Method for detecting a living being

Abstract

The invention relates to a method for detecting a living being, characterized in that it comprises: - the acquisition of a plurality of images in a time window, each image comprising a series of data to which are associated a position, a power, and a velocity; - the construction of a first database of instantaneous data from the plurality of images, said instantaneous data being based on said position, said power, and said velocity of each data point of each image acquired in said time window; - the construction of a second database of temporal data from said plurality of images, said temporal data being calculated from a periodic signal of each image recorded over a moving time window; - the application of said first database and said second database as input to a machine learning model; - the execution of said machine learning model.- the establishment of a classification by said machine learning model regarding the presence / absence of a living being, - based on said classification, the transmission of corresponding information to a management unit.

Description

Title of the invention: Method for detecting a living being

[0001] The present invention relates to a method for detecting a living being. It finds a particular but non-limiting application in motor vehicles.

[0002] In the field of motor vehicles, a method for detecting a living being inside a vehicle known to those skilled in the art comprises analyzing the respiratory rate from the variation of a power or phase signal over time and comparing it with a respiratory model of a living being. This makes it possible in particular to trigger an alarm system when an occupant, in particular a child, is forgotten in the vehicle.

[0003] A disadvantage of this state of the art is that the dependence on a clear respiratory signal pattern can be masked by any movement of the body of the living being. Furthermore, the body movement patterns of living beings are difficult to model due to their variability and the detection algorithms used are generally based on a simple verification of the signal-to-noise ratio to detect movements of living beings with high energy (i.e. with a high signal-to-noise ratio). However, this type of detection can be confused with external effects (vehicle vibrations, vehicle jolts, etc.) which generate energy comparable with a movement of a living being and can easily lead to false alarms, called true negative cases.

[0004] In this context, the present invention aims to propose a method for detecting a living being which makes it possible to resolve the mentioned drawback.

[0005] To this end, the invention proposes a method for detecting a living being, characterized in that said detection method comprises: - the acquisition of a plurality of images in a time window, each image comprising a series of data to which a position, a power and a speed are associated, - constructing a first database of instantaneous data from the plurality of images, said instantaneous data being based on said position, said power and said speed of each data of each image acquired in said time window, - constructing a second database of temporal data from said plurality of images, said temporal data being calculated from a periodic signal of each image taken over a moving time window, - applying said first database and said second database as input to a machine learning model, - the execution of said machine learning model, - the establishment of a classification by said machine learning model as to the presence / absence of a living being, - based on said classification, the transmission of corresponding information to a management unit.

[0006] Thus, as will be seen in detail later, the detection method is based on the introduction of a temporal characteristic. This temporal characteristic takes into account the changes in the radar images over time. The objective is to have a characteristic that reflects any periodic or oscillation pattern of a signal (e.g. respiratory) and to use it to improve the accuracy of the detection. One of the advantages over conventional detection methods is that this characteristic is not used alone but combined with other characteristics based on a single radar image and, therefore, the final decision does not depend solely on a single characteristic as in the case of conventional detection methods.

[0007] According to non-limiting embodiments, said detection method may further comprise one or more additional characteristics taken alone or in all technically possible combinations, among the following.

[0008] According to a non-limiting embodiment, the instantaneous data are: - average values ​​of said position and / or said power and / or said speed of said data of each image, or - variances of said data of each image.

[0009] According to a non-limiting embodiment, said machine learning model is a supervised model.

[0010] According to a non-limiting embodiment, said supervised model is a support vector machine or a random forest.

[0011] According to a non-limiting embodiment, said periodic signal is a breath or a heartbeat.

[0012] According to a non-limiting embodiment, said living being is of the adult, child, or animal type.

[0013] According to a non-limiting embodiment, said time window is approximately five seconds.

[0014] According to a non-limiting embodiment, the time data is a maximum power or a phase variation of said periodic signal.

[0015] According to a non-limiting embodiment, the construction of said second database comprises: - buffering the time data to obtain a plurality of time values, - the removal of the continuous component of said time values, - the application of a Fourier Transform on said time values ​​for obtain frequency values, - from said frequency values, the extraction of a maximum value in a given frequency range.

[0016] According to a non-limiting embodiment, said given frequency range is a breathing frequency range or a heartbeat frequency range.

[0017] According to a non-limiting embodiment, the construction of said second database further comprises the normalization of said frequency values ​​before said extraction of said maximum value.

[0018] According to a non-limiting embodiment, said plurality of images is obtained from waves received by a wave transmission / reception entity of a sensor.

[0019] According to a non-limiting embodiment, said moving time window is three images per second.

[0020] According to a non-limiting embodiment, said detection method is a method for detecting a living being inside a vehicle.

[0021] According to a non-limiting embodiment, the data series is a series of points.

[0022] According to a non-limiting embodiment, the images are images acquired at the interior of a vehicle.

[0023] According to a non-limiting embodiment, said periodic signal is representative of a type of living being.

[0024] According to a non-limiting embodiment, the sensor is a radar sensor.

[0025] According to a non-limiting embodiment, the images are radar images.

[0026] According to a non-limiting embodiment, the moving time window at the same duration than the time window.

[0027] There is further provided a system for detecting a living being, characterized in that said detection system comprises: - (a) a sensor configured to acquire a plurality of images in a time window, each image comprising a series of data to which a position, a power and a speed are associated, - (b) a processing unit configured to: - constructing a first database of instantaneous data from the plurality of images, said instantaneous data being based on said position, said power and said speed of each data of each image acquired in said time window, - constructing a second database of temporal data from said plurality of images, said temporal data being calculated from a periodic signal of each image taken over a moving time window, - applying said first database and said database as input to a machine learning model, - run said machine learning model, - (c) said machine learning model configured to be executed and establish a classification as to the presence / absence of a living being, - said processing unit being further configured to, depending on said classification, transmit corresponding information to a management unit.

[0028] The invention and its various applications will be better understood upon reading the following description and examining the accompanying figures:

[0029] [Fig-1] illustrates a diagram of a method for detecting a living being, according to a mode non-limiting embodiment of the invention,

[0030] [Fig.2] illustrates a non-limiting example of a radar image acquired by a step radar image acquisition of the detection method of [Fig.l], said radar image comprising a plurality of points,

[0031] [Fig.3] illustrates a first graph which illustrates a phase variation of a signal periodic corresponding to the breathing or heartbeat of a living being and a second graph which illustrates noise in the absence of a living being, according to a non-limiting embodiment,

[0032] [Fig.4] illustrates a first graph which illustrates a power level of a signal pe periodic corresponding to the breathing of a child, and a second graph which illustrates a power level of a water bottle, according to a non-limiting embodiment,

[0033] [Fig.5] illustrates a spectrum of the first graph of [Fig.4], according to a mode of realization non-limiting use,

[0034] [Fig.6] illustrates a spectrum of the second graph of [Fig.4], according to a mode of realization non-limiting use,

[0035] [Fig.7] illustrates a diagram of a step in the construction of a tem database pores of said detection method of [Fig.l], according to a non-limiting embodiment,

[0036] [Fig.8] illustrates a system for detecting a living being configured to implement the detection method of [Fig.l], according to a non-limiting embodiment.

[0037] Identical elements, by structure or by function, appearing in different figures retain, unless otherwise specified, the same references.

[0038] The method 1 for detecting a living being o according to the invention is described with reference to Figures 1 to 7. In a non-limiting embodiment, the living being o is located in a passenger compartment 20 of a vehicle 2, namely inside the vehicle 2. This non-limiting embodiment is taken in the remainder of the description. In [Fig. 2], the living being o is located on a rear seat in a non-limiting example. Thus, the detection method 1 makes it possible in particular to detect a living being o in a vehicle 2. In a non-limiting embodiment, the vehicle 2 is a motor vehicle. By motor vehicle means any type of motorized vehicle. This embodiment is taken as a non-limiting example in the remainder of the description. In the remainder of the description, the vehicle 2 is thus otherwise called motor vehicle 2.

[0039] As illustrated in [Fig.l], the detection method 1 comprises the following steps.

[0040] In a step E1, a plurality of images I is acquired by a sensor 21 in a time window w. In a non-limiting embodiment, the sensor 21 is a radar sensor. The images I are therefore in this case radar images I. This non-limiting embodiment is taken as a non-limiting example in the remainder of the description. In a non-limiting embodiment, the time window w is approximately five seconds. It is less than 10 seconds according to the EuroNCAP specification for the automotive application. As illustrated in [Fig.2], each radar image I comprises a series of data p with which a position Po, a power Pu and a speed v are associated. The position Po is defined by the Cartesian parameters x, y and z or by the parameters of radius (or distance), azimuth and elevation.In a non-limiting embodiment, the data p are points. This non-limiting embodiment is taken as a non-limiting example in the remainder of the description. It will be noted that the acquisition (in milliseconds) of a single radar image is done in an acquisition window called a “frame” in English.

[0041] The plurality of radar images I is obtained from radar waves r' received by a radar wave transmitting / receiving entity 210 of the radar sensor 21 configured to transmit radar waves r (which are millimeter waves) and receive in return radar waves r' (which are millimeter waves) which have been reflected on a target object, such as a living being o.

[0042] From the radar waves r', a Fourier Transform FFT is applied by a processing unit (not shown) of the radar sensor 21 to calculate the reflection distances on the target object, for example according to a time of flight algorithm known to those skilled in the art ("Time Of Flight" in English). This process known to those skilled in the art is called "Range Processing" in English. It is not shown in the figures.

[0043] From the data obtained by the Fourier Transform, the processing unit of the radar sensor 21 calculates peaks of the reflection distances by removing the peaks which have not moved over the time window w. This process known to those skilled in the art is called “Clutter Removal” in English. It is not illustrated in the figures.

[0044] We thus obtain a radar image I with points p and the information on these points p, namely their position Po, their power Pu and / or their speed v.

[0045] In the context of the vehicle application, the radar images I are images acquired inside the motor vehicle 2, namely in its passenger compartment 20.

[0046] In a step E2, a first database DS of instantaneous data Fi is constructed by a processing unit 4 from the plurality of radar images I. The instantaneous data Fi (otherwise called instantaneous characteristics Fi) are based on the position Po, the power Pu and / or the speed v of each point p of each radar image I acquired in said time window w.

[0047] In non-limiting embodiments, the instantaneous data Fi are: - the average values ​​of the position data Po, power Pu and speed v of the points p of each radar image I, - the variances Va of the position data Po, power Pu and speed v of the points p of each radar image I. In addition to instantaneous data Fi, time data Ft are used as described below. The idea of ​​using time data Ft (otherwise called time characteristics Ft) comes from the fact that a radar image I of a target object, here a living being o, can show periodic patterns, in particular in the case of respiration or heart rate. The time data Ft are thus calculated from a periodic signal S. In non-limiting embodiments, the periodic signal S is thus respiration or heart rate (corresponding to heartbeats).

[0048] These periodic patterns can be detected in the phase Ph of the signal received by the radar sensor 21 as shown in drawing (a) of [Fig.3] which is compared to an empty environment (here the passenger compartment 20 of the motor vehicle 2) in drawing (b) of [Fig.3]. The periodic signal S illustrated in drawing (a) of [Fig.3] illustrates a periodic pattern of the phase Ph of a living being, here an adult, over a moving time window w'. In [Fig.3], on the abscissa is represented the signal emitted in an emission window called "Chirp" in English, the emission being done periodically and on the ordinate is represented the phase Ph in degrees (°). The emission "Chirps" (in microseconds) make it possible to construct the radar image I of each acquisition window. In an acquisition window, there are several emission "Chirps".

[0049] A same periodic pattern can be extracted from the variation of the power over time which is due to the periodic movement of the breathing or heart rate of a living being o. Indeed, the power changes periodically due to the back and forth movement of the breathing or heart rate of the living being o which is moving closer to and further away from the radar wave transmitting / receiving entity 210 of the radar sensor 21.

[0050] Drawing (a) of [Fig.4] illustrates a non-limiting example of a child in different positions inside the motor vehicle 2, the different positions posl, pos2, pos3, pos4, pos5, pos6 being illustrated by different shaded colors. The diagram is compared to a true negative diagram (drawing (b)) which uses a water bottle which is shaken in the motor vehicle 2, these different positions posl', pos2', pos3', pos4', pos5', pos6' are also illustrated by different shaded colors. The abscissa represents the maximum power Pk of the radar image I. The ordinate represents the number of radar images I. Drawing (a) illustrates the power variation of a periodic signal S on a moving window w' corresponding to the breathing or a heartbeat of a living being o which is here a child, at different positions here in the passenger compartment 20 of the motor vehicle 2. Drawing (b) thus illustrates the power variation of an inert object, here a bottle of water, at different positions here in the passenger compartment 20 of the motor vehicle 2. As can be seen, the power levels are similar in both cases, but the variation is different over time, namely the periodic pattern.

[0051] Thus, in a step E3, a second database DS' of temporal data Ft is constructed by the processing unit 4 from the plurality of radar images I. The temporal data Ft are calculated from a periodic signal S of each radar image I recorded over the moving time window w'. In a moving time window w', there will thus be a periodic signal S of several radar images I. It will be noted that the moving time window w' has the same duration as the time window w.

[0052] In a non-limiting embodiment, the moving time window w' is three radar images I per second. This makes it possible to have a variation of the temporal data Ft over time sufficient to be able to differentiate between cases of true positives (living being o) and cases of true negatives (inert object in motion).

[0053] It will be noted that step E3 can be carried out in parallel with step E2 or before step E2.

[0054] In a non-limiting embodiment, the time data Ft are a maximum power Pmax or a phase variation A of said periodic signal S.

[0055] As illustrated in [Fig.7], in a non-limiting embodiment, the construction of the second database DS' comprises: - in a sub-step E30, placing the time data Ft in a buffer memory B to obtain a plurality of time values ​​Vt, - in a sub-step E31, the removal of the continuous component DC of said time values ​​Vt, - in a sub-step E32, the application of a Fourier Transform FFT on said time values ​​Vt to obtain frequency values ​​Vf, - in a sub-step E34, from said frequency values ​​Vf, the extraction of a maximum value Max in a given frequency range I_Bth.

[0056] Thus, for each radar image I, the maximum power information Pmax or phase variation A is calculated according to a history using the moving time window w'.

[0057] In a non-limiting embodiment, the buffer memory B has a size for ten radar images I, which corresponds to approximately three seconds if the moving time window w' is three radar images I per second.

[0058] The application of the Fourier Transform FFT makes it possible to move from the time domain to the frequency domain.

[0059] Thus, the instantaneous signal (phase variation or maximum power) of each acquisition window ("frame") is stored for a certain number of frames in the moving time window w', and the periodicity is calculated by performing a frequency analysis of this stored signal.

[0060] It will be noted that the removal of the continuous DC component makes it possible to remove the average value of the periodic signal S and to keep only the sinusoidal signal. The process of removing a continuous DC component being known to those skilled in the art, it is not described here.

[0061] In non-limiting embodiments, the given frequency range I_Bth is a breathing frequency range or a heartbeat frequency range.

[0062] In other words, for each radar image I, we extract the maximum value Max that we find on all the harmonics of the given frequency range I_Bth. The maximum value Max is a power. The frequency of the maximum value Max represents the frequency and therefore the periodicity of the maximum power Pmax or of the phase variation A seen previously.

[0063] In an illustrated non-limiting embodiment, the construction of said second database DS' further comprises: - in a sub-step E33, the normalization Norm of said frequency values ​​Vf before said extraction of said maximum value Max. This makes the calculations simpler. The normalization process being known to those skilled in the art, it is not described here.

[0064] To start the execution of the ML machine learning model (described later), it is necessary to wait until the buffer memory B is filled with the time data Ft. Also, in a non-limiting embodiment, we have a normalized frequency value equal to 0.5. This value makes it possible to start on a neutral value which does not indicate whether we have a living or inert object. It should be noted that a normalized frequency value equal to 0 corresponds to an inert object (something which therefore does not breathe).

[0065] [Fig.5] illustrates a non-limiting example of normalization of the first graph (drawing (a)) of [Fig.4]. On the abscissa we find the frequency band of respiration in Bpm and on the ordinate the power Pk. [Fig.5] highlights the difference in the case of a child and a bottle of water whose normalization (of the second graph (drawing (b) of [Fig.4]) is illustrated in [Fig.6]. As can be seen, the highest value Max in the breathing frequency band I_Bth in the case of a child is about 1, while for a water bottle it is about 0.2.

[0066] In a step E4, the first database DS and the second database DS' are applied by the processing unit 4 as input to a machine learning model ML. Thus, there will be a first input with all the instantaneous data Fi (of all the points p of the radar image I) acquired in the time window w, and a second input with all the time data Ft which have been accumulated in the buffer memory B, namely all the time data Ft of all the periodic signals S of each radar image I acquired over the moving time window w'.

[0067] In a non-limiting embodiment, the ML machine learning model is a supervised model. In non-limiting alternative embodiments, the supervised model is a support vector machine or a random forest. Since the operation of these supervised models is known to those skilled in the art, they are not described here.

[0068] In a step E5, the ML machine learning model is executed.

[0069] In a step E6, the ML machine learning model establishes a classi classification Cl as to the presence / absence of a living being o. It will be noted that the periodic signal S is in particular representative of a type of living being o. In non-limiting embodiments, the type of living being o includes that of a child, that of an adult, or that of an animal. Consequently, the temporal data Ft which are calculated from the periodic signal S and which are used as input to the machine learning model ML are also representative of the type of living being o. Thus, at the output of the machine learning model ML, in a non-limiting embodiment, there will be a classification Cl associated with the radar images I taken inside the motor vehicle 2 which will indicate whether they correspond to a living being o or not and if so to what type (adult, child, animal).

[0070] It will be noted that the machine learning model ML will have been previously trained with at least one instantaneous database Fi (different from the first database DS) of true positive cases and true negative cases and labels associated with the associated true positive cases and true negative cases, and a temporal database Ft (different from the second database DS') of true positive cases and true negative cases and labels associated with the associated true positive cases and true negative cases.

[0071] Thus, the temporal data Ft make it possible to differentiate a living being o from an inert object but also the type of living being o (adult, child, animal) because the periodic patterns (of breathing or heart rate) represented by the signals periodic S are different for these different types of living beings o.

[0072] In a step E7, depending on the classification C1, the processing unit 4 transmits corresponding information D (absence / presence of a living being o) to a management unit ECU. In a non-limiting embodiment, the management unit ECU is an electronic control unit 2 of the motor vehicle 2.

[0073] Depending on decision D, in a non-limiting embodiment, the ECU management unit can trigger an alarm to warn that there is a living being o in the motor vehicle 2.

[0074] Thus, the method 1 for detecting a living being o is implemented by a detection system Sys of a living being o illustrated in [Fig.8].

[0075] The detection system Sys comprises a radar sensor 21 configured to acquire a plurality of radar images I in a time window w, each radar image I comprising a series of points p with which a position Po, a power Pu and a speed v are associated (illustrated function f21(21,1, (p(Po, Pu, v), w)).

[0076] The radar sensor 21 comprises a radar wave transmission / reception entity 210. In non-limiting embodiments, the radar wave transmission / reception entity 210 comprises, as illustrated in [Fig.8]: - two 2100 transmitting antennas configured to transmit radar waves r, - three receiving antennas 2110 configured to receive in return radar waves r' which have been reflected on a target object, such as a living being o.

[0077] The detection system Sys further comprises a processing unit 4 configured to: - constructing a first database DS of instantaneous data Fi from the plurality of radar images I, said instantaneous data Fi being based on said position Po, said power Pu and said speed v of each point p of each radar image I acquired in said time window w (illustrated function f41(4, DS, Fi(I)), - constructing a second database DS' of temporal data Ft from the plurality of radar images I, the temporal data Ft being calculated from a periodic signal S of each radar image I taken over a moving time window w' (illustrated function f42(4, DS', Ft(S)), - apply this first database DS and this second database DS' as input to an ML machine learning model (illustrated function f43(4, DS, DS', ML)), - execute said ML machine learning model (illustrated function f44(4, ML)).

[0078] The processing unit 4 is further configured to, depending on said classification C1, transmit corresponding information D to a management unit ECU (illustrated function f45(4, ECU, D(C1))).

[0079] The detection system Sys further comprises said ML machine learning model which is configured to: - receive as input the data Fi, Ft ​​respectively from the first database DS and the second database (illustrated function f51(ML, DS, DS')), - be executed (illustrated function f52(ML)), - establish a classification Cl as to the presence / absence of a living being o (illustrated function f53(ML, Cl, S(o)).

[0080] Of course, the description of the invention is not limited to the embodiments described above and to the field described above. Thus, in a non-limiting embodiment, the detection method 1 can be applied for monitoring the breathing or heart rate of elderly people in hospitals. In this case, the management unit ECU is an electronic control unit of a medical monitoring device. And, in this case, depending on the corresponding information D transmitted by the processing unit 4, in a non-limiting embodiment, the management unit ECU can trigger an alarm to warn that the living being o is no longer breathing or that its heart is no longer beating. Thus, in another non-limiting embodiment, the radar image I is a thermal map. Thus, in another non-limiting embodiment, the sensor 21 is an ultrasonic sensor.

[0081] Thus, the invention described presents in particular the following advantages: - it allows for more reliable and more precise detection of living beings, - it allows to identify the type of living being o, - the combination of instantaneous data Fi with temporal data Ft makes it possible to improve the detection of true positive cases and to reduce the cases of true negatives and thus to have a more robust differentiation between a living being o and an inert object; the detection method 1 is thus more robust. Thus, the detection rate and robustness with respect to false alarms are improved.

Claims

Claims

1. Method for detecting (1) a living being (o), characterized in that said detection method (1) comprises: - the acquisition of a plurality of images (I) in a time window (w), each image (I) comprising a series of data (p) with which a position (Po), a power (Pu) and a speed (v) are associated, - the construction of a first database (DS) of instantaneous data (Fi) from the plurality of images (I), said instantaneous data (Fi) being based on said position (Po), said power (Pu) and said speed (v) of each data (p) of each image (I) acquired in said time window (w), - the construction of a second database (DS') of temporal data (Ft) from said plurality of images (1), said temporal data (Ft) being calculated from a periodic signal (S) of each image (I) recorded over a moving time window (w'),- applying said first database (DS) and said second database (DS') as input to a machine learning (ML) model, - executing said machine learning (ML) model, - establishing a classification (Cl) by said machine learning (ML) model as to the presence / absence of a living being (o), - based on said classification (Cl), transmitting corresponding information (D) to a management unit (ECU).,

2. Detection method (1) according to claim 1, according to which the instantaneous data (Fi) are: - average values ​​of said position (Po) and / or of said power (Pu) and / or of said speed (vO) of said data (p) of each image (I), or - variances of said data (p) of each image (I).

3. A detection method (1) according to claim 1 or claim 2, wherein said machine learning (ML) model is a supervised model.

4. Detection method (1) according to the preceding claim, wherein said supervised model is a support vector machine or a random forest.

5. A detection method (1) according to any one of the preceding claims preceding, according to which said periodic signal (S) is a breath or a heartbeat.

6. Detection method (1) according to any one of the preceding claims, according to which said living being (o) is of the adult, child, or animal type.

7. A detection method (1) according to any preceding claim, wherein said time window (w) is approximately five seconds.

8. Detection method (1) according to any one of the preceding claims, according to which the time data (Ft) are a maximum power (Pmax) or a phase variation (A) of said periodic signal (S).

9. Detection method (1) according to any one of the preceding claims, according to which the construction of said second database (DS') comprises: - placing in a buffer memory (B) the time data (Ft) to obtain a plurality of time values ​​(Vt), - removing the continuous component (DC) from said time values ​​(Vt), - applying a Fourier Transform (FFT) on said time values ​​(Vt) to obtain frequency values ​​(Vf), - from said frequency values ​​(Vf), extracting a maximum value (Max) in a given frequency range (I_Bth).

10. Detection method (1) according to the preceding claim, wherein said given frequency range (I_Bth) is a breathing frequency range or a heartbeat frequency range.

11. Detection method (1) according to claim 9 or claim 10, according to which the construction of said second database (DS') further comprises: - the normalization of said frequency values ​​(Vf) before said extraction of said maximum value (Max).

12. Detection method (1) according to any one of the preceding claims, according to which said plurality of images (I) is obtained from waves received (r) by a wave transmission / reception entity (210) of a sensor (21).

13. A detection method (1) according to any preceding claim, wherein said moving time window (w') is three images (I) per second.

14. A detection method (1) according to any preceding claim, wherein said detection method (1) is a method of detecting a living being (o) inside a vehicle (2).

15. Detection system (Sys) of a living being (o), characterized in that said detection system (Sys) comprises: - (a) a sensor (21) configured to acquire a plurality of images (I) in a time window (w), each image (I) comprising a series of data (p) to which a position (Po), a power (Pu) and a speed (v) are associated, - (b) a processing unit (4) configured to: - build a first database (DS) of instantaneous data (Fi) from the plurality of images (I), said instantaneous data (Fi) being based on said position (Po), said power (Pu) and said speed (v) of each data (p) of each image (I) acquired in said time window (w), - build a second database (DS') of temporal data (Ft) from said plurality of images (I),said temporal data (Ft) being calculated from a periodic signal (S) of each image (I) recorded over a moving time window (w'), - applying said first database (DS) and said database (DS') as input to a machine learning model (ML), - executing said machine learning model (ML), - (c) said machine learning model (ML) configured to be executed and establish a classification (Cl) as to the presence / absence of a living being (o), - said processing unit (4) being further configured to, as a function of said classification (Cl), transmit corresponding information (D) to a management unit (ECU).,

Images