FR2742706A1 - Monitoring process for observing watchful state of driver to guard against tiredness - Google Patents

Abstract

The monitoring process involves using a number of detectors (1) which are able to detect movements caused by the driver and characteristic of his active control of the vehicle. These may include the position of the accelerator pedal, the speed of the vehicle and rotary movements of the steering wheel. Such signals are fed to an electronic processing box (2) comprising a data collection unit (3) and several real time processors. By monitoring these functions, the electronic system passes through a learning stage in which the normal degree of activity of the driver is determined. The functions are monitored within certain windows of time, the process being repeated at intervals with a new set of measurements. When the level of activity falls below a certain threshold, determined by comparison with the learnt normal levels, the alarm signal is activated.

Description

 The present invention relates to a method for detecting the evolution of the driving behavior of a driver of a motor vehicle, as well as the device for implementing this method.

 It is known that a significant proportion of single-vehicle traffic accidents are attributed to driver failure resulting from hypovigilance, drowsiness and / or fatigue.

 If the driver of a motor vehicle could be alerted to a decrease in performance before he committed a driving fault, a large number of accidents could then be avoided.

 Various solutions have already been proposed to warn the driver of a motor vehicle of a particularly dangerous driving.

 One of these solutions consists in using a hypovigilance detector that is part of a system on the vehicle and analyzing the instantaneous movements of the steering wheel and comparing them with previously observed movements in hypovigilance. The system is based on the following principle: a perfectly awake driver performs a continuous and low amplitude trajectory correction. On the other hand, a hypovigilant or even sleepy driver performs this correction irregularly. The driver no longer corrects his trajectory continuously, but from time to time, which forces him to make large movements of the steering wheel. This known system is thus based on a measurement of the steering wheel angles when they follow a period of time during which the steering wheel has remained stationary, a measurement of this time interval being also carried out, and an alarm is produced in case of detection of hypovigilance of the driver.

 This known system is based solely on the analysis of the movements of the steering wheel so that it is unreliable because the electrical signal for measuring the rotation angles of the steering wheel can be disturbed by other events. This results in a large number of false alarms. In addition, this known system does not take into account the specificity of the driver or that of his driving environment.

 Another solution for preventing the driver of a motor vehicle from a dangerous driving activity is to use an on-board driver fatigue detector and based on a measurement of the duration of the journey and the driving conditions during this time, which are represented by vehicle speed and braking activity. An alarm signal is generated to indicate to the driver that he must rest when the driving time reaches a certain threshold adapted to the driving conditions encountered.

 The solution which is known above is therefore only concerned with the detection of accumulated and supposed fatigue of the driver during a journey and does not take into account the specificity of the driver.

 Other solutions are also known to avoid the failure of a driver, but which require the wearing of specific equipment (helmet, glasses ...). These solutions are rather gadgets and in any case unreliable.

 The present invention provides a method and an onboard device capable of detecting any evolution of the behavior of the driver that can lead to dangerous driving whatever the cause.

For this purpose, the method of the invention for detecting the evolution of the driving behavior of a motor vehicle driver is characterized in that it consists of
measure a certain number of parameters characteristic of the activity or the driving state of the driver, such as the position of the accelerator pedal of the vehicle, the speed of the vehicle, the rotational movements of the steering wheel of the vehicle
- calculate, from the measured values of the parameters, the characteristic quantities or characters of the pipe
- Processing by a specific software these characters calculated so that the software can perform a function of progressive learning of the characteristics of the behavior of the driver as and when driving and a function of detecting any change in calculated characters indicating a change significant driver driving activity; and
- emit an alarm when a significant change in the behavior of the driver is detected.

 The parameters are measured a certain number of times during a given interval or measuring window and this measurement is repeated for a new measuring window offset from the first, preferably with overlap between the two successive measuring windows.

 The calculated characters include statistical quantities such as the average of n values measured during a measurement window, the variance characterizing the dispersion of the values measured around the mean, a symmetry coefficient reflecting the symmetry of the distribution of the n values around the mean and / or the flattening coefficient characterizing the shape of the distribution of values on the measurement window and / or frequential quantities.

 The various characters mentioned above constitute, for each measurement, a vector of characters defining in the space of dimension d defined by d characters a point of this space marked by its d coordinates which are the values of the characters calculated at a time t and defined learning classes materialized respectively by closed geometric zones of the dimension space d and membership functions, based on the principle of fuzzy logic, to determine the class to which a given point of this space belongs and to know thus the driving state of the driver at each moment t.

 The aforementioned membership functions are created using neural networks capable of performing a learning phase by storing a set of training examples consisting of pairs each comprising a given input vector applied to a network. of neurons and a desired output value of the network for this input vector and, once the training phase is complete, to switch to a usage phase so that at any input vector the neural network provides an output value reflecting the input / output relationship he has learned.

Preferably, the above-mentioned neural networks are networks with a Gaussian activation function with two successive operation layers, each of which calculates an approximate value of the membership function by presenting the input vector x at N Gaussian functions Ti of the first layer for calculating N values of Gaussian functions Ti (x) of this vector and then performing in the second network layer the sum of these N values by applying to them multiplicative coefficients ci so that the output O (S) of the network is expressed by

 with (S) which is the approximate value of the membership function of a class representing the driving state of the driver.

 As many networks are used as there are known classes representing the various states of the driver, with the value at the output of a network for any input vector S translating a degree of belonging of this vector to a class and setting a threshold value of the output value below which an input vector is considered as not belonging to a class.

 A learning set of a new class is created as input vectors, whose membership value is less than the aforementioned threshold value, are rejected, using a criterion to determine whether several rejected input vectors represent the same state of conduct of the driver so as not to group in the same learning set points represented by input vectors not belonging to the same class.

 Preferably, the criterion used is the spatial and / or temporal proximity between the points represented by the rejected input vectors so as to progressively constitute groups of points judged to be similar as the rejects progress.

 Two simultaneous conditions are set for a group of points to become a learning set: the driver must have left the driving state represented by the group and the latter must have a sufficient size representing a duration of observation of the behavior of the group. driver, for example at least one minute.

 A learning set for a reference class is created using the proximity criteria in space and time between the points represented by the input vectors so that the group that fills the first learning conditions becomes the reference class used to learn the corresponding network.

 Advantageously, the alarm is emitted when a significant variation on several membership values is detected or when variations in the reference state are detected.

 According to an alternative embodiment, the entire process of creating a learning set of a new class and / or a learning set of a reference class is repeated every time the vehicle travels.

 When the driver's driving activity is unstable, observed from the beginning of the journey, meaning that the driver is driving in a degraded state or that driving conditions are difficult, and preventing the constitution of the reference class the reference class is used on a previously performed path and it is checked whether the input vectors belong to this reference class.

 The invention also proposes a device for detecting the evolution of the driving behavior of a motor vehicle driver for carrying out the method as defined above and which comprises sensors making it possible to measure a certain number of characteristic parameters of driver's driving activity, such as the position of the accelerator pedal, the angle of rotation of the steering wheel and the speed of the vehicle; an electronic unit comprising a unit for acquiring analog signals from the sensors, consisting of analog-to-digital converters for sampling these analog signals and a parallel port for the acquisition of binary signals, electronic cards comprising several processors real-time processing of signals that can operate in parallel to ensure rapid processing of the acquired data and specific software implemented on the processors to perform the progressive learning functions of the characteristics of the driving activity of the driver as and when driving, detecting any significant change in the driver's driving activity and issuing an alarm.

 Advantageously, the processors are connected so as to communicate with each other, with one of the processors controlling the acquisition of the signals by the aforementioned acquisition unit and the other processors periodically performing different tasks distributed over them and including in particular the calculating the characters of the conduct on each acquired signal followed by the formation of the input vector, calculating the output of each existing network, launching and realizing the learning of a new class.

The invention will be better understood and other objects, characteristics, details and advantages thereof will appear more clearly in the following explanatory description made with reference to the accompanying diagrammatic drawings given solely by way of example illustrating a preferred embodiment of the invention. embodiments of the invention and in which
FIG. 1 is a block diagram of the device of the invention for detecting the evolution of the driving behavior of a motor vehicle driver
FIG. 2 represents a particular configuration of signal processing processors used in the device of FIG. 1
FIGS. 3A to 3E show characteristic curves of statistical magnitudes calculated from numerical quantities supplied to the device of FIG. 1
FIGS. 4A and 4B represent a point defined by its statistical magnitudes or characters respectively in a two-dimensional space and a three-dimensional space
FIG. 5 represents the principle of measurement of characteristic parameters of the driving activity of the driver according to the invention;
FIG. 6 represents the principle of the algorithm of the invention making it possible to define the behavior of the driver over time
FIG. 7 represents an example of a fuzzy set of trapezoidal shape making it possible to define membership functions determining to which class belongs a given point of space
FIG. 8 is a Gaussian activation function type neural network architecture for calculating an approximate value of the shape of a fuzzy set or membership function.
FIG. 9 is a curve of membership signals to a reference class
FIG. 10 is a curve of membership signals evolving symmetrically for the creation of two classes and
- Figure 11 shows the operating mode of the device of the invention.

The device of the invention, embarked on a motor vehicle has been developed on the basis of the following two observations
- The driving activity of a motor vehicle driver is specific to the driver and his driving environment. Indeed, the action of the driver on the controls of the vehicle depends on one hand on his physical condition (tired, nervous, ...) and mental (serene, preoccupied, ...) and on the other hand related factors its driving environment (weather conditions, time of day, traffic density, ...). It is therefore not possible to model a priori this driving activity it must be specifically learned for each driver.

The invention does not seek to detect a particular state of the driver (fatigue or hypovigilance for example), because we do not know a priori the events on the driving activity
- it is necessary to avoid the arrival of a dangerous state of the driver and it is the reason for which it is better to detect a development towards this state before it settles and has detrimental consequences. Insofar as the invention does not seek a particular state, such as hypovigilance or fatigue, it will focus on highlighting any change in driver behavior of conduct whatever the cause.

 These two findings constitute the two guidelines guiding the development of the device of the invention.

 The device of the invention as shown in FIG. 1 generally comprises a certain number of sensors 1 making it possible to measure parameters characteristic of the driver's activity or driving state and constituted, for example, by a measurement sensor of the position of the accelerator pedal of the vehicle, a sensor for measuring the angle of rotation of the vehicle steering wheel, a vehicle speed sensor and a sensor for activating the commodos and any other sensor capable of providing a signal useful for the description of a change in driving activity. Preferably, three sensors are used such as those for measuring the position of the accelerator pedal, measuring the angle of rotation of the steering wheel and measuring the speed of the vehicle.

The device also comprises an embedded electronic box 2 containing
a unit 3 for acquiring the analog signals coming from the various sensors 1 and comprising analog / digital converters (not shown) for sampling the analog signals of the sensors 1 and a parallel port (not shown) for the acquisition of signals binary of these converters
electronic cards 4 comprising several processors 5 of the DSP type (FIG. 2) for signal processing and capable of operating in parallel to ensure rapid processing of the data acquired. Processors 5 may be connected together as shown in Figure 2 for purposes to be defined later.

 The electronic box 2 is preferably constituted by the on-board computer of the vehicle of which the central unit 6 controls the signal acquisition unit 3 and transmits the digitized data to the electronic cards 4.

 The device also includes a specific software for detecting changes in driver behavior, implemented on the processors 5 to process the data acquired.

This software performs the following functions
- progressive learning of the characteristics of the driver as and when driving so that the device adapts to the specificities of the driver and its environment
- detection of any significant change in driving activity, whatever the cause (hypovigilance, fatigue, drowsiness, hunger, ... of the driver) and emission of an alarm signal AL. This characteristic of the software gives it the ability to predict and thus anticipate possible driver failures.

 The output signals of the accelerator pedal position sensors 1, the vehicle speed and the rotational movements of the steering wheel from which the driving activity of the driver is determined are continuously applied to the analog-to-digital converters. of the unit 3 and digitized by them to be then used to calculate characteristic digital quantities of the pipe.

 These quantities may be statistical or frequential. These latter characterize the frequencies contained in each received signal.

 Statistical magnitudes may include mean, variance, and symmetry and flattening coefficients. These are explained in FIGS. 3A to 3E by considering a set of n numerical values or samples of the analog-digital converters taken over an interval [a, b]. FIG. 3A shows the average value m and the standard deviation cr in the case of a Gauss curve or bell curve, this average being an indicator of the central value of the samples. The variance characterizes the dispersion of the samples around the mean. The symmetry coefficient reflects the symmetry of the distribution of n values around the mean as represented in FIG. 3B as opposed to the non-symmetric distribution of these values over the interval [a, b] (FIG. 3C). Finally, the flattening coefficient characterizes the form of the distribution of the numerical values over the interval [a, b], for example flat or uniform as represented in FIG. 3D or else modal as represented in FIG. 3E, that is to say say which favors certain values of the interval.

 The quantities thus calculated are called characters and are gathered in a vector of character or vector form. These characters define, according to their number d, a space of dimension d called space of representation.

Thus, every point of this space is identified by its coordinates, which are in fact the values of the characters calculated at a time t. Figures 4A and 4B show two examples of representation spaces. FIG. 4A shows that if two characters are sufficient to characterize the conduct activity of the conductor, the vectors T, are represented in a plane while the use of three characters to characterize this driving activity leads to a representation of the vectors form T, in a three-dimensional space as shown in Figure 4B.

The quantities or characters are calculated from the signals of the sensors 1 according to the "sliding window" principle shown schematically in FIG.
F1, F2 are of fixed length with an overlap between the successive windows F1 and F2. For example, the duration of the measurement or calculation windows is 1 minute and 20 seconds and the windows are offset from At, for example 8 seconds. The overlap between two successive windows is then in this case 90%, but it is understood that this quantity can be easily modified by the user.

We thus obtain a measure of the driving activity, or form ~ ce, all At seconds. This sequence of points (~ r; t) t> 0 is then supplied to the software or evolution detection algorithm as input data.

 The general principle of the algorithm, to satisfy the two functions stated above relating to the progressive learning of the driving characteristics of the driver and to the detection of significant changes in the driver's driving activity, is based on the method next: at the beginning of the driving, the driver is in an initial state called reference state, which is the subject of an apprenticeship. Subsequently, any deviation from this reference is interpreted as an evolution.

This can correspond to the appearance of a new state which is then in turn subject to learning. As a result, the knowledge of the driving activity of the driver in question is increased.

 This method is based on a fuzzy pattern recognition algorithm, and will be explained.

 Pattern recognition is a technique that aims to develop classification methods. Let O be an object and M the number of possible classes. A classification problem can be expressed in a general way as follows: to which of the M classes does the object belong? To answer this question, one must have previously identified and memorized what differentiates the different classes and, in this case, this differentiation is performed by the characters of the conduct. We must then know the value of the characters, that is to say the value of the vector form S t. The decision to determine to which of the M classes the vector forms ~ ct belongs is then taken by comparing the values obtained for the vector shape to those which characterize each possible class. The quality of the decision also depends on the ability of the characters chosen to differentiate the different classes. It is therefore clearer already what learning is: it is a procedure that aims at memorizing the characteristics of the shape vector so that it can be recognized if it were to appear again.

 Figure 6 summarizes the application of the principle of the algorithm to the detection of changes in driver behavior. In this case, the object is the form tt which contains the values of the characters calculated at time t. This form therefore represents the activity of the driver at this moment.

Each known state of the driver (reference, etc.) is associated with a class. This class is materialized by a closed geometrical area of the Ed representation space. The decision rule is then the following: the state of the driver at time t is the class associated with the geometric zone in which X is located . There may be points defined by the form vector that do not belong to any class and they represent states that the device has not yet learned and will be used to learn these new states.

 The succession of measurements S t represents the behavior of the driver over time. As shown in FIG. 6, detecting an evolution then amounts to detecting a characteristic path of this temporal sequence of measurements, either between two known classes, # 1, O 2 O 0) 1, 3, or of a known class O; 02; 0) 3 to an untagged area of space. In the first case defined by case 1 in FIG. 6, the driver moves from a known state 2, for example the reference, to another known state 3. In the second case, case 2 of FIG. a known state and is moving towards a state of conduct in which it had not been found before and which, consequently, had not been memorized.

 To follow these developments between states, we need benchmarks. It is therefore necessary to be able to locate the points within the classes, that is to say with respect to their centers and their borders. For this, each class is modeled by a fuzzy set, hence the name "fuzzy pattern recognition". Figure 7 shows a fuzzy trapezoidal shape whose boundaries are gradual, that is to say progressive. They are described by a continuous function m (x) with values in the interval [0,1] and this function is called membership function. As is apparent from Figure 7, this membership function is 1 for any point that belongs totally to the class (this object is then called prototype), 0 for those who do not belong to it at all, and 0 to 1 for the others, according to their degree of resemblance to the prototypes of the class. The value of the membership function is called the membership degree and it will be used as a measure of proximity to the prototypes.

 After explaining the principle of the algorithm, we will now detail its implementation to which we distinguish three tasks: the recognition of known states, the detection and learning of unknown states and finally the generation of alarms.

Recognition of known states
Each state known to the driver, as we have seen previously, will therefore be represented by a fuzzy set. But as we do not have knowledge a priori about the driver, we do not know the shape of this fuzzy set, that is to say, its membership function. To calculate an approximate value, we use a particular architecture of neural networks: networks with Gaussian activation function, whose principle will now be explained.

 A neural network is a tool for mapping an input space, for example the d-dimensional space, and an output space, for example the interval [a, b]. This network therefore plays the same role as a function f, which establishes a relationship between a starting set and an arrival set. In addition, a neural network is capable of learning by memorizing examples. Thus, if we do not know the mathematical expression of f but we have pairs (x, f (x)), the network can approach the form of f using these examples.

The network plays the role of a black box: for every point x presented as input, we get f (x) at the output, without ever being aware of the exact expression of f.

 There are therefore two phases in the use of a network: the learning phase and the testing phase.

 The learning phase is that during which the network stores the set of examples, called learning set. The learning is done "with teacher", which means that it is clear to the network what is the desired output for a given input. The training examples are therefore input x / desired output pairs f (x). This learning phase then consists of configuring the network, that is to say, arranging the inside of the black box, so as to minimize, for each example, the error between the actual output (approximate value of f (x)) and the desired output f (x). Once the learning phase is over, the network can go into test phase or phase of use. He no longer needs examples and he generalizes what he has learned. For any object x presented as input, it provides an output value that reflects the input / output relationship that it has learned, that is, the relation xf (x).

 There are several types of networks. Those which are preferably used in the present invention are the two-layer networks, i.e. which perform two successive operations on the input data, in this case the vector forms, to obtain the output. Networks with Gaussian activation function are of this type and they implement, in their first layer, Gaussian functions.

 A Gaussian function, or Gauss function, is a symmetrical "bell" function as represented in FIG. 3A, whose form depends on two parameters: the average, or central value, value for which the function is maximal, and the difference -type, or square root of the variance, which defines the "flattening" of the curve around the mean.

 A Gaussian activation function network calculates the value there is one network per class, so that each class can have a different form. Suppose that for any known class, we have a set of prototypes, that is to say a set of examples for which the membership is 1.

This set of prototypes is used as a learning set: for each prototype the network is asked for an output equal to 1. Once the learning is finished, any point of the representation space Ed can be presented to the network. The output of this network is then 1 if the point is located in the entourage of the prototypes, 0 if it is very far from it, and between 0 and 1 in the other cases, according to its distance to the prototypes. The output value of the network for any input x therefore reflects a degree of membership of x to the class. The closer this value is to 1, the closer x is to the class. The closer it is to 0, the further it gets away from it. By setting a threshold on this value, we can highlight the points that are too far from the class to be considered as part of it.

This is called the rejection threshold. All points whose membership value falls below this threshold are considered as not belonging to the class. A point for which the membership value of each class is below the threshold does not belong to any class: it is likely to give rise to a new class associated with the new state it represents.

 For a sequence of successive measurements (x presented at the input of a network, the corresponding output sequence is obtained.This sequence is called the membership signal and it represents the conduct behavior of the driver vis-à-vis the state The example of a membership signal is shown in Figure 9. As shown, a decrease in the value of the membership signal indicates that the driver is moving away from this state, while an increase indicates that Figure 10 shows an evolution from a state i to a state j which results in a significant decrease of the signal i, possibly associated with an increase of the signal j if j is a known state.

Detection and learning of unknown states
We have seen previously how to measure the belonging of a point to a known class. The problem of detection and learning of reference and unknown states must now be addressed. To begin, it is assumed that the reference is known and learned. The associated network provides for every point the membership value for that class. Any point whose membership value is less than the rejection threshold (it is called the rejected point) represents a new state and will participate in its creation. We must therefore build the learning set of this new class as the points are rejected. This process is called the rejected point aggregation process.

 But rejected points can represent different states. There must be a criterion to determine whether or not these points represent the same state. It is important not to group together points that do not belong to the same class in a single learning set. For this, we propose to take into account both the spatial proximity and the temporal proximity between the points. Indeed two points x and x are likely to represent the same state if they are close in the representation space Ed. This means that the characters computed at instants t1 and t2 have similar values: they reflect the same driving activity . They must also be close in time, that is to say that t1 and t2 are not too far apart because the driver has a certain inertia. It is therefore unlikely that he will change his driving state at each time interval AT. Consequently, points representing the same state are, by hypothesis, also characterized by temporal proximity.

In other words, a learning set must consist of successive or "almost" successive points. By combining spatial proximity and temporal proximity, several aggregation criteria were obtained that were applied and compared.

 We gradually build groups of points deemed "similar" as and when releases.

Each group represents a state of the driver who has shown up on his driving activity but has not yet been taught. As soon as a group satisfies the learning conditions developed below, it becomes a set of prototypes of the state it represents and learning a new network is initiated. We are then in the presence of two known states of the driver. The process continues as long as the driving journey lasts.

For a group of points to become a learning set, two conditions must be fulfilled simultaneously
- first, the driver must have left the state
represented by the group. Indeed, as long as the
group fills up, do not interrupt this
process, in order to get the whole learning
the most complete possible, that is to say the most
representative of the state considered
- the group must be of sufficient size. He must
for example represent at least a minute
of behavior, this value not being
indicative and being set by the user of the
device of the invention.

 The formation of the learning set for the reference class is performed according to the same principle. There is no rejected point, since at this stage no class has yet been created. All measurement points therefore participate in the aggregation process, according to the same criteria of proximity in space and time. The group which first fulfills the learning conditions then becomes the reference class and is used to learn the corresponding network.

Generating an alarm
The purpose of the device of the invention is to detect any change in the behavior of the driver to generate an alarm. This is triggered when a significant variation is detected on one or more membership signals. The amplitude of a significant variation is an adjustable parameter which is preferably around 0.5 and this value corresponds to half of the maximum variation of the signal which, as previously seen, is 1.

 The alarm can be triggered on the variations of the reference state. It can also take into account the variations of all the signals. This alarm consists in sending either an alert message on the screen of the onboard device, or an audible or visual alarm, or a combination thereof.

 A complete functional block diagram of the driver behavior monitoring device is shown in FIG. 11 and summarizes well the three tasks of the algorithm explained above, namely the recognition of known states, the detection and the learning of unknown states and the production of alarms. It is therefore useless to detail this synoptic diagram, while pointing out that M denotes the number of known classes, that is to say the number of existing networks at the instant considered.

 Various alternative embodiments of the method and the device described above may be performed without departing from the scope of the present invention.

 Thus, it is possible to calculate other characters than those exemplified from the moment when any numerical value, the value of which changes with a change in the driving activity of the driver is appropriate, regardless of its size.

 Other driving signals than those listed above may be used, for example signals from a vehicle brake sensor, a vehicle operating sensor. All signals related to the driving of the vehicle can be used to calculate the vector of characters and it is sufficient that the information they bring is translated in digital form.

 The learning sets that have been previously described consist only of prototypes. It is possible to include points that are a little distant, and for which the space-time aggregation criterion considered makes it possible to determine a membership value of less than 1. In this case, the learning is carried out in requesting at the output of the network this value less than 1.

 Several space-time aggregation criteria have been considered. Some only consider proximity in space, others only proximity to time, others both.

 The process described above with regard to the learning of the reference and the detection of evolutions can be renewed in full at each trip.

Each time we search for a reference state that corresponds to the conditions of the trip (meteorology, traffic, ...). However, if the driver is driving in a degraded state, we must be able to report it immediately. A degraded state usually results in some instability of the driving activity. The criterion of proximity in space is no longer satisfied and the reference can not be created. This problem can be detected very quickly because it results in the creation of a large number of groups of points, without any of them reaching the required minimum size, which corresponds to about a minute of observation. In this case, it is possible to use a reference on a previously performed path. The signal of belonging to this reference will show immediately that the driver has left this state. In addition, this signal may have a maximum value of less than 1, but its pace remains the same and evolutions can be detected in a similar manner. Finally, the fact of not being able to learn the reference is in itself a reason to send an alarm to the driver.

 The device of the invention as shown in FIG. 1 is implanted so as to operate in real time.

Knowing now the principle of operation of this device, its effective implementation can be described in more detail.

 The device has, as computing power, several signal processing processors (DSP) connected as indicated for example in Figure 2 and can therefore communicate fully.

A processor 5 is first responsible for controlling the data acquisition, that is to say the sampling of the signals (analog / digital conversion). The various tasks to be performed are then distributed over the processors so as to obtain the fastest possible processing of the data acquired. The tasks that must be performed at each moment At are as follows
- calculation of characters (statistical quantities and
frequencies) on each acquired signal, then
shape vector determination
- calculation of the output of each existing network
- application of the aggregation process if necessary
- launching and carrying out the learning of a
new class when the time comes.

 These different tasks will be distributed over the various processors so as to obtain a distribution of calculations allowing real-time operation of the device. In this context, a real-time operation means that any new data acquired must be able to undergo all the processing and calculations provided for in the seconds following its acquisition.

The advantages of the solution according to the invention to prevent any failure of the driver of a motor vehicle are as follows
- it is embeddable and non-binding for the
driver
- it adapts to any driver and any situation
of driving - it is more reliable, more robust and more relevant
other solutions currently known - it is not restricted to the detection of
hypovigilance, but applies to any evolution of
driver behavior with repercussions
on the driving activity of the latter; and - it allows to use as input any signal related to
driving activity, provided that the
corresponding characters are values
digital.

Claims (16)

 1. A method for detecting the evolution of the driving behavior of a driver of a motor vehicle, characterized in that it consists of:  measure a certain number of parameters characteristic of the activity or the driving state of the driver, such as the position of the accelerator pedal of the vehicle, the speed of the vehicle, the rotational movements of the steering wheel of the vehicle  - calculate, from the measured values of the parameters, the characteristic quantities or characters of the pipe  - Processing by a specific software these characters calculated so that the software can perform a function of progressive learning of the characteristics of the behavior of the driver as and when driving and a function of detecting any change in calculated characters indicating a change significant driver driving activity; and  - emit an alarm when a significant change in the behavior of the driver is detected.  2. Method according to claim 1, characterized in that it consists in measuring the aforementioned parameters a certain number of times during a given time interval or measurement window and to repeat such a measurement for a new measurement window offset relative to at the first, preferably with overlap between the two successive measurement windows.  3. Method according to claim 1 or 2, characterized in that the calculated characters comprise statistical quantities such as the average of n values measured during a measurement window, the variance characterizing the dispersions of the values measured around the mean, a coefficient of symmetry reflecting the symmetry of the distribution of n values around the mean and / or the flattening coefficient characterizing the shape of the distribution of the values on the measurement window and / or the frequential quantities.  4. Method according to one of the preceding claims, characterized in that the various characters mentioned above constitute, for each measurement, a vector of characters defining in the space of dimension d defined by d characters a point of this space marked by its d coordinates which are the values of the characters calculated at a time t and in that it consists in defining learning classes materialized respectively by closed geometric zones of the dimension space d and learning functions, based on the the principle of fuzzy logic, to determine the class to which belongs a given point of this space and thus to know the state of conduct of the driver at each instant t.  5. Method according to claim 4, characterized in that it consists in creating the learning functions using neural networks capable of performing a learning phase by storage of a set of training examples constituted by couples each comprising a given input vector applied to a neural network and a desired output value of the network for that input vector and, once the learning phase is complete, to move to a phase of use of so that at any input vector the neural network provides an output value reflecting the input / output relationship that it has learned.  6. Method according to claim 5, characterized in that the above-mentioned neural networks are networks with Gaussian activation function with two successive operation layers, each of which calculates an approximate value of the membership function by presenting the vector d. entry x to N Gaussian functions Fi of the first layer to calculate N values of Gaussian functions Ti (S) of this vector and then performing in the second network layer the sum of these N values by applying them multiplicative coefficients ci so that the output (T,) of the network is expressed by

 with (S) which is the approximate value of the membership function of a class representing the driving state of the driver.  7. Method according to claim 5 or 6, characterized in that it uses as many networks as there are known classes representing the different states of the driver, with the output value of a network for an input vector any S translating a degree of membership of this vector to a class and in that it consists in setting a threshold value of the output value below which an input vector is considered as not belonging to a class.  8. Method according to claim 7, characterized in that it consists in creating a learning set of a new class as input vectors, whose membership value is less than the value. threshold, are rejected, using a criterion for determining whether or not several rejected input vectors represent the same driving state of the driver so as not to group together in the same learning set points represented by vectors d entry not belonging to the same class.  9. A method according to claim 8, characterized in that it consists in using as criterion aforesaid the spatial and / or temporal proximity between the points represented by the rejected input vectors so as to progressively constitute groups of points judged to be similar to the as and when releases occur.  10. The method of claim 8 or 9, characterized in that it consists in setting two simultaneous conditions for a group of points to become a training set the driver must have left the state of driving represented by the group and the latter must have a sufficient size representing a duration of observation of the behavior of the driver, for example at least one minute.  11. Method according to one of claims 8 to 10, characterized in that it consists in creating a learning set for a reference class using the proximity criteria in space and time between the points represented by the input vectors so that the group that fills the first learning conditions becomes the reference class for learning the corresponding network.  12. Method according to one of claims 9 to 11, characterized in that it consists in emitting an alarm when a significant variation on several membership values is detected or that changes in the reference state are detected.  13. Method according to one of claims 8 to 12, characterized in that it consists in completely renewing the process of creating a learning set of a new class and / or a set of learning a reference class for each journey of the vehicle.  14. The method of claim 13, characterized in that it consists, during instability of the driving activity of the driver, observed from the beginning of the journey, meaning that the driver takes the road in a degraded state or that the driving conditions are difficult, and preventing the constitution of the reference class, to use the reference class on a path made previously and to check if the input vectors belong to this reference class.  15. Device for detecting the evolution of the driving behavior of a motor vehicle driver for carrying out the method as defined in any one of claims 1 to 14, characterized in that it comprises sensors (1) for measuring a number of parameters characteristic of the driving activity of the driver, such as the position of the accelerator pedal, the angle of rotation of the steering wheel and the speed of the vehicle; an electronic control unit (2) comprising a unit (3) for acquiring analog signals from the sensors (1), consisting of analog-to-digital converters for sampling these analog signals and a parallel port for acquiring binary signals; electronic cards (4) comprising a plurality of processors (5) for real-time processing of signals that can operate in parallel to ensure rapid processing of the acquired data and specific software implemented on the processors (5) to perform the learning functions progressive characteristics of the driving activity of the driver as and when driving, detection of any significant change in the driving activity of the driver and issuance of an alarm (AL).  Device according to claim 15, characterized in that the processors (5) are connected so as to communicate with each other, with one of the processors (5) controlling the acquisition of the signals by the acquisition unit (3). ) and the other processors (5) periodically performing different tasks distributed thereon and comprising in particular the calculation of the characters of the conduct on each acquired signal followed by the formation of the input vector, the calculation of the output of each network existing, launching and realizing the learning of a new class.