BE1026660B1 - A human body fall detection system based on ground vibration signals - Google Patents

Abstract

The invention relates to a method for detecting a fall in the human body based on ground vibration signals, the method comprising the steps: collection of ground vibration signals by three geophones, filtering processing, noise reduction processing, processing endpoint segmentation and alignment processing is performed on the collected vibration signals, then the signals are superimposed and the characteristics of the signal are extracted; The extracted characteristic signals are composed in training sets for learning the hidden Markov model, the currently used ground vibration signals are used as test data which are processed by processing training data. The hidden Markov model is used to calculate the probability of a fall and the EoA positioning algorithm is used to follow the coordinates over a short period, if the moving range of the target is below the threshold, the secondary confirmation is a fall, otherwise it is not a fall. A human body fall detection system is also described accordingly. The invention uses the hidden Markov model to calculate the vibration signal as the probability of a fall, and then combines the EoA positioning algorithm for secondary confirmation, which allows precise recognition of the fall and a false positive rate. low, and does not require the wearing of equipment, protects the risk of invasion of privacy and improves the user experience.

Description

Description

A human body fall detection system based on ground vibration signals

Technical field of the invention

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

Prior art

The trend of the aging of the world population is irreversible: according to UN statistics, there would be more than 960 million people aged over 60 years in the world in 2017, and this number should double to reach 2 billion in 2050. For seniors, falling could be a deadly event they can face every day. According to the World Health Organization, a third of people over the age of 65 fall at least once a year, and 60% of falls occur at home. Therefore, we urgently need to invent a practical and effective method and system for automatically detecting the fall of an elderly person. The existing fall detection methods are mainly based on vision, on the mobile phone and on the radio signal. The vision-based detection method can effectively detect the fall event, but the privacy problem is complex: its deployment is difficult in sensitive environments such as bathrooms where the floor slides a lot and where there are a lot of falls. In addition, the camera also has a visual blind spot and cannot operate in a dark environment. The portable device-based fall detection system can detect if the user is falling using the device's inertial sensor, but this kind of portable device is not user friendly, especially people elderly at home who cannot keep the device at all times. Once the device is removed, the detection is invalid. Radio communication technology based fall detection method does not require the user to wear the device, but the high false positive rate has long been criticized and the multipath propagation effect of the signal makes it work difficult in a complex and dynamic domestic environment.

BE2019 / 5278

Statement of the invention

In order to solve the above technical problem, the present invention proposes to use geophones to collect the vibration signals from the ground and to implement a more precise method and system, with low false positive rate, without risk of '' invasion of privacy, without user training and without wearing equipment, by computer processing. The method and the system specifically adopt the following technical solutions:

A method of detecting a fall in the human body based on vibration signals from the ground, comprising the following steps:

S1: collection of ground vibration signals by three geophones;

S2: filtering processing, noise reduction processing and endpoint segmentation processing are performed on the collected vibration signals to determine the starting point and ending point of an effective vibration signal;

S3: alignment processing is performed on the actual vibration signals after processing of endpoint segmentation;

S4: superimposition of the effective vibration signals of the three geophones, then extraction of the characteristics of the signals;

S5: the extracted characteristic signals are composed in training sets for learning the hidden Markov model;

S6: the ground vibration signals currently in use are used as test data which is processed by processing training data;

S7: calculation of the probability of a fall with the hidden Markov model, if said probability is greater than a threshold, a secondary confirmation is made, otherwise it is not considered as a fall;

S8: application of an EoA positioning algorithm to follow coordinates over a short period, if the moving range of the target is less than a threshold, the secondary confirmation is a fall, otherwise it is not a fall.

In a preferred embodiment, in step S3, the actual vibration signal after endpoint segmentation processing is aligned by the general cross-correlation method, and said specific operation of the alignment processing is to calculate the offset between the two effective vibration signals, the signal

BE2019 / 5278 of the actual effective vibration is then moved, only the full part shared between the two vibration signals is taken after the movement.

In a preferred embodiment, in step S3, the formula is adopted:

not

C (a, b) = £ a (i) -b (i) P (A, B) = arg max (C (A [i, i + n -1], B)), ie 1,2 ,. .., 2n ^{! = 1} > i and ⁽ , ^B) = ^{P (} ) ^{- n} calculates the offset O (A, B) between the two effective vibration signals, where a and b represent two effective vibration signals with a length signal 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 correlation between the effective vibration signal a and the effective vibration signal b; A represents the filling to zero on both sides of the effective vibration signal a whose length is equal to n, and thus obtains the first signal of length 3n; B represents the effective vibration signal b of length n; P (A, B) represents the signal position of the first signal A having the highest correlation with the second signal B of length n; O (A, B) is the offset calculated between the first signal A and the second signal B.

In a preferred embodiment, in which in step S4, the effective vibration signals from the three geophones are superimposed, processed by a weighted filter, the signal is then decomposed into several layers of components of different frequency by transforming into discrete wavelet, and calculates the energy corresponding to certain time intervals, finally, the energy characteristics in the time domain and in the frequency domain are obtained using the formula Feature = 10 log ₁₀ E, where E is the energy determined initially from each time interval at different frequency levels, and Feature is a corresponding feature after logarithmic transformation.

In a preferred embodiment, in which the state of the hidden Markov model of step S5 is equal to 7, the initial state is equal to 1 and the probability function observed at the output of each state is composed by several Gaussian mixing functions, wherein the number of Gaussian mixing functions is equal to the number of layers of the discrete wavelet transform decomposition in step S4, and each Gaussian mixing function has three Gaussian components.

BE2019 / 5278

In a preferred embodiment, in which in 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 at the threshold, the secondary confirmation is carried out by step S8, otherwise it is directly determined that it is not a fall.

In a preferred embodiment, in which in step S8, the positioning algorithm EoA (Full Name Arrival Energy) proposed by the present invention is used for positioning, which is based on the attenuation model of the signal Amp (d) = Am ^ e ~ ^axd , in which Am ^ represents the initial amplitude of the vibration signal, and d represents the distance over which the vibration propagates, & represents the attenuation coefficient based on the medium of propagation, the following formula is given:

—2u = ^c 1 ^{ln E} ac —2u = ^C 2 where d _lt d ₂ , d ₃ are respectively the distances from the source of vibration to the three geophones A, B and C, & (^ ¹ ) is the logarithm of the signal-energy ratio received by geophones A and B, the position of coordinates of the source of vibration can be solved by the above formula.

In a preferred embodiment, in which in step S8, the positioning algorithm EoA calculates the coordinate position of the suspected fall after the test data has been recognized as a suspicious fall by the hidden Markov model, then continues to calculate the coordinates of the target over a short period, if the movement exceeds the defined threshold, this indicates that the user can move normally and that it is not a fall, otherwise the secondary confirmation is transmitted, and confirmed as a fall.

In a preferred embodiment, in which in step S2, the collected vibration signals are subjected to filtering processing and noise reduction processing using the Butterworth filter, and the DC components and the low frequency noise are eliminated by high pass filtering with the cutoff frequency of 10 Hz.

BE2019 / 5278

In a preferred embodiment, in which in step S2, when processing the endpoint segmentation, the entire segment of the vibration signal is first subjected to frame processing, then the variance of each signal frame is used as a criterion: when the variance of a certain signal frame exceeds the given threshold, the effective vibration signal is considered to appear, and a signal of a certain length before and after this signal frame is removed as a vibration signal after processing endpoint segmentation.

A human body fall detection system is characterized in that it includes three geophones and a processor, and the processor performs the human body fall detection method based on the vibration signals from the ground.

The invention real-time monitoring of the fall as a function of ground vibrations, calculates the vibration signal as the probability of falling by the hidden Markov model, then performs secondary confirmation with the positioning algorithm EoA. Recognition of the fall is precise and the rate of false positives is low. Damage caused by the fall can be dealt with in time, and the problem of poor robustness of the prior art, problems with wearing equipment, problems with accuracy and risks of invasion of privacy are resolved, and the user experience is improved. The detection method of the invention is new, practical and reliable, and can meet the requirements of use of various living environments, and has wide application and low cost.

Brief description of the drawings

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

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

Figure 3 is a flow chart showing the effect simulation before the actual vibration signal alignment processing after the end point segmentation processing of the present invention.

Fig. 4 is a flow chart showing the effect simulation after the actual vibration signal alignment processing after the end point segmentation processing of the present invention.

BE2019 / 5278

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

Detailed description of exemplary embodiments of the invention

The present invention will be described in more detail below with reference to the drawings and embodiments. It is understood that the specific embodiments described here are merely illustrative of the invention and are not intended to limit the invention.

The preferred embodiments of the present invention are described in more detail below with reference to the accompanying drawings.

As indicated in FIG. 1, the present invention proposes a fall detection method based on the hidden Markov model and the EoA algorithm, comprising the following steps:

Step S1: place a geophone (or any other sensor capable of detecting vibrations) in a corner of the operating environment, detect a vibration signal acting on the ground when the human body falls, convert it into an analog electrical signal , then converting said analog electrical signal into a digital signal to be processed. Three geophones can be adopted in particular: Figure 2 is an external view of the geophone.

Step S2: uses the Butterworth filter to perform filtering processing and noise reduction processing on the collected vibration signals using bandpass filtering with a frequency band greater than 10 Hz. More specifically, said embodiment uses a high-pass filter with a cut-off frequency of 10 Hz to filter the DC components and low frequency noise.

The endpoint segmentation processing is also called endpoint detection processing, and is executed by a cutting algorithm in a computer program to determine the starting point and ending point of the actual vibration signal. The processing process consists first of all in subjecting the entire segment of the vibration signal to frame processing, then the variance of each frame of the signal is used as a criterion: when the variance of a certain frame of the signal exceeds the threshold given, the effective vibration signal is

BE2019 / 5278 considered as appearing and a signal of a certain length before and after this frame of the signal is withdrawn as a vibration signal after endpoint segmentation, the vibration signal after processing of endpoint segmentation is also called an effective vibration signal. The given threshold can be customized according to the needs of the user or based on the value of the training set of the sample.

Step S3, the actual vibration signal after endpoint segmentation processing is aligned by the general cross correlation (GCC) method, said operation of the alignment processing consists in calculating the offset between the two signals effective vibrations. The current effective vibration signal is then moved, only the full part shared between the two vibration signals is taken after the movement. Alignment processing in this embodiment can align all of the effective vibration signals, which is beneficial for the classification accuracy of the machine learning algorithm, and the simulation effect diagrams before and after processing The alignment signals are illustrated in FIGS. 3 and 4. After the GCC processing, the cut-off signal is further aligned in the time series to more precisely determine the fall signal.

In said step S3 of this example, the formula is adopted:

not

C (a, b) = £ a (i) -b (i) P (A, B) = arg max (C (A [i, i + n -1], B)), ie 1,2 ,. .., 2n ^{! = 1} > i and O (A, Bj = P (A, Bj - n calculates the offset O (A, Bj between the two effective vibration signals, where a and b represent two effective vibration signals with a signal 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 correlation between the effective vibration signal a and the effective vibration signal b; The first signal A represents the zero filling on both sides of the effective vibration signal a whose length is equal to n, and thus obtains the first signal of length 3n; The second signal B represents the effective vibration signal b of length n; P (A, Bj represents the signal position of the first signal A having the highest correlation with the second signal B of length n ; O (A, Bj is the calculated offset between the first signal A and the second signal B.

BE2019 / 5278

Step S4, after a vibration, the signals from the three geophones are superimposed, processed by a weighted filter, the discrete wavelet transform is then used to extract the energy intensity from different frequency and time domains of the vibration signal as characteristics of the extracted signal. Preferably, in step S4, the energy characteristics in the time domain and in the frequency domain are obtained by the formula Feature = 10 log ₁₀ E, where E is the energy determined initially from each time interval at different frequency levels, and Feature is a signal of functionality after logarithmic transformation.

Step S5, the extracted characteristic signals are composed in a training set intended to form the hidden Markov model and, when the training is carried out, only the vibrating signals of the falling ground coming from other people (such as volunteers or of participants) is used as a training set.

More precisely, the number of transition states of the hidden Markov model is equal to 7 and the initial state to 1. The probability of observation function at the output of each state is composed of several Gaussian mixture models containing several components the number of Gaussian mixture models and the number of layers of the discrete wavelet transform decomposition in step S4 is the same.

Initially, the relevant parameters are initialized according to the training set, i.e. initializes the state probability, the state transition probability and the output observation probability, and a model of initial hidden Markov is obtained. Then, the baum-welch algorithm is used to optimize the parameters of the model according to the training set. After each optimization, all the samples of the training set will be calculated one by one by the viterbi algorithm in order to obtain the degree of correspondence between the training set and the model (i.e. say the probability of falling. In reality, it is not the probability. In order to improve calculation performance, the calculation process uses a logarithmic operation. So the result is negative) and records the degree. Then, from the second call to the baum-welch algorithm for optimization, it is necessary to compare the difference in the degree of total correspondence between

BE2019 / 5278 the drive unit and the model (i.e. the probability of falling) between the present and the last optimization. If it is below the threshold, it means that the drive converges and jumps out of the loop. A formed hidden Markov model is obtained. For the following test samples, it suffices to calculate the degree of correspondence between the samples and the model (i.e. the probability of falling) by the viterbi algorithm.

Step S6, the currently used ground vibration signals are used as test data which is processed by processing training data; in actual use, the test data is collected in the same way as steps S1, S2, S3 and S4.

Step S7, calculating test data as probability of falling based on the hidden Markov model obtained in step S5 and, if the probability is greater than the threshold, the process goes to step S8 for secondary confirmation , otherwise it is directly determined that it is not a fall.

Step S8, the EoA positioning algorithm is used to calculate the target position for the short period following the suspected fall event and the target tracking. If the movement exceeds the threshold, the warning is disarmed, otherwise the secondary confirmation succeeds and the fall is confirmed.

More precisely, the positioning calculation is carried out on the basis of the signal attenuation model Amp (d) = Am ^ e ~ ^axd , as illustrated in FIG. 5. Where Amp ₀ represents the initial amplitude of the vibration signal and d represents the distance over which the vibration propagates, & is the attenuation coefficient based on the propagation medium. For three geophones A, B and C, the amplitudes of the received signal are the

following: A (d _± ) = Amp ₀ e ~ ^axdl B (d ₂ ) = Amp ₀ e ~ ^axd2 C (d ₃ ) = Amp ₀ e ~ ^axds

where d _lt d ₂ , d ₃ are respectively the distances from the source of vibration to the three geophones A, B and C. Then, the signal-energy ratio received by the geophones A and B is as follows:

Eab - A ² (d1) Ampoe ^axdl 2 _. -__ ^v = (___LX_______) = _e ~ 2ax (d ₁ -d ₂ ) B ² (d ₂ ) Amp _o e ~ ^axd z

BE2019 / 5278 after taking the logarithm lnE _AB = -2a (d ₁ - d ₂ ), then pass the following formula:

d ^ d ₂ d ^ d ₃

Zn E _AB 2a ~ ^Ci lnE _AC 2a = ^C2 where, on the basis of the signal-energy ratio obtained in advance, we can obtain c _{1 (} c ₂ , thus obtaining the position of the vibration source and performing the positioning.

The present invention makes it possible to obtain the hidden Markov model formed in combination with the EoA algorithm, then to use the hidden Markov model to perform fall detection, detect the vibration signal in real time through the geophone and when the fall event occurs or other actions (such as falling things, etc.) that generate a high energy vibration signal. The system then detects the vibration signal, extracts the vibration signal by performing noise reduction and filtering processing, endpoint detection, GCC alignment and extraction of the characteristics of the vibration signal. The generated characteristics of the signal are used as input to the hidden Markov model and the results returned by the hidden Markov model are obtained. The suspected fall signal is then confirmed by the EoA algorithm and the conclusion is reached.

This embodiment also provides a fall detection system based on the hidden Markov model and the EoA algorithm, which adopts said fall detection method based on the hidden Markov model and the EoA algorithm.

The foregoing is a further detailed description of the present invention in connection with the specific preferred embodiments, and the specific embodiments of the present invention are not limited to the description. For ordinary technicians concerned with the present invention, it is possible to carry out a certain number of extrapolations or simple replacements, without departing from the conception of the invention, which should be considered to fall within the scope of protection. of the present invention.

Claims (11)

1. A method for detecting a fall in the human body based on vibration signals from the ground, characterized in that the method comprises the following steps: S1: collection of ground vibration signals by three geophones; S2: filtering processing, noise reduction processing and endpoint segmentation processing are performed on the collected vibration signals to determine the starting point and ending point of an effective vibration signal; S3: alignment processing is carried out on the effective vibration signals after processing of endpoint segmentation; S4: superimposition of the effective vibration signals of the three geophones, then extraction of the characteristics of the signals; S5: the extracted characteristic signals are composed in training sets for learning the hidden Markov model; S6: the ground vibration signals currently used are used as test data which is processed by processing training data; S7: calculation of the probability of a fall with the hidden Markov model, if said probability is greater than a threshold, a secondary confirmation is made, otherwise it is not considered as a fall; S8: application of an EoA positioning algorithm to follow coordinates over a short period, if the target's moving range is less than a threshold, the secondary confirmation is a fall, otherwise it is not a fall. 2. The detection method according to claim 1, in which, in step S3, the effective vibration signal after processing of endpoint segmentation is aligned by the general cross-correlation method, and said specific operation of the processing of alignment consists in calculating the offset between the two effective vibration signals, the current effective vibration signal is then moved, only the complete part shared between the two vibration signals is taken after the movement. BE2019 / 5278 3. Effective detection method according to claim 2, in which, in step S3, a formula is adopted: not C (a, b) = £ a (i) -b (i) P (A, B) = arg max (C (A [i, i + n -1], B)), ie 1,2 ,. .., 2n ^{! = 1} > i and 0 (A, B ~) = P (A, B ~) - n calculates the offset 0 (A, B ~) between the two effective vibration signals, where a and b represent two effective vibration signals with signal 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 correlation between the effective vibration signal a and the effective vibration signal b; A represents the filling to zero on both sides of the effective vibration signal a whose length is equal to n, and thus obtains the first signal of length 3n; B represents the effective vibration signal b of length n; P (A, B ~) represents the signal position of the first signal A having the highest correlation with the second signal B of length n; O (A, B ~) is the offset calculated between the first signal A and the second signal B. 4. Detection method according to any one of claims 1 to 3, in which in step S4, the actual vibration signals from the three geophones are superimposed, processed by a weighted filter, the signal is then broken down into several layers of different frequency components by discrete wavelet transform, and calculates the corresponding energy at certain time intervals, finally, the energy characteristics in the time domain and in the frequency domain are obtained using the formula Feature = 10 log ₁₀ E, where E is the energy determined initially from each time interval at different frequency levels, and Feature is a corresponding characteristic after logarithmic transformation. 5. Detection method according to claim 1, in which the state of the hidden Markov model of step S5 is equal to 7, the initial state is equal to 1 and the probability function observed at the output of each state is composed by several Gaussian mixing functions, in which the number of Gaussian mixing functions is equal to the number of layers of the decomposition by BE2019 / 5278 transformed into discrete wavelets in step S4, and each Gaussian mixing function has three Gaussian components. 6. The detection method according to claim 1, in which in step S7, the probability that the test data is a fall signal is calculated on the basis of the hidden Markov model obtained in step S5, and s' it is greater than the threshold, the secondary confirmation is carried out by step S8, otherwise it is directly determined that it is not a fall. 7. Detection method according to claim 1, in which in step S8, the positioning algorithm EoA (Full Name Arrival Energy) proposed by the present invention is used for positioning, which is based on the model d signal attenuation Amp (d) = Am ^ e ~ ^axd , in which Am ^ represents the initial amplitude of the vibration signal, and d represents the distance over which the vibration propagates, & represents the attenuation coefficient based on the medium of propagation, then the following formula is given: —2u = ^C 1 ^{ln E} ac —2u = ^C 2 where d _lt d ₂ , d ₃ are respectively the distances from the source of vibration to the three geophones A, B and C, d ₂ ~) is the logarithm of the signal ratio- energy received by geophones A and B, the position of coordinates of the vibration source can be solved by the formula above. 8. The detection method according to claim 7, in which in step S8, the positioning algorithm EoA calculates the coordinate position of the suspected fall after the test data has been recognized as a suspected fall by the model of Markov hidden, then continues to calculate the coordinates of the target over a short period, if the displacement exceeds the defined threshold, this indicates that the user can move normally and that he BE2019 / 5278 is not a fall, otherwise the secondary confirmation is transmitted, and confirmed as a fall. 9. A fall detection method according to any one of claims 1 to 3, in which in step S2, the collected vibration signals are subjected to filtering treatment and noise reduction treatment using the Butterworth filter, and the DC components and low frequency noise are eliminated by high pass filtering with the cutoff frequency of 10 Hz. 10. A fall detection method according to any one of claims 1 to 3, wherein in step S2, during the processing of endpoint segmentation, the entire segment of the vibration signal is first subjected to frame processing, then the variance of each frame of the signal is used as a criterion: when the variance of a certain frame of the signal exceeds the given threshold, the effective vibration signal is considered to appear, and a signal of a certain length before and after this frame of the signal is removed as a vibration signal after processing endpoint segmentation. 11. A fall detection system of the human body characterized in that it comprises three geophones and a processor, the processor executes the fall detection process of the human body on the basis of the vibration signals from the ground according to any one of claims 1 to 10.