WO2020000191A1

WO2020000191A1 - Method for driver identification based on car following modeling

Info

Publication number: WO2020000191A1
Application number: PCT/CN2018/092903
Authority: WO
Inventors: Mathieu MOZE; Francois Aioun; Franck Guillemard; Donghao XU; Chenfeng TU; Huijing ZHAO
Original assignee: Psa Automobiles Sa; Peking University
Priority date: 2018-06-26
Filing date: 2018-06-26
Publication date: 2020-01-02
Also published as: EP3826895A4; EP3826895A1

Abstract

A method for driver identification based on car following modeling is provided. The method comprising :defining, at a processor, driver classes associated to drivers based on driver state parameters and driver trusted signature parameters in an initialization mode considering driving sequence; obtaining, at the processor, a set of parameters estimation of the driver state and the driver trusted signature that discriminates the most between all the drivers and the less between sequences generated by the same driver in the initialization mode; providing , at the processor, a car-following sequence composed of sequences of leading vehicle's relative motion states to the ego vehicle in a normal usage mode; and selecting, at the processor, driver identification from measurements by computation of class belonging probability in the normal usage mode based on the driver classes defined in the initialization mode.

Description

METHOD FOR DRIVER IDENTIFICATION BASED ON CAR FOLLOWING MODELING

FIELD OF THE INVENTION

The present invention is directed to the domain of automobile vehicles, and more particularly, to a method for driver identification based on car following modeling.

BACKGROUND OF THE INVENTION

The driving style of each driver is different. Heterogeneities exist in the way a driver operates on steering wheel, gas and brake pedals etc. in performing certain behaviors, which turns out different driving styles correlating with road and scene vehicles. It is possible to treat such heterogeneities as a kind of signature thus leading to identification of the driver.

Many researches have been conducted in classifying driving style or understanding driving state of such as sporty, normal or comfortable, evaluating driving skill or recognizing identity of the driver, which are crucial for the potential applications such as situation-based or personalized assistance.

On the other hand, car following is an essential component of a driver’s behavior, where heterogeneity has been studied as an important facet that is a consequence of human factors in this driving process. The level of heterogeneity in the car-following behaviors of different drivers is substantial as well as of vehicle/driver combinations, which is called inter-driver heterogeneity. The internal stochasticity of an individual driver, called intra-driver heterogeneity, is another rational cause for the randomness of car-following behaviors. Many car following models have been developed, where the process of a driver’s behavior is generally described as a transformation from some perceived information about the driving situation, such as the speed and distance of a leading vehicle relative the ego (i.e. the follower) vehicle, and the ego vehicle’s speed, to control actions for acceleration or deceleration. Hence, researches focusing on modeling the heterogeneity in car following behaviors would be available for driver identification.

SUMMARY OF THE INVENTION

The present invention realizes that driver identification could lead to multiple improvements regarding attractiveness and safety of cars through adaptive algorithms and HMI to the driver style (attractiveness) and driver monitoring by comparing actual and usual driving style (safety) . From measurement of ego vehicle acceleration and velocity, ego vehicle distance to the leader (vehicle in front) and relative velocity of the leader to the ego vehicle, this invention enables driver identification from stochastic distribution of his driving behavior.

According to one aspect of the present invention, a method for driver identification based on car following modeling is proposed, the method comprising:

defining, at a processor, driver classes associated to drivers based on driver state parameters and driver trusted signature parameters in an initialization mode considering driving sequence;

obtaining, at the processor, a set of parameters estimation of the driver state and the driver trusted signature that discriminates the most between all the drivers and the less between sequences generated by the same driver in the initialization mode;

providing, at the processor, a car-following sequence composed of sequences of leading vehicle’s relative motion states to the ego vehicle in a normal usage mode; and

selecting, at the processor, driver identification from measurements by computation of class belonging probability in the normal usage mode based on the driver classes defined in the initialization mode.

In accordance with the foregoing technical concept, the present invention may further include any one or more of the following alternative forms.

In some alternative forms, the computation of class belonging probability further comprises a calculation of acceleration estimation associated to a class from the car-following sequence at a given instant.

In some alternative forms, the computation of class belonging probability further comprises a calculation of class belonging probabilities for all instants contained in the car-following sequence by comparing the measured and class estimated accelerations.

In some alternative forms, a signature is determined by compilation of instantaneous signatures for a sequence of measurements at different instants.

In some alternative forms, the driver identification is performed by comparing its Euclidean distances to the driven trusted signatures defined in the initialization mode based on a determined sequence signature.

In some alternative forms, the car-following sequence comprises information of ego vehicle’s acceleration and velocity, ego vehicle’s distance to the leading vehicle and relative velocity of the leading vehicle to the ego vehicle.

In some alternative forms, the set of parameters estimation in the initialization mode is obtained by minimizing the loss function with gradient descent.

In some alternative forms, the initialization mode is performed offline and no computational issue is involved.

In accordance with another aspect of the invention, a device for vehicle arranged and operable to carry out the above method is provided.

In accordance with another aspect of the invention, a processing means programmed and operable to execute instructions for carrying out the above method is provided.

The present invention enables driver identification of a vehicle when this vehicle is behind another vehicle and the driver regulates its speed and position according to its normal driving. It uses information from available sensors: ego vehicle acceleration and velocity, ego vehicle distance to the leader (vehicle in front) and relative velocity of the leader to the ego vehicle. That is, it does not need dedicated sensors to perform driver identification such as cameras.

The proposed invention delivers identification of a driver from multiple already known drivers (previously identified) and a level of confidence associated to the identification in the initialization mode, and thus a stochastic approach can be used based on probability for the driver to belong to a previously defined class. Each class is supposed to represent a unique driver.

These and other aspects of the present invention will be apparent from and elucidated with reference to the embodiments described hereinafter.

BRIEF DESCRIPTION OF THE DRAWINGS

Throughout the course of the following detailed description, reference will be made to the drawings, and in which:

Fig. 1 schematically illustrates notations used in an initialization mode according to the present invention;

Fig. 2 schematically illustrates car following sequence input according to the invention;

Fig. 3 schematically illustrates calculation of acceleration estimation from input sequence;

Fig. 4 schematically illustrates calculation of class belonging probabilities for all instants contained in input sequence;

Fig. 5 schematically illustrates driver identification from class belonging probabilities;

Fig. 6 schematically illustrates signatures of three drivers from experimental dataset;

Fig. 7 schematically illustrates states PCA for three drivers from experimental dataset; and

Fig. 8 schematically illustrates twelve sequences signatures and three driver’s signatures, and dissimilarity of the first sequence is computed for each driver signature.

DESCRIPTION OF EMBODIMENTS

Although the invention may be susceptible to embodiment in different forms, there is shown in the drawings, and herein will be described in detail, specific embodiments with the understanding that the present disclosure is to be considered an exemplification of the principles of the invention, and is not intended to limit the invention to that as illustrated and described hereinafter. Therefore, unless otherwise noted, features disclosed herein may be combined together to form additional combinations that were not otherwise shown for purposes of brevity.

As mentioned, the driving style of each driver is different. Heterogeneities exist in the way a driver operates on steering wheel, gas and brake pedals etc. in performing certain behaviors, which turns out different driving styles correlating with road and scene vehicles. It is possible to treat such heterogeneities as a kind of signature thus leading to identification of the driver.

Recent researches are addressed focusing more on modeling the heterogeneity in car following behaviors. Diver identification based on car-following sequences can be regarded as a time series classification problem. Assuming that a driver can be characterized by a specific probability distribution over driver states during car following, it is a natural choice to represent a car-following sequence using the bag-of-words scheme based on a dictionary of driver states.

In view of the above mentioned, in this invention, two different modes are presented:

· Normal usage mode: Driver identification from measurements by computation of class belonging probability

· Initialization mode: Classes definition associated to each driver.

Initialization Mode

Initialization mode consists in identifying driver state parameters ω _q, σ _q, β _1, q, β _2, q, β _3, q, β _4, q and ΔT _q that will be supposed to be contained in vector θ _q in the following and driver trusted signature parameters C _d, d=1...N _d, from N _S driving sequences associated with the N _d enrolled drivers contained in the set to sort from. Note that each driver d _i, i=1...N _d, generates

sequences such that

Notation is depicted on Fig. 1.

The goal of this initialization mode is to find the optimal set of parameters θ _q ^* and C _d ^* that discriminates the most between all the drivers and the less between sequences generated by the same driver. The discriminating function used for this purpose is called loss function and is presented now.

Let denote

the function that associates a given set of driver state parameters θ _q to a signature

considering driving sequence S _i.

The idea is thus to minimize the loss function

where

is a function that returns the driver trusted signature parameter C _d of the driver that generated S _i, and [x] ₊=max (x, 0) , D (a, b) denotes Euclidean distance between a and b and (h) is the relative distance of ego to leader.

The problem for parameters estimation is thus:

○ ΔT _max≥ΔT _q≥0, q=1, ..., Q,

○ σ _q≥ε, q=1, ..., Q,

○ ω _q≥0, q=1, ..., Q

○

○ C _q≥0, q=1, ..., Q

○

○ and C _d [q] ≥0, d=1, ..., N _d, q=1, ..., Q

where X [y] denotes the y ^th coordinate of any vector, ΔT _max denotes the upper limit of a driver’s response time (10s in practice) and ε is a small positive real number to keep any σ _q from getting too close to 0 which would lead to numerical problems (0.0001 in practice) .

It would be appreciated that minimization is performed through a classical gradient descent method.

Please note that this initialization mode is performed offline and that no computational issue is involved here. Therefore, it would be more convenient and simplify in respect to other methods which rely on spatial positioning or require network.

Normal usage Mode

This mode supposes existence of Q predefined classes or states, each one corresponding to a particular state. These states are abstract concepts and cannot be related to any objective specificity of a driver.

The input time series is a car-following sequence (S) , which is composed of sequences of leading vehicle’s relative motion states to the ego vehicle and containing (h) and

respectively relative distance and velocity of ego vehicle to leader, ego vehicle’s velocity (v) and ego vehicle’s acceleration (a) . The sequence (S) is supposed to correspond to a car following phase, i.e. where the ego vehicle is behind a leader. Moreover, the proposed invention uses only available information for autonomous and assisted driving from usual dedicated sensors such as radars and GPS and more classical ones such as speed meter. This car following sequence input in a normal usage mode is depicted on Fig. 2.

Therefore, the driver identification can be performed by the following steps.

Step 1: Computation of class belonging probability

Each class q is represented by parameters: ω _q, σ _q, β _1, q, β _2, q, β _3, q, β _4, q and ΔT _q. Computation of these parameters is performed during initialization mode.

At a given instant k, a parameter

associated to class q is calculated using the relation:

This calculation is depicted on Fig. 3 for all instants contained in input sequence (S) .

Then at same instant k, a probability for the driver to belong to a particular class q is computed comparing the measured and class estimated accelerations with the relation:

This relation applied for each state permit to obtain a kind of instantaneous signature of current (instant k) driver state by compilation of these probabilities for every state q=1...Q. This signature is thus represented by the vector:

x _k= [x _1, k, ..., x _q, k, ..., x _Q, k] .

This calculation is depicted on Fig. 4 for all instants contained in input sequence (S) and all states.

For a sequence (S) of measurements at T _S different instants, a signature X _S is determined by compilation of instantaneous signatures such that

Please note that vectors x _k and X _S represent probabilities and that the sum of their components is equal to 1.

Please also note that practical experiment showed convergence of X _S when T _S＞20s. In this way, this invention can be used for analyzing and evidencing intra-driver heterogeneity during long-term driving.

Step 2: Driver identification from state belonging probabilities

Once a sequence signature has been determined, driver identification is performed by comparing its Euclidean distances to given trusted signatures C _d, d=1...N _d, each representing a particular driver taken among N _d and previously computed during the initialization mode.

Identified driver denoted d ^*is thus finally given by the relation:

where D (a, b) denotes Euclidean distance between a and b.

Please note that adding a new driver to the set of previously computed trusted signatures is not straightforward and would require to recompute every signatures, meaning recording of a sequence associated to each already enrolled driver.

Final driver estimation is presented on Fig. 5.

Experimental results

The method was applied on three different drivers whose driving sequences were obtained considering situations for which:

· The leading vehicle is the same during the whole sequence;

· The relative distance to the leading vehicle is within 40 meters;

· The sequence is longer than 25 seconds.

The number of different states was fixed to Q=22.

Initialization mode permitted to produce signature for each driver. These are presented on Fig. 6, in which the driver profile distribution is shown from driver 1 to driver 3 in sequence.

In order to analyze the results, a PCA is performed. PCA (Principal Components Analysis) is a statistical procedure that uses an orthogonal transformation to convert a set of observations of possibly correlated variables into a set of values of linearly uncorrelated variables called principal components. Fig. 7 presents such a representation into a 3 dimensional space for the drivers.

From Fig. 7, it can be noted that sample driver 2 is misclassified as driver 1 on a particular (and unique) sequence.

Fig. 8 presents the states distribution obtained on 4 different sequences of 3 different drivers. The first line is the distribution density of each driver (signature) calculated from these associated 3 sequences.

The first graph (line 1, column 1) is used to compute its dissimilarity from each driver profile. As expected, the first distribution density has the lowest dissimilarity, which is correct and validates the method.

It will be understood that there are numerous modifications of the illustrated embodiments described above which will be readily apparent to one skilled in the art, such as many variations and modifications as for the structure of the device for guiding and the assistant component. These modifications and variations fall within the scope of the claims, which follow.

Claims

A method for driver identification based on car following modeling, comprising:

defining, at a processor, driver classes associated to drivers based on driver state parameters and driver trusted signature parameters in an initialization mode considering driving sequence;

obtaining, at the processor, a set of parameters estimation of the driver state and the driver trusted signature that discriminates the most between all the drivers and the less between sequences generated by the same driver in the initialization mode;

providing, at the processor, a car-following sequence composed of sequences of leading vehicle’s relative motion states to the ego vehicle in a normal usage mode; and

selecting, at the processor, driver identification from measurements by computation of class belonging probability in the normal usage mode based on the driver classes defined in the initialization mode.
The method according to claim 1, characterized in that the computation of class belonging probability further comprises a calculation of acceleration estimation associated to a class from the car-following sequence at a given instant.
The method according to claim 2, characterized in that the computation of class belonging probability further comprises a calculation of class belonging probabilities for all instants contained in the car-following sequence by comparing the measured and class estimated accelerations.
The method according to claim 3, characterized in that a signature is determined by compilation of instantaneous signatures for a sequence of measurements at different instants.
The method according to claim 4, characterized in that the driver identification is performed by comparing its Euclidean distances to the driven trusted signatures defined in the initialization mode based on a determined sequence signature.
The method according to anyone of the preceding claims, characterized in that the car-following sequence comprises information of ego vehicle’s acceleration and velocity, ego vehicle’s distance to the leading vehicle and relative velocity of the leading vehicle to the ego vehicle.
The method according to anyone of the preceding claims, characterized in that the set of parameters estimation in the initialization mode is obtained by minimizing the loss function with gradient descent.
The method according to anyone of the preceding claims, characterized in that the initialization mode is performed offline and no computational issue is involved.
A device for vehicle arranged and operable to carry out the method according to anyone of claims 1 to 8.
A processing means programmed and operable to execute instructions for carrying out the method according to anyone of claims 1 to 8.