Method for identifying or authenticating a motor vehicle occupant

Abstract

The invention relates to a method for identifying and / or authenticating an occupant (20) of a motor vehicle (10) using a computer, comprising the following steps: - acquiring data and establishing at least two biometric fingerprints of said occupant based on said data, said at least two biometric fingerprints being of different types, - comparing each of the at least two biometric fingerprints with an associated reference fingerprint in order to deduce at least two scores (SCOvocal, SCOfacial), - merging the scores, and - identifying and / or authenticating said occupant based on the result of the merging. According to the invention, a step is provided for determining at least one noise parameter related to the data acquisition noise, and the merging of the scores is performed based on said noise parameter. Figure for the abstract: Fig. 2

Description

Title of the invention: Method for identifying or authenticating an occupant of a motor vehicle Technical field of the invention

[0001] The present invention relates generally to the field of identification of persons based on biometric data.

[0002] It relates more particularly to a method for identifying and / or authenticating an occupant of a motor vehicle by a computer, in which the following steps are provided: - data acquisition and establishment of at least two biometric prints of said occupant based on said data, said at least two biometric prints being of different types, - comparison of one and the other of the at least two biometric prints with a reference print associated with them in order to deduce scores, - score merging, and - identification and / or authentication of said occupant depending on the result of the merger.

[0003] It also relates to a motor vehicle comprising the technical means necessary for implementing this method. State of the art

[0004] A current area of ​​research in the automobile industry consists of automatically adapting the different functions offered by a vehicle to its user.

[0005] In this context, it is necessary to be able to identify or authenticate any person, for example in order to adjust the vehicle appropriately or to ensure that this person is authorized or not to use particular functions of the vehicle. For example, it is desired to be able to ensure that the passenger who orders the total opening of his window is authorized to do so. It is also desired to be able to offer personalized functions to the vehicle's passengers (automatic seat adjustment, audio settings, contactless payment, etc.).

[0006] Voice biometrics, which is a voice identification technology, can then be used. It relies on algorithms capable of identifying voice characteristics specific to each person, in order to construct a voice print specific to each person.

[0007] Facial biometrics is another identification technology (by processing images of a person's face) that can be used. It similarly relies on algorithms capable of identifying specific characteristics of each face, in order to construct a facial imprint specific to each person.

[0008] When a passenger wishes to identify himself, an image of his face and a voice signal are acquired to create his voice and facial prints. These two prints can then each be compared to reference prints stored in a database and each associated with a registered user, in order to deduce scores. The score obtained by comparison between the facial print and the reference print considered then makes it possible to conclude whether or not the user is recognized. In the same way, the score obtained by comparison between the voice print and the reference print considered then makes it possible to conclude whether or not the user is recognized.

[0009] One idea is not to use these biometrics separately but to merge the scores obtained via the two biometric methods in order to deduce a new score allowing a more reliable conclusion as to whether or not the user is recognized.

[0010] Unfortunately, this method is not infallible. Thus, it often happens that it does not allow a person to be identified or that it identifies a person incorrectly.

[0011] Indeed, this type of technology, once installed in a vehicle, is faced with difficulties linked to the context in which the user's voice or face is recorded (presence of noise, shadowy areas, sudden overexposure, etc.).

[0012] Furthermore, this type of technology turns out to be very consuming in practice in terms of computing power, so much so that it is often expensive to implement. Presentation of the invention

[0013] In order to overcome the aforementioned drawbacks of the state of the art, the present invention proposes to merge in a simple but efficient manner the scores obtained by two biometric technologies, so that the context of the data recording (face, voice, etc.) has little influence on the identification result.

[0014] More particularly, according to the invention, an identification method is proposed as defined in the introduction, in which a step of determining at least one noise parameter relating to the data acquisition noise is provided, and in which the merging of the scores is carried out as a function of said noise parameter.

[0015] Thus, thanks to the invention, the acquisition noise is taken into account within the score fusion step, so that this fusion provides reliable results regardless of the type of noise and the noise level.

[0016] Other advantageous and non-limiting characteristics of the method according to the invention, taken individually or in all technically possible combinations, are the following: - the fusion of the scores is carried out according to a method which is selected according to said noise parameter; - said method is selected from a linear combination fusion method and a power product fusion method; - said method is selected by choosing a function or mapping based on said noise parameter, calculating the value of a balance point between false acceptance rate and false rejection rate given by each method, using the chosen function or mapping, and selecting the method for which the calculated value is the lowest; - the merging of the scores is carried out using a mathematical equation which uses a variable determined as a function of said at least one noise parameter; - said function or mapping makes it possible to calculate the value of the equilibrium point as a function of said variable; - one of the acquired data is a voice signal emitted by said occupant and one of the two biometric prints is of the voice type; - said at least one noise parameter relates to the sound intensity of the ambient noise at the time of the acquisition step; - said at least one noise parameter relates to a type of noise detected at the time of the acquisition step among several predefined types of noise.

[0017] The invention also proposes a motor vehicle comprising at least one seat for accommodating an occupant, data measurement means making it possible to establish two biometric prints of distinct types, and a computer adapted to implement an identification method as mentioned above.

[0018] Of course, the various features, variants and embodiments of the invention may be combined with each other in various combinations to the extent that they are not incompatible or mutually exclusive. Detailed description of the invention

[0019] The description which follows with reference to the appended drawings, given as non-limiting examples, will make it clear what the invention consists of and how it can be implemented.

[0020] In the attached drawings:

[0021] [Fig-1] represents several graphs which illustrate the variations of different pa meters used in the context of the invention;

[0022] [Fig.2] is a schematic representation of the different stages of a process identification in accordance with the invention.

[0023] In [Fig. 2], a part of a motor vehicle 10 and a user of this motor vehicle are shown, namely here the driver 20.

[0024] This could be any type of vehicle, for example a truck, bus, train or even an airplane.

[0025] This is a car which typically has a chassis which delimits a passenger compartment.

[0026] In this passenger compartment, the vehicle includes various equipment making it possible to identify at least the driver 20, or even each of the occupants of the vehicle. In the remainder of this description, for reasons of clarity, we will only focus on the identification of the driver 20.

[0027] These devices are designed to enable identification to be carried out using at least two different biometric identification methods. These will be facial and voice identification, but other methods could be used in addition or as a variant (iris recognition, fingerprints, hand veining, etc.).

[0028] One of these devices is a microphone 11 placed and oriented in such a way that it is able to record the voice of the driver 20.

[0029] Another of these devices is an image sensor, and more precisely here a camera 12 placed and oriented in such a way that it is able to capture images of the face of the driver 20 when the latter is driving the vehicle.

[0030] The motor vehicle 10 further comprises a computer (not shown) designed to implement a driver identification method, as will be described below with reference to [Fig.2].

[0031] This calculator comprises a processor, a memory and a data exchange interface, connected for example to a CAN type vehicle network.

[0032] Thanks to this interface, the computer is adapted to receive different information, and in particular the signals acquired (and possibly pre-processed) by the microphone 11 and the camera 12 (see [Fig.2]).

[0033] Thanks to its memory, the computer stores a computer application, consisting of computer programs comprising instructions whose execution by the processor allows the computer to implement the identification method described below.

[0034] The computer also stores a database 13 of reference biometric fingerprints. This database 13 records, for each user of the vehicle who has registered with this identification service (this is referred to as “enrollment”), at least one reference voice print and at least one reference facial print.

[0035] These fingerprints are characteristics that allow a person to be reliably identified.

[0036] The reference voice print is constructed here during the enrollment phase, which consists of asking the driver to pronounce (and possibly repeat several times) a key phrase, then extracting information from these recordings to build the reference voiceprint. This enrollment phase is carried out only once, for example when the driver drives the vehicle for the first time. The key phrase is either imposed by the vehicle manufacturer (for example "Hello Renault"), or chosen by the driver (for example a secret code made up of numbers).

[0037] The process of constructing a voice print can be briefly described here as follows.

[0038] This process firstly involves an acoustic analysis which consists of extracting relevant and characteristic information from the recorded voice signal. To this end, sets of acoustic coefficients are calculated at regular time intervals (i.e. for successive observation windows), on signal blocks of fixed lengths. These sets of coefficients together constitute an acoustic matrix which forms a digital signature of the driver's voice.

[0039] Each set of coefficients is for example calculated using inverse Fourier transforms of the logarithm of the signal power spectral density. The cepstral coefficients resulting from such an analysis indeed characterize the shape of the spectrum well. The cepstral coefficients used are typically the MFCC (from the English "Mel Frequency Cepstral Coefficients"). They have the advantage of being little correlated with each other.

[0040] The process can also be supplemented by “mel-filterbank” type filtering, which allows the richness of voiced sounds to be emphasized.

[0041] The acoustic fingerprint generation process thus makes it possible to generate, based on the recorded voice signal, a basic voice signature characteristic of the driver's voice.

[0042] The reference facial print is constructed in a similar manner during the enrollment phase.

[0043] These two reference fingerprints are therefore recorded in a database 13 in such a way that each record in this database is associated with a particular user and includes an identifier of this user, the two reference fingerprints and rights and preferences associated with this user (right to open windows, musical preference, etc.).

[0044] The computer then stores an algorithm which makes it possible to identify the driver 20.

[0045] This algorithm will be described in detail later in this presentation. It can only be specified here that, when the already enrolled driver wishes to identify himself, he must pronounce the key phrase and place himself in front of the camera. From then on, the computer can generate “temporary fingerprints” from the recordings made of it, according to a process identical to that used to construct the reference fingerprints. These temporary prints are then compared with each of the reference prints recorded in the database 13. The calculator then determines a SCOvocai, SCOfaciai score for each reference print, which is higher the closer the temporary print is to this reference print.

[0046] If one of the SCOvocai, SCOfaciai scores is higher than a SCOmin threshold, it is possible that the user has been identified (the threshold used here will be considered the same in both cases since we will consider that the scores have been normalized).

[0047] In the context of the invention, the SCOvocai score resulting from the comparison between the temporary print and the reference print used in the context of voice recognition is merged with the SCOfaciai score resulting from the comparison between the temporary print and the reference print used in the context of facial recognition.

[0048] Thus, it will be the result of this fusion (hereinafter called the merged score SCOf) which will be compared with the minimum threshold SCOf>min in order to determine whether the driver is identified or not.

[0049] Before describing in more detail how this identification method is implemented with reference to [Fig.2], we can explain the calculations on which this method is based. These calculations will indeed be useful for understanding how to configure the computer so that it can implement this identification method.

[0050] The use of the SCOvocai, SCOfaciai score from facial or voice recognition and its comparison with a SCOmin threshold gives results that are not always reliable. It is indeed possible to find oneself in the following four situations: - the pre-enrolled user is correctly identified (we speak of good acceptance “BA”), - the pre-enrolled user is not identified (we speak of false rejection “FR”), - the non-enrolled user is wrongly identified (we speak of false acceptance “FA”), - the non-enrolled user is rightly not identified (we speak of good rejection “BR”).

[0051] To illustrate these situations, it is possible to carry out a “battery of tests” during which a number N of speakers placed in the driver’s seat of a test vehicle (only half of which are already enrolled) are asked to try to be identified by the computer via each of the two voice and facial recognition processes. Here, this battery of tests is carried out with no noise in the passenger compartment (engine off, vehicle stationary, etc.) and with average brightness.

[0052] The first graph G1 illustrated in [Fig.l] represents two histograms which illustrate the distribution (on the ordinate) of the scores (on the abscissa) using the facial recognition process. Thus, during this battery of tests, the scores are distributed in two Gaussian curves centered on scores of 0.5 and 1.

[0053] The leftmost Gaussian curve is an illustration of the BRR rate of good rejections BR and the right one is an illustration of the BAR rate of good acceptance BA, in using the facial recognition process.

[0054] The second graph G2 represents the same parameters, obtained however in the context of voice recognition.

[0055] It is understood that the ideal would be, on these graphs, for the two Gaussian curves to be disjoint, which would mean that the methods used make it possible to avoid any false rejection and any false acceptance. But this is not the case. The overlap zone between the curves is in fact an area illustrating that such false rejections and false acceptances exist.

[0056] We also understand from these graphs Gl, G2 that the higher the SCOmin threshold, the fewer false acceptances FA there will be but the more false rejections FR there will be, and vice versa.

[0057] At this stage, several useful notions can be defined within the framework of the method described below.

[0058] The false rejection rate FR will be considered equal to the number of false rejections FR over the total number of tests carried out during this battery of tests with pre-enrolled (or “legitimate”) speakers.

[0059] The FAR rate of false FA acceptances will be considered equal to the number of false FA acceptances over the total number of tests carried out with non-enrolled users (or “imposters”).

[0060] The EER (Equal Error Rate) equilibrium point will correspond to the point where the FRR and FAR rates are equal. This equilibrium point is preferred since it avoids favoring false acceptances to the detriment of false rejections, and vice versa.

[0061] It is understood that by varying the SCOmin threshold used, the values ​​of the FAR and FRR rates will vary. We can then construct a curve to illustrate this variation.

[0062] Here, the curve chosen will be a so-called DET curve (from the English “Detection Error Tradeoff”) which will allow the FRR rate to be written as a function of the FAR rate (i.e. FRR = f(FAR)).

[0063] Alternatively, one could use a so-called ROC curve (from the English “Receiver Operating Characteristic”).

[0064] On graph G3, two DET type curves are then illustrated, referenced DET1, DET2 and corresponding respectively to the two graphs Gl, G2.

[0065] It can be seen in this graph G3 that the DET1 curve corresponding to the facial recognition process is closer to the origin than the DET2 curve corresponding to the voice recognition process. Indeed, the Gaussian curves represented on the graph G1 have a smaller overlap area than those represented on the graph G2, which means that the facial recognition process provides better results (in other words, the facial recognition process is more reliable than the voice recognition process).

[0066] The results of these two recognition processes, i.e. their SCO facial and SCOvocai scores, are merged here. Thus, the two recognition processes can be combined so that identification is done on the basis of these two processes.

[0067] The fusion will allow these scores to be combined in order to obtain a single one (the SCOf merged score). This merged score can then be compared with a threshold in order to determine more reliably whether an individual is identified or not.

[0068] Preferably, the score fusion method may not be the same depending on the recording conditions of the user's voice and face. Thus, the most effective method can be chosen.

[0069] To understand how this choice is made and how the chosen method is then configured, we can continue to work on the results of the battery of tests carried out and illustrated on graphs G1 and G2.

[0070] The first fusion method that may be chosen is, preferably, fusion by linear combination.

[0071] This type of fusion is expressed in the form of the following mathematical equation:

[0072] [Math.l] SC O fl — OS CO facialj +

[0073] In this equation: - i is an integer between 1 and N, which therefore makes it possible to identify each corresponding facial and vocal test, - SCOfi is the score resulting from the fusion, also called fusion statistic, - a is a weight expressed as a number between 0 and 1, inclusive, - SCOfaciaij is the ith of the N normalized scores obtained via the facial recognition system, and - SCOvocai,i is the ith of the N normalized scores obtained via the voice recognition system.

[0074] We therefore understand that the fusion consists of combining the scores obtained with the two facial and voice recognition technologies, giving a more or less significant weight to each score.

[0075] The result of the merger would again make it possible to establish a histogram of the type illustrated on graphs G1 and G2, on which the overlap of the Gaussian curves would be minimized.

[0076] It is therefore also possible to plot a DET type curve, noted DETs on graph G3, illustrating the variation of the FRR rate of false rejections as a function of the FAR rate of false acceptances obtained by merging the two recognition technologies. facial and vocal birth. We note that this DETs curve is closer to the origin than the DET1 and DET2 curves because this fusion method is more reliable than each of the two facial and vocal recognition technologies considered in isolation.

[0077] On this same graph G3, we have represented the line D^ illustrating the EER equilibrium points, which we recall correspond to the points where the FRR rate of false rejection FR is equal to the FAR rate of false acceptances FA.

[0078] It should be noted that by varying the value of the weight a, it is possible to modify the DETs curve.

[0079] We recall here that the closer a point on this DETs curve is to the origin, the better the result of the fusion will be. The idea then consists of optimizing the fusion by looking for the weight a which minimizes the DETs curve at the level of the DEER line.

[0080] On graph G4, the value of the point of intersection between the DETs curve and the DEER line is illustrated (this value being hereinafter called “equilibrium rate EERfl”), by varying the weight a between 0 and 1. It can be seen that in this example, the weight giving the equilibrium rate EERfl a minimum value EERflmin is 0.62.

[0081] It can therefore be considered that under the conditions considered for recording voices and faces, fusion by linear combination gives optimal results for this optimal weight value aopt.

[0082] As explained above, the battery of tests which led to the plotting of the histograms of graphs G1 and G2 is carried out in the passenger compartment of the test vehicle, in the absence of external noise.

[0083] The idea here will consist of repeating the previous operations while varying the external noises, in order to find the optimal weight values ​​aopt in varied noise conditions.

[0084] More precisely, according to an advantageous characteristic of the invention, the aforementioned process is repeated using the results of at least one other “battery of tests” carried out in the presence of a given noise in the passenger compartment.

[0085] As an example, the noise used here is that generated by the powertrain in the passenger compartment of the test vehicle.

[0086] In practice, the process is repeated for several noise levels (for example with different powertrain speeds and / or by varying the intensity of the speaker's voice).

[0087] To this end, we can introduce the concept of “SNR ratio” corresponding to the ratio between the acoustic power of the voice signal used for voice biometrics and the acoustic power of the noise in the passenger compartment (or the ratio of their respective sound intensities).

[0088] So, for each SNR ratio considered, it is possible to carry out a new battery of tests, then to plot a curve of the type represented on graph G4 in order to find a new optimal weight aopt, which is illustrated on graph G5.

[0089] On this graph G5, one of the optimal weights corresponds to the case where there is no noise, another corresponds to an SNR ratio of OdB, another to an SNR ratio of 5dB and a last one to an SNR ratio of 10dB. Of course, the greater the number of test batteries, the greater the number of optimal weights aopt found will be, and the more reliable the result of the process will be.

[0090] It is therefore appropriate to multiply the test batteries with different SNR ratios, the number of test batteries being at least equal to 4 and preferably greater than 8.

[0091] Here, the idea is to carry out the test batteries for SNR ratios between -20 and +40 dB, for example by varying this ratio by 5 dB between two test batteries.

[0092] On graph G6, we then illustrated the values ​​of the optimal weights aopt found as a function of the tested SNR ratios. An interpolation process then makes it possible to find the function allowing us to write: aopt=fl(SNR).

[0093] In other words, thanks to this graph, it will be possible to find, based on the SNR ratio recorded at the time of recording a voice identification signal, the corresponding optimal aopt weight.

[0094] Using graph G6, it is also possible to plot the minimum values ​​EERfmin of the corresponding EERf equilibrium rates, as a function of the optimal weight values ​​aopt (graph G7). An interpolation process then makes it possible to find the function allowing us to write: EERflmin=flEER(aopt).

[0095] The entire above-mentioned process can be repeated for other types of noise qk. For example, one can then consider the noise caused by the movement of the vehicle (aerodynamic and road noise), then the noise caused by the opening of the driver's window, then the noise generated by the music emitted by the vehicle's multimedia system, then the noise generated by the voices of passengers in the vehicle.

[0096] It is thus possible to multiply the batteries of tests in order to plot curves such as those illustrated on graphs G6 and G7 for each of these other types of noise qk. We will preferably consider at least three different types of noise. Here, the number k of types of noise considered will be equal to five.

[0097] More generally, we can therefore find the functions allowing us to write, for each type of noise k:

[0098] aoptjk=flk(SNR)

[0099] EERflmin>k=flEERjk(aoptjk)

[0100] We then understand that thanks to the set of graphs G6, G7 plotted, it will be possible to find, depending on the type of noise and the SNR ratio recorded at the time of recording the identification voice signal, the corresponding optimal weight aopt (using graphs G6). It will also be possible to find the minimum value of the corresponding balance point (using graphs G7).

[0101] The second fusion method that may be chosen is, preferably, fusion by product of weighted powers.

[0102] This type of fusion is expressed in the form of the following mathematical equation:

[0103] [Math.2] SCO, 2 = SCOf^ / SCO^, *

[0104] In this equation: - SCOf2 is the fusion score, and - [3 is a weight expressed as a number between 0 and 1, inclusive.

[0105] The same process is applied for this fusion method as for the first method, which makes it possible to find curves of the type illustrated on graphs G6 and G7, for the different types of noise mentioned.

[0106] We can therefore find the functions allowing us to write, for each type of noise k:

[0107] |3optjk=f2k(SNR)

[0108] EERf2min>k=f2EERjk([3optjk)

[0109] As explained above, the fusion will provide results that are all the more reliable as the minimum value EERflmink, EERf2min>k of the equilibrium rate is low.

[0110] Therefore, when a driver seeks to identify himself, we can choose the best of the two fusion methods by comparing the minimum values ​​EERflmink, EERf2 min,k*

[0111] Thus, if the value of the EERflmink function is lower than that of the EERf2 min>k function, we will select the fusion by linear combination. Otherwise, we will select the other.

[0112] We can also calculate the optimal weight aopt, [3opt corresponding, so that the chosen fusion method presents the best possible performances, taking into account the type of noise and the SNR level.

[0113] It will be noted here that the batteries of tests, carried out with the different noise levels and for the different types of noise, are carried out offline, that is to say at the design stage of the vehicle range, on a test vehicle substantially identical to those which will subsequently be marketed and equipped with the computer programmed to carry out the driver identifications.

[0114] The calculations to determine the curves of the aforementioned functions aoptjk, EERflmin>k, [3optjk, EERf2 min>k are also carried out offline, so that the computing power required necessary to implement them has no impact on the final price of the vehicle.

[0115] In fact, the vehicle's computer will not have to redetermine these functions, but will be able to simply use them to identify the driver of the vehicle.

[0116] For this, the equations of the functions aopt>k, EERflmink, [3optjk, EERf2min>k (or maps resulting from these functions) can be recorded in the memory of the computer of the motor vehicle 10, which is of the same range as the test vehicle used offline to establish these functions.

[0117] We can then describe the process that will be implemented by the vehicle's computer. As we will see, this process will only involve simple calculations, so that the computing power of the computer will not have to be very high, to the benefit of its cost price.

[0118] This driver identification method comprises several steps illustrated in [Fig.2],

[0119] A background task is to continuously record sound signals in the vehicle interior using the microphone 11.

[0120] The first step E0 then consists of detecting at each instant whether the driver 20 pronounces the key phrase (the same as that used during the enrollment phase) and then recording either the voice signal corresponding to the pronunciation of this key phrase, or data characterizing this signal.

[0121] Here, in practice, a fingerprint of this signal is generated during a step E2 using the same algorithm as that used during the enrollment phase to create the reference voice fingerprint. This so-called temporary fingerprint is then stored in a register of the computer's memory.

[0122] In parallel, another step E30, triggered by the first, consists of taking a photo or a short video sequence of the face of the driver 20 then recording in the memory of the computer either this image or this sequence, or data from this image or this sequence.

[0123] Here, in practice, a biometric fingerprint of the driver's face is generated during a step E32 using the same algorithm as that used during the enrollment phase to create the reference facial fingerprint. This so-called temporary fingerprint is then stored in a register of the computer's memory.

[0124] During steps E4 and E34, the computer proceeds in a conventional manner to an attempt to identify the driver by voice biometrics and by facial biometrics. For this, the pair of temporary facial and voice prints is compared with each pair of reference voice and facial prints stored in the database 13.

[0125] During each comparison, the calculator calculates two SCOvocai, SCOfaciai scores allowing the evaluation of the similarities between the two temporary prints and the two selected reference voice and face prints (steps E6 and E36).

[0126] Before merging this pair of SCOvocai, SCOfaciai scores (step E40), it is planned here to search for the merging algorithm which will potentially be the most effective taking into account the recording conditions of the key phrase, then to parameterize this merging algorithm in an optimal manner.

[0127] For this, during a step E8, it is planned to determine the type of noise qk recorded by the microphone at the time of acquisition of the key phrase (or immediately before or after).

[0128] In practice, a noise classification process is applied to the voice signal captured during the utterance of the key phrase. Many technologies are known, so this step will not be described in detail here. It can only be explained that, in general, several types of noise will be detected but only the predominant one will be selected.

[0129] Then, during step E10, the calculator determines the value of the SNR ratio, that is to say the value of the power ratio between the voice signal and the background noise. Here again, many technologies are known, so this step will not be described in detail here.

[0130] Therefore, during steps El2, El4, E16 and El8, based on the functions or maps stored in its memory, the computer can determine the values ​​of the functions aoptjk, EERflmin>k, [3optjk, EERf2min>k.

[0131] To do this, it begins by selecting the functions or maps corresponding to the type of noise detected. Then, taking into account the SNR ratio, the calculator calculates the values ​​of the weights aopt>k, [3optjk, based on the function or map corresponding to graph G6 of [Fig.l]. It then deduces the values ​​of the functions EERflmink, EERf2min>k, based on the function or map corresponding to graph G7 of [Fig.l].

[0132] During step E20, the calculator compares the values ​​of the functions EERflmink and EERf2mink.

[0133] The idea here is to select the fusion algorithm that will give the best results given the type of noise and the SNR ratio, by looking for the one for which the minimum value of the equilibrium rate is closest to zero.

[0134] Then, during step E40, if the minimum value EERflmink is lower than that EERf2min>k, the scores are merged by linear combination (E40a). Otherwise, the scores are merged by power product (E40b).

[0135] Whatever method is used, the merger will be done taking into account the optimal value of the weight aoptjk, [3optjk.

[0136] The result of the fusion will be a fusion score which can then be compared to a threshold in order to determine whether or not the driver could be identified.

[0137] The present invention is in no way limited to the embodiment described and shown, but those skilled in the art will be able to provide any variant in accordance with the invention.

[0138] Typically, it would be possible to apply the same type of method as that described above to the facial recognition process. Thus, one could carry out batteries of tests with different brightness levels and for distinct disturbances. Indeed, facial recognition is used in a restrictive capture context where successive shadow zones, sudden overexposure or underexposure may appear when entering and then exiting a tunnel, directly impacting the quality of the video capture.

[0139] It should also be noted that if the aforementioned process is used to identify a driver, it can also be used to authenticate him. Indeed, the process used will make it possible, by adjusting the threshold used, to ensure that the user is correctly identified and that the latter is not an imposter trying to hack the system.

[0140] As a further variant, it has been specified here that we always seek to position ourselves on the DEER line so that the rates of false rejections and false acceptances are identical. Of course, as a variant, we could have positioned ourselves on another line in order to favor the reduction of false rejections (to the detriment of false acceptances) or towards a.

Claims

Claims

1. Method for identifying and / or authenticating an occupant (20) of a motor vehicle (10) by a computer, in which steps are provided: - acquiring data and establishing at least two biometric fingerprints of said occupant (20) as a function of said data, said at least two biometric fingerprints being of different types, - determining at least one noise parameter (qk, SNR) relating to the data acquisition noise, - comparing each of the at least two biometric fingerprints with an associated reference fingerprint in order to deduce at least two scores (SCOvocai, SCOfaciai), - merging the scores (SCOvocai, SCOfaciai) as a function of said noise parameter (qk, SNR), and - identifying and / or authenticating said occupant (20) as a function of the result of the fusion, characterized in that: - the fusion of the scores is carried out according to a method which is selected according to said noise parameter (qk,SNR), - said method is selected by: H choosing a function or a mapping as a function of said noise parameter (qk, SNR), H calculating the value (EERf) of a point of equilibrium between false acceptance rate (FAR) (FA) and false rejection rate (FRR) (FR) given by each method, using the chosen function or mapping, H selecting the method for which the calculated value is the lowest, - the fusion of the scores (SCOvocai, SCOfaciai) is carried out using a mathematical equation which uses a variable (a, |3) determined as a function of said at least one noise parameter (qk, SNR), - said function or mapping makes it possible to calculate the value (EERf) of the point of equilibrium as a function of said variable (a, |3).,

2. The method of claim 1, wherein said method is selected from a linear combination fusion method and a power product fusion method.

3. Method according to one of claims 1 and 2, in which one of the acquired data is a voice signal emitted by said occupant (20) and one of the two biometric prints is voice type.

4. Method according to claim 3, wherein said at least one noise parameter (SNR) relates to the sound intensity of the ambient noise at the time of the acquisition step.

5. Method according to one of claims 3 and 4, in which said at least one noise parameter (qk) relates to a type of noise detected at the time of the acquisition step among several predefined types of noise.

6. Motor vehicle (10) comprising at least one seat for receiving an occupant (20) and data measurement means making it possible to establish two biometric prints of distinct types, characterized in that it further comprises a computer (30) programmed to implement an identification and / or authentication method in accordance with one of the preceding claims.

Images