EP2207154A1 - Method and device for detecting a fall - Google Patents

Abstract

The method involves classifying characteristics of a user from user's activity indicators obtained from acceleration measurements acquired by an accelerometer (20) i.e. three-axis accelerometer, of a fall detection device (10), where the device is in the form of pendant, wrist-watch strap or belt. Actuation of an electromagnetic or radiofrequency alarm signal is delayed so as to avoid false alarms, where the signal is sent from a transmission unit (18) via a base connected to a telecommunication network e.g. Internet or global system for mobile communication (GSM) network. An independent claim is also included for a user fall detection device comprising a case enclosing a supply unit.

Description

The present invention relates to a method of detecting a fall of a person and the device associated with this method.

We know that in the elderly, one in three people is at risk of falling, each fall being likely to cause risks for the person and a further loss of autonomy: the psychological after-effects left by a long period of consecutive immobility a fall and without assistance being an important factor in this loss of autonomy and generally leading to the cessation of home support.

We also know that in certain professions, a person may be required to work alone, and sometimes even alone in dangerous conditions.

Also, whether it is an elderly person, disabled, or a patient in a health facility, in a health care facility or at home, or whether it is a person working in a environment at risk or in isolation, it is necessary that said person can itself easily alert the rescue during a malaise or any incident.

Thus, known devices of the prior art in the form of medallion, bracelets, or belts, allowing their user to send a call for help by means of a voluntary decision, such as pressing a button push, in case of distress or discomfort.

These devices of the prior art make it possible to transmit an alert signal to a remote assistance center, or to any other designated person, directly or via a base connected to a telecommunication network such as the Internet or GSM. Or other.

A major disadvantage of these devices of the prior art is to require a voluntary action by the user to trigger the sending of the alert signal.

Excluding, in the event of a fall following an accident or a malaise, the user is generally surprised and does not have the necessary time to react and trigger the alert before or during the fall. And after the fall, the person may be unconscious or in a state or position preventing him from operating the alert device.

It is to meet the needs of those people who may fall and be distressed in certain situations that user-developed devices that can monitor a user's behavior and identify a movement have been developed. suspect corresponding to a fall to automatically trigger the sending of an alert signal.

However, several difficulties arise in the development of these fall detection devices.

First of all, these devices are difficult to accept by their user, and moreover, in the case of the fall detection of elderly persons of particular interest to the invention, nearly half of the falls occurring at night and during the toilet, said device must be worn almost permanently.

A fall detection device must therefore be the least inconvenient in the habits or activities of the user, and also, said device must be discreet so that the user does not refuse to wear it vis-à-vis his entourage .

Also, said device must be easily attached to a simple place.

An interesting position retained by several devices of the prior art is the wrist of the user, said devices being in the form of bracelets.

Indeed, this location makes it possible to respond optimally to the constraints set out above in the manner of a watch or other jewel.

However, unlike a device placed for example on the chest of the user, the detection at the wrist is more difficult because we do not know the position of the trunk of the user.

In more detail, the sensor adopted by most devices of the prior art for the fall detection is the accelerometer for measuring shocks and movements.

The devices of the prior art have encountered major difficulties in positioning an accelerometer at the wrist of the user.

First, the many possible movements of the arm can be interpreted as a fall, and secondly, there is no obvious and reliable reference point in space and in relation to the ground at the position of the body.

Because of these difficulties, many devices of the prior art make detection errors and trigger false alarms.

Without, a fall detection device must absolutely avoid triggering alerts inadvertently, the accumulation of false alerts to remote assistance centers, or other people, discrediting the detection and may lead to unnecessary interventions.

To overcome these errors and ensure the detection of a fall, other prior devices, such as that described in the French patent FR-2874727 of the same applicant, furthermore include means for measuring other physical quantities, such as shock sensors, inertial units, temperature sensors, or others, in order to seek to correlate the acceleration measurements with other clues such as heart rate and blood pressure, body temperature, or even the ambient sound level.

This multiplication of the measuring means increases the size of the detection device, which goes against the criteria of discretion of the device in the life of the user, and especially the cost of the device.

In addition, the multiplication of sensors also decreases the electrical autonomy of these devices, this is particularly troublesome in cases, such as for the elderly, where the device must be worn almost permanently.

Finally, another obstacle encountered by the devices of the prior art in the context of the supervision of a senior concerns the large proportion of falls occurring in the bathroom, said detection devices must therefore at least be resistant to the water.

Also, it is an object of the present invention to overcome the disadvantages encountered with the devices of the prior art by proposing a method of detecting a fall of a user limiting the detection errors and the power consumption of the measuring means and of calculation used for the realization of a portable detection device permanently, discrete, compact, inexpensive and waterproof.

For this purpose, the subject of the invention is a method for detecting a fall of a user, said user wearing a detection device comprising a housing enclosing supply means, means for transmitting an alert signal, an accelerometer, and means for processing the measurements provided by the accelerometer comprising calculation means and storage means, said method being characterized in that it consists in classifying the behavior of the user on the basis of indicators of its activity resulting from acceleration measurements acquired by the accelerometer and to postpone the triggering of an alert signal to avoid false alarms.

• figure 1 : une représentation schématique sous forme de diagramme d'un dispositif de détection pour la mise en oeuvre d'un procédé de détection selon l'invention,
• figure 2 : une représentation des différents solides pouvant être utilisés pour l'étape d'extraction de la gravité du procédé de détection selon l'invention,
• figure 3 : une représentation du treillis d'une chaîne de Markov pouvant être utilisée pour l'étape d'extraction de la gravité du procédé de détection selon l'invention,
• figure 4 : une représentation schématique de l'étape d'extraction de la gravité du procédé de détection selon l'invention,
• figure 5 : une représentation d'un automate de Markov à états cachés utilisé pour l'étape de classification comportementale du procédé de détection selon l'invention.

The present invention is now described in detail according to a preferred embodiment, particularly of the method and the associated device, not this description being made with reference to the appended drawings in which the different figures represent:

• figure 1 : a diagrammatic representation in the form of a diagram of a detection device for the implementation of a detection method according to the invention,
• figure 2 a representation of the different solids that can be used for the gravity extraction step of the detection method according to the invention,
• figure 3 a representation of the lattice of a Markov chain that can be used for the gravity extraction step of the detection method according to the invention,
• figure 4 : a schematic representation of the gravity extraction step of the detection method according to the invention,
• figure 5 : a representation of a hidden state machine Markov used for the behavioral classification step of the detection method according to the invention.

The method of detecting a fall of a user according to the invention is carried out by means of a detection device carried by a user.

Per user, the invention includes a person, such as an elderly person, a disabled person, a patient of a medical center, or an isolated worker, said person being likely to fall, to feel unwell, and to fall unconscious because of his difficulties, his disability or the conditions of his activity.

Such a detection device 10 used for the implementation of the method according to the invention is illustrated in FIG. figure 1 .

Said detection device 10 comprises a closed housing 12, sealed, and connected to fixing means 14.

In a preferred embodiment, the housing 12 is composed of two half-parts of polymer material obtained by injection / molding, the sealing of said housing being achieved through an assembly by ultrasonic welding of the two half-parts.

The fastening means 14 support the housing and attach the assembly thus formed to a body portion of said user.

Preferably, the housing 12 and its attachment means 14 will take the form of a bracelet, a pendant or a device attached to the belt.

Said housing 12 encloses power supply means 16, means for transmitting an alert signal 18, an accelerometer 20, and processing means 22 for measurements provided by the accelerometer.

Advantageously, the power supply means 16 take the form of at least one commercial battery, such as a so-called "button" battery of a few hundred milliampere hours to give an order of ideas.

The transmission means 18 are capable of sending an electromagnetic, radiofrequency or other waveform warning signal to a remote assistance center or to any other designated person, directly or via a base connected to a network of telephones. telecommunication such as Internet, GSM or other.

Preferably, the accelerometer 20 is a three-axis accelerometer, and the processing means 22 comprise calculation means 24 and storage means 26 in the form of a random access memory and a mass memory.

Of course, the detection device 10 comprises means of voluntary call 28, such as a push button, allowing the user to trigger itself an alert signal, the automatic detection of fall being placed as a complement to the voluntary call of the user.

In order to increase the autonomy and to minimize the bulk and the weight of the detection device 10, the method according to the invention aims at to perform a reliable treatment, and economic in terms of electrical energy consumption, measurements made by the accelerometer to deduce the supposed behavior of the wearer.

For this purpose, said method consists in classifying the behavior of the user on the basis of indicators of his activity resulting from the acceleration measurements acquired by the accelerometer 20 and deferring the triggering of an alert signal so as to avoid false alerts.

In a preferred embodiment of the detection method according to the invention, only acceleration measurements are used to deduce the activity of the carrier.

Of course, the invention also covers a drop detection method that would use measurements of another physical quantity in parallel with the steps of the method as will now be described.

Thus, the method according to the invention focuses solely on the perception of accelerations.

The acceleration measurements made by the accelerometer 20 of the detection device 10 comprise several components.

First of all, the user and therefore the accelerometer 20 attached to a part of his body are subjected to a constant acceleration to the ground due to gravity.

To the perception of gravity will be added the component due to the activity of the wearer.

This user-specific acceleration component is due to accelerations experienced or initiated by the part of the body at which the accelerometer is fixed as well as the whole body.

Thus, in the case of an embodiment in the form of a bracelet, the perceived accelerations are due to movements of the forearm, the arm and finally the whole body, thus revealing the activity, more or less important and disordered, of the user.

In the case of a fall, the accelerations are close to a relative weightlessness that will be perceived until the shock with the ground, then, if the user is unconscious following the fall, inactivity will be perceived.

It is thus clear that the acceleration measurements and the observation of their variations will make it possible to deduce, in addition to the bottom direction, the level of activity and the coherence of this activity of the user.

Also, in the present method, the previous acceleration measurements make it possible to establish variables relating to the gravity and the nature of the activity of the wearer, defining a state of perception. Each new acceleration measure modifies said variables and changes the perception state that is used to trigger an alert or not.

• normalisation des mesures d'accélération en données d'accélération,
• extraction de la composante gravité desdites données,
• transformation et agrégation desdites données en indicateurs, tant qualitatifs que quantitatifs, de l'activité de l'utilisateur,
• classification du comportement de l'utilisateur à l'aide desdits indicateurs,
• déclenchement d'un signal d'alerte en cas d'observation d'un comportement supposé anormal de l'utilisateur.

Also, in a preferred embodiment, the fall detection method provides for the processing means 22 to carry out a processing chain of the measurements acquired by the accelerometer 20 composed of the following steps:

• standardization of acceleration measurements in acceleration data,
• extracting the severity component of said data,
• transformation and aggregation of said data into indicators, both qualitative and quantitative, of the user's activity,
• classification of the user's behavior using said indicators,
• triggering an alert signal in case of observation of a supposedly abnormal behavior of the user.

The measurements acquired by the accelerometer take the form of a vector w in three dimensions, with w = g + v , where g corresponds to the gravity vector and v to the own acceleration, that is to say the acceleration not attributable to gravity and therefore corresponds to a movement or movement of the user.

In order to deduce indicators of the user's activity, it is therefore necessary to distinguish the vector v from the vector g .

Since this is a unique equation with two unknowns, discrimination with certainty of the gravity vector g is impossible.

• critère de constance du vecteur g : la norme du vecteur g est de 1, de préférence l'unité choisie étant g égal à 9,81 m.s^-2,
• critère de proximité de la direction de la gravité : le vecteur g à l'instant t+1 est très proche du vecteur g à l'instant t, en supposant que l'intervalle de temps entre t et t+1 soit faible,
• critère de marginalité de l'accélération propre v, l'accélération propre v étant souvent transitoire et proche du vecteur nul en moyenne.

Also, in order to extract the vector g with a good probability, it is planned to use the following three criteria simultaneously:

• constancy criterion vector g: the norm of the vector g is 1, preferably the selected unit being g equal to 9.81 ms ^-2,
• proximity criterion of the direction of gravity: the vector g at time t + 1 is very close to vector g at time t, assuming that the time interval between t and t + 1 is small,
• criterion of marginality of the eigenvacculation v , the eigenvacculation v being often transient and close to the null vector on average.

Subsequently, the solution adopted by the invention to arrive at an estimate of the gravity vector with a good confidence rate is to use a sphere of unit radius discretized in several vertices to model all the possible directions of the unit vector g , and a stochastic calculation to determine the probability carried by each vertex that the vector g is exactly in the direction indicated by this vertex.

Different meshes of the sphere are thus possible, such as a geodesic mesh, but for reasons of efficiency, a uniform mesh using the Platonic solids is chosen preferably.

So, and as illustrated by the figure 2 , it can be chosen a tetrahedron, an octahedron, a hexahedron (cube), an icosahedron or a dodecahedron, having respectively 4, 6, 8, 12 and 20 vertices, the dodecahedron being the solid giving the best results because of the density of its mesh.

The general principle of the invention for determining the gravity vector g is therefore to determine for each vertex the probability that said gravity vector is exactly in the direction indicated by said vertex.

Once these probabilities are determined, and in the case of the uniform mesh, It suffices to carry out the vector sum of each of the directions represented by the vertices of the solid, weighted by the probability carried by this vertex, then we obtain a vector indicating the probable direction of Earth's gravity and whose norm gives us an indication of the degree of trust that can be attached to this direction.

In summary, the determination of the probabilities carried by a set of certain directions of the vector g makes it possible to deduce the probable direction of this vector as well as the degree of confidence in this direction.

The vector thus obtained being of the form: α. g , it must be normalized in order to obtain the probable vector g, the degree of confidence then being a function of α.

The solution adopted to determine the probability carried by each vertex of the selected solid is in the field of stochastic and Bayesian inference, in particular Markov chains and more precisely H.M.M. (Hidden Markov Model) or hidden Markov chains.

More precisely, the set of vertices of the Platonic solid used represents the states of a Markov chain used within a hidden Markov model and whose coefficients of the transition matrix Tr are function of the distance (linear or angular) between two vertices, and the emitting function Em is a function of the own acceleration vector v , deduced by subtracting the vector g designated by each vertex to the observed accelerometric vector w. Thus, for each vertex n, we have v _n = w - g _n .

By way of example, if the selected solid is a dodecahedron, which thus defines a twenty-state Markov chain representing the possible directions of the gravity vector, the transition matrix Tr can be simplified by taking into account only the direct neighbors. of each vertex, thus defining the elemental lattice of the Markov chain where each vertex, as illustrated in figure 3 , in addition to being connected to itself, is connected to each of its neighbors. We can also simplify the emission function Em by defining the probability of emission as a function of a supposedly Gaussian distribution of the norm of the vector v .

Thus, the probabilities carried by a set of certain directions of the gravity vector allow us to deduce the most probable direction for this vector, while determining its degree of confidence, the gravity detector behaves substantially like a two-axis probabilistic gyroscope.

The implementation of this Markov chain for the extraction of the gravity component from the accelerometric measurements, ie the implementation of an algorithm in the calculation means 24 of the detection device 10, will now be described in one embodiment. preferred embodiment of the method according to the invention and with regard to the figure 4 .

First, the raw acceleration measurements provided by the accelerometer 20 undergo a normalization step 40.

Thus, said normalization 40 consists of applying offsets and scaling coefficients to the raw measurements in order to obtain a representation of said measurements in the form of a standardized three-dimensional vector.

A simple criterion used to determine the parameters of this normalization is that, in the absence of proper acceleration or when it is very close to zero, and whatever the position of the detection device, the norm of the vector measured w , which is then substantially equal to g , must be as close as possible to one.

Subsequently, in order to obtain a good compromise between the useful capacities of the calculation means, ie the necessary computing power, and the accuracy of the result of the extraction of gravity, the Platonic solid preferably used is a hexahedron, a cube.

The chosen solid being the cube, we have a vector P of eight coefficients, representing the set of probabilities carried by each of its vertices. The sum of the probabilities of this vector to be of one, it is consequently initialized with the coefficient 1/8.

The application 42 of the transition matrix introduces a probabilistic blur on the knowledge of the direction of the vector g . The transition matrix Tr has been implemented by taking into account only the direct neighbors of each vertex of the selected solid, it is therefore defined by a single rotation coefficient θ. The coefficient of the transition matrix Tr from a vertex to itself is then 1- θ, and that of transition from one vertex to one of its neighbors is θ / 3, the others being considered null (or very close to zero). This rotation coefficient θ then represents the probability that, from one instant to another, the direction of gravity has moved from a vertex to a neighboring vertex.

An intelligent numbering of the vertices of the solid makes it possible to quickly know all the neighboring vertices of each one. By using the exclusive or, the implementation of the product of the vector v by the matrix Tr can then be reduced to: for any i varying from 0 to 7, v '[i] = (1- θ) * (v [i] ) + (θ / 3) * (v [i ^ 1] + v [i ^ 2] + v [i ^ 4]).

The application 44 of the emission function Em provides the probability that an observation w will occur assuming that the direction of the vector g is exactly on each of the vertices n of the solid. The corresponding acceleration vector v _n corresponding can then easily be computed by a simple subtraction. Since the natural accelerations are considered as marginal, it is assumed that the probability of occurrence of this vector follows a Gaussian distribution according to its norm. This function is implemented as a table of discretized values and the probabilities are then obtained by linear interpolation.

Now that the probabilities of the vertices of the solid are updated over the course of the new measurements, the evaluation function 46 can determine the most probable direction for the gravity vector g , performing the weighted vector sum. The vector thus obtained, of the form α. g , denotes the most likely direction for g , and its norm α is proportional to the concentration of probabilities around that direction.

It is therefore necessary to normalize this vector, the role of the unitarization function 48, in order to obtain the vector g , and the confidence function 50 provides the confidence rate in this direction, which is a function of the norm α.

The Markov chain making it possible to obtain an estimate of the vector g at the output of the unitarization function 48 introduces a form of inertia to the change, it is therefore necessary to delay the observed acceleration vector w with the help of the delay function 52 to obtain, by the subtraction function 54, the own acceleration vector v delayed.

After the gravity extraction step, the following information is available: an estimate of the terrestrial gravity vector g , an indicator of the confidence rate of this vector, a function of α, and therefore, by subtraction, an estimate of the vector d clean acceleration v .

The detection method according to the present invention then provides for the transformation and aggregation of said data into indicators, both qualitative and quantitative, of the user's activity.

• la confusion : cet indicateur est calculé comme étant inversement proportionnel au minimum des taux de confiance dans la direction du vecteur g. Par exemple, le taux de confusion peut-être très élevé dans le cas de mouvements erratiques persistants mais aussi et surtout lors d'une chute,
• la dispersion : cet indicateur, similaire à un écart type, est calculé comme la moyenne de la norme des différences des vecteurs d'accélération propres v consécutifs, il donne une bonne indication de la cohérence de l'activité de la personne. Par exemple, plus la dispersion est grande, plus le comportement est erratique, et plus le choc est brutal, plus cet indicateur est élevé.
• l'activité : cet indicateur est calculé comme la moyenne de la norme des vecteurs d'accélération propres v, il permet de savoir si le porteur est conscient et de connaître le niveau général de son activité.

Thus, it is planned to use at least the following three indicators:

• confusion: this indicator is calculated as being inversely proportional to the minimum of the confidence rates in the direction of the vector g . For example, the confusion rate can be very high in the case of persistent erratic movements but also and especially during a fall,
• dispersion: this indicator, similar to a standard deviation, is calculated as the average of the norm of the differences of the eigenvectors of acceleration v consecutive, it gives a good indication of the coherence of the activity of the person. For example, the greater the dispersion, the more erratic the behavior, and the more the shock is brutal, the higher this indicator is.
• activity: this indicator is calculated as the average of the standard of v own acceleration vectors , it makes it possible to know if the wearer is conscious and to know the general level of his activity.

These indicators, resulting from an aggregation over a given time interval, are then thresholded and composed so as to deduce various observation parameters relating to the activity of the user.

The detection method according to the invention then provides a classification step of the behavior of the user made from these observation parameters.

The behavior classifier is implemented in the form of a hidden state machine Markov (H.M.M.), whose hidden part represents the state of the carrier and the observable part is provided by the previously calculated indicators.

A simplified version of the possible states, or equivalence classes, of the user's behavior is represented on the figure 5 a rest state 56, an active state 58, a falling state 60 and a state of unconsciousness 62.

Each of these states (56,58,60,62) carries the probability that the user is in this state, depending on the previous state and new observations.

Thus, in a normal behavior cycle, probabilities oscillate between active state 58 and rest 56.

During a suspicious transient event likable to a fall, a part of the probability of the active state 58 will pass to the falling state 60, a fall being the consequence of an unfortunate activity, and the probability carried by the Falling state 60 is transferred in its quasi-totality to the state of unconsciousness 62.

More precisely, even in the absence of a suspicious activity, part of the probability of the active state 58 passes to the falling state 60, but in the absence of suspicious activity, the confrontation with the parameters of observation discredits this assumption and reduces the probability of falling state 60.

In summary, once all the states are fixed, all that remains is to determine the transition matrix Tr and the em emission matrix to completely define the hidden state machine Markov and that to learn the parameters of said automaton using a sufficiently representative corpus, in particular from Viterbi or Baum-Welch training, simulated annealing or genetic algorithm.

The probabilities of being in one of the four states (56,58,60,62) of the automaton evolve over the observations, and it is enough to examine the probability of the state of unconsciousness 62 to decide the sending an alert using the transmission means 18 as soon as it passes a predetermined threshold corresponding to a supposed abnormal behavior of the user.

Concretely, the fall detection method that has just been described can be implemented in processing means 22 taking for example the form of a microcontroller.

To give an order of ideas, the storage means 26 require only a useful random access memory of about 200 bytes and a useful mass memory of about 6 kilobytes. Also, with a 16-bit microcontroller, the computing times achieved by the processing means 22 accelerometer data 4 megahertz range between 0.3 milliseconds and 1.4 milliseconds with an average of about 0.6 milliseconds.

In addition, in order to limit the consumption of electrical energy, the accelerometric acquisitions are performed at 25 Hertz in normal mode and can go down to 1 Hertz in the case of prolonged rest or during the user's sleep.

Also, it can be seen that, by making it possible to reliably simplify the calculations of a fall detection made from the measurements of an accelerometer, the method according to the invention makes it possible to limit the useful capacities of the processing means, hence a significant decrease in their consumption of electrical energy and a significant extension of the autonomy of the detection device vis-à-vis those obtained with the detection methods of the prior art.

More precisely, the implementation of the detection method as just described in said processing means 22 makes it possible to obtain a detection device 10 whose autonomy is at least greater than one year for means power supply 16 providing at most five hundred milliamperes-hours.

In an alternative embodiment, especially in the case where the detection device is placed at the level of the trunk of the user, a two-axis accelerometer may be sufficient to perform a reliable and economical fall detection using the method according to the invention. invention, the extraction of gravity using a circle discretized polygon and thus simplifying the elementary lattice of the Markov chain. This variant is obviously covered by the invention.

In addition, the invention also covers the applications of the detection method which has just been described to movements other than a fall, as well as to persons or objects in situations different from those set out above.

Claims (10)

A method of detecting a fall of a user, said user carrying a detection device (10) comprising a housing (12) enclosing supply means (16), means for transmitting an alert signal (18) , an accelerometer (20), and processing means (22) for the measurements provided by the accelerometer comprising calculation means (24) and storage means (26), said method providing for the realization by the processing means ( 22) of a processing chain of the measurements acquired by the accelerometer (20) composed of the following steps: - standardization of acceleration measurements in acceleration data, extraction of the gravity component of said data, transformation and aggregation of said data into indicators, both qualitative and quantitative, of the user's activity, - classification of the behavior of the user using said indicators, - triggering of an alert signal in the event of observation of a supposedly abnormal behavior of the user, said method being characterized in that the extraction of the gravity component consists in estimating the gravity vector g by using a sphere of unit radius discretized in several vertices to model all the possible directions of the unit vector g, and a stochastic calculation for determine the probability carried by each vertex that the vector g is in the direction indicated by this vertex. A method of detecting a fall of a user according to claim 1, characterized in that the sphere of unit radius is discretized using a uniform mesh using the platonic solids. A method of detecting a fall of a user according to claim 2, characterized in that the platonic solid used is a cube. Method for detecting a fall of a user according to one of claims 1 to 3, characterized in that the set of vertices of the discretized sphere, or of the platonic solid used, represents the states of a Markov chain used in the within a hidden Markov model, and in that the coefficients of the transition matrix Tr are a function of the distance between two vertices and the emission function Em is a function of the own acceleration vector v deduced by subtracting the vector g denoted by each vertex with the accelerometric vector observed w. A method of detecting a fall of a user according to claim 4, characterized in that the use of the Markov chain provides an estimate of the gravity vector g , an indicator of the confidence rate in the direction of this vector, and an estimate the own acceleration vector v. A method of detecting a fall of a user according to one of claims 1 to 5, characterized in that the normalization step consists in applying to the raw acceleration measurements offsets and scaling coefficients in order to obtain a representation of said measurements in the form of a normalized vector in three dimensions, and in that a criterion used to determine the parameters of this normalization consists in considering that in the absence of its own acceleration or when it is very close to zero, and whatever the position of the detection device (10), the norm of the measured vector w, which is then substantially equal to g, must be as close as possible to one. Method for detecting a fall of a user according to one of claims 5 or 6, characterized in that the step of transforming and aggregating the acceleration data over a given time interval results at least in the three indicators following: confusion: this indicator being calculated as being inversely proportional to the minimum of the confidence rates in the direction of the vector g. - the dispersion: this indicator being calculated as the average of the norm of the differences of the consecutive vectors of acceleration v consecutive, - activity: this indicator is calculated as the average of the norm of the eigenvectors of acceleration v . A method of detecting a fall of a user according to claim 7, characterized in that the step of classifying the behavior of the user is implemented in the processing means (22) in the form of a state Markov automaton hidden, the hidden part of which represents the state of the carrier and the observable part is provided by said indicators, and in that a simplified version of the possible states, or equivalence classes, of the behavior of the user are: a state of rest (56), an active state (58), a fall state (60) and a state of unconsciousness (62), each bearing the probability that the user is in that state. A method of detecting a fall of a user according to claim 8, characterized in that it is intended to trigger the sending of an alert using the transmission means (18) as soon as the probability of the state of unconsciousness (62) passes a predetermined threshold corresponding to a supposedly abnormal behavior of the user. Device for detecting (10) a fall of a user comprising a housing (12) enclosing supply means (16), means for transmitting an alert signal (18), an accelerometer (20) , and means for processing (22) the measurements provided by the accelerometer comprising calculation means (24) and storage means (26), the method of detecting a fall according to one of the preceding claims being implemented in said processing means (22).

Images