EP1979854A1

EP1979854A1 - Method for identifying a person's posture

Info

Publication number: EP1979854A1
Application number: EP07704329A
Authority: EP
Inventors: Stéphane BONNET; Christelle Godin
Original assignee: Commissariat a lEnergie Atomique CEA
Current assignee: Commissariat a lEnergie Atomique et aux Energies Alternatives CEA
Priority date: 2006-02-02
Filing date: 2007-02-02
Publication date: 2008-10-15
Also published as: FR2896896A1; US20090012921A1; FR2896896B1; US9026477B2; WO2007088205A1

Abstract

The invention relates to a classification method including a first step (6) for the classification of any type of event using first rules and a subsequent second step (8) for the classification of events that were not identified by the first classification, using a training corpus (9) consolidated with all of the events identified by the first classification. The method is adaptive if the second classification rules are amended according to the new examples that may have been determined by the first rules.

Description

METHOD FOR IDENTIFYING POSTURES OF A PERSON

DESCRIPTION

The subject of this invention is the remote identification of postures taken by a person by using the signals given continuously from a sensor it carries. It involves the classification of the states taken successively by the sensor and the person. A notable difficulty is to exploit abundant but not very descriptive signals since the postures are complex, subject to variations and transient states and they do not lend themselves to direct measurements. In the art of classification, each event to be classified is defined a number of characteristics called parameters that are used for classification and may come from sensors or other information. Events must be divided into categories called classes based on the values that the parameters take. Some classification methods are performed automatically, by a computer having logical or digital tools of comparison and decision. There are two main processes. In the first, the classification uses explicit rules that make membership in a class depend on the value of a particular parameter or group of parameters. For example, an event is assigned to a class if a parameter reaches a value greater than a threshold, or if the sum of one group of parameters is greater than a threshold, whose value is defined by those skilled in the art. Such methods make it possible to identify certain postures as will be seen, but their effectiveness is not sufficient and many portions of the sensor signals can not be identified.

The other group of known methods uses what is called a learning base comprising a certain number of events. The computer must then define the classification rules, which remain implicit to the operator according to the values of the parameters. In a first variant called unsupervised learning, the computer itself distributes the events of the base into classes according to the similarities or distances between the events, and the definition of the classes remains unknown. In another variant, called supervised learning, the operator indicates the class of each event of the learning base, and the task of the computer is to define a numerical or logical function which respects the classification decided by the operator according to the event parameters. It is also possible to combine the two variants to obtain a third so-called semi-supervised learning where the membership of a class is defined by the operator only for some of the events of the learning base (in practice, in small number compared to others).

The disadvantage common to all these methods is the difficulty of establishing good rules, able to minimize the proportion of badly classified events. With explicit rules defined by the operator, many events remain indeterminate in practice because it is difficult to establish precise and detailed rules by observation or intuition; in learning processes, automatically defined rules will be imprecise if the examples in the learning base are poorly representative of future events, and sometimes the lack of operator control over these implicit rules will not allow it to be corrected. . In addition, the manual constitution of a learning base takes time and may involve errors. Finally, such a table will not be scalable.

The progress made by the invention, in its most general form, is to improve the classification results in the application mentioned.

In its most general form, the invention relates to a method of identifying postures of a person, comprising the following steps: - fixing a sensor on the person; recording at least one sensor signal during a time window; extracting partitioning parameters from said signal; dividing, from the partitioning parameters, the time window into segments satisfying a criterion of homogeneity of posture or change of posture, and extracting, for each of the segments, a first set of parameters from said signal ; classifying each of the segments in one of the event classes associated with said events or in a class that is not associated with an event, by applying first decision rules from the first parameter set associated with the segments;

- comprising a calibration operation for creating a learning base, comprising the following steps for each segment belonging to an event class:

extracting a second set of parameters from said signal; adding to said learning base data item listing the second set of parameters and the class of events associated with the segment in question,

and using the learning base to classify segments of said record or other subsequent record that have not been associated with an event by applying second decision rules using the learning base from a second set of parameters associated with these segments.

Some processes use multiple sensors, placed in as many different locations of the body, to detect unambiguous postures; but the permanent wearing of all these sensors is unpleasant.

Other methods use a single sensor and typically look for transient states indicative of a change of posture on the signals; but ambiguities remain, for example between a sitting position and a standing position.

No method exists in the intended application, according to which two fundamentally different classification steps are used, since using generally different parameters, to classify signal segments characteristic of different postures. The rational separation of the signal into homogeneous segments, which can be done automatically, is itself not solved. As a result, a much larger number of segments remains undetermined with prior methods. The advantage of the invention comes from the fact that certain parts of the signals can be analyzed in a simple way to extract certain characteristics directly contributing to their identification (the first group of parameters), then exploited to extract other characteristics (the second group of parameters) to help identify other parts of the signals. The second decision rules, which call on the experience acquired, and all the more so that they can be made evolutive by welcoming new results, are therefore of particular interest here. Among some particularly important embodiments, there will be mentioned the distribution of the segments into two categories, one being stable at the postures and the others at transitional states between two stable postures; and a rational division of the signal into segments, as has been mentioned. The invention will now be described in more detail by means of the figures: FIG. 1 is a general view of the process, FIG. 2 is a view of a neural network,

FIG. 3 is a schematic view of an application of the method,

FIG. 4 is a diagram detailing an example of a signal sample of this application, FIG. 5 illustrates another signal sampling,

and FIG. 6 is a detailed view of a particular method.

FIG. 1. The invention implements an operator in the first steps of the method and then a computer. In a step of beginning 1, the operator distinguishes two subsets of parameters of the events in step 2, and elaborates first decision rules in step 3 by means of the subset of the first parameters. It finally creates a learning base that can be either empty or partially filled. The computer then receives the information of a new event as a signal to be processed in step 4; this signal includes the parameters of the event, obtained by measurements or other means. The computer extracts the first parameter values of this signal in step 5. It then tries to classify the event among the different classes using the first rules. defined in Step 3, but not necessarily in the presence of a questionable event (step 6). In this case, the computer extracts from the signal the values of the second parameters in step 7 and resumes the classification attempt in the following step 8 by applying second decision rules in step 10. A classification status is finally given to step 11.

If, however, the first step of classification of step 6 has been successful, the method is carried out differently and the extraction of the second parameters, carried out at a step 12, then serves to fill the learning base by adding these parameters and the result of the first classification, completing or amending the second decision rules used in step 10 according to the classification result of the present event and the values of its second parameters. There is a presumption that future events will be classified with increasing safety.

Concrete means of developing the second decision rules include various known genres. An illustration of the use of a neural network in FIG. 2 will be described here. The network comprises a layer of input neurons P1, P2,..., Pn for receiving the values of the second parameters of the events. in digital form.

The classification decision in a given class is given, also in digital form, by the content of an output neuron D. The neural network comprises at least one hidden layer of neurons y-, which perform numerical combinations of the parameters and transmit them to the next layer of the network, here immediately to the output neuron D. Numerical weights w are applied to the transmitted values before they reach the neurons of the hidden layer or the output neuron. Here, the neurons of the hidden layer perform a hyperbolic tangent activation and the output neuron performs a linear activation. The decision function can then be written according to the formula (1)

where sgn is the sign function, tanh the hyperbolic tangent function, w-, the weight connecting the output of the neuron y-, to the output neuron D, w ₀ and w _0: particular weights called biases or thresholds connecting a fictitious neuron of output of value equal to 1 to the output neuron (for w ₀ ) or to the neurons y-, of the hidden layer (for w ₀ :) and w _{: i} the weights connecting the input neurons P ₁ to the neurons of the layer hidden

Neural networks allow, subject to choosing the number of neurons in the hidden layer or hidden layers, to approach any function once the correct weights have been found. The elaboration of the second decision rules therefore amounts to adjusting the weights w so that, for each or almost every example of the learning base 9, the result of the decision by the neural network and the result known, of the classification. When step 12 is implemented, the computer adjusts the weights so that the largest number of events in the learning base - enriched with the new event - is properly evaluated by the second decision rules.

Other types of numerical classifiers are known in the art: linear separators, decision trees, SVM (vector support machine or wide margin separator); the invention also applies to them.

According to the invention, there is concern to identify the posture of a person at any time by remote monitoring, especially for medical purposes. The person is equipped with a motion sensor of a known kind and may include a triaxial accelerometer and a triaxial magnetometer to detect the orientation and movement of the wearer according to the directions of the Earth's magnetic field and gravity. Referring to FIG. 3, where the carrier has the reference 15 and the sensor the reference 16, it is seen that the sensor 16 is attached to the torso of the person 15 and that the main measurement axes are the anterior axis. posterior AP directed towards the front, the medio-lateral axis ML directed to the right and the vertical axis VT directed upwards; if necessary, a preliminary calibration procedure is performed on the person 15 to align the measurement axes of the sensor and the anatomical axes of the person 15, by determining a passage matrix that connects them. This matrix is supposed to be invariable, which is true as long as the sensor 16 is not moved on the person 15. Moreover, information fusion algorithms make it possible to give the orientation of the sensor 16 in space from its measurements by expressing them by rotations of Euler angles or quaternions. In the example detailed below, the angle of inclination of the person in the sagittal plane (vertical and anteroposterior) using measurements of the accelerometer and at the yaw angle to determine the azimuth (angle of the anteroposterior axis with a fixed reference, here magnetic North) with the magnetometer.

The invention is implemented to distinguish between certain kinds of activity of the person 15 and some of his postures. In the diagram of Figure 4, expressed as energy of activity as a function of time, there is successively a standing state, a standing / sitting transition, a sitting state, a sitting / standing transition and a standing state in states 19 to 23 Transition states always correspond to strong activities and energy peaks, and stable position states are often longer and less active, but exceptions exist, such as running, during which the wearer in a stable standing position will have great energy. In addition, the energy can vary greatly during the same state. A race period 24 during the first standing state 19 is thus shown. It is therefore not possible to directly deduce the succession of states from a diagram such as that of FIG. 4, which may, however, be used for a rational division into segments of unequal duration but of homogeneous characteristics. In the complete example of the method of classification which is given below with the support of FIG. 6, the classes are nine, ten or eleven: three stable states (sitting, standing and lying), and six states. transfer from one to another of the previous steady states (standing towards sitting, sitting towards standing, standing towards bed, lying towards standing, sitting down to bed and lying down towards sitting); we can add a new class "unknown", in which we pour the events that could not be in one of the previous classes, or a class "ARTEFACT" which designates the movements limited by two similar postures and do not correspond not to a defined event. Various stages are undertaken. In a first step (steps E1 to E5), the recording is divided into portions or segments according to a set of several criteria. This leads to an alternation of homogeneous zones in the sense of postures (called "O-segments") and active zones ("1- segments") which potentially represent postural transfers. Each temporal segment designates an event whose class is to be determined. At the end of this first phase, all time segments have the UNKNOWN class.

Secondly (steps E6 to E10), a first classification is made according to the first rules explicitly notified by seeking to assign - and in the safest possible way - to certain events (associated with respective time segments) a class of posture. . Due to the positioning of the movement sensor to the torso, it is easier to distinguish the horizontal and vertical postures occupied by the person. We therefore seek above all to distinguish horizontal postures (LAYER) from vertical postures

(SITTING or STANDING) according to a set of several criteria. A delicate point remains the case of postures strongly bent forward.

Finally, it is determined on the basis of other criteria the membership of certain segments of the signal with a small variation of activity (O-segments, to which a stable posture can be associated) of sufficient length of time to the class ASSISTS or STANDING, and in fact, it is also determined on the basis of a temporal coherence the membership of certain segments of the signal with a strong variation of activity (1- segments, to which we can associate a transfer from one stable posture to another, or a movement in the same posture), to transitions between postures and intra-postures (transfers).

At the end of this second phase, some time segments have a well-identified class and others still have the UNKNOWN class. The former are put into a learning base by the system and are used to define second decision rules using their classification results and some of their characteristics.

Third (step E12), the learning base is implemented in an attempt to best classify the events still remaining in the unknown class. The manner in which the first part of the process is conducted will be described below.

After obtaining the signals from the sensor 16 (step E1 of FIG. 6), the analysis is first made on short sampling data, of the order of one second (E2); we try to determine areas of homogeneous posture without trying to determine this posture.

The criteria for dividing the signals into unequal length segments (E3) are two and based on the activity index IA provided by the signals 3a (of the triaxial accelerometer) and on the variations of inclination of the torso of the person 15 in the sagittal plane. The corresponding parameters are calculated as follows. For the activity criterion, the signal standard 3a of the accelerometer is representative of the energy of the movement in progress and gives objective information on the physical activity of the wearer. The average value is calculated for each duration of one second of the signal of the quantity:

N _aœ HF [n] = 4aHF _ΛP [nf + aHF _ML [nf + aHF _vτ [nf to obtain the activity index

where n is an index, F _e a sampling frequency, F _AP , F _ML and F _V τ the measured values of the signal components, F the signal, at acceleration and H denotes a band-pass filtering to conserve the "High frequencies" (here 0.5 to 10 Hz), representative of human activity, in removing the DC component (acceleration of gravity) and noise.

The activity index IA based on the signal 3a is compared with a threshold - represented by the horizontal line 25 of FIG. 5, of value 0.04 - to determine whether a time range is associated with an active phase or a phase resting. The values of the index IA are positive and generally between 0 and 1. It is also possible to determine an index of activity of the yaw activity of the wearer. To reduce the correlation of the 3m signal (of the triaxial magnetometer) with the signal 3a, it is proposed to calculate an index based solely on the temporal variations of the ML (mediolateral) component of the magnetometer:

IM [n] = h _ML [n + \] - h _ML [n]

With the 1-second average value of the ML component of the magnetometer:

For the second criterion, the low frequency signal (at 0.5 Hz) is used.

1 'accelerometer to determine the pitch of the person that is to say its inclination in the sagittal plane.

The inclination in the sagittal plane (pitch) is trivially given by:

Pitch [n] = arcsin (aLF _AP ), where, similarly, a designates the acceleration, L a low-pass filtering of the signal, and F the signal. A negative pitch is a person leaning forward; Positive pitch, to a person leaning backward.

Thresholds of variation of these amounts of activity and inclination of the torso are used to detect posture transfers. Consecutive sampling portions of non-varying signals reaching these thresholds are grouped together. Segments of unequal duration are thus obtained from uniform sampling portions, and more precisely with an alternation of time segments of value 0, (O-segments) corresponding to zones homogeneous in the sense of the posture (still unknown) and short-term time segments of value 1 (1-segments), corresponding to areas with a large movement that may be a postural transfer. The distinction between these two categories of segments is made (E4). The underlying idea is that a postural transfer corresponds to a certain level of activity and a certain variation of inclination of the torso.

The thresholds used during this step are relatively critical in the sense that the distribution of the data obtained will condition the sensitivity of the classification of postures. In unfavorable cases, a small number of segments with the same 0-segment can be obtained which can contain several different postures or on the contrary a high number of segments and unspecific 1-segments of a postural transition. The two criteria were finally applied as follows: a probable posture transfer corresponds to a minimum variation of 20 degrees for 3 seconds simultaneous to a minimum variation of 0.08 (standardized units) for 3 seconds. Any sampling portion meeting these two criteria will therefore belong to one segment, the others to one segment.

These thresholds are low enough to make sure you do not miss a postural transfer. In return, we identify many 1-segments that are not true postural transfers (artifacts).

In a second pass (E5), some 1- segments can be disabled to increase specificity and thus reduce over-segmentation. After various tests, it was decided to invalidate a 1-segment if the average level of activity remains high on both sides of this 1-segment. Indeed, it will be assumed later that a SAT or LATE posture is characterized by a low activity, where IA and IM are less than 0.04 and a strong activity is relative to a STANDING position. The average of the activity index IA is thus calculated on each of the 0-segments. If the activity index is above the threshold before and after the potential transfer, then the 1-segment is invalidated.

The following is an example of implementation of the decision tree that are the first decision rules (E6-E10). In all criteria following values, the values used are averages for the segment under consideration.

In the time window of signal analysis, we search (E6) the first 0-segment likely to represent a horizontal posture (E7).

The criterion chosen is accVT <thl) or accMlJ> th3), where ace is this time used to designate the measured acceleration. This allows to detect cases where the person is strongly bent in the sagittal plane (person lying on the stomach / back) or in the lateral plane (person lying on the side).

We will then detect the next 0-segment likely to represent a vertical posture (E8). The criterion chosen is (accVT> th4). All O-segments and 1-segments within the range are assigned to the LAYER class.

The thresholds used for the accelerometers and the rankings obtained when these inequalities are observed are given in Table 1.

Table 1

A difficulty here is not to confuse a horizontal posture type LAYER with the case where the person leans strongly forward. A temporal criterion (duration of the horizontal posture bent forward) is implemented in this case to detect these situations, and the posture of the LAYER type is assumed if a threshold of 300 seconds is reached.

At the end of these first steps, the classification HORIZONTAL (LAYER) and VERTICAL (DEBOUT / ASSIS) is supposed to be carried out.

We then focus on the classification

SITTING / STANDING (E9) and identification of transfers

(ElO) by developing a set of decision rules based on both accelerometer and magnetometer based criteria. We prefer to identify postures and transfers.

After various tests, the following decision criteria have been retained: For the STANDING class, a) The 0-segment must have a temporal duration greater than a given time (10 seconds for example). b) The phases of walking are identified by analysis of the signal 3a filtered at high frequencies. If the

0-segment contains a significant part of the walking phase, the whole segment the range is classified STANDING. The step is detected by the presence of a large energy peak at about 1 Hz. c) Similarly, if the activity level IA or IM is greater than a threshold (0.04), it is assigned to the STANDARD class.

The criteria a and b, or a and c, must both be verified.

For the ASSIS class, a) It is assumed that the 0-segment must have a time duration greater than a given time (30 seconds for example). b) If the 0-segment corresponds to a posture inclined towards the back (of 20 degrees at least), one assigns the segment to the class ASSIS. c) Similarly, if the activity level IA or IM is below a threshold (0.02), the segment is assigned to the ASSIS class.

The criteria a and b, or a and c, must be respected.

For 1-segments:

Certain 1-segments are determined in fact in intra-postural movement (artifact) or inter-postural (transfer). This assignment is determined by the nature of the preceding and following postures, when they have been determined.

A 1-segment between two different postures is a transfer, a 1-segment between two similar postures is an artifact and is not classified.

The segments then classified are poured by the system into a learning base which collects their classification results and second parameters which, related to the results of the classification. classification, are used to develop second decision rules by the system (by adjusting the weights in the case of a neural network); these second decision rules generally remain implicit, or unknown to the operators, who do not program them, and they are different from the first rules since they call on different parameters. In addition, the first decision rules use the acceleration measurements in the example in question, whereas the second rules instead use magnetometry measurements. It should be noted that other sensors could also be useful for second decision rules, such as gyrometers to measure the speed of rotation of the person, including yaw.

The second rules are applied to classify the remaining events in UNKNOWN class (E12). The parameters on which the algorithm operates are conventionally calculated on the signals acquired by the sensor 16 and on the intermediate signals (tilt angle, yaw angle, activity index, etc.). The calculated parameters are, for example, for each segment:

• signal statistics, • duration,

• frequency content (distribution of energy peaks in the frequency domain),

• even the segment itself (signature).

At the end of this phase, a number of events (0-segment or 1-segment) are not yet classified and remain in UNKNOWN class. An idea of the invention is to use all or some of the already well-classified events to build a machine learning base to move to a supervised learning algorithm (IS).

The learning base is completed as a result of the results of the newly identified segments. The second decision rules (weight in the neural network) are advantageously updated. The computer holding the learning base therefore analyzes each segment, already identified or not, in a similar way according to the second group of parameters using the learning base. For each of the unknown segments, it calculates similarity scores with elements of each of the classes, or distances calculated according to usual mathematical norms in a space with multiple dimensions. The parameters considered can be, with regard to the statistics of the signals, average values, energies, variances; as far as the segment itself is concerned, similarities are sought with previous segments as a whole; this research is done essentially for the 1-segments likely to represent transient states based on the frequent and almost exact repetition of the same gestures to move from one posture to another. The signature of the segment is a component of the signal that reflects a function of evolution of the posture. Another interesting criterion would be the recognition of a significant rest phase with a yaw angle (average azimuth considered on a segment) similar to a reference azimuth

(corresponding to a fixed orientation seat in the residence of the wearer), therefore sitting class, again taking advantage of repetitions of attitudes related to the habits of the person, regardless of the signals themselves; these examples also show that a ranking parameter will often be the yaw angle or its temporal variation.

When a new segment has been poured into the learning base, the following calculations can be made also with reference to it.

We now give some possibilities of generalization. The rules of the first group may be chosen from among the following, in more complex embodiments of the method: recognition of a phase of activity corresponding to a step (positive and fairly regular acceleration in the AP axis), therefore of standing class;

- Recognition of a rest phase and, without transition, a phase of walking (only the acceleration in the axis AP changing), so classification of the rest phase in standing class; - recognition of a rest phase having a pitch angle or elevation, (around the axis ML from the VT axis) strongly negative corresponding to a seated person inclined backwards, so seated class ; - Recognition of a long rest phase with a yaw angle (rotation around the axis VT) varying slightly, therefore of seated class; recognition of an activity phase having a greatly varying yaw angle and a pitch angle or moderate variation of elevation, corresponding to a change of direction of the wearer, therefore of standing class; recognition of a phase of strong activity of little duration between two identical postures, therefore of identical class: one is in the presence of an interior movement in the same standing or sitting class; recognition of a phase of strong activity of little duration between two different postures, therefore corresponding to a transition, and ranking in standing / sitting or sitting / standing.

This information is given either by the accelerometers (for the movement), by the magnetometers (for the orientations), or by a combination of the two.

The parameters used for the second decision rules can be those used to apply the first classification rules (by combining them differently) or, more typically and more fruitfully, other parameters. The second classification rules are determined and can be amended to tend to match the evaluations they give of the identified states with the evaluations obtained with the first classification rules for the same phases. If the first decision rules are associated with the production of a confidence index of the ranking result that they give, the events that have been classified in the first step can be paid to the learning base only if the index has reached a value above a certain threshold: events classified in the first step but with insufficient safety will have no influence on second decision rules.

The method of the invention lends itself well to supervised learning of the second decision rules. All the decision rules are then defined by the operator: the first decision rules can be those whose application is the simplest, and the second decision rules those which require calculations, which interact with each other or which tolerate exceptions or are not known perfectly and may be amended later.

In other cases, it may be useful, on the contrary, to correct the first decision rules, when they give a result different from the second decision rules or when they contradict each other: the result given by the second decision rules will allow identify the first decision rules that will have given an opposite result; these rules may be amended (when they depend on a threshold for example), reduced by giving them less importance, or even deleted. The correction may be decided according to the number and the confidence index of the results, or the respective reliability of the rules.

The second decision rules can themselves be controlled by third decision rules specified by the operator and used to detect classification errors of the second decision rules. When errors are detected, the second decision rules can be corrected to tend to match their ranking result with the desired result. It is still possible to pour the events concerned into the learning base with the values of their parameters and the result corrected by the third decision rules. Another improvement of the method would be to delete the first decision rules or only some of them, gradually or not, when it is judged that the second decision rules have reached a sufficient reliability, especially when the learning base has reached a sufficient size or that the second decision rules are stabilized to a state where they classify all events reliably.

For all these refinements where the decision rules are likely to evolve, the use of confidence indexes provides help not only by pointing out dubious results, but by revealing the rules that give them and are therefore less useful or likely to to be corrected.

Claims

A method of identifying a person's postures, for identifying events corresponding to a particular posture or a change of posture and associated with respective classes, comprising the steps of:

- fixing a sensor on the person;

recording at least one sensor signal during a time window; extracting partitioning parameters from said signal; dividing, from the partitioning parameters, the time window into segments satisfying a criterion of homogeneity of posture or change of posture, and extracting, for each of the segments, a first set of parameters from said signal; ranking each of the segments in one of the event classes associated with said events or in a class that is not associated with an event, applying first decision rules from the first set of parameters associated with the segments; - comprising a calibration operation for creating a learning base, comprising the following steps for each segment belonging to an event class:

2. A method of identifying postures of a person according to claim 1, characterized in that the sensor (16) comprises a triaxial accelerator and a triaxial magnetometer.

3. A method of identifying postures of a person according to claim 1 or 2, characterized in that at least some of the segments identified by the first and second decision rules are incorporated into the learning base by their parameters and the associated postures.

4. A method of identifying postures of a person according to claim 1 to 3, characterized in that it comprises steps of sampling the signal in portions of uniform durations and grouping groups of portions to give the segments.

5. A method of identifying postures of a person according to claim 4, characterized in that the grouping of the groups of portions is carried out according to a criterion of maximum variation of activity level of the person and a criterion of maximum variation of an inclination of the person.

6. A method of identifying postures of a person according to claim 4 or 5, characterized in that it comprises a division of the segments into two alternating categories.

7. A method of identifying postures of a person according to claim 1 to 6, characterized in that the first decision rules comprise a step of distinction between horizontal postures and vertical postures according to a measurement of orientation of gravity acceleration by the sensor.

8. A method of identifying postures of a person according to claim 7, characterized in that the first decision rules comprise a step of distinction between a standing posture and a sitting posture according to, at least, duration criteria. and level of activity.

9. A method of identifying postures of a person according to claim 7 or 8, characterized in that the first decision rules comprise a step of identifying states. transients immediately between pairs of states of different postures separating said transient states, and transient state elimination immediately between pairs of states of similar postures.

10. A method of identifying postures of a person according to any one of claims 1 to 9, characterized in that the second group of parameters comprises, for each of the segments, statistical parameters on the signals, a duration of the segment , a frequency content of the segment.

11. A method of identifying postures of a person according to any one of claims 1 to 10, characterized in that the second group of parameters comprises a segment signature, that is to say at least one component of the signal reflecting a function of evolution of the posture.

12. A method of identifying postures of a person according to any one of claims 1 to 11, characterized in that the second decision rules depend on similarity scores or distance between the segments to identify and the segments paid into the training base according to the postures associated with said paid segments.

13. A method of identifying postures of a person according to any one of claims 1 to 12, characterized in that the sensor comprises gyrometers and the second decision rules comprising a comparison of a mean azimuth in the segments to a reference azimuth.