METHOD FOR CHARACTERIZING A GALLEY

Description

GENERAL DOMAIN

[0001] The present invention relates to a method and system for characterizing a person's gait. STATE OF THE ART

[0002] The methods for characterizing a person's gait are of two types: simple step counting methods, and step detection methods, which are more complex because they aim to determine very specific characteristic moments of the step, such as the moment the heel hits the ground.

[0003] The majority of known step detection methods are based on a succession of ad hoc processing steps applied to a motion signal measured on a target subject.

[0004] For example, we know the Pan-Tomkins preprocessing, described in the article "A real-time QRS detection algorithm", by J. Pan and W. J Tompkins. This preprocessing includes band-pass filtering, derivation, squaring, integration and peak search applied to the motion signal measured on the target subject.

[0005] A major drawback of these methods is that they require the calibration of many parameters (filter sizes, threshold values, peak selection criteria, etc.) which are often difficult to estimate or learn, and are often chosen empirically. In addition, these methods rely on a a priorivery strong on the shape that a step should have, which clearly limits their effectiveness on target subjects whose step is of an unconventional shape, for example people suffering from a gait disorder. Documents EP 2439492 A1 (HONEYWELL, 2012-04-11) and US 2007 / 021689 A1 (STERGIOU NIKOLAOS ET AL, 2007-01-25) are relevant for understanding the state of the art in relation to the present invention. SUMMARY OF THE INVENTION

[0006] An aim of the invention is to characterize the gait of a subject without requiring a preliminary calibration step. This objective may include detecting steps or determining information characteristic of the subject's gait.

[0007] Another aim of the invention is to be able to implement a gait characterization method which can work equally well on subjects suffering from gait disorders as on healthy subjects.

[0008] According to a first aspect of the invention, there is therefore proposed a method for characterizing gait according to claim 1.

[0009] The gait characterization method proposed above can be completed using the following characteristics taken alone or in combination when technically possible.

[0010] The method may comprise a step of identifying, in the measurement signal, at least one reference instant characteristic of a step of the target subject, as a function of the reference signal from which the selected local maximum of interest was produced.

[0011] The method may comprise a step of counting the number of selected local maxima of interest corresponding to a given reference signal.

[0012] The method may comprise a step of counting the number of local maximums of interest linked to each class of reference signals, each class of reference signals having been generated from information characteristic of the reference subjects.

[0013] At least one reference signal can be compared with several portions of the measurement signal temporally offset from each other, so as to produce a sequence of similarity coefficients ordered according to the temporal positions of the portions of the measurement signal; a similarity coefficient of a sequence is then selected as a local maximum depending on the result of a comparison between the value of the coefficient with the value of the coefficient which precedes it in the sequence, and on the result of a comparison between the value of the coefficient with the value of the coefficient which follows it in the same sequence.

[0014] Each reference signal comprising at least one predetermined reference instant characteristic of a step of a reference subject, at least one reference signal can be compared with several portions of the measurement signal temporally offset from each other, and the reference instant characteristic of a step of the target subject can be identified from the predetermined reference instant of the reference signal from which the selected local maximum of interest was produced.

[0015] A local maximum may not be selected when it has a value below a predetermined threshold.

[0016] Local maxima can be selected in decreasing order of the value of the similarity coefficients.

[0017] As defined in claim 1, a local maximum cannot be selected when the portion of the measured signal from which it was produced temporally overlaps another portion of the measured signal from which an already selected local maximum was produced. Such an overlap makes it possible to accelerate the search for local maxima and consequently reduces the implementation time of the gait characterization method.

[0018] The method may comprise calculating a ratio between dimensional characteristics of the two compared signals. In this case, the identification step is not implemented for a local maximum produced from the compared signals whose calculated dimension ratio crosses a predetermined threshold.

[0019] The dimensional characteristics can be amplitudes or variances of the compared signals.

[0020] The comparison step may include a correlation producing a Pearson linear coefficient, the similarity coefficient depending on the absolute value of the Pearson linear coefficient.

[0021] The plurality of reference signals may include reference signals of different durations.

[0022] The measurement signal may be a signal representative of the movement of a foot of the target subject.

[0023] The measuring signal can be or depend on at least the following signals: an acceleration signal of the target subject along a direction normal to the ground, and / or an acceleration signal of the target subject along a direction normal to an upper surface of a foot of the target subject, and / or an angular velocity signal about an axis of rotation of a foot of the target.

[0024] The method may include acquiring the measurement signal by at least one motion sensor attached to the target subject.

[0025] At least one reference signal may be acquired during a walking sequence comprising a straight line walk of at least 5 meters from a reference subject.

[0026] According to a second aspect, there is further provided a gait characterization system according to claim 7.

[0027] The system, according to this second aspect, may further comprise at least one motion sensor adapted to provide the measurement signal to the receiving interface.

[0028] This system may further comprise a memory in which the plurality of reference signals are stored. DESCRIPTION OF FIGURES

[0029] There figure 1 schematically represents a gait characterization system and the foot of a subject seen in profile. The figure 2is a flowchart of steps of a preliminary phase to a gait characterization method, according to an embodiment of the invention. The Figures 3a and 3b represent two measurement signals acquired by implementing the preliminary phase of the figure 3 . There figure 4 is a flowchart of steps of a gait characterization method according to an embodiment of the invention. The Figure 5 illustrates a reference signal and a target signal compared during the implementation of the detection method of the figure 4 .

[0030] In the figures, identical elements are designated by identical references. DETAILED DESCRIPTION

[0031] In reference to the figure 1 , a gait characterization system comprises a measuring device 1, a signal processing device 2 and a server 3.

[0032] The measuring device 1 comprises at least one motion sensor 10 and a belt 12.

[0033] The belt 12 is adapted to be fastened around a subject's foot, for example between the heel and the toes as shown in the figure 1 The motion sensors 10 are attached to the belt 12 so as not to hinder the walking of the foot to which the belt 12 is attached.

[0034] For example, three motion sensors 10 are provided in the measuring device 1, in order to measure movements of the foot along three axes X, Y and Z together defining a measurement reference attached to the device 1.

[0035] For example, at least one angular velocity sensor around one of the aforementioned axes and / or at least one acceleration sensor along one of the aforementioned axes is used.

[0036] The measuring device may also comprise at least one magnetometer. For example, three magnetometers are provided, each magnetometer being configured to acquire a magnetic field component along one of the respective X, Y, Z axes.

[0037] The measuring device 1 also comprises internal signal processing means 14, and a communication interface 16.

[0038] The signal processing means 14 are configured to develop at least one output movement signal from at least one of the elementary signals provided by each of the different movement sensors.

[0039] The communication interface 16 is configured to communicate with the signal processing device 2. The communication interface is in particular capable of transmitting to the signal processing device 2 each output signal produced by the internal processing means of the measuring device 1.

[0040] The communication interface 16 is preferably wireless so as not to hinder the subject's walking when the measuring device 1 is attached to his foot.

[0041] In one embodiment, the measuring device 1 is a device marketed by the company Xsens, which is attached to a foot using a hook-and-loop strap, and the output signals are sampled at 100 Hz.

[0042] The signal processing device 2 comprises at least one processor 20, a first communications interface 22, a second communications interface 24 and an internal memory 26.

[0043] The processor 20 is configured to implement a gait characterization method encoded in the form of program code instructions, based on signals received via the two communications interfaces 22 and 24.

[0044] The communication interface 22 is adapted to receive the output signals emitted by the measuring device 1, for example by means of the Wi-Fi protocol or the Bluetooth protocol.

[0045] The communication interface 24 is adapted to communicate with the remote server 3, via a public network such as the Internet.

[0046] The signal processing device 2 may, for example, be in the form of a mobile terminal, a laptop computer, or a PC.

[0047] The server 3 comprises a communication interface 34 for communicating with the signal processing device 2, a processor 30 and a memory 32 adapted to store a set of measurement signals, and data relating to these signals, in the form of a database.

[0048] We will now detail the steps of a gait characterization process implemented using the system of the figure 1 . Preliminary phase of collecting reference signals

[0049] In reference to the figure 2 , a preliminary phase of collecting reference signals is implemented.

[0050] In a first step 102 of this preliminary phase, each reference signal is measured on a reference subject.

[0051] The acquisition 102 of a signal is implemented by means of the acquisition system illustrated in the Figures 1 and 2 . To do this, a measuring device 1 is attached to one foot of a reference subject (see two devices 1, one for the left foot and the other for the right foot). The reference subject is then asked to walk according to a predetermined clinical protocol.

[0052] During the walking of the reference subject, the movement sensors 10 acquire signals representative of a movement parameter (acceleration and / or speed, for example angular speed) of the foot along the X, Y, Z axes, which are then processed by the internal processing means of the measuring device 1 to generate at least one reference acceleration signal of the corresponding foot of the reference subject.

[0053] The internal processing means can also use the magnetic field measurements acquired by the magnetometers, if present in the measuring devices 1.

[0054] Each of the two measuring devices 1 (left and right feet) acquires nine raw signals (3D accelerations, 3D angular velocities and 3D magnetic fields) in the reference frame linked to the sensor defined by the X, Y and Z axes, these signals possibly being able to be recalibrated or calibrated in a reference frame fixed relative to the ground via the internal processing means 14 of the measuring devices 1, in order to obtain the reference signals.

[0055] Preferably, some of the components relevant to characterizing a gait are used as reference signals. Three components relevant to characterizing the gait, and represented on the figure 1 , are: an acceleration signal of the reference subject measured along a direction G normal to the ground, an acceleration signal of the reference subject measured along the Z axis linked to the device 1, which axis is normal to the upper surface of a foot of the reference subject, an angular speed signal of rotation of a foot of the reference subject around the Y axis during a straight line walk of the reference subject (this Y axis being perpendicular to the Z axis and parallel to an axis of inclination of the foot at the ankle).

[0056] Each reference signal is then transmitted to the processing device 2, via the interface 22, then to the server 3 via the interfaces 24 and 30.

[0057] In a step 104, the measured reference signal is stored in the memory 32 of the server 3. A reference signal can be stored in the internal memory 26 of the processing device 2, instead of or in addition to the memory 32 of the remote server 3.

[0058] In a step 106, each stored reference signal is the subject of an annotation, aimed at determining at least one reference instant characteristic of the step of the reference subject.

[0059] This annotation step can be implemented empirically by a practitioner who scrutinizes the gait of the reference subject during the acquisition 102 of the reference signal.

[0060] Furthermore, at least one piece of information characteristic of the corresponding reference subject, previously determined, can be associated with each reference signal.

[0061] A characteristic gait information can for example identify a particular class of subject, among several predetermined classes.

[0062] A class can thus relate to a type of gait (normal, limping, etc.), a severity of the impairment to the gait, a gait speed, the age of the reference subject (young, old, etc.), etc.

[0063] For example, this information may relate to a complaint expressed by subjects (sensory, motor, balance, cognitive or hearing disorders, vertigo, dizziness or pain), to a noted and characterized deficit (lateralization, chronic or acute occurrence) of the subject (proprioceptive, vestibular, visual, motor: hemiparesis, paraparesis, Parkinsonian syndrome...). This information may include the anatomical level of the lesion: central or peripheral nervous system or in the sensory organs (central vestibule, peripheral vestibule, peripheral nerves, spinal cord, trunk, cerebellum, joints of the lower limbs, Cerebral hemisphere: cortex, basal ganglia or white matter). This information may also relate to the etiology of the injury (vascular, inflammation, trauma, infection, tumor before or after surgery, degenerative, toxic, metabolic, psychosomatic...).

[0064] The same steps 102, 104, 106 are implemented repeatedly for several reference subjects.

[0065] At the end of the preliminary phase, a plurality of reference signals are thus stored in the memory 32, and each reference signal is annotated with at least one instant characteristic of the step of the reference subject on which the reference signal was measured, and / or at least one piece of information characteristic of the reference subject.

[0066] It is possible to store a signal whose duration covers at least part of a step or several steps of the corresponding reference subject.

[0067] However, it is preferable to store a reference signal having a duration that covers at least part of a single step of the reference subject, so as to limit the memory consumption by the reference signal, and to avoid potential redundancy of information. The or each characteristic instant stored in addition to the reference signal itself relates to this single step.

[0068] Furthermore, using a single step of the reference subject or a part of this step increases the sensitivity and specificity of this method. In addition, it allows better results when the step of the target subject is of unconventional shape.

[0069] The use of at least part of a single step of the reference subject corresponds, for example, to the use of one or more parts of the phases of the step such as the support phase, the initial contact, the swing, etc.

[0070] In practice, this annotation is materialized by the storage of data making it possible to locate the characteristic instants in the corresponding reference signal, this data being logically linked with this reference signal in memory 32 or 26.

[0071] The reference subjects from which the reference acceleration signals are measured preferably have different gait profiles. In general, the reference subjects from which the reference acceleration signals are measured can be healthy subjects or patients (subjects suffering from a medical condition).

[0072] For example, at least one of the reference signals is measured on a subject with a gait disorder (three examples of reference signals measured on a patient with a hip problem are illustrated in Figure 3b); and at least one of the reference signals is measured on a healthy subject, i.e. a subject who does not suffer from gait disorders (such as the three reference signals represented in Figure 3a ).

[0073] Thus, the invention also relates to a method for collecting information on gait such as walking style or possible pathology. Thus, it is for example possible to distinguish slight limps, particular ways of lifting the foot, bowed legs, a swaying during walking. Even if a common structure can possibly be detected between the reference signals represented on the Figures 3a and 3b , it should be noted that their amplitudes are very different (on the Figure 3b , the subject in fact tends to drag his feet, which modifies the structure and amplitude of the step compared to the reference subject of the Figure 3a ).

[0074] Furthermore, the different reference signals collected are not necessarily of the same duration, nor associated with an identical number of characteristic step instants. We understand, in fact, that, from one reference subject to another, a step can be of different duration.

[0075] The collected reference signals thus constitute a library of steps covering a wide range of gaits, which will be used to characterize the step of a subject qualified in the sequence as a “target” subject, this target subject preferably being different from the reference subjects. Target subject's approach characterization phase

[0076] In reference to the figure 4 , a method for characterizing the gait of the target subject comprises the following steps.

[0077] In a step 202, the measuring device 1, previously attached to a foot of the target subject, acquires at least one movement signal of the target subject, this signal being referred to below as a “target signal”.

[0078] The method according to the invention can be implemented on a target signal having been previously measured.

[0079] A target signal is for example a foot acceleration signal along the X, Y and / or Z axis, or a foot angular velocity signal around the X, Y and / or Z axis.

[0080] The target signal and the reference signals are therefore measured on different subjects. In order to improve detection, the acquisition protocol of the reference signals is representative of the conditions under which the target signal is generated.

[0081] The target signal is then transmitted to the processing device 2.

[0082] In a step 204, a comparison between the target signal and a reference signal is implemented.

[0083] This comparison 204 can be implemented by the processor 20 of the processing device 1; in this case, the processing device 1 sends a request to the server 3 to download a reference signal for this comparison 204. Alternatively, the comparison 204 is implemented by the server 3, the processing device then serving only as a relay for transmitting the target signal between the measuring device 1 and the server 3.

[0084] The comparison 204 is more precisely implemented for different time portions of the target signal temporally shifted relative to each other in the target signal. It is then understood that as many comparisons are implemented as there are time portions defined in the target signal.

[0085] Each time portion of the target signal has the same duration T as the reference signal to which this portion will be compared.

[0086] The different portions of the target signal can be shifted from each other by a constant time step.

[0087] Different portions of the target signal may overlap.

[0088] The comparison 204 of a given temporal portion of the target signal with a reference signal produces a coefficient representative of a probability of similarity between the two compared signals, this coefficient being called in the following “similarity coefficient”.

[0089] The comparison 204 comprises, for example, a correlation of the temporal portion of the target signal with the reference signal; in this case, the similarity coefficient depends on the linear Pearson coefficient between the two correlated signals.

[0090] The similarity coefficient ρ ( x,y ) between two signals x and y can more precisely be expressed as the absolute value of the Pearson linear coefficient, as follows: ρ x y = cov x y std x std y where cov(x, y) denotes the covariance of these signals x and y, std ( x ) And std ( y ) denote the respective standard deviations of the x and y signals.

[0091] In this case, a high value of the similarity coefficient will indicate a strong similarity between the temporal portion of the target signal and the reference signal being compared. On the other hand, a low value of this similarity coefficient will indicate a weak similarity between the two signals being compared.

[0092] For N temporal portions of the target signal, N similarity coefficients are therefore produced, each having a specific value, via N comparisons 204. The N comparisons carried out cover the entire duration of the target signal.

[0093] The set of N similarity coefficients produced from the target signal and the same given reference signal is hereinafter called a coefficient sequence.

[0094] The coefficients of a sequence are classified in order of temporal position of the N different temporal portions of the target signal considered during the comparison step 204. A direction of travel of a sequence of coefficients therefore corresponds to a direction of temporal travel of the target signal.

[0095] The comparison step 204 is repeated for each reference signal previously stored during the preliminary phase. It will be noted that, each reference signal being of duration T potentially distinct from that of the others, the time portions of the target signal which are compared to it can also be redefined differently for each reference signal.

[0096] By crossing each reference signal with the target signal, the total number of sequences of similarity coefficients produced at the end of these repetitions is equal to the number of target signals acquired simultaneously on the target subject, multiplied by the number of reference signals memorized.

[0097] Each sequence can have a different number of coefficients, depending on the step chosen to shift the time portions relative to each other in the corresponding target signal, and depending on the length (duration) of the reference signal considered.

[0098] In a step 206, the processor implements an identification of local maxima in a given sequence of similarity coefficients.

[0099] In a given sequence, a current coefficient constitutes a local maximum if the next coefficient in the sequence and the previous coefficient in the sequence have values ​​that are less than that of the current coefficient. The detection of local maxima typically proceeds by a series of comparisons between values ​​of adjacent coefficients in a sequence.

[0100] Each local maximum is then a similarity coefficient produced from a portion of the signal that resembles the reference signal more than adjacent portions of the target signal.

[0101] A given sequence can thus include zero, one or more local maxima.

[0102] Step 206 is repeated for each sequence of coefficients produced.

[0103] In a selection step 208, the processor selects, from among several identified local maxima, certain local maxima of interest.

[0104] The local maxima of interest are selected on the basis of the value of the similarity coefficients: a local maximum is not selected as a local maximum of interest if its value is lower than a predetermined threshold λ, so as to eliminate local maxima with low similarity probability values.

[0105] For example, in the case of similarity coefficients defined as the absolute value of Pearson linear coefficients, only local maxima with a value greater than or equal to 0.5 are retained as local maxima of interest.

[0106] Preferably, the predetermined threshold λ is between 0.6 and 0.8.

[0107] The selection 208 can advantageously be implemented in decreasing order of value of the similarity coefficients; in this way, the number of comparisons with the threshold λ is reduced.

[0108] Another criterion for selecting 208 a local maximum is the overlap of the portion of the corresponding target signal with the temporal portion of the same target signal from which another similarity coefficient was produced, constituting not only another local maximum, and in addition having already been selected as a local maximum of interest among the local maxima of the sequence.

[0109] In other words, knowing that a first local maximum is produced from a comparison between a first portion of the target signal and a reference signal, then selected as the local maximum of interest, a second local maximum produced from a comparison between the reference signal and a second portion which overlaps the first portion is not selected as a local maximum of interest, and a third local maximum produced from a comparison between the reference signal and a third portion which does not overlap the first portion is indeed selected as a local maximum of interest.

[0110] Preferably, the step 208 of selecting the local maxima of interest is implemented in order of decreasing value of coefficient value; it is thus certain that the local maxima having the largest values ​​in the sequence will be selected as the local maximum of interest.

[0111] Since as many sequences as reference signals can be produced, the selection order 208 can be done after sorting all the local maxima listed in the set of sequences.

[0112] At the end of the selection step 208, at least one local maximum of interest is available, which is associated with a specific portion of the target signal; and when a plurality of local maxima of interest are produced, it is certain that the associated specific portions of the target signal are temporally disjoint in the target signal (i.e. do not overlap).

[0113] In an identification step 210, the processor identifies in the target signal at least one instant characteristic of a step of the target subject on the basis of the selected local maxima of interest.

[0114] This identification step 210 comprises the downloading from the server 3 to the processing device 2, of the characteristic reference instant(s) annotated in the reference signal considered here, during the preliminary phase of collecting the reference signals. Alternatively, the identification step is implemented by the server 3.

[0115] A landmark instant characteristic of the target subject is identified in the target signal by examining a local maximum of interest selected in step 208.

[0116] As a reminder, the reference signal leading to the obtaining of a given local maximum of interest was annotated with at least one characteristic instant during the preliminary phase and / or with at least one characteristic of the corresponding reference subject.

[0117] For example, we have represented in Figure 5 a reference signal p comprising two characteristic instants: a step start instant ta and a step end instant tb.

[0118] The characteristic instant ta is reported in the time portion of the target signal at the time position you + your, where ti denotes the start time of the time portion in the target signal.

[0119] Similarly, the characteristic instant tb is reported in the time portion of the target signal at the time position you + your, where ti denotes the start time of the time portion in the target signal.

[0120] Each predetermined reference instant during the preliminary phase is thus mapped into the target signal.

[0121] The processor 20 proceeds in the same manner for any other selected local maximum of interest.

[0122] The identification step 210 thus makes it possible in a particular application to count the steps of the target subject.

[0123] In one embodiment, the identification 210 of characteristic reference instant(s) is not necessarily implemented for all the local maxima selected during the selection step 208. For each local maximum of interest, the processor calculates in this case a ratio between dimensional characteristics of the two compared signals on the basis of which the maxima was identified 206, then selected 208.

[0124] If this dimension ratio crosses a predetermined threshold µ, then the characteristic instant identification step 210 is not carried out in the time portion of the target signal.

[0125] For example, if the numerator of the ratio is the dimensional characteristic relating to the target signal and if the denominator of the ratio is the dimensional characteristic relating to a reference signal, then it is checked whether the ratio is less than the threshold µ, and if this is the case, then the step 210 of identifying the characteristic instant is not carried out in the time portion of the target signal.

[0126] Otherwise, the identification step 210 according to the preceding methods is implemented on this time portion of the target signal.

[0127] The dimension ratio can be for example an amplitude ratio between the compared signals or a variance ratio between the compared signals.

[0128] Several ratios can be calculated for a pair of compared signals; it can then be planned not to trigger the benchmark instant identification step if at least one of the ratios is greater than a predefined threshold.

[0129] Preferably, the calculation of the dimension ratios is implemented after the identification 206 of the local maxima, very preferably after the selection of the local maxima of interest, so as to limit the number of ratios to be calculated.

[0130] Steps 202, 204, 206, 208 and 210 can be repeated for several target signals acquired simultaneously via different sensors of the measuring device 1 (for example acceleration signals along different axes among those mentioned above).

[0131] Other processing than the identification step 210 can be implemented once local maxima have been selected in the selection step 208.

[0132] The processor 20 can, for example, count 212 the number of times that a reference signal has made it possible to generate a local maximum of reference interest.

[0133] The processor 20 then downloads the characteristic information of the reference subjects (identified during the preliminary phase) which are associated with the reference signals having generated at least one of the local maxima of interest.

[0134] The processor 20 can for example calculate a percentage of representation of each characteristic information in the set of local maxima of interest and therefore of the target signal.

[0135] It is entirely possible that a reference signal (or several reference signals carrying the same characteristic gait information) is represented in a majority manner in the local maxima of interest.

[0136] The majority characteristic information can thus be attributed to the target subject.

[0137] The processor 20 can also, for each class of reference signals generated from information characteristic of the reference subjects, count 212 the number of local maximums of interest linked to this class and therefore having been associated with the target signal.

[0138] The processor 20 can then, for example, calculate a percentage of representation of each characteristic information in the target signal.

[0139] The majority characteristic information can thus be attributed to the target subject. Example of algorithmic implementation of the detection process

[0140] The detection method can be implemented in the form of a computer program comprising code instructions capable of being executed by the processor 20.

[0141] An algorithm consisting of a possible embodiment of this computer program will be detailed below.

[0142] In the following, we consider that the measuring device 1 provides a k-uplet of discrete signals measured on the same subject. This k-uplet can be represented by a vector of length |p| and dimension k.

[0143] We introduce the following notations: x is a vector representing a k-tuple of discrete target signals measured on the target subject. p is a vector representing a k-tuple of discrete reference signals measured on the same reference subject. |P| is the number of vectors p. |x| (respectively |p|) is the length of the vector x (respectively p), counted in number of samples. x ( k )< (respectively p ( k )< ) is the kth< component of the vector x x i : j k is the temporal portion of x(k) between the samples of indices i and j; we therefore have x 1 : x k = x k .

[0144] The similarity coefficients, noted r, are calculated between the reference signals and the target signal during the comparison step 204, for all possible time portions of length |p|, which makes a total of 3 × |P| × |x| similarity coefficients: ∀ k , ∀ p ∈ P , ∀ 1 ≤ x ≤ x r k p i = ρ p k x i : i + p − 1 k where r(k, p, i) is the similarity coefficient resulting from the comparison between the kth component of the reference signal p and the kth component of the target signal at time i, and where ρ is a linear Pearson coefficient in absolute value.

[0145] During step 206, a coefficient r ( k,p,i ) is designated as a local maximum if it is larger than its direct temporal neighbors. The set L of local maxima is thus the set of similarity coefficients r ( k,p,i ) cumulatively verifying the following two conditions: r k p i > r k , p , i − 1 r k p i > r k , p , i + 1

[0146] The set L contains all acceptable positions for steps of the target subject; each coefficient r ( k, p, i ) can be interpreted as the likelihood of finding, in the target signal at time i, a step similar to that of the reference signal of index p.

[0147] The selection step 208 is coded as a greedy process. At each iteration of the algorithm, the largest value is chosen r ( k*, p*,i* ) of coefficient in the set L: If the step p * positioned at the moment i * overlaps a step that has already been detected, we reject it and start the process again; If the step p * can be positioned instantly i * without overlapping previous detections, it is considered detected and the indices between i* And i * + | p *| - 1 are prohibited in future iterations.

[0148] The greedy algorithm of selection step 208 ends when all indices are forbidden, or when all local maxima of the set L have been considered, or when the values ​​of the coefficients remaining in the set L are all smaller than the predetermined threshold λ. This threshold λ acts as an acceleration parameter of the algorithm; in fact, the algorithm is avoided from running for too long.

[0149] The greedy algorithm of step 208 produces a set S of local maxima of interest.

[0150] The step 210 of identifying the characteristic instants is then implemented on the basis of this set 5 and the threshold µ. As indicated above, the threshold µ is used to reject certain characteristic instants, in the event of too large a dimensional difference between a portion of the target signal and a portion of the reference signal associated with a local maximum of interest included in the set S.

[0151] For example, the predetermined threshold µ is less than or equal to 0.2 and preferably is between 0.05 and 0.2, more preferably between 0.1 and 0.2. Results

[0152] The inventors were able to obtain particularly satisfactory results after establishing a signal database containing 295 signals measured from 74 subjects, including 23 healthy subjects and 51 suffering from various orthopedic disorders of the knee, hip or ankle, some of the subjects having carried out the protocol several times.

[0153] The clinical protocol for acquiring reference signals, to which the reference subjects wearing measuring device 1 were subjected, is as follows: stand still for 6 seconds, walk straight for 10 meters at a comfortable speed, turn around, return to the starting point, stand still for 2 seconds.

[0154] All signals in the database were manually annotated with two characteristic landmark times: a step start instant ta , at which the heel leaves the ground, a step end instant tb , at which the toes are placed on the ground and stabilized.

[0155] In total, 9357 steps were memorized (4674 on the right foot and 4683 on the left foot).

[0156] The operation of the detection method was tested using the following evaluation method. We randomly chose |P| steps from the database that constitute a learning base. We also chose steps from the database that constitute a test base. In order to avoid over-fitting, all signals measured on subjects present in the learning base are removed from the test base. We applied the gait characterization method to each signal in the test base using the |P| signals in the learning base as reference signals. We then calculated a precision score and a recall score for the characteristic instants of the start and end of steps identified in each signal in the test base by applying the method according to the invention, considering that a step is correct if the average of the detected instants of the start and end of steps falls at an instant where a step is actually present in view of the annotations collected by the practitioners during the preliminary collection phase described above.The precision and recall scores were then averaged for all signals in the test base.

[0157] In the results presented below, this process was repeated 100 times and the mean and standard deviation of the precision and recall scores were considered.

[0158] Table 1 below shows the precisions and recalls (mean, and standard deviation in parentheses over 100 simulations) for different numbers of |P| of reference signals. Table 1 Value of |P| Precision Reminder 5 96.6 (5.04) 96.6 (1.04) 10 96.1 (3.76) 96.5 (3.42) 15 95.7 (3.73) 96.4 (3.25) 20 96.0 (2.81) 97.0 (1.76) 25 95.8 (2.67) 97.0 (1.43)

[0159] The threshold λ was set to 0.6 to reduce the computation time. The first observation is that all these scores seem quite similar: all average scores are between 95% and 97%, even when only 5 reference signals are used. All standard deviations are smaller than 5%, which suggests a high stability of the method.

[0160] While the number of reference signals does not seem to significantly affect the average performance, it mainly influences the robustness of the method. It can be observed that the more the number of reference signals increases, the more the standard deviations decrease, and therefore the more robust the method is. This observation makes sense because it is intuitive that the more reference signals we have, the more likely we are to consider all pathologies and step types, which tends to standardize the results for all signals in the test base. When |P| = 5, we observe an asymmetry between the standard deviations of precision and recall. Due to the small number of reference signals, we force the correspondence between small pathological steps and larger healthy steps, which can cause over-detection and tends to lower precision.

[0161] Table 2 below shows the precisions and recalls (mean, and standard deviation in parentheses over 100 simulations) for different values ​​of the predetermined threshold λ. Table 2 Value of λ Precision Reminder 0,60 95,23 (4,50) 97,63 (1,61) 0,65 93,97 (7,52) 96,74 (4,58) 0,70 95,96 (4,96) 97,33 (3,22) 0,75 96,44 (4,32) 97,19 (1,99) 0,80 97,03 (1,02) 96,59 (1,89) 0,85 97,83 (2,92) 95,26 (1,93) 0,90 98,49 (4,74) 90,36 (2,51)

[0162] When the predetermined threshold λ increases, only reference signals that are highly correlated with the models are selected: this increases the precision. On the contrary, when the predetermined threshold λ decreases, the recall increases and the precision decreases. These results confirm the usefulness of the parameter λ: by increasing λ to an appropriate value (of the order of 0.6 to 0.8), it is possible to increase the precision (and the robustness of the precision) while maintaining the recall constant. In addition, λ has an impact on the computational cost. For example, using λ = 0.8 instead of λ = 0 (equivalent to the algorithm without λ) it is possible to calculate the result about 2 times faster. It is therefore interesting to use the largest value of λ possible. The best performances are obtained for λ = 0.8, which is a good compromise between recall and precision.

[0163] Table 3 below shows the precisions and recalls (mean, and standard deviation in parentheses over 100 simulations) for different values ​​of the predetermined threshold µ. Table 3 Value of µ Precision Reminder 0,05 96,23 (2,29) 97,42 (0,93) 0,10 95,92 (5,66) 97,04 (1,60) 0,15 97,03 (4,02) 96,59 (1,89) 0,20 96,72 (6,30) 95,78 (1,89) 0,25 97,68 (2,67) 95,25 (1,31) 0,30 97,84 (4,94) 93,88 (2,73)

[0164] The parameter µ mainly influences the recall. Indeed, when µ is too large, all measurements whose amplitude is too different from that of the reference signals are not considered. This has a double effect: if one of the reference signals corresponds to a pathological patient whose steps have a low amplitude, then it will not be able to detect measurements on healthy patients. The opposite situation can also occur. Indeed, when µ increases, the normalization effect provided by the Pearson correlation coefficient is neutralized. Table 3 shows that, preferably, µ should not be greater than 0.2 so that the recall is not too low.

[0165] This step, based on 100 simulations, led to testing 21,000 different configurations. Among these configurations, it is possible to mention the configuration: |P| = 5 reference signals, λ = 0.75 and µ = 0.1, with an average recall of 98.40% and an average precision of 98.44%. This choice of parameter is an example and the parameters can be modified depending on the content of the step database and the sequences to be analyzed.

[0166] The good performance of the method according to the invention is related to the reference signals that make up the library. It is remarkable that this method requires only a small number of reference signals, which tends to show that the algorithm does not need a large library to perform accurately. For example, the optimized method includes 5 reference signals: 1 reference signal belonging to a healthy subject, 3 reference signals corresponding to neurological diseases, and 1 reference signal associated with orthopedic diseases.

[0167] The method according to the invention also has the advantage of being able to accurately detect step limits and adapt to different step durations. For 90% of the steps, the start, arrival and time duration errors are less than 0.25 seconds (in absolute value). These results are satisfactory compared to the uncertainty annotations of experts and specialists.

[0168] The detection method according to the invention was compared with a Pan-Tomkins preprocessing method cited in the introduction, configured with the parameters suggested in the article “Automatic step detection in the accelerometer signal,” by H. Ying, C. Silex et al.

[0169] The results provided by this prior method are of low precision (of the order of 25.6% to 51.20% depending on the groups of sequences used). Furthermore, this method does not allow the detection of a characteristic instant of a step such as the present invention could do (e.g. instant of start of step or instant of end of step).

[0170] The Pan-Tomkins method was optimized on the entire database by choosing optimal values ​​for several parameters (e.g. fmin, fmax, integration window, neighborhood size, and noise level) to maximize the F-measure. When optimized, the algorithm yields 97.82% recall and 95.72% precision. Detailed results are shown in Table 4. While these scores are comparable to the method according to the invention on healthy subjects, it should be noted that the Pan-Tomkins method has difficulty handling subjects with disorders such as neurological or orthopedic disorders. In particular, on these subjects, over-detection occurs, thus reducing accuracy.This shows a limitation of classical step detection methods based on signal processing: if the signals to be studied have different properties (noise, frequency content, amplitudes), it is difficult to find a single processing adapted to all the signals. Table 4 Invention Pan-Tomkins Reminder Precision Reminder Precision Healthy Subjects 98,93 98,98 99,14 97,09 (3,78) (2,22) (2,43) (1,71) Subjects with orthopedic disorders 97,54 98,77 98,78 94,87 (5,09) (2,92) (2,12) (2,09) Subjects with orthopedic disorders 98,55 98,05 96,80 95,49 (4,55) (3,05) (3,02) (3,52) Total 98,40 98,44 97,82 95,72 (4,56) (2,89) (2,72) (3,07)

[0171] Ultimately, the detection method implemented in accordance with the invention achieves robust performance, even using a small number of reference signals.

[0172] The detection method according to the invention can find multiple applications. The detection method can in particular serve as a preliminary phase in the diagnosis or prognosis of gait disorders. However, no diagnostic method is claimed. Since it uses data that has been previously measured, it does not require interaction with the target subject to be implemented.

Claims

1. A method of gait characterization comprising the steps of: • reception of a motion measurement signal from a previously measured target subject, • comparison (204) of at least a portion of the measurement signal with a plurality of reference signals, the plurality of reference signals comprising at least one motion signal previously measured on a reference subject different from the target subject, each comparison of the measurement signal with a reference signal producing a similarity coefficient representative of a probability of similarity between the compared signals, in which a first similarity coefficient forming a first local maximum is produced by comparing a first portion of the measurement signal with a reference signal, and a second similarity coefficient forming a second local maximum is produced by comparing a second portion of the measurement signal with the same reference signal as the first similarity coefficient, the second portion being different from the first portion, selecting (208) at least one local maximum of interest from the similarity coefficients produced, in which : ∘ a local maximum among the produced similarity coefficients is not selected as the local maximum of interest when its value is below a predetermined threshold (λ), so as to eliminate local maxima with low similarity probability values, ∘ the first local maximum is selected, then the second local maximum is not selected when the second portion temporally overlaps the first portion, wherein the plurality of reference signals comprises : • at least one movement signal previously measured on a reference subject suffering from a gait disorder, and • at least one motion signal previously measured on a healthy reference subject.

2. A method according to the preceding claim, further comprising a step of identifying (210), in the measurement signal, at least one reference time characteristic of a step of the target subject, as a function of the reference signal from which the selected local maximum of interest was produced.

3. Process according to the preceding claim, in which • each reference signal comprises at least one predetermined reference time (ta ) characteristic of a reference subject step, • at least one reference signal is compared with several portions of the measurement signal that are temporally offset from one another, • the landmark time characteristic of a target subject step is identified (210) from the predetermined landmark time of the reference signal from which the selected local maximum of interest was generated.

4. A method according to one of claims 2 to 3, comprising calculating a ratio between dimensional characteristics of the two compared signals, and wherein the identification step (210) is not implemented for a local maximum produced from the compared signals whose calculated dimension ratio crosses a predetermined threshold.

5. Process according to one of the preceding claims, in which : • at least one reference signal is compared (204) with several portions of the measurement signal temporally offset from one another, so as to produce a sequence of similarity coefficients ordered according to the temporal positions of the portions of the measurement signal, • a similarity coefficient of a sequence is selected as a local maximum (206) according to the result of a comparison between the value of the coefficient with the value of the coefficient that precedes it in the sequence, and the result of a comparison between the value of the coefficient with the value of the coefficient that follows it in the same sequence.

6. Method according to one of the preceding claims, wherein the measurement signal is representative of the movement of a foot of the target subject.

7. Gait characterization system, comprising : • a receiving interface configured to receive a motion measurement signal from a target subject, and • at least one processor configured to : o comparing the measurement signal with a plurality of reference signals, the plurality of reference signals comprising at least one motion signal previously measured on a reference subject different from the target subject, each comparison of the measurement signal with a reference signal producing a similarity coefficient representative of a probability of similarity between the compared signals, wherein a first similarity coefficient forming a first local maximum is produced from a first portion of the measurement signal, and a second similarity coefficient forming a second local maximum is produced from a second portion of the measurement signal different from the first portion, ∘ select at least one local maximum from the generated similarity coefficients, wherein the first local maximum is selected and then the second local maximum is not selected when the second portion temporally overlaps the first portion, wherein the plurality of reference signals comprises at least one movement signal previously measured on a reference subject suffering from a gait disorder, and at least one movement signal previously measured on a healthy reference subject.

8. A system according to the preceding claim, in which the processor is configured to count the number of local maximums of selected interest corresponding to a given reference signal, each reference signal being associated with information characteristic of the corresponding reference subject.