EP2421438A1

EP2421438A1 - System and method for determining the activity of a person lying down

Info

Publication number: EP2421438A1
Application number: EP10717610A
Authority: EP
Inventors: Pierre Jallon
Original assignee: Commissariat a lEnergie Atomique CEA; Movea SA; Commissariat a lEnergie Atomique et aux Energies Alternatives CEA
Current assignee: Movea SA; Commissariat a lEnergie Atomique et aux Energies Alternatives CEA
Priority date: 2009-04-24
Filing date: 2010-04-26
Publication date: 2012-02-29
Also published as: US20120101785A1; US8898041B2; WO2010122173A1

Abstract

The invention relates to a system for determining the activity of a person lying down, including at least one motion sensor (CM) with at least one measurement axis, provided with an attachment means (MF) for rigidly connecting said motion sensor (CM) to a user, characterised in that said system includes: a filter (FILT) for selecting, for at least one measurement axis of the motion sensor (CM), a high-frequency signal (HF) and a low-frequency signal (BF); a first means (CALC1) for calculating a first variable (x(n)) representing a temporal variation of the low-frequency signal (BF); a second means (CALC2) for calculating a second variable (y(n)) including the Euclidean norm, according to at least one measurement axis of the high-frequency signal (HF); and an analysis means (AN) suitable for determining an activity of a user lying down according to time using a hidden Markov model with N states respectively corresponding to N activities, said analysis means (AN) being suitable for combining: joint probability density functions of first and second variables, said probability density functions being defined for each state of the model in question, and the probability of changing between two consecutive states.

Description

SYSTEM AND METHOD FOR DETERMINING THE ACTIVITY OF A

EXTENDED PERSON

The present invention provides a system and method for determining the activity of an elongated person.

Systems and methods relating to hidden Markov model-based motion analysis are known, as described, for example, in the documents "Gesture Recognition Using the XWand" by Daniel Wilson and Andy Wilson and "Motion-based gesture recognition with an accelerometer "(bachelor's thesis) by PV Borza.

The paper "A hidden Markov model-based stride technology segmentation applied to equine inertial sensor trunk movement data," Thilo Pfau's Journal of Biomechanics 41 (2008) 216-220, Marta Ferrari, Kevin Parsons, and Alan Wilson, discusses analysis of the horse's walk.

An object of the invention is to allow to determine the activity of a person lying down.

According to one aspect of the invention, there is provided a system for determining the activity of an elongate person, comprising at least one motion sensor with at least one measurement axis, provided with fixing means for integrally binding said sensor motion to a user, the motion sensor producing a time signal S (n). The system includes:

a filter for selecting, for at least one measurement axis of the motion sensor, a high frequency signal and a low frequency signal,

first means for calculating a first variable x (n), representing a temporal variation of the low frequency signal,

second means for calculating a second variable y (n), comprising the Euclidean norm, according to at least one measuring axis, of the high frequency signal, and

analysis means adapted to determine an activity of an elongated user as a function of time by using a hidden Markov model with N states corresponding respectively to N activities, these analysis means being adapted to combine: - joint probability densities of a first and second variable, these density of probabilities being defined for each state of the considered model, and

- probabilities of passage between two successive states.

According to a preferred embodiment, the probability densities associated with each state comprise the product of at least a first probability density of said first variable by at least a second probability density associated with said second variable. It is recalled that a hidden Markov model is defined by two random processes: a first one which is called a "state" in the present application and which is not observed, or, in other words, which is hidden, and a second which is the observation whose probability density at a given moment depends on the value of the state at the same instant. According to this aspect of the invention, the state takes discrete values.

Such a system makes it possible to determine the activity of a person lying down, with precision and at a reduced cost.

In one embodiment, a joint probability density comprises a product of at least one probability density of obtaining the first variable and at least one probability density of obtaining the second variable, said probability densities. being defined by the following expressions:

where: P _BF (X) is the probability density of the first variable x corresponding to the state under consideration; P _HF (Y) is the probability density of the second variable y corresponding to the considered state; k represents the degree of freedom of the high frequency component (HF) equal to the number of measurement axes taken into account of said motion sensor (CM); σ _x represents a square root of the vahance of the first variable x, in the state of the hidden Markov model considered; σ _y represents a square root of the vahance of the second variable y in the state of the hidden Markov model considered; and F is the gamma function satisfying r - \ = 4π, r (l) = 1 and

Thus, for each state i, we define probability densities P _x j (n) and Pyj (n), such that

σ _y ; _\ being a quantity proportional to the time average of the variable y (n), in state i. For example, σ ^ is the time average of the variable y (n) divided by k.

Thus, the actual probability densities of the observed signals are approximated by probability densities broadly adapted to most movements. In one embodiment, said analysis means are adapted to determine elongated user activity, as a function of time using a hidden Markov model at at most five states among a rest activity, a small agitation activity. , tremor activity, agitation activity, and high agitation activity. Such a system makes it possible to analyze the activity of a mobile element with improved accuracy. Indeed, taking into account the high frequency component makes it possible to use additional information, which allows to detect small movements of the sensor, or, in other words, oscillations, or vibrations, such as tremors.

According to one embodiment, said analysis means are adapted to determine an elongated user's activity, from a set of predetermined pairs of values of first and second variances, defining classes of movement.

Thus, the unobserved process or hidden Markov model state is a discrete order 1 Markov process taking values in the set {0,1,2,3,4}. It is characterized by the probabilities of transition or transition from one state to another: p (State = j \ State = i).

The observed process of the hidden Markov model is the multidimensional signal (x (ή), y (ή)), whose joint probability density depends on the state (the hidden process) at a given moment. Thus, for each state, the joint probability density of the observed signal is defined by the following relation:

P {x (n), y (n) \ State = i) = P _ιEm {x (n), y (n)) = Σ Σtf _{! Βflf, m + (ffW + 1) n} P _{x (σM} Mn) ) P _y ( _σ u) {y (n))

B = O »1 = 0

where a _{state J} is a weighting coefficient for modeling a state by several elementary movements. This coefficient of determination is determined experimentally.

P _ιEm {χ, y), also denoted Pi (x (n), y (n)) represents the probability density associated with the state i, at the instant n, of x (n) and y (n) . It comprises the product of the previously defined probability densities P _{x ι} (x (n)) and P _{y ι} (y (n)).

If we consider a set of observed data θ (n), gathering the observed data x (n) and y (n), we can write that Pι (x (n), y (n)) = Pi (θ (n) = p (θ (n) / E (n) = i), E (n) representing the state at time n.

P (X, y \ State = i) being a probability density, the following condition must be realized by the weights:

"My," W

/, 2- ₁ State, m + (m _maI + l) n ^"" ^ "= 0 m = 0 The use of such classes, allows to define, for each pair of classes of movement, an elementary movement. A state of the model can then be described by several elementary movements.

In one embodiment, said pairs of values of first and second vahances are adapted to comply with the following conditions:

_ ≤ ¹⁰ m_ <10

"Max ^X "W> 6

Thus, the number of defined elementary movements can describe most of the movements of an elongated person and remains small enough to allow real-time treatments.

In one embodiment, the said classes of motion are eighteen in number and are defined by: σj0] = 5.10- ³ , σJl] = 1.8 × ² -2, σ _x [2] = 3.5 × 10 ^-2 , σ _c [3] = 5, 510- ^2, σ _c [4] = 8.10- ^2, σ _x [5] = 1.10- \ σ _y [0] = 1.10- ² σJl] = 3.10- \ σj2] = 8.10- ² These values are particularly well suited to determining the activity of a person lying down.

According to one embodiment, said real coefficients are defined by the following table:

These values are particularly well suited to determining the activity of a person lying down.

At each instant n, we can then determine a state of the person, knowing Pj (x (n), y (n)) by the expression: E (n) = arg max (P _{x ι} (x (n)) P _{y ι} (y (n)) = arg HIaX (P ₁ (x (n), y (n))

in which the arg max function represents maximum argument. If, at time n, the person is in the state i, E (n) = i. But the determination of the state E (n) at instant n, solely from the observed data, x (n) and y (n), and associated probability densities P _xi (x (n)) and P _y, i (y (n)) associated with these data, is generally unsatisfactory. Experience shows that it is necessary to take into account a priori, and for example the state E (n-1) determined during the time n-1. Consider the magnitude θ (n), gathering the observed data x (n) and y (n), we can write that P _i (x (n), y (n)) = P _i (θ (n) = = /), Where Ε (n) represents the state at time n. If Ε (0: N) denotes the series of states between the instant n = O and the instant n = N, and if Θ (O: N) denotes the observed data between the moment n = O and the moment moment n = N, the probabilities of the sequence of states Ε (0: N) corresponding to the sequence of states E (O), E (I) ... E (N) is written:

p (E (0: N) \ θ (0 - N -I)) ap (E (0)) p (θ (0) / E (0)) f [p (E (n) / E (n - V)) p (Θ (n) / E (n))

B = I For example, for the sequence E (0: N) = {i, i, i,, i}, this probability is written:

P (E (O) = i) p (θ (0) \ E (0) = p (E (n) = i \ E (n -1) = i)) p (θ (n) \ E (n) ) = i)) (1)

The estimated sequence of states E (0: N) is the one with the highest probability. In practice, rather than considering the set of possible sequences and for each calculate its probability, we can advantageously use a Viterbi algorithm to estimate this sequence.

P (E (O)) denotes the probability associated with the initial state E (O). One can for example choose an equiprobable distribution of each of the possible states when n = 0.

p (θ (0) / E (0)) represents the observation probability of the data θ (0) at the instant E (O). This corresponds to the probability Pi (x (n = 0), y (n = 0)) with E (n) = i. p (E (n) / E (n -I)) represents the probability of being in a state E (n) at the instant while one was in a state E (n-1) in the moment n-1. p (θ (n) / E (n)) represents the probability of observing the quantities θ (n) while in the state E (n). This corresponds to the probability Pi (x (n), y (n)) with E (n) = i.

The probabilities p (E (ή) / E (n-1)) correspond to probabilities of transition from a state E (n-1) to a state E (n). These probabilities are indicated in the following table by adopting the notations E (n-1) = j and E (n) = i, The series of states E (O) ... E (N) maximizing the expression (1 ) can be obtained using for example the Vitterbi algorithm, well known to those skilled in the art. So,

1) by establishing, for each state E (n): - the probability of observing the quantities θ (n) while we are in the state E (n), denoted p (θ (n) / E ( not))

the probability of passing from a state E (n-1) to a state E (n), denoted

2) by establishing the probability associated with each state E (O), 3) by obtaining observed quantities θ (n) at each instant n between n = 0 and n = N, we can obtain the series of states most probable E (O) .... E (N).

Recall that in the description, θ (n) = {x (n), y (n)}, x (n) and y (n) are respectively low and high frequency components of the signal S (n) measured by the motion sensor at the moment n.

In one embodiment, the probabilities P, of said hidden Markov model, of passage between two successive states respectively representing a type of posture are such that:

According to one embodiment, the system comprises display means.

Thus, a visualization of the results can be performed in real time or in deferred time.

In one embodiment, said motion sensor comprises an accelerometer, and / or a magnetometer, and / or a gyrometer. According to one embodiment, the system comprises fastening means adapted to be fixed to the wrist, torso or head of the person.

It is also proposed, according to another aspect of the invention, a method for determining the activity of a person lying down, in which one:

- Filtering, for selecting, for each measurement axis of a motion sensor, high frequencies greater than a first threshold, and low frequencies below a second threshold less than or equal to said first threshold;

a first variable representing, between a temporal variation, at least one measurement axis, said low frequencies is calculated;

a second variable is calculated that is equal to, at least the measurement axis, said high frequencies; and

a sequence of states is determined by combining:

- joint probability densities of a first and second values, these densities of probabilities being defined for each state, and - probabilities of passage between two successive states.

According to a preferred embodiment, the joint probability densities are established by determining, for each state:

first probability densities of a first value, and second probability densities of a second value.

The invention will be better understood from the study of some embodiments described by way of non-limiting examples and illustrated by the accompanying drawings in which: - Figures 1 and 2 illustrate a system, according to one aspect of the invention; and FIG. 3 illustrates an exemplary recording of a system according to one aspect of the invention.

FIG. 1 illustrates a system for determining the posture of a person comprising at least one CM motion sensor at less a measuring axis, provided with fastening means comprising for example a resilient element, for integrally bonding the motion sensor CM to a user. The CM motion sensor can be, an accelerometer, a magnetometer, or a gyrometer, with one, two, or three measurement axes.

The system comprises a FILT filter for selecting, for each measurement axis of the CM motion sensor, high RF frequencies higher than a first threshold S1, and low frequencies below a second threshold S2 less than or equal to the first threshold S1. The system also comprises a first module for calculating a first value x representing the variation, between two different instants, and for example successive, according to at least one measurement axis, of said low frequencies BF.

The system also comprises a second calculation module CALC2 of a second value y equal to the average of the variances, according to each measurement axis, of the high frequencies HF.

The establishment, at each instant (n), of the first value x (n) and the second value y (n) is more precisely explained below.

The signal measured by an accelerometer at an instant n can be decomposed into two summed components: S (n) = SG (n) + SP (n),

SG (n), representing the projection of the gravitational field, and SP (n) representing the own acceleration of the accelerometer carrier.

By applying a low pass filter to the signal S (n), the cutoff frequency being less than 1 Hz, and for example equal to 0.5 Hz, an SG ^* (n) estimate of SG (n) is obtained.

Thus, an estimate SP ^* (n) of SP (n) is such that SP ^* (n) = S (n) - SG ^* (n)

It is then possible to define a first value, or low frequency component, and denoted x (n), representing a variation of the signal SG ^* between different instants, and for example successive ones, that is to say SG ^* (n-1 ) and

SG ^* (n). Thus, it can be said that x (n) represents a temporal variation of the low frequency signal of the accelerometer. We can advantageously choose: x (n) = ^ SG _k ^* (n) - SG _k ^* (n - 1),

the index k representing the axis (or axes) according to which the component Sk (n) of the signal S (n) has been measured.

We can determine x (n) using only one component (for example the vertical axis), or several components Sk (n) of the signal S (n).

It is also possible to define a second value, or high-frequency component y (n) of the signal S (n), this component being determined from SP ^* (n). According to a preferred embodiment, y (n) = \\ SP * (n) f, this norm being calculated by considering one or more axes.

Thus, it can be said that y (n) comprises the average of the variances, according to at least measurement axis, of the measured high frequency signal.

AN analysis means make it possible to determine an activity of an elongated user as a function of time by using a hidden Markov model with N states corresponding respectively to N activities.

Thus, for each state i, we define probability densities P _x j (n) and

P _y, i (n), such that

k representing the degree of freedom of the high frequency component (HF) equal to the number of measurement axes taken into account of said motion sensor (CM); σ _x representing a square root of the variance of the first variable x, in the state of the hidden Markov model considered; σ _{yt \} being a quantity proportional to the time average of the variable y (n), in the state i. For example, σ _{y! 1} is the time average of the variable y (n) divided by k; and r being the gamma function satisfying r - \ = 4π, r (l) = 1 and

The system also includes an AFF display screen. For example, the system includes an accelerometer with a measurement axis and a fastener for attaching the accelerometer to the wrist of the user.

The AN analysis means are adapted to determine an elongated user's activity as a function of time using a hidden Markov model at at most five states among a rest activity, a small agitation activity, a tremor activity , a stirring activity, and a high agitation activity.

The analysis means AN are adapted to determine an activity of the elongated user, from a set of predetermined pairs of values of first and second variances, defining motion classes. For example, the pairs ((σjm ^ σjn])) of values of first and second variances are adapted to meet the following conditions:

The analysis means AN are adapted to determine the probability density of an iState state of the hidden Markov model by the following relation:

P ₁ EIaI (*> ^ - Σ Σ ^a state, m ₊ (m _mm + 1) » ^P BF (σ _x [m]) U) ^P H _F {σ _γ []]) G) n = 0 m = _Where : a _state, ^{is a} real coefficient between 0 and 1;

jmax

Σ ^a _{> Status, j} = 1 and; "= Im _1112x + I) (^ _x + I) -I

J = O The classes of motion are, for example, eighteen in number and are defined by: σJθ] = 5.1O- ³ , σ _{; c} [l] = 1, 8.1O- ² , σ _x [2] = 3, 5.1O- ² , σ _{; c} [3] = 5., 51O- ² , σ _{; c} [4] = 8.1O- ² , σ _Λ [5I = LlO- ¹ , σ _y [θ] = l.Kr ² , σJl] = 3.1 (r ² , σj2] = 8.1 (r ² .

The actual coefficients at _State are defined by the following table:

The probabilities P, of said hidden Markov model, of passage between two states respectively representing a type of posture are such that:

State δ O .007 0. 0010 0. 0010 0. 0010 0.99

(strong agitation)

For example, the motion sensor CM may comprise an accelerometer, and / or a magnetometer, and / or a gyrometer.

FIG. 2 illustrates an embodiment variant, in which, unlike the embodiment of FIG. 1, for which the motion sensor CM and the other elements are included in a box BT, elements of the system can be outsourced, for example in a laptop OP.

The analysis module AN determines, from the input signals and the hidden Markov model as defined, the sequence of states (postures) most likely, according to conventional methods, for example by calculating for the whole sequences of possible states the associated probability taking into account the observed signal and keeping the most probable sequence, as described for example in the document "An introduction to hidden Markov models" LR Rabiner and BH Juang, IEEE ASSP Magazine,

January 1986, or in the book "Inference in Hidden Markov Models" of

Cappé, Moulines and Ryden by Springer, from the series "Springer series in statisctics".

Figure 3 illustrates an example of a user registration of a system of Figures 1 and 2, including a triaxial accelerometer, and the result provided by the system.

This recording portion of an elongated user was performed over a period of about ten minutes.

The system indicates that the elongated user was in a resting activity (étati) from 6h 04min 53s to 6h 05min 18s, in a tremor activity (state3) from 6h 05min 18s to 6h 07min 57s, in a rest activity ( etati) from 6h 07min 57s to 6h 08min 10s, in a stirring activity

(state4) from 6h 08min 10s to 6h 08min 22s, in an activity of rest (étati) from 6h 08min 22s to 6h 08min 32s, in a tremor activity (state3) from 6h 08min 32s to 6h 08min 57s, in an activity d 'agitation (state4) of 6h

08min 57s to 6h 09min 33s, in a small activity agitation (state2) of 6h 09min 33s to 6h 09min 47s, in an activity of rest (étati) from 6h 09min 47s to 6h 1 1 min 49s, in an activity of agitation (state4) of 6h 1 1 min 49s with 6h 13min 07s, in a activity of rest (étati) from 6h 13min 07s to 6h 13min 19s, in a stirring activity (state4) from 6h 13min 19s to 6h 13min 35s, in a rest activity (étati) from 6h 13min 35s to 6h 14min 22s, in a tremor activity (state 3) from 6h 14min 22s to 6h 14min 59s, in a shaking activity (state4) from 6h 14min 59s to 6h 15min 18s, and in a rest activity (étati) from 6h 15min 18s to 6h 15min 40s .

The present invention makes it possible, at reduced cost and with improved accuracy, to determine in real time or deferred the posture of a person, by accurately determining the changes in posture. Such an invention makes it possible to analyze the activity of a sleeping person, or to detect a nocturnal epileptic seizure, with improved accuracy and reduced cost.

Claims

1. System for determining the activity of an elongated person, comprising at least one motion sensor (CM) with at least one measurement axis, provided with fixing means (MF) for integrally connecting said motion sensor (CM) to a user, characterized in that it comprises:

a filter (FILT) for selecting, for at least one measurement axis of the motion sensor (CM), a high frequency signal (HF) and a low frequency signal (BF); ,

first calculation means (CALC1) of a first variable (x (n)) representing a temporal variation of the low frequency signal (BF);

second calculation means (CALC2) of a second variable (y (n)) comprising the Euclidean norm, according to at least one measurement axis of the high frequency signal (HF); and

analysis means (AN) adapted to determine an activity of an elongated user as a function of time by using a hidden Markov model with N states corresponding respectively to N activities, said analysis means (AN) being adapted to combine:

- joint probability densities of a first and second variable, these density of probabilities being defined for each state of the considered model, and

- probabilities of passage between two successive states.

The system of claim 1, wherein the probability densities (P (x, y)) associated with each state (i) comprise the product of at least a first probability density of said first variable (x) by at least one second probability density associated with said second variable (y).

The system of claim 2, wherein a joint probability density (P (x, y)) comprises a product of at least one probability density (P _x j (n)) for obtaining the first variable ( x) and at least one probability density (P _y j (n)) for obtaining the second variable (y), said probability densities (P _x j (n), P _y, i (n)) are defined by the following expressions:

in which: k represents the degree of freedom of the high frequency component (HF) equal to the number of measurement axes taken into account of said motion sensor (CM); σ _x representing a square root of the variance of the first variable x, in the state of the hidden Markov model considered; σ _{y, \} being a quantity proportional to the time average of the variable y (n), in the state i. For example, cr ,, ,, is the time average of the variable y (n) divided by k; and r being the gamma function satisfying r - \ = 4π, r (l) = 1 and

The system of one of the preceding claims, wherein said analyzing means (AN) is adapted to determine an elongated user's activity as a function of time using a hidden Markov model at at most five states among resting activity (étati), small agitation activity (state2), tremor activity (state3), agitation activity (state4), and high agitation activity (stateδ).

The system of claim 4, wherein said analyzing means (AN) is adapted to determine elongated user activity from a set of pairs {{σ _x [m]; σ ["])) predetermined values of first and second variances, defining motion classes {σ _x , σ _y ).

6. System according to claim 5, wherein said pairs ((σjm]; ^ ^])) of values of first and second variances are adapted to comply with the following conditions:

The system of claim 6, wherein said analyzing means (AN) is adapted to determine the joint probability density of a state (iState) of said hidden Markov model by the following relation:

P (x (n), y (n) \ State = i) = P _state {x (n), y (n)) = Σ ΣΩW _{, m + (BW +1) π} P _{x {σ, M)} { x (n)) P _ΛσM) {y (n))

where: iState represents a state of the hidden Markov model; a _State] is a weighting coefficient; and v Σ vΣ a state, m + (m _mn + 1) n = 1 n = 0 »1 = 0

8. The system of claim 7, wherein said motion classes are among eighteen and are defined by: σj0] = 5.10- ³ σJl] = l, 8.10- ^2, σ _x [2] = 3 , 5.10- ^2, σ _c [3] = 5, 510- ^2, σ _c [4] = 8.10- ² σj5] = 1.10- \ σjθ] = l.lθ- \ σ [l] = 3.1 (T ² , σ [2] = 8.1 (T ² .

9. The system of claim 8, wherein said real coefficients {a _ιState} ) are defined by the following table:

10. The system of claim 9, wherein the probabilities (P (State / Etati)) of said hidden Markov model, of passage between two states respectively representing a type of posture are such that:

1. System according to one of the preceding claims, comprising display means (AFF).

12. System according to one of the preceding claims, wherein said motion sensor (CM) comprises an accelerometer, and / or a magnetometer, and / or a gyrometer.

13. System according to one of the preceding claims, comprising fixing means (MF) adapted to be fixed to the wrist, torso, or head of the person.

14. A method for determining the activity of an elongated person, characterized in that one

- filtering, for selecting, for each measurement axis of a motion sensor (CM), high frequencies (HF) higher than a first threshold (S1), and low frequencies (BF) below a second threshold ( S2) less than or equal to said first threshold (S1); - calculating a first variable (x (n)) representing, between a temporal variation, according to at least measurement axis, said low frequencies (BF);

calculating a second variable (y (n)) equal to standard, according to at least measurement axis, of said high frequencies (HF); and - a sequence of states is determined by combining:

- joint probability densities of a first and second values, these density of probabilities being defined for each state, and

- probabilities of passage between two successive states.

The method of claim 14, wherein the joint probability densities are established by determining, for each state: