Method for identifying an individual near a motor vehicle

Abstract

Method for identifying an individual near a motor vehicle Method for identifying an individual near a motor vehicle, using sensors arranged on said vehicle, the method comprising the following steps when said individual approaches said motor vehicle: a) detecting his gait by means of the sensors, analyzing at least one parameter related to the gait, comparing it to previously established and recorded gait models and calculating a gait recognition score, b) detecting the face of said individual by means of the sensors, analyzing at least one parameter of said face, comparing it to previously established and recorded face models and calculating a facial recognition score, c) normalizing said scores according to reference scores,(d) merging said normalized gait recognition and facial recognition scores to identify whether or not the individual is associated with said motor vehicle. Figure for abstract: Fig. 1,

Description

Title of the invention: Method for identifying an individual near a motor vehicle Technical field

[0001] The present invention relates to a method for identifying an individual near a motor vehicle, in particular with a view to authenticating him in order to determine whether or not the individual is associated with said motor vehicle. Prior art

[0002] Connected vehicle access is a technology deployed in motor vehicles. It has emerged thanks to the evolution of wireless networks, Bluetooth connectivity and Internet of Things (IoT) technologies. Connected vehicle access is mainly used for simple functions, such as opening vehicle doors remotely using a smartphone or remote control.

[0003] Renault, through its Virtual Key project based on Bluetooth Low Energy (BLE) technology, offers this service for customer comfort. The Virtual Key solution is based on the creation of a physical box containing a BLE module that interacts with a wireless remote control from the Renault range. Access to the openings is via this remote control, which already has all the pairing mechanisms with the vehicle, and in particular the module that operates the vehicle's openings. This technology is present in the prior art.

[0004] It is also possible to manage access to your car with your smartphone rather than your key, in a way that is interoperable with the largest number of smartphone references present in the world. Access to vehicles via connected objects is therefore a mature means of accessing the vehicle.

[0005] Access to the vehicle without a connected device could be envisaged, so that the person who owns the vehicle no longer needs a key or a code to open their vehicle. The latter would recognize legitimate identified persons as they approach and can thus unlock the openings and then welcome them without any constraints for them.

[0006] Application KR 20220144186 describes a biometric authentication method that uses an individual's walking pattern as a means of identification, as well as a two-dimensional composite multiplier neural network model for classification.

[0007] Application US 2022 / 0201389 discloses a method for locating a primary user performing a remote control operation of a vehicle and uses amplitude modulated (AM) ultrasound communication via an ultrasonic beamforming device.

[0008] Application EP 4 139 177 describes a system for assessing the risk posed by a person approaching a vehicle equipped with surveillance cameras.

[0009] Application US 2019 / 0051069 discloses a user recognition system for automated recognition in an autonomous vehicle which comprises an environmental sensor, in particular to recognize a stop request gesture from the user.

[0010] It is known from international application WO 2017 / 176618 to use gesture biometrics and cardiac biometrics to determine whether a user is authorized to use the vehicle.

[0011] Application EP 3 342 097 describes a device managing biometric information, allowing in particular the user to record new biometric information, such as a fingerprint and that from a connected device.

[0012] Application US 2020 / 0193005 discloses a method for unlocking a vehicle when the person approaching the vehicle corresponds to an authorized person. Images are acquired and the person's step and facial features are determined and compared to stored data. Statement of the invention

[0013] There is thus a need to further improve the means for selecting and identifying an individual near a motor vehicle in order to facilitate possible access to said vehicle. Summary of the invention Identification process

[0014] The present invention meets this need thanks to, according to one of its aspects, a method for identifying an individual near a motor vehicle, using a plurality of sensors arranged on said vehicle, the method comprising at least the following steps when said individual approaches said motor vehicle:

[0015] a) detecting one's gait by means of at least one of said plurality of sensors, analyzing at least one parameter linked to the gait, comparing it to previously established and recorded gait models and calculating a gait recognition score,

[0016] b) detecting the face of said individual by means of at least one of said plurality of sensors, analyzing at least one parameter of said face, comparing it to previously established and recorded face models and calculating a facial recognition score,

[0017] c) normalizing said scores according to reference scores, and

[0018] d) merging said normalized gait recognition and facial recognition scores to obtain an identification score to identify whether or not the individual is associated with said motor vehicle.

[0019] Thanks to the invention, it is possible to select and detect a pedestrian who is heading towards the vehicle and who could potentially be the owner of the vehicle to be identified. The vehicle is equipped with a sophisticated system which allows it to estimate the trajectory of the user and to secure reliable biometric identification.

[0020] The invention makes it possible to identify an individual based in particular on their way of walking, i.e. their gait (in English "Gait"). This gait can be considered as a distinctive characteristic of the person, because it is unique and can be used to authenticate them. The identification and authentication process involves a thorough analysis of how the person moves, including how they position their feet, the movements of their arms, their walking speed and their rhythm.

[0021] The fusion of the two recognition techniques makes it possible to improve the accuracy of the identification and authentication of the person, by using the information from each technique to compensate for any limitations of one or the other. The fusion of these two techniques thus makes it possible to obtain more reliable and more precise recognition of the person, by reducing the risks of errors or uncertainties in the identification.

[0022] The fusion guarantees the security and protection of vehicles by ensuring that only authorized persons have access to them. To achieve this, a system of identification and final authentication of the person is put in place before allowing them to enter their vehicle. This system uses facial and gait recognition technologies, the fusion of which makes it possible to verify the identity of the person by comparing these elements to a previously completed database. Once the double recognition has been successfully carried out, access to the vehicle is authorized. This security procedure guarantees the confidentiality and security of the vehicles, preventing unauthorized persons from having access to them.

[0023] The owner of a vehicle can access it freely, without needing any object connected with it. Thanks to the invention, the user can enter his vehicle in a completely natural way, without encountering any constraints or difficulties. In other words, there is no need to carry a key, a badge or any other unlocking device, because the access system is fully automated and personalized.

[0024] The method according to the invention makes it possible to minimize the power consumed necessary for the implementation of its steps.

[0025] When the vehicle's electronic device performs data processing, this can generate significant power consumption. Consequently, to preserve battery life, it is important to reduce this consumption as much as possible. In the case of the invention, in order to optimize power consumption, data processing is only activated when it is really necessary. Thus, each step of the method is advantageously triggered only if the previous one has validated its need for activation. This approach therefore makes it possible to save battery energy by avoiding unnecessary processing. The vehicle's electronic device, in particular the multimedia system, only calls upon its resources at the appropriate time, selectively, depending on the needs of the current task. This technique aims to optimize battery life, which is crucial.

[0026] Furthermore, by optimizing power consumption, the energy efficiency of the multimedia system is improved, which can also have significant environmental and economic implications. Indeed, a reduction in energy consumption makes it possible to reduce the environmental impact of energy production and transport activities.

[0027] Step b) of detecting the face of the individual advantageously uses the analysis of at least one image of said individual, taken by at least one of said plurality of sensors.

[0028] In a preferred embodiment, at least one machine learning algorithm, in particular using at least one neural network, is used to establish the gait models and the face models.

[0029] Step a) can begin when an individual is detected near the vehicle at a predefined minimum distance, in particular between 2 and 5 meters.

[0030] The method according to the invention may further comprise a step of authenticating the individual in order to authorize or not authorize access to the vehicle, in particular using an authentication score. Gait biometrics

[0031] Gait biometrics is a biometric technology that focuses on the unique characteristics of a person's gait. It is based on the fact that each person has a unique way of walking. Gait recognition biometrics is a part of behavioral biometrics, in which subjects can be identified with their walking pattern. The theory behind this recognition system is that each person has a unique gait. It is common for a familiar person to be recognized from a distance by their gait. This is one of the few recognition methods capable of identifying people from a distance.

[0032] A person's gait is as unique as the timbre of their voice. With this knowledge, gait recognition technology can be implemented based on machine learning (ML) algorithms. Nowadays, identifying individuals by their gait is being deployed for the following reasons:

[0033] - gait recognition works well at a distance,

[0034] - gait recognition can be performed from low resolution video and with simple instrumentation,

[0035] - recognition of the approach can be done without the cooperation of people,

[0036] - gait recognition may work well while others features such as faces and fingerprints are not available,

[0037] - gait characteristics are generally difficult to imitate.

[0038] Gait analysis advantageously focuses on a range of personal characteristics at least as follows and in a non-exhaustive manner:

[0039] - walking speed: walking speed is the distance covered by a individual during a given period of time,

[0040] - stride length: stride length corresponds to the distance traveled by a foot between two successive contacts with the ground,

[0041] - time of contact of the foot with the ground: the time of contact of the foot with the ground is the time a foot is in contact with the ground during a stride,

[0042] - swing time: swing time is the length of time a foot is in the air during a stride,

[0043] - stride length: stride length is the distance traveled by both feet during a stride,

[0044] - range of motion: the range of motion is the distance traveled by a body member during normal walking,

[0045] - symmetry of gait: symmetry of gait is the similarity between the left and right body movements while walking,

[0046] - ankle angle: the ankle angle is the angle formed between the foot and the leg while walking,

[0047] - pelvic movement: pelvic movement is the amplitude of the movement of the pelvis while walking,

[0048] - walking cadence: walking cadence is the number of steps taken by an individual during a given period of time,

[0049] - knee angle: the knee angle is the angle formed by the leg and the thigh during walking,

[0050] - size: listing that the person sought is very tall or very short can help with selection,

[0051] - hip height: hip height is the distance between the hip and the ground while walking.

[0052] These characteristics can then be analyzed using at least one machine learning (ML)-based pattern recognition algorithm to create a unique model of the individual's gait.

[0053] Such an algorithm may use a deep neural network architecture to identify patterns and relationships between gait data and the individual's characteristics. Once the individual's gait model has been created, it may be used to recognize that person by comparing the characteristics of their gait to those of the model, in order to identify them if their gait recognition score is above a predetermined reference threshold. Facial biometrics

[0054] Facial recognition is an identification and authentication method that uses facial features to uniquely identify a person. It is based on the analysis of physical characteristics of the face such as interocular distance, jaw shape, nose-to-mouth distance, cheek and eyebrow shape.

[0055] The authentication process, for its part, consists of verifying that the identified person is indeed present and that consequently the system is not in the presence of fraud by photo, video or mask, in particular.

[0056] Facial biometrics is notably based on deep machine learning techniques offering precise and robust facial recognition. To do this, Machine Learning algorithms analyze a large number of facial images in order to identify common characteristics and build a model that will detect and extract the same characteristics in new images.

[0057] This model can then be used to compare an unknown facial image to previously established and stored facial models, in order to identify the person when their facial recognition score is above a predetermined reference threshold.

[0058] In the invention, facial biometrics outside the vehicle. To do this, it must be able to operate in a variety of environmental conditions, such as sunlight, wind, rain or snow. It captures the face tracking of a person in a real-time video, even when this person is moving towards the vehicle. Merger

[0059] The normalized gait recognition and facial recognition scores are merged.

[0060] Each biometric advantageously behaves as a detection source reporting whether a single hypothesis is true or false, and the same for the opposite hypothesis. This means that these detectors produce a binary response from a decision threshold: either they indicate that there is a detection, in our biometric case, or an identification, i.e. the hypothesis is true, or they indicate that there is no detection, i.e. the hypothesis is false. They can therefore generate the following four situations:

[0061] - “true positive” is a situation in which a test or model predicts correctly the existence of a hypothesis. In other words, a true positive is when the test result is positive and the condition being tested actually exists.

[0062] - the “false positive” occurs when a test hypothesis is wrongly rejected, i.e. say that it is reported as positive when it is actually negative.

[0063] - the “false negative” occurs when a test hypothesis is wrongly accepted, that is- that is, it is reported as negative when it is actually positive.

[0064] - the “true negative” occurs when a test hypothesis is correctly accepted as negative. This means that the test correctly indicates the absence of the targeted condition.

[0065] This explains why, in statistics and machine learning, various metrics such as precision, recall, and Fl score can be used to evaluate model performance measures.

[0066] We define in a known manner:

[0067] - the false positive rate (FPR) is the probability that the detection falsely reports the presence of a condition that is actually absent. This is the proportion of false positive errors in detection tests.

[0068] - the false negative rate (FNR) is the probability that the detection fails to correctly identify a condition that is actually present. This is the proportion of false negative errors in detection tests.

[0069] In the invention, the biometric detection system can be considered bimodal because it uses two biometrics that combine their detection scores. Therefore, score fusion applies.

[0070] A system of bimodal fusion biometric detection machine learning algorithms, and more optionally, may be used. Said system is advantageously made learning so that it improves itself. This makes it possible to minimize the energy consumption of said system.

[0071] In its general definition, score fusion is the process of combining the results of multiple classifiers or detectors to produce a more accurate and reliable estimate. The rationale for score fusion is as follows: different Models or classifiers use different features or different processing methods to represent the same object or phenomenon. Merging the scores from each model or classifier results in a more robust, complete, and accurate estimate.

[0072] Score-level fusion is the most common approach because scores generated by different detectors can be easily consulted and combined. This fusion is advantageously preceded by a score normalization step.

[0073] The established gait and facial models are preferably pre-recorded in one or more reference databases. These models can be used to establish statistics.

[0074] Distributions and histograms of scores from tests of legitimate applications and non-legitimate or imposter applications are advantageously calculated.

[0075] In a preferred embodiment, the fusion of said normalized gait recognition and facial recognition scores is performed by linear combination.

[0076] The following function is applied to the sets of gait detection scores #1 and face detection #2 to obtain the fusion statistic S:

[0077] [Math.l] S | S«.i +

[0078] With i:1..N, N the number of scores from detections #1 and #2, if the N normalized scores from detection #1, s2 the N normalized scores from detection #2, “G [0,1],

[0079] In this preferred embodiment of the invention, the fusion scores are thus constructed by linear combination.

[0080] This fusion is advantageously optimized by seeking the a minimizing the associated Equal Error Rate (or EER):

[0081] [Math.2] oc G [0.1], Kmin | Minimum EER^^

[0082] The gait recognition score and the facial recognition score are advantageously compared to predetermined reference thresholds. Sensors

[0083] The method according to the invention may use at least one video sensor, in particular one or more cameras, the method also using in particular a radar sensor and / or an audio sensor. These sensors not only make it possible to detect movements near the vehicle, but also to capture videos and images. This combination of sensors offers considerable potential for identifying a person outside the vehicle, providing precise information about its physical appearance, position and movements.

[0084] Vehicle exterior video sensors are monitoring devices that capture images and videos of the exterior environment. These sensors can be located at various locations on the vehicle, such as the windshield, sides, and rear of the body. The primary purpose of a vehicle's exterior video sensors is to improve driver safety and visibility to detect obstacles and potential hazards to avoid collisions.

[0085] External vehicle radar sensors are sensors that use radio waves to detect objects and obstacles in the external environment. The radar sensor works by establishing its RCS (Radar Cross Section), which is a measure of an object's ability to reflect back radio waves emitted by a radar. RCS is determined by several factors, such as the object's size, shape, texture, and composition. In the context of automotive radar sensors, RCS is used to assess the ability of an "object," potentially a pedestrian, to be detected by the radar and to help determine the object's distance, speed, and direction relative to the vehicle. These sensors can be placed in various locations on the vehicle, such as the front or rear bumper. The primary purpose of external vehicle radar sensors is to improve safety by detecting obstacles and potential hazards on the road.These radar sensors are particularly useful in situations of limited visibility, such as at night, in fog, or in the rain. Radar sensors can detect the presence of moving or stationary objects near the vehicle, such as cars, pedestrians, animals, or obstacles. This type of sensor can therefore be used to enhance driver assistance functions.

[0086] Any type of video, radar or other sensors can be used. Specific identification

[0087] The method according to the invention may comprise a specific identification or re-identification step which consists of identifying an individual who has already been previously identified and that the time interval between these two identifications is sufficiently short so that one can legitimately think that his fundamental characteristics have not changed between these two identifications. These fundamental characteristics considered as a priori knowledge can then considerably help other identification systems. It is thus possible to find a specific person in a scene by using information previously collected during his first identification. This technique is based on the exploitation of highly discriminating criteria, in particular:

[0088] - skin color: this information makes it possible to differentiate people according to their skin tone, which can be useful for identifying a person in a scene where several individuals look alike,

[0089] - gender: a characteristic that allows us to distinguish people according to their sex male or female,

[0090] - estimation of the person's height: a criterion which can help to identify a person among a group taking into account their estimated size,

[0091] - the general shape of the person's morphology: this characteristic allows to identify individuals by taking into account their general silhouette, their build or their posture, pregnant woman,

[0092] - the clothes worn: this information can help differentiate people in depending on their clothing,

[0093] - worn accessories such as hats, bags or glasses: these objects distinctive features can help identify a person in a scene,

[0094] - the person's distinctive movements or gestures: characteristics such as a Limping, swaying, or a specific gesture can also help identify a person in a scene.

[0095] Thanks to the combination of these discriminating criteria, the re-identification technique makes it possible to sort a specific person in a scene, even if they are not in the same position or the same context as when they were first identified.

[0096] These criteria are advantageously associated with classes of a model based on automatic deep learning techniques allowing detection of discriminating criteria via different classes that one seeks to extract from the person to be identified. Other recognitions

[0097] The method according to the invention may further comprise a step of recognizing the individual's voice by means of at least one of said plurality of sensors, said voice being compared to previously recorded voice models. Identification and authentication are thus carried out by voice biometrics, for example by asking the user to pronounce a secret phrase which would have previously enrolled him.

[0098] At least one of the sensors can thus be an external microphone to capture the individual's voice so that it can be identified by voice biometrics.

[0099] The method according to the invention may further comprise a step of recognizing a particular gesture of the individual. The user may perform a particular gesture that only he knows and which would serve as additional authentication to secure access to the vehicle. Estimation of the trajectory of individuals

[0100] The method according to the invention may comprise a prior step of predicting the trajectory of the individuals, if the detection is positive, indicating that the individual is approaching the vehicle, steps a) to d) are carried out.

[0101] This step consists of estimating the trajectory of a pedestrian towards the vehicle. This step allows the more complex steps described previously to be triggered. This step must also take place with a minimum of energy consumed because it is functional in standby mode.

[0102] The problem of predicting the trajectory of a pedestrian is to predict where and in which direction the pedestrian will be in the future using information about the pedestrian and the environment. Predicting the trajectory of pedestrians is complex due to the uncertainty of its interaction with the environment.

[0103] There are two types of known methods for predicting pedestrian trajectories: methods that consist of establishing a model of the pedestrian's kinematics to predict their trajectory, and recently methods based on prediction based on deep learning. However, as this step may be the first to be implemented during what will be the vehicle's standby cycle, it must consume a minimum of energy resources. The invention advantageously uses a simple kinematic trajectory estimation rather than deep learning processing which would be too costly in terms of resources at this stage.

[0104] The pedestrian kinematics model is a physical method for predicting the trajectory of a pedestrian. It uses information such as the variation of the pedestrian's position, its initial velocity, its acceleration and its direction.

[0105] The method according to the invention advantageously calculates the estimation and confirmation of a trajectory. Only establishing that a pedestrian is heading towards the vehicle can be sufficient to implement steps a) to d) of the method. Learning models

[0106] In one embodiment of the invention, the models of both gait and facial biometrics are updated in delayed time. Each time an individual is correctly identified and authenticated, the video data that enabled this decision is advantageously used to create new feed data, such as validation or development data. This data is different from the training data used to calculate the values ​​of the reference model and the test data used to measure the accuracy of the model. The updated model can be evaluated on the test data once training is complete to measure its final performance.

[0107] Two machine learning algorithms may be used in parallel with those used to establish the gait models and the face models, in order to enrich future detections. The performance of the gait and face biometric detection models of the invention is continuously improved by using this new feed data from the experiences of already authenticated users. Device

[0108] According to another of its aspects, the invention relates to a device for identifying an individual near a motor vehicle, the device comprising a plurality of sensors arranged on said vehicle and being configured to carry out at least the following steps when said individual approaches said motor vehicle:

[0109] a) detecting one's gait by means of at least one of said plurality of sensors, analyzing at least one parameter linked to the gait, comparing it to previously established and recorded gait models and calculating a gait recognition score,

[0110] b) detecting the face of said individual by means of at least one of said plurality of sensors, analyzing at least one parameter of said face, comparing it to previously established and recorded face models and calculating a facial recognition score,

[0111] c) normalizing said scores according to reference scores, and

[0112] d) merging said normalized gait recognition and facial recognition scores to obtain an identification score to identify whether or not the individual is associated with said motor vehicle.

[0113] The device advantageously comprises at least one video sensor, in particular one or more cameras. The device may further comprise a radar sensor and / or an audio sensor.

[0114] The characteristics stated in relation to the method apply to the device and vice versa. Motor vehicle

[0115] According to another of its aspects, the invention relates to a motor vehicle comprising a powertrain and at least one identification device according to the invention.

[0116] The characteristics stated in relation to the method apply to the vehicle and vice versa. Brief description of the drawings

[0117] The invention may be better understood by reading the detailed description which follows, a non-limiting example of its implementation, and by examining the attached drawing, in which

[0118] [Fig-1] [Fig. 1] illustrates an example of implementation of the invention,

[0119] [Fig.2] [Fig.2] represents a flowchart illustrating the example of implementation of the invention of [Fig.l],

[0120] [Fig.3] [Fig.3] illustrates an example of implementation of the invention on a motor vehicle,

[0121] [Fig.4] [Fig.4] illustrates the recognition of the gait according to the invention,

[0122] [Fig.5] [Fig.5] illustrates the face recognition according to the invention,

[0123] [Fig.6a] [Fig.6b] [Fig.6c] [Fig.6d] [Fig.6e] [Fig.6f] Figures 6a to 6f are histograms and curves illustrating the principle of fusion,

[0124] [Fig.7] [Fig.7] illustrates an example of implementation of a part of the invention,

[0125] [Fig.8] [Fig.8] represents a flowchart illustrating the example of implementation of the invention of [Fig.7],

[0126] [Fig.9] [Fig.9] illustrates another example of implementation of the invention, and

[0127] [Fig. 10] [Fig. 10] illustrates an alternative embodiment of [Fig.9]. Detailed description

[0128] [Fig.l] illustrates an example of implementation of the invention.

[0129] In a step 0, the vehicle is parked, its sensors are ready to operate. An individual approaches the vehicle. In a step 1 where the individual is at a distance greater than 5 meters, the basic estimation of the pedestrian trajectory begins, as described previously.

[0130] In step 2, if the individual approaches from a distance of less than 5 meters, specific identification or re-identification is triggered.

[0131] In a step 3, the individual's gait is detected by means of a plurality of sensors, visible in [Fig.3]. At least one gait-related parameter is analyzed and compared to previously established and recorded gait patterns. A gait recognition score is calculated. These steps are also performed for the individual's facial biometrics, with a facial recognition score being calculated. These scores are normalized.

[0132] In steps 4 and 5, said normalized gait recognition and facial recognition scores are merged in order to identify whether or not the individual is associated with said motor vehicle, and to authenticate it. Identification and authentication scores are calculated.

[0133] In a step 6, the specific identification or re-identification can be triggered again.

[0134] Step 7 corresponds to an optional step of recognizing the individual's voice by means of at least one of said plurality of sensors, said voice being compared to previously recorded voice models, and to an optional step of recognizing a particular gesture of the individual.

[0135] In a step 8, access to the vehicle may be granted to the individual if he or she has been correctly identified and authenticated.

[0136] [Fig.2] is a flowchart illustrating the invention and broken down as follows:

[0137] 1) Original situation in which the vehicle is parked.

[0138] 2) Estimation of the trajectory of a pedestrian heading towards the vehicle.

[0139] 3) If a pedestrian is detected heading towards the vehicle, the processing continues; otherwise, return to trajectory estimation.

[0140] 4) If it is the first passage there is no possible comparison for a new one identification.

[0141] 5) Otherwise, the specific identification module establishes the current classes instantaneous.

[0142] 6) If there is consistency with the previously established classes, the person therefore has the same discriminating characteristics as in the previous identification, processing can continue; otherwise, return to trajectory estimation.

[0143] 7) Biometric processing by gait and facial recognition.

[0144] 8) Merging scores.

[0145] 9) Reporting of the identification score and the authentication score.

[0146] 10) If there is identification and authentication, the person is the one expected; otherwise, return to trajectory estimation.

[0147] 11) the specific identification module establishes the instantaneous current classes and passes them to the same pre-processing module to serve as reference classes.

[0148] 12) Optional biometrics requiring user intervention such as voice or gesture biometrics.

[0149] 13) If there is identification and authentication, the person is the one expected; otherwise, return to trajectory estimation.

[0150] 14) The identified and authenticated person is legitimate to access his vehicle.

[0151] [Fig.3] shows the sensors and their location on the motor vehicle.

[0152] [Fig.4] illustrates the biometrics of gait, and [Fig.5] [Fig.5] illustrates the facial biometrics, as described above. The internal architecture of the neural networks used is indicative and is based on the needs and expected performance.

[0153] [Fig.6a] depicts the possible responses provided by the sensors, as described previously. The established gait and facial patterns are preferably pre-recorded in one or more reference databases. These patterns can be used to establish statistics.

[0154] Distributions and histograms of scores from tests of legitimate applications and non-legitimate or imposter applications are advantageously calculated.

[0155] [Fig.6b] illustrates an example of a histogram of a first biometric system #1 that could be assigned to facial biometrics and its associated DET curve, with 'tar' for 'target' representing the histogram of the scores of the legitimate tests, and 'nontar' for 'non target' representing the histogram of the scores of the imposter tests. The red median line intercepts the EER.

[0156] [Fig.6c] illustrates an example of a histogram of a first biometric system #1 that could be assigned to the biometrics of gait and its associated DET curve, with the same notations as for [Fig.6b].

[0157] By grouping the two DET curves on the same graph, in [Fig.6d], we see that the DET curve of detection system #1 is lower than that of detection system #2 with EER#1 < EER#2. This comes from the fact that the statistics of detection system #1 have an overlapping surface between its two histograms smaller than that of detection system #2. Detection system #1 is therefore more selective than detection system #2.

[0158] Figure 6e shows the EER of the fusion of the two biometrics. In the example illustrated, a minimum EER is observed for a ^min equal to 0.62. The fusion by linear combination is therefore optimal for this a_min. As shown in Figure 6f, the optimal fusion produces a DET curve lower than the two DET curves of detection systems #1 and #2 with a lower minimum EER. The optimal fusion is therefore the most selective. This fusion with this ^min generates a DET curve lower than each of the two DET curves of detection systems #1 and #2 but also lower than all those resulting from this same fusion with a different K. This fusion with this ^min is therefore the most selective.

[0159] [Fig.7] illustrates an example of trajectory detection, using the front left side radar. The ego vehicle is parked. The front side radar of the ego vehicle receives the echo of the target person. If the trajectory of the person simply converges towards the radar of the ego vehicle, this triggers the more complex identification steps according to the invention.

[0160] [Fig.8] is a flowchart illustrating the processing for a radar and divided as follows:

[0161] 1) The radar in functional standby has a sampling rate of F=5 Hz.

[0162] 2) The radar receives the "Radar Cross Section" or (RCS radar cross section)

[0163] 3) The radar establishes the polar coordinates of the visible targets for a maximum of 10.

[0164] 4) For all targets.

[0165] 5) If the target is less than 5 meters away it is of interest, otherwise it is of no interest.

[0166] 6) Calculation of the variation of its distance from the vehicle.

[0167] 7 and 8) If the distance of the target to the vehicle decreases, it approaches the vehicle; Otherwise it moves away from the vehicle and is of no interest.

[0168] 9) As the target approaches the vehicle, more complex processing is activated to identify the person.

[0169] [Fig.9] illustrates another example of implementation of the invention, using machine learning algorithms (3') in parallel with those used to establish the gait models and the face models. In step 8', these data are transmitted in order to enrich the future detections.

[0170] In this example, since this processing is deferred, it is possible to carry it out via the cloud, as illustrated in [Fig. 10].

[0171] The invention is not limited to the examples which have just been described.

[0172] In particular, the arrangement and number of sensors on the vehicle may be different.

[0173] The neural networks used may also be different.

Claims

Claims

1. Method for identifying an individual near a motor vehicle, using a plurality of sensors arranged on said vehicle, the method comprising at least the following steps when said individual approaches said motor vehicle: a) detecting his gait by means of at least one of said plurality of sensors, analyzing at least one parameter related to the gait, comparing it to previously established and recorded gait models and calculating a gait recognition score, b) detecting the face of said individual by means of at least one of said plurality of sensors, analyzing at least one parameter of said face, comparing it to previously established and recorded face models and calculating a facial recognition score, c) normalizing said scores according to reference scores,and (d) merging said normalized gait recognition and facial recognition scores to obtain an identification score to identify whether or not the individual is associated with said motor vehicle.,

2. Method according to claim 1, in which step b) of detecting the face of the individual uses the analysis of at least one image of said individual, taken by at least one of said plurality of sensors.

3. A method according to claim 1 or 2, wherein the fusion of said normalized gait recognition and facial recognition scores is performed by linear combination.

4. Method according to any one of the preceding claims, in which step a) begins when an individual is detected near the vehicle at a predefined minimum distance, in particular between 2 and 5 meters.

5. Method according to any one of the preceding claims, using at least one video sensor, in particular one or more cameras, the method also using in particular a radar sensor and / or an audio sensor.

6. Method according to any one of the preceding claims, in which at least one machine learning algorithm, in particular using at least one neural network, is used to establish the gait models and the face models.

7. A method according to any preceding claim, wherein the gait recognition score and the facial recognition score are compared to predetermined reference thresholds.

8. Method according to any one of the preceding claims, further comprising a step of authenticating the individual in order to authorize or not authorize access to the vehicle, in particular using an authentication score.

9. A method according to any preceding claim, further comprising a step of recognizing the individual's voice by means of at least one of said plurality of sensors, said voice being compared to previously recorded voice models.

10. Method according to any one of the preceding claims, further comprising a step of recognizing a particular gesture of the individual.

11. Method according to any one of the preceding claims, comprising a prior step of predicting the trajectory of the individuals, if the detection is positive, indicating that the individual is approaching the vehicle, steps a) to d) are carried out.

12. Device for identifying an individual near a motor vehicle, the device comprising a plurality of sensors arranged on said vehicle and being configured to carry out at least the following steps when said individual approaches said motor vehicle: a) detecting his gait by means of at least one of said plurality of sensors, analyzing at least one parameter related to the gait, comparing it to previously established and recorded gait models and calculating a gait recognition score, b) detecting the face of said individual by means of at least one of said plurality of sensors, analyzing at least one parameter of said face, comparing it to previously established and recorded face models and calculating a facial recognition score, c) normalizing said scores according to reference scores,and (d) merging said normalized gait recognition and facial recognition scores to obtain an identification score,

13.

14.

15. to identify whether or not the individual is associated with said motor vehicle. Device according to the preceding claim, comprising at least one video sensor, in particular one or more cameras. Device according to claim 12 or 13, further comprising a radar sensor and / or an audio sensor. Motor vehicle comprising a powertrain and at least one identification device according to any one of claims 12 to 14.

Images