration signal, and a high-pass filter with a cut-off frequency of 10 Hz is used to filter out DC components and low-frequency noise.

As a preference, in the step S2, in the end-point segmentation processing, the entire segment of the vibration signal is divided into frames first, and then the variance of each frame signal is used as the criterion. When the variance of a certain frame signal exceeds the given value When the threshold is set, it is considered that an effective vibration signal appears, and the signal of a certain length before and after the frame signal is taken as the vibration signal after the end point is cut.

A human fall detection system is characterized by comprising three geophones and a processor, and the processor executes the above-mentioned human fall detection method based on ground vibration signals.

The invention realizes real-time monitoring of falls based on ground vibration, calculates the probability that the vibration signal is a fall through the hidden Markov model, and then combines the positioning algorithm EoA to perform secondary confirmation. The fall recognition is accurate and the false alarm rate is low, so Injuries caused by falls can be treated in time, which solves the problems of poor robustness, troublesome wearable devices, accuracy problems, and privacy hazards in the prior art, and improves user experience. The detection method of the present invention is novel, convenient and reliable, and can meet various requirements. It is widely used and low-cost for the use of living environment.

BRIEF DESCRIPTION

Fig. 1 is a flow chart of the human body fall detection method of the present invention.

Figure 2 is an external view of the geophone of the present invention.

FIG. 3 is a schematic diagram of the simulation of the effect of the present invention before the effective vibration signal alignment processing after the end point is segmented.

FIG. 4 is a schematic diagram of the simulation of the effect of the present invention after the effective vibration signal alignment processing after the end point is segmented.

Figure 5 is a schematic diagram of the EoA algorithm of the present invention.

detailed description

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

The preferred embodiments of the present invention will be further described in detail below in conjunction with the accompanying drawings.

As shown in Figure 1, the present invention provides a fall detection method based on hidden Markov model and EoA algorithm, including the following steps:

Step S1, arranging a geophone (or other sensor capable of detecting vibration) in the corner of the use environment, detecting the vibration signal acting on the ground when the human body falls and converting it into an analog electrical signal, and then converting the above analog electrical signal For digital signals that can be processed, three geophones can be specifically used. Figure 2 shows the appearance of the geophone.

In step S2, a Butterworth filter is used to perform filtering and noise reduction processing on the collected vibration signal using a band-pass filter with a frequency band greater than 10 Hz. More specifically, this embodiment uses a high-pass filter with a cut-off frequency of 10 Hz to filter out the DC component and Low frequency noise.

The end-point segmentation processing is also called end-point detection processing. It is executed by the cut-off algorithm in the computer program to determine the starting point and the end point of the effective vibration signal. The processing process is to first divide the entire segment of the vibration signal into frames, and then use each The variance of the frame signal is used as the criterion. When the variance of a frame signal exceeds a given threshold, it is considered that a valid vibration signal appears, and the signal of a certain length before and after the frame signal is taken as the vibration signal after the end segment is cut. The vibration signal is also called effective vibration signal. The given threshold can be customized according to the needs of the user, or can be based on the numerical value in the training library of the sample as a reference value.

Step S3: align the effective vibration signals after the end segment is cut by the general cross correlation method (GCC). The specific operation of the alignment processing is to calculate the offset between the two effective vibration signals, and then Move the current effective vibration signal, and only take the integral part common between the two effective vibration signals after moving. The alignment processing described in this embodiment can align all effective vibration signals, which is beneficial to the improvement of the classification accuracy of the machine learning algorithm. The simulation effect diagrams before and after the alignment processing are shown in Figures 3 and 4, which shows that , After the cut signal is processed by GCC, it is further aligned in the time series to facilitate more accurate judgment of the fall signal.

In step S3 in this example, the formula

And O(A,B)=P(A,B)-n to calculate the offset O(A,B) between two effective vibration signals, where a and b represent two effective vibration signals of length n , A(i) represents the amplitude of the i-th point of the effective vibration signal a, b(i) represents the amplitude of the i-th point of the effective vibration signal b, and C(a,b) represents the effective vibration signal a and the effective The correlation degree of the vibration signal b; the first signal A represents the zero-filling of the length n on both sides of the effective vibration signal a, and then a first signal with a length of 3n is obtained; the second signal B represents a vibration signal b with a length of n ; P(A, B) represents the signal position of length n with the highest correlation between the first signal A and the second signal B; O(A, B) is the calculated value between the first signal A and the second signal B The offset.

Step S4, superimpose the aligned signals from the three geophones after one vibration, process the pre-emphasis filter, and then perform discrete wavelet transform to extract the energy intensity of the vibration signal in different frequency domains and different times as the extracted signal feature. Preferably, in the step S4, the energy characteristics in the time domain and the frequency domain are obtained by the formula Feature=10log ₁₀ E, where E is the energy of each time interval at different frequency levels obtained preliminarily, and Feature is the passed-pair Characteristic signal after digital conversion.

In step S5, the extracted characteristic signals are formed into a training set for training the hidden Markov model; during training, only falling ground vibration signals from other people (such as volunteers or participants) are used as the training set.

Specifically, the number of transition states of the hidden Markov model is 7, and the initial state is 1, and the output observation probability function of each state is composed of several Gaussian mixture functions containing multiple Gaussian components. The number of Gaussian mixture functions is Step S4 Discrete wavelet transform has the same number of layers.

At the beginning, the relevant parameters are initialized according to the training set, namely the initial state probability, state transition probability and output observation probability, and an initial hidden Markov model is obtained. Then use the baum-welch algorithm to optimize the parameters of the model based on the training set. After each optimization, all samples in the training set will be calculated one by one through the viterbi algorithm to match the model (that is, the probability of falling. In fact, the result is It is not probability. In order to improve the calculation performance, the calculation process takes the logarithmic operation, so the result is a negative value), and records it; then from the second call to the baum-welch algorithm for optimization, compare the optimization results What is the difference between the total matching degree of the training set and the model (that is, the probability of falling) and the difference after the last optimization. If it is less than the threshold, it means that the training has converged and jumped out of the loop. The trained hidden Markov model is obtained. For the subsequent test samples, you only need to use the viterbi algorithm to calculate the degree of matching with the model (that is, the probability of falling).

In step S6, the ground vibration signal during actual use is used as test data, and the test data is processed by the method of processing training data; in actual use, the test data is collected in the same manner as steps S1, S2, S3, and S4.

Step S7: Calculate the probability that the test data is a fall based on the hidden Markov model obtained in Step S5. If the above probability is greater than the threshold, proceed to Step S8 for a second confirmation; otherwise, it is directly confirmed that it is not a fall;

Step S8: Calculate the target position when the suspicious fall event occurs and within a short period of time after the suspicious fall event occurs through the EoA positioning algorithm, and perform coordinate tracking on the target. If a movement exceeding the threshold occurs, the warning is cancelled; otherwise, the second confirmation is successful and the confirmation is fall.

Specifically, the positioning calculation is performed based on the signal attenuation model Amp(d)=Amp ₀ e ^-α×d , as shown in Figure 5. Among them, Amp ₀ represents the initial amplitude of the vibration signal, d represents the distance of vibration propagation, and α is the attenuation coefficient based on the propagation medium. Given three geophones A, B, and C, the received signal amplitude is as follows:

Among them, d ₁ , d ₂ , and d ₃ are the distances from the three geophones respectively. Then the signal energy received by the A and B detectors is as follows:

After taking the logarithm, it is ln E _AB =-2α(d ₁ -d ₂ ), then the following formula is adopted:

Among them, c ₁ , c ₂ can be obtained based on the α obtained in advance and the obtained energy ratio, so that the position of the vibration source can be obtained and the positioning can be realized.

In the present invention, the trained hidden Markov model is combined with the EoA algorithm, and then the hidden Markov model can be used for fall detection. The vibration signal is detected in real time by the geophone. When a fall event occurs or other actions (such as things Falling, etc.) will generate a vibration signal with greater energy. At this time, the system detects the vibration signal, takes out the vibration signal and filters the vibration signal for denoising, endpoint detection, GCC alignment and signal feature extraction, and the vibration signal The generated signal features are used as the input of the hidden Markov model, and the result returned by the hidden Markov model is obtained. The signal suspected of falling is confirmed by the EoA algorithm for a second time, and finally a conclusion is drawn.

This example also provides a fall detection system based on the hidden Markov model and the EoA algorithm, and adopts the fall detection method based on the hidden Markov model and the EoA algorithm as described above.

The above content is a further detailed description of the present invention in conjunction with specific preferred embodiments, and it cannot be considered that the specific implementation of the present invention is limited to these descriptions. For a person of ordinary skill in the technical field to which the present invention belongs, without departing from the concept of the present invention, several simple deductions or replacements can be made, which should be regarded as falling within the protection scope of the present invention.

Claims

A human fall detection method based on ground vibration signals is characterized in that it comprises the following steps:

S1: Three geophones collect ground vibration signals;

S2: Perform filtering and noise reduction and end-point segmentation processing on the collected vibration signal to determine the starting point and ending point of the effective vibration signal;

S3, aligning the effective vibration signal after the end point is cut;

S4, superimpose the effective vibration signals of the three geophones, and then perform signal feature extraction;

S5, forming a training set of the extracted feature signals to train a hidden Markov model;

S6: Use ground vibration signals during actual use as test data, and process the test data by processing training data;

S7: Calculate the probability of falling based on the Hidden Markov Model, and if the above is greater than the threshold, perform a second confirmation, otherwise it is not recognized as a fall;

S8: Use the EoA positioning algorithm to track the coordinates in a short period of time. If the target moving range is less than the threshold, it is confirmed as a fall for the second time, otherwise it is not a fall.
The detection method according to claim 1, characterized in that, in the step S3, the effective vibration signal after the end point segmentation is aligned by the overall cross-correlation method, and the specific operation of the alignment processing is to calculate two effective vibration signals. Offset between the vibration signals, and then move the current effective vibration signal. After the movement, only the integral part shared between the two vibration signals is taken.
The effective detection method according to claim 2, characterized in that, in the step S3, the formula

And O(A,B)=P(A,B)-n to calculate the offset O(A,B) between two effective vibration signals, where a and b represent the effective vibration of two signal lengths n Signal, a(i) represents the amplitude of the i-th point of the effective vibration signal a, b(i) represents the amplitude of the i-th point of the effective vibration signal b, and C(a,b) represents the effective vibration signal a and The correlation degree of the effective vibration signal b; A means that the length of the part of the effective vibration signal a on both sides is filled with zero, and then a first signal with a length of 3n is obtained; B is the effective vibration signal b of length n; P(A ,B) represents the signal position of length n with the highest correlation between the first signal A and the second signal B; O(A,B) is the calculated offset between the first signal A and the second signal B .
The detection method according to any one of claims 1 to 3, characterized in that, in the step S4, the effective vibration signals from the three geophones are superimposed, processed by a pre-emphasis filter, and then passed through a discrete wavelet The transform decomposes the signal into multiple layers of different frequency components, and calculates the corresponding energy at a certain time interval. Finally, the energy characteristics in the time domain and frequency domain are obtained by the formula Feature=10log 10 E, where E is the preliminary calculation For the energy of each time interval at different frequency levels, Feature is the corresponding feature after logarithmic conversion.
The detection method according to claim 1, wherein the number of states of the hidden Markov model in step S5 is 7, the initial state is 1, and the output observation probability function of each state is composed of multiple Gaussian mixture functions The composition, where the number of Gaussian mixture functions is equal to the number of layers decomposed by discrete wavelet transform in step S4, and each Gaussian mixture function has 3 Gaussian components.
The detection method according to claim 1, wherein in the step S7, the probability that the test data is a fall signal is calculated based on the hidden Markov model obtained in step S5, and if it is greater than the threshold, the second step is performed through step S8. Confirmation times, otherwise it is directly regarded as a fall.
The detection method according to claim 1, characterized in that, in the step S8, the EoA positioning algorithm (full name Energy-of-Arrival) proposed by the present invention is used for positioning, which is based on the signal attenuation model Amp(d) =Amp 0 e -α×d , where Amp 0 represents the initial amplitude of the vibration signal, d represents the distance the vibration propagates, and α is the attenuation coefficient based on the propagation medium, then the following formula:

Among them, d 1 , d 2 , d 3 are the distances from the vibration source to the three detectors A, B, C, lnE AB = -2α (d 1 -d 2 ) is the ratio of the signal energy received by the A and B detectors Log the value, the coordinate position of the vibration source can be solved by the above formula.
The detection method according to claim 7, characterized in that, in the step S8, the EoA positioning algorithm calculates the coordinate position of the suspected fall after the test data is recognized as a suspicious fall by the hidden Markov model, and then Continue to calculate the coordinates of the target within a short period of time. If the displacement exceeds the set threshold, it indicates that the user can move normally and is not considered as a fall; otherwise, the second confirmation is passed and it is confirmed as a fall.
The fall detection method according to any one of claims 1 to 3, characterized in that, in the step S2, a Butterworth filter is used to perform filtering and noise reduction processing on the collected vibration signal, and a cutoff frequency of 10hz is used. High-pass filtering filters out DC components and low-frequency noise.
The fall detection method according to any one of claims 1 to 3, wherein in the step S2, in the end point segmentation processing, the entire segment of the vibration signal is first subjected to frame processing, and then each frame The variance of the signal is used as the criterion. When the variance of a certain frame signal exceeds a given threshold, it is considered that a valid vibration signal appears, and the signal of a certain length before and after the frame signal is taken as the vibration signal after the end segment.
A human fall detection system, which is characterized by comprising three geophones and a processor, and the processor executes the human fall detection method based on ground vibration signals according to any one of claims 1 to 10.