Method for training a classifier capable of determining a type of gesture

Abstract

A method for training a classifier capable of determining the type of gesture made by at least one part of a user's body in the vicinity of a vehicle, said determination being performed from a time-series signal or at least one feature of such a signal, said signal being obtained using a gesture sensor integrated into the vehicle, wherein the training is performed from time-series training signals or from at least one feature from each of a plurality of such training signals, said training signals or features representing the gesture of said part of the user's body relative to the vehicle, from data related to the position and / or orientation of the sensor on the vehicle and / or data related to the vehicle, and from data related to the type of gesture of said part of the user's body. Abstract figure: Figure 3

Description

Title of the invention: Method for training a classifier capable of determining a type of gesture Technical field

[0001] The present disclosure relates to a method for training a classifier capable of determining a type of gesture performed by at least one part of a user's body near a vehicle, from time series type signals or from at least one characteristic derived from such signals, said signals or said characteristic representing the gesture of said part of the user's body relative to the vehicle. Prior art

[0002] A kick sensor may allow a user to activate or control certain vehicle functions such as opening the trunk of a car by making a movement or gesture with their foot. This is a form of hands-free interaction designed to provide convenience and ease of use.

[0003] Instead of using a physical button on the car or on its key fob, the user can activate the mechanism by simply kicking under a specific area of ​​the rear bumper at the trunk, where the sensor is located. The sensor detects the movement and triggers the unlocking or opening of the trunk.

[0004] The purpose of a kick sensor is to allow the user to access the cargo area of ​​the vehicle without requiring them to put their belongings down or physically interact with the car, such as when their hands are full with grocery bags.

[0005] While kick sensors can be convenient and provide hands-free access, they also have some drawbacks, particularly regarding sensitivity. Kick sensors can sometimes be too sensitive or not sensitive enough, and can then misinterpret other movements or objects and lead to unintended activations. For example, if the user or a passerby accidentally brushes against the sensor area while walking past the vehicle, this can trigger the trunk to open or close.

[0006] It would therefore be beneficial to have means capable of classifying the signals from such sensors, i.e. of determining whether the detected gesture is a valid gesture authorizing the opening of the trunk, or whether, on the contrary, the detected gesture should not lead to the opening of the trunk.

[0007] Such a classification can be carried out using a model or classifier that is trained using a training database comprising both signals belonging to the class of valid signals allowing the trunk to be opened and signals belonging to the class of invalid signals that should not lead to such an opening. This requires a large number of signals to obtain a correctly trained classifier. Furthermore, each vehicle configuration requires the training of a separate classifier, each time requiring a training database specific to the vehicle configuration. In total, the number of signals required to train these numerous classifiers becomes extremely large and it is expensive to record real signals to obtain all the signals necessary for this training. Summary

[0008] The present disclosure improves the situation.

[0009] For this purpose, the present document relates to a method for training a classifier capable of determining a type of gesture performed by at least one part of the body of a user in the vicinity of a vehicle, said determination being performed from a time series type signal or from at least one characteristic of such a signal, said signal being derived in use from a gesture sensor integrated into the vehicle, in which the training is performed from:

[0010] - of time series type training signals or from at least one ca characteristic resulting from each of a plurality of such drive signals, said drive signals or characteristic representing the gesture of said part of the user's body relative to the vehicle,

[0011] - data related to the position and / or orientation of the sensor on the vehicle and / or vehicle-related data,

[0012] - data related to the type of gesture of said part of the user's body,

[0013] said method comprising the steps:

[0014] - generating at least a portion of said drive signals or characteristics at using a driven generator,

[0015] - training said classifier at least in part on the basis of the training signals or characteristics generated by the driven generator.

[0016] The gesture may be a blow. The said body part may be a foot or a hand.

[0017] A time series signal is a sequence of data collected and recorded in chronological order. Each data point in the series is associated with a specific moment in time, which makes it possible to analyze and understand the variations that occur over time. The duration of the signal can be between 0.5 and 2 seconds for example, which corresponds to the duration of a gesture allowing the opening of a vehicle door. The value of the signal can be dependent on the distance and / or the position between the target, here said body part, and the corresponding sensor.

[0018] The signals generated with the driven generator are hereinafter called syn signals thetics, as opposed to real signals captured, i.e. obtained using a sensor, and recorded during acquisition phases where a user performs gestures near a vehicle.

[0019] Training the classifier using synthetic data eliminates the need for long and costly phases of acquiring real data in order to properly train the classifier. Of course, the classifier can be trained using real data and synthetic data, or using synthetic data only.

[0020] It is also possible to generate representative features of such signals and train the classifier based on such features. A feature of a time series signal refers to a specific attribute derived from the time series data. Features are extracted to capture relevant information or patterns in the time series.

[0021] Such characteristics may include: - the average: it represents the average value of the signal over a given period, - variance: it measures the dispersion or extent of the signal values ​​around the mean, - the trend: it indicates the general direction or pattern of the signal over time, whether it is increasing, decreasing or remaining relatively constant, - seasonality: it captures recurring patterns or cycles in the time series, generally observed at fixed intervals, - autocorrelation: it quantifies the relationship between a signal and its shifted versions, indicating the degree of similarity between past and present values, - the Fourier transform: it decomposes the time series signal into its constituent frequencies, revealing periodic components or dominant frequency components, - statistical moments: Higher-order statistical moments, such as skewness and kurtosis, provide information about the shape and distribution of the signal, - the wavelet transform: it decomposes the time series into different scales, which allows the analysis of localized frequency patterns, - entropy: it measures the complexity or irregularity of the signal, which gives an idea of ​​its information content.

[0022] The data linked to the type of gesture of said part of the user's body may in particular include an indication of whether or not the gesture is a valid gesture, i.e. authorizing the opening of the corresponding opening of the vehicle. This allows to separate the data representing a valid gesture from the data representing an invalid gesture, to train the classifier. Alternatively, the data related to the gesture types may be data relating to the type of opening to be opened depending on the gesture. For example, a first type of gesture may allow the opening of a first opening and a second type of gesture may allow the opening of a second opening. According to another variant, the data related to the gesture types may be data relating to the type of opening of the opening to be carried out depending on the gesture. For example, a first gesture may allow a partial opening of the opening and a second gesture may allow a complete opening of the opening.

[0023] The vehicle-related data may include data related to the mechanical configuration of the vehicle in proximity to the sensor, in use, including a bumper shape, a shape and number of exhaust pipes, etc.

[0024] The data relating to the position and / or orientation of the sensor on the vehicle may include indications relating to the position or orientation of the sensor, in use, relative to the front or rear bumper, relative to the chassis of the vehicle or relative to the ground, and / or indications relating to the type of paint on the vehicle (which may have an influence on the reflection of the signal).

[0025] Thus, after training the generator, it is capable of generating synthetic signals associated with all kinds of configuration of the vehicle and the sensor(s), without requiring specific training of the generator for each vehicle configuration or for each modification of the arrangement of the sensor on the vehicle.

[0026] For this purpose, according to the present document, the generator is trained with data (signal or signal characteristics) associated with such metadata (representing the configuration of the vehicle and / or the arrangement of the sensor relative to the vehicle), so that the trained generator is then capable of generating synthetic signals associated with other values ​​of this metadata.

[0027] The generator can also take into account physical models, simulations and / or prior knowledge of physical phenomena, to generate signals or characteristics.

[0028] The generator may also generate said signals or said characteristics from data related to the morphology of the user. The data related to the morphology of the user may include indications related to the size of the user, indications related to the position and length of certain limbs or joints of the user.

[0029] The signal may be a radar signal comprising an in-phase component and a quadrature-phase component.

[0030] Such a signal is generally called a dual-phase component signal (or 1 / Q, for In-Phase / Quadrature). This type of signal is commonly used in radar and communication systems to represent modulated signals.

[0031] The in-phase and quadrature phase representation makes it possible to describe a complex signal using two orthogonal components: an in-phase component (I) and a quadrature phase component (Q). The in-phase component is generally associated with the real part of the complex signal, while the quadrature component is associated with the imaginary part.

[0032] By combining these two components, the I / Q signal makes it possible to represent both the amplitude and phase of a modulated signal. It provides a more complete representation than simply using the instantaneous value of the signal, because it captures both phase and amplitude information. This makes it possible to perform operations such as demodulation, modulation, object detection, or distance and speed measurement in radar systems.

[0033] The in-phase and quadrature representation also facilitates the use of signal processing techniques such as Fourier transform, digital modulation / demodulation, carrier suppression and other complex signal processing operations.

[0034] Said radar signal may be a signal from an HFA radar (for High-Frequency Active Radar, in English). This is a type of radar system which operates in the high frequency range of the electromagnetic spectrum, generally between 3 and 30 MHz.

[0035] At least a portion of said training signals or said characteristics may be obtained using a gesture sensor.

[0036] The generator may be a generative adversarial network (GAN).

[0037] A GAN is composed of two main parts: the generator part and the discriminator part, which are deep neural networks. The generator part takes a random noise vector as input and generates a new data. Its objective is to learn to generate data that is as similar as possible to the real data set.

[0038] On the other hand, the discriminator part acts as a binary classifier that takes as input either a data generated by the generator or a real data from the training data set. It tries to distinguish the generated data from the real data.

[0039] The two parts of the GAN compete in a competitive learning process: the generator part seeks to deceive the discriminator part by producing synthetic data indistinguishable from real data, while the discriminator part seeks to improve its ability to distinguish between real and generated data.

[0040] As training progresses, the generator part learns to generate increasingly realistic data, while the discriminator part improves in its ability to distinguish the generated data from real data. Ideally, at the end of training, the generator part produces synthetic data that is indistinguishable from real data for the discriminator part.

[0041] Each data item may be composed of several signals, for example two signals coming from two radar antennas located at a first location on the bumper, for example in the upper part, and two signals coming from two other radar antennas located at a second location on the bumper, for example in the lower part.

[0042] The generator part and / or the discriminator part may comprise successive convolution layers.

[0043] The generator part can for example comprise 4 ConvlDTranspose type layers, a Dropout type regularization can be applied to each layer, the activation function can be of the LeakyRelu type, and the optimizer can be of the Nadam type.

[0044] The discriminator part can for example comprise 3 ConvlD type layers, a BatchNormalization type regularization can be applied to each layer, the activation function can be of the LeakyRelu type, and the optimizer can be of the Rmsprop type.

[0045] Alternatively, the generator may be a diffusion model or a variational autoencoder (VAE).

[0046] The classifier may be a boosting type classifier (for example of the XGBoost or AdaBoost type), a random forest-based model, or a logistic regression model, not directly using the signal as it can be provided by a gesture sensor, but rather using at least one characteristic of such a signal, or a recurrent neural network, directly using the signal as it can be provided by a gesture sensor.

[0047] This document also relates to a method for controlling an opening of a vehicle from at least one gesture made by at least one part of the body of a user in the vicinity of a vehicle, said method comprising the steps:

[0048] - capture at least one time series type signal, said signal being representative of the gesture,

[0049] - classify said signal or at least one characteristic from said signal, using of a classifier trained according to the method according to the invention,

[0050] - control said opening according to the type of gesture made by at least one part of a user's body, as determined by the classifier.

[0051] Said class can illustrate the fact that the signal is a valid signal or not, that is to say authorizing or not the opening of the corresponding opening.

[0052] Said opening may be a trunk or a door of a vehicle, for example of a motor vehicle.

[0053] The present document also relates to a computer program comprising instructions for implementing the method according to the invention, when this program is executed by a processor.

[0054] The present document also relates to a non-transitory recording medium readable by a computer on which is recorded a program for implementing the method according to the invention, when this program is executed by a processor. Brief description of the drawings

[0055] Other characteristics, details and advantages will appear on reading the detailed description below, and on analyzing the attached drawings, in which:

[0056] [Fig-1] is an illustration of a user opening a car trunk with a foot gesture;

[0057] [Fig.2] illustrates an example of an electrical signal from the contactless gesture sensor following the detection of a gesture;

[0058] [Fig.3] is a diagram schematically illustrating the method according to one embodiment of the present document;

[0059] [Fig.4] is a diagram illustrating the projection into a two-dimensional space of real signals and synthetic signals generated with the driven generator; and

[0060] [Fig.5] is a confusion matrix illustrating the performance of the trained classifier. Description of the embodiments

[0061] Reference is now made to [Fig. 1] which shows a user 10 opening the trunk 12 of a car 14 by the action of his foot 16. For this, a contactless sensor 18 is arranged at the rear of the car below the trunk 12 at the bumper. It is by passing his foot 16 in front of the sensor 18 following a kicking gesture, that a trunk opening system 20 detects the kick and controls the opening of the trunk 12. The system 20 includes the sensor 18 and a calculation unit 22 which comprises a classification algorithm or classifier, for example of the boosting type, in order to determine whether the gesture detected by the sensor 18 is indeed a kicking gesture requiring the opening of the opening.

[0062] The contactless sensor 18 may be, according to one example, of two types. It may have one detection antenna (such as a UWB sensor) or two detection antennas (for example, a 24 GHz sensor). In the case where the sensor 18 has two detection antennas, the sensor 18 may have a first antenna 18a directed towards the rear of the car 14 (axis A1), and a second antenna 18b directed towards the underside of the car 14 (axis A2). These two antennas combined make it possible to detect the movement of the foot 16. The sensor 18 receives from each antenna 18a, 18b an electrical signal corresponding to the detected movement. This signal comprises a real part I and an imaginary part Q which have a time shift P between them.

[0063] A method for determining a gesture event uses the computing unit 22 to analyze the signals received by the contactless sensor 18 and determine whether there has indeed been a kick event. The classification carried out by the method makes it possible to avoid false positives or false negatives. For this, an electrical signal corresponding to a gesture detected by the sensor 18 is obtained. [Fig. 2] illustrates an example of a signal obtained. In particular, this figure includes a curve I illustrating the real part of the signal and a curve Q illustrating the imaginary part of the signal.

[0064] This signal is then classified using a classification algorithm (or classifier), this classification algorithm being for example of the boosting type. This classifier determines whether the kick event belongs to one of the following two classes: positive kick event (the gesture is indeed a kick), and negative event not corresponding to a kick.

[0065] A covering, for example a door or a trunk, can then be operated using an actuator, for example an electric motor or a cylinder, depending on the result of the classification.

[0066] As previously indicated, in order to obtain sufficient performance of the classifier, it is necessary to train it beforehand using a training database containing both signals belonging to the class of valid signals allowing the trunk to be opened and signals belonging to the class of invalid signals which must not cause such opening. A large number of signals in each class is necessary to obtain a classifier offering good performance. In addition, each vehicle configuration requires a separate classifier, which implies the creation of a large training database specific to each vehicle configuration. Overall, the number of signals required for training these classifiers becomes extremely high and it is expensive to acquire all the real signals necessary for this training.

[0067] As illustrated in [Fig. 3], the method according to the present document thus proposes to generate synthetic signals using a generator and then to use at least in part such signals to carry out the training of the classifier. Once trained, this classifier can be used, during an inference or production phase, to efficiently classify signals representative of gestures requiring the opening of an opening and then control the opening concerned if necessary.

[0068] The generator may be a GAN, itself trained from real signals and additional data, such as data linked to the type of gesture of said body part. of the user, data related to the position and / or orientation of the sensor on the vehicle, data related to the vehicle and / or data related to the morphology of the user.

[0069] The duration of the signals used to drive the generator can be homogenized so that all the signals have the same duration, for example 1 second. This can be achieved by padding or truncation.

[0070] [Fig.4] is a diagram illustrating the projection into a two-dimensional space of real signals and synthetic signals generated with the trained generator. It can be seen that the real signals and the synthetic signals are indistinguishable from each other, so that the synthetic signals are representative of real signals and can thus constitute a good quality database for training the classifier.

[0071] [Fig.5] is a diagram classically called confusion matrix, used to evaluate the performance of the classification model trained only from synthetic data. This diagram allows to visualize the model's predictions by comparing the actual and predicted results for each class (valid gesture or invalid gesture).

[0072] It can be seen that the number of true positives and true negatives is high and that the number of false positives and false negatives is low, which makes it possible to validate the good performance of the classifier trained on said synthetic data.

Claims

Claims

1. Method for training a classifier capable of determining a type of gesture performed by at least one part of the body of a user in the vicinity of a vehicle, said determination being carried out from a time series type signal or from at least one characteristic of such a signal, said signal being derived in use from a gesture sensor integrated into the vehicle, in which the training is carried out from: - time series type training signals or from at least one characteristic derived from each of a plurality of such training signals, said training signals or characteristics representing the gesture of said part of the user's body relative to the vehicle, the training signals or characteristics each being associated with - data relating to a position and / or an orientation of the sensor on the vehicle and / or data relating to the vehicle,- data related to the type of gesture of said part of the user's body, said method comprising the steps: - generating at least part of said training signals or characteristics using a trained generator, - training said classifier at least in part on the basis of the training signals or characteristics generated by the trained generator.,

2. A method according to the preceding claim, wherein the drive signal is a radar signal comprising an in-phase component and a quadrature-phase component.

3. Method according to the preceding claim, wherein at least a part of said training signals or said characteristics is obtained using a gesture sensor.

4. Method according to one of the preceding claims, in which the generator is a generative adversarial network.

5. Method for controlling an opening of a vehicle from at least one gesture made by at least one part of the body of a user near a vehicle, said method comprising the steps: - capturing at least one time series type signal, said signal being representative of the gesture, - classify said signal or at least one characteristic from said signal, using a classifier trained according to the method according to one of the preceding claims, - control said opening according to the type of gesture performed by at least one part of the body of a user, as determined by the classifier.

6. Computer program comprising instructions for implementing the method according to one of the preceding claims, when this program is executed by a processor.

7. Non-transitory recording medium readable by a computer on which is recorded a program for implementing the method according to one of claims 1 to 6, when this program is executed by a processor.

Images