DE102020215781A1 - Method and device for determining a driving state of at least one moving target vehicle in the vicinity of an ego vehicle - Google Patents

Abstract

Method for determining a driving status of at least one moving target vehicle (TV) in the vicinity of an ego vehicle (1), comprising the steps:- sensory detection of a transverse position of the at least one target vehicle (TV) in relation to lanes (10a. ..10n);- Sensory detection of a transverse speed of the at least one target vehicle (TV) in relation to the ego vehicle (1);- Sensory detection of distances in the longitudinal and transverse direction between the target vehicle (TV) and at least one surrounding vehicle (SV) and sensory detection of longitudinal and transverse speeds of the at least one surrounding vehicle (SV);- sensory determination of a position (ỹL) of at least one lane marking (10a... 10n) of a lane in which the target vehicle (TV) is traveling; wherein- when determining the driving state at a current point in time (t), a driving task (B) and a transverse position of the target vehicle (TV) are determined from the data recorded by sensors, with the determination of the driving task (B) and the transverse position of the target vehicle (TV ) an estimated driving task (B) and an estimated lateral position (y) at a previous point in time (t-1) are taken into account.

Description

The invention relates to a method for determining a driving state of at least one moving target vehicle in the vicinity of an ego vehicle. The invention also relates to a device for determining a driving state of at least one moving target vehicle in the vicinity of an ego vehicle. The invention also relates to a computer program product.

State of the art

In a driver assistance system or a system for automatic driving, the current behavior of the drivers of the surrounding vehicles is determined on a more abstract level from sensor data. It is thus recognized, for example, whether a lane change is taking place into the gap in front of the ego vehicle, or whether the vehicle in front of the ego vehicle is in the process of changing to another lane (“driving task”).

In an ACC system, this information can be used to decide which vehicle the ego vehicle will adjust its speed to. In a known system, a method for probabilistic inference about the driving task of each target vehicle (target vehicle, TV) already exists for determining this information. The method assumes that there are nine discrete driving tasks to be distinguished, such as "keep in lane x", "change lanes from lane x to lane y", etc. The input for the method are essentially measurements of lateral position and speed of the observed vehicle, measurements of the position of lane markings, and measurements of the position and speed of other vehicles. In each time step, the output is a table of probabilities for the possible driving tasks of the observed vehicle, with the method being carried out for one observed vehicle in each case. The influence of the driving situation on the observed vehicle can be taken into account, e.g. by relative position and speed to the vehicle in front of the observed vehicle.

• - Sensorisches Erfassen einer Querposition des wenigstens einen Zielfahrzeugs bezogen auf vom Zielfahrzeug benutzte Fahrspuren;
• - Sensorisches Erfassen einer Quergeschwindigkeit des wenigstens einen Zielfahrzeugs in Relation zum Ego-Fahrzeug;
• - Sensorisches Erfassen von Abständen in Längs- und Querrichtung zwischen dem Zielfahrzeug und wenigstens einem Umgebungsfahrzeug, und sensorisches Erfassen von Längs- und Quer-Geschwindigkeiten des wenigstens einen Umgebungsfahrzeugs;
• - Sensorisches Ermitteln einer Position wenigstens einer Fahrspurmarkierung einer Fahrspur, die das Zielfahrzeug befährt; wobei
• - bei der Ermittlung des Fahrzustands in einem aktuellen Zeitpunkt eine Fahraufgabe und eine Querposition des Zielfahrzeugs aus den sensorisch erfassten Daten ermittelt werden, wobei zur Ermittlung der Fahraufgabe und der Querposition des Zielfahrzeugs eine geschätzte Fahraufgabe und eine geschätzte Querposition in einem vorherigen Zeitpunkt berücksichtigt werden.

D. Barber, "Bayesian Reasoning and Machine Learning", Cambridge University Press, 2012, ISBN-10: 0521518148 discloses techniques related to graph-based models.
Disclosure of Invention
It is an object of the present invention to provide an improved method for determining a driving state of a driving road user in the vicinity of an ego vehicle.
According to a first aspect, the object is achieved with a method for determining a driving state of at least one moving target vehicle in the vicinity of an ego vehicle, having the steps:

• - Sensory detection of a transverse position of the at least one target vehicle based on lanes used by the target vehicle;
• - Sensory detection of a transverse speed of the at least one target vehicle in relation to the ego vehicle;
• - Sensory detection of distances in the longitudinal and transverse directions between the target vehicle and at least one surrounding vehicle, and sensory detection of longitudinal and transverse speeds of the at least one surrounding vehicle;
• - Sensory determination of a position of at least one lane marking of a lane in which the target vehicle is traveling; whereby
• - When determining the driving state at a current point in time, a driving task and a transverse position of the target vehicle are determined from the data recorded by sensors, with an estimated driving task and an estimated transverse position being taken into account at a previous point in time to determine the driving task and the transverse position of the target vehicle.

As a result, an inference of the driving status of the observed road user is realized using a probabilistic graphic model. In this way, for example, a vector with a defined number of probabilities of driving tasks can be determined and output to a subordinate function of the ego vehicle.

This provides a physically motivated structure for determining the driving condition of the moving target vehicle, with the driving condition being calculated back from measured values based on a model. The fact that the transverse position changes as a result of the transverse speed of the observed moving target vehicle is exploited. The driving task of the observed person changes Target vehicle depending on the progress of its driving task, with a transverse speed of the target vehicle depend on the driving task and the degree of its achievement. As a result, an improved driving function of the ego vehicle is supported, for example on the basis of a control unit controlled by a result of the proposed method.

• - eine Erfassungseinrichtung; und
• - eine mit der Erfassungseinrichtung funktional verbundene Ermittlungseinrichtung; wobei
• - die Erfassungseinrichtung ausgebildet ist, ein sensorisches Erfassen einer Querposition des wenigstens einen Zielfahrzeugs bezogen auf vom Zielfahrzeug benutzte Fahrspuren; ein
• - sensorisches Erfassen einer Quergeschwindigkeit des wenigstens einen Zielfahrzeugs in Relation zum Ego-Fahrzeug; ein
• - sensorisches Erfassen von Abständen in Längs- und Querrichtung zwischen dem Zielfahrzeug und wenigstens einem Umgebungsfahrzeug, und ein sensorisches Erfassen von Längs-und Quer-Geschwindigkeiten des wenigstens einen Umgebungsfahrzeugs; ein
• - sensorisches Ermitteln einer Position wenigstens einer Fahrspurmarkierung einer Fahrspur, die das Zielfahrzeug befährt, durchzuführen; und wobei
• - die Ermittlungseinrichtung ausgebildet ist, bei der Ermittlung des Fahrzustands in einem aktuellen Zeitpunkt eine Fahraufgabe und eine Querposition des Zielfahrzeugs aus den sensorisch erfassten Daten zu ermitteln und zur Ermittlung der Fahraufgabe und der Querposition des Zielfahrzeugs eine geschätzte Fahraufgabe und eine geschätzte Querposition in einem vorherigen Zeitpunkt zu berücksichtigen.

According to a second aspect, the object is achieved with a device for determining a driving task of at least one moving target vehicle in the vicinity of an ego vehicle, having:

• - a detection device; and
• - a determination device functionally connected to the detection device; whereby
• - the detection device is designed, a sensory detection of a transverse position of the at least one target vehicle based on lanes used by the target vehicle; a
• - Sensory detection of a lateral speed of the at least one target vehicle in relation to the ego vehicle; a
• - sensory detection of distances in the longitudinal and transverse directions between the target vehicle and at least one surrounding vehicle, and sensory detection of longitudinal and transverse speeds of the at least one surrounding vehicle; a
• - carry out sensory determination of a position of at least one lane marking of a lane in which the target vehicle travels; and where
• - the determination device is designed to determine a driving task and a transverse position of the target vehicle from the data recorded by sensors when determining the driving state at a current point in time, and to determine the driving task and the transverse position of the target vehicle, an estimated driving task and an estimated transverse position at a previous point in time to consider.

According to a third aspect, the object is achieved with a computer program product with program code means for carrying out the proposed method when it runs on a proposed electronic device or is stored on a computer-readable data carrier.

Advantageous developments of the method are the subject of the respective dependent claims.

An advantageous development of the proposed method provides that a partial model is used for a transition probability between the driving tasks of the observed driving road user. In this way, the driving condition of the observed road user can be estimated even better.

A further advantageous development of the proposed method provides that a partial model for the lateral speed of the target vehicle is used, which is a manually created model or a parameterized neural network. In this way, different variants for realizing the partial model for the transverse speed of the observed driving road user are advantageously provided.

A further advantageous development of the proposed method provides that a partial model for kinematics of the target vehicle is used. This can usually be expressed using a simple mathematical relationship, as a result of which the driving condition of the observed road user can be determined even more precisely.

A further advantageous development of the proposed method provides that a neural network trained with estimated transverse positions at the previously defined point in time is used for the partial model for the transition probability between the driving tasks. This enables a so-called expectation-maximization method to be carried out, with which provisional models with unknown variables are estimated in order to be able to train better versions of the models.

A further advantageous development of the proposed method provides that discretized parameters are used for the partial models. In this way, a formula-based inference using empirical formulas of the driving status of the observed driving road user is advantageously supported.

A further advantageous development of the proposed method provides that continuous parameters are used for the partial models. This advantageously provides an alternative calculation method for determining the driving status of the driving road user.

A further advantageous development of the proposed method provides that information is extracted from the measurement data recorded by sensors, with a method for approximate inference being used with the continuous parameters. As a result, calculation methods that are known per se become possible, with which the method can be implemented computationally on a control unit.

A further advantageous development of the proposed method provides that only measured values recorded by sensors are used for a model-based recording of a situation surrounding the ego vehicle. As a result, it is possible to use an uninformative prior, with which the inference can be carried out more quickly.

In a further embodiment of the method, this includes the additional step of providing a warning signal and/or a control signal for controlling the ego vehicle, the warning signal and/or the control signal being determined based on the determined driving state of the target vehicle.

By determining the driving status of the target vehicle, its future behavior can be predicted and the movement of the target vehicle can be better predicted. This makes it possible to correspondingly optimize the driving behavior of the ego vehicle or to issue warning functions if this should be necessary due to the behavior of the target vehicle. Consequently, based on the determined driving behavior, a control signal can be provided for the ego vehicle, which is optimally adapted to the behavior of the target vehicle. In this way, lane changes, speed adjustments or braking maneuvers can be carried out at an early stage, which increases both driving comfort and the safety of road users. By providing a warning signal, a driver of the ego vehicle can also be made aware of dangers at an early stage, which also increases the safety of all road users.

The invention is described in detail below with further features and advantages on the basis of several figures. All of the features described or illustrated form the subject matter of the invention, either alone or in any combination, regardless of how they are summarized in the patent claims or their back-reference, and regardless of their wording or representation in the description or in the figures.

Disclosed features and advantages of the proposed device result in an analogous manner from disclosed features and advantages of the proposed method and vice versa.

• 1 eine prinzipielle Situation, in der das vorgeschlagene Verfahren und die vorgeschlagene Vorrichtung verwendbar sind;
• 2 eine prinzipielle Darstellung eines konventionellen Modells zum Erkennen einer Fahraufgabe eines beobachteten fahrenden Verkehrsteilnehmers;
• 3 eine prinzipielle Darstellung eines vorgeschlagenen Modells zum Erkennen einer Fahraufgabe eines beobachteten fahrenden Verkehrsteilnehmers;
• 4 ein Blockschaltbild einer vorgeschlagenen Vorrichtung; und
• 5 einen prinzipiellen Ablauf einer Ausführungsform des vorgeschlagenen Verfahrens.

In the figures shows:

• 1 a principle situation in which the proposed method and device can be used;
• 2 a basic representation of a conventional model for recognizing a driving task of an observed driving road user;
• 3 a basic representation of a proposed model for recognizing a driving task of an observed driving road user;
• 4 a block diagram of a proposed device; and
• 5 a basic sequence of an embodiment of the proposed method.

Description of Embodiments

In the following, the term "automated vehicle" is used synonymously with the terms "fully automated vehicle", "autonomous vehicle" and "partially autonomous vehicle".

• - beobachtetes Zielfahrzeug
• - beobachtetes fahrendes Zielfahrzeug
• - beobachteter Verkehrsteilnehmer
• - beobachteter fahrender Verkehrsteilnehmer

A "target vehicle" within the meaning of the invention can also be understood in the following, for example, as follows:

• - observed target vehicle
• - Observed moving target vehicle
• - Observed road user
• - Observed driving road user

For the sake of simplicity, the term “target vehicle” is primarily used in the following.

In the following, the term “inference” is understood to mean an activity in which a driving task of a vehicle is determined or estimated from various data.

1 shows a plan view of a traffic situation in which the proposed method can be applied. One recognizes an ego vehicle 1 and several surrounding vehicles SV driving in front of and to the side of ego vehicle 1 (for example motor vehicle, truck, motorcycle, cyclist, etc.). The ego vehicle 1 uses sensors to detect positions of the other driving surrounding vehicles SV and in this way determines a driving task of at least one observed driving target vehicle TV (target vehicle) in order in this way, for example, to have its own automated driving function (for example automatic distance control or distance cruise control, ACC) and/or to be able to better control a driver assistance system. The observed target vehicle TV is preferably driving directly in front of the ego vehicle 1, although each of the surrounding vehicles SV driving in the vicinity of the ego vehicle 1 can be selected as the observed target vehicle TV.

A well-known method is algorithmically based on a simple forward model, as in 2 shown, which describes the probability of being on a lateral axis for each driving task, as well as a probability distribution for the lateral speed. In 2 each node represents a sub-model, each sub-model having a probability distribution for the random variable specified in the node, conditional on the random variables connected by incoming edges.

The model represents a "forward structure" that is used based on measured values recorded by sensors in order to mathematically determine a probability distribution of a driving task B ("in reverse direction"). A mathematical implementation in several variants is conceivable, with a variant with discretized internal parameters or variables of the model (backward calculation (i.e. against the direction of the arrow) with sum formulas) and a variant with continuous internal variables of the model being described in particular below.

The parameters or variables of the model have the following meanings: B ... driving task P ... probability t ... Current point in time arranged in a defined time grid t-1 ... Previous time arranged in a defined time grid _vy ... Lateral speed of the target vehicle TV y ... Lateral position of the target vehicle TV yL ... Lane marking position S... Situation (driving situations of other driving road users SV)

• - Das Teilmodell für die Übergangswahrscheinlichkeiten zwischen verschiedenen Fahraufgaben B berücksichtigt nicht, wie weit die aktuelle Fahraufgabe, bspw. ein Spurwechsel, bereits fortgeschritten ist.
• - Die Wahrscheinlichkeitsdichten von Querposition und -geschwindigkeit, gegeben die Fahraufgabe B, sind vergleichsweise unspezifisch - es müssen Gauß-Verteilungen verwendet werden. Insbesondere sind Querposition und -geschwindigkeit als voneinander unabhängig modelliert, gegeben eine Fahraufgabe B. Intuitiv würde man beispielsweise bei einer Spurhalteaufgabe und einer Position des Zielfahrzeugs TV links der Spurmitte eher eine Geschwindigkeit nach rechts erwarten und umgekehrt, also eher zur Spurmitte hin. Im Modell kann die erwartete Verteilung der lateralen Geschwindigkeit aber nicht von der Position abhängen.
• - Der eigentlich physikalisch gegebene Zusammenhang von Position und Geschwindigkeit über der Zeit ist nicht im Modell berücksichtigt.

Hatched nodes of the model of 2 represent measured variables recorded by sensors (e.g. by means of radar sensors, cameras, etc.), the model making assumptions that greatly simplify reality compared to reality in order to obtain a model structure in which an analytical calculation of the driving task probabilities of a target vehicle TV is possible. The measured values depend on the true values. The analytical calculability is important so that execution on inexpensive computing hardware in the vehicle is possible. The simple model structure is also advantageous for the application, because many parameters can be determined by hand, and in the event of an error, relationships can be traced become gene. However, the existing method described has weaknesses in practice because of the underlying model, which is greatly simplified compared to reality. For example, lane changes of the target vehicle TV are often recognized very late, ie when they are already well advanced. Reasons for these weaknesses can be:

• - The partial model for the transition probabilities between different driving tasks B does not take into account how far the current driving task, e.g. a lane change, has already progressed.
• - The probability densities of transverse position and speed, given driving task B, are comparatively unspecific - Gaussian distributions must be used. In particular, lateral position and speed are modeled as independent of each other, given a driving task B. Intuitively, for example, for a lane keeping task and a position of the target vehicle TV to the left of the lane center, one would expect a speed to the right and vice versa, i.e. closer to the center of the lane. In the model, however, the expected distribution of the lateral velocity cannot depend on the position.
• - The physical relationship between position and speed over time is not taken into account in the model.

These structural weaknesses of the model mean that the conclusions that can be drawn using the inference method to determine the probabilities P of the driving tasks B are only weak. The model has few opportunities to extract a valid portion of information from measurement data that may contain errors or be inconsistent.

• - Im Teilmodell für die Übergangswahrscheinlichkeiten zwischen Fahraufgaben B wird der Fortschritt der jeweiligen Fahraufgabe B berücksichtigt.
• - Berücksichtigt wird, dass die Quergeschwindigkeit, die der Fahrer eines TV wählt, von Fahraufgabe B, tatsächlicher Position und Position der Spurmarkierungen abhängt
• - Berücksichtigt wird eine Kinematik der Querbewegung des Zielfahrzeugs TV.

A method is proposed that can largely eliminate the above-mentioned structural disadvantages of the conventional method by introducing a more complex forward model that can largely eliminate the weaknesses mentioned. The essential properties of the proposed method are:

• - In the partial model for the transition probabilities between driving tasks B, the progress of the respective driving task B is taken into account.
• - It is taken into account that the lateral speed selected by the driver of a TV depends on driving task B, the actual position and the position of the lane markings
• - A kinematics of the transverse movement of the target vehicle TV is taken into account.

The inference about the probabilities of the driving tasks can only be done with an approximate method because of the increased model complexity. There is an increased computing effort, but execution on a more powerful embedded system (embedded hardware) is still possible. Some of the parameters of the model, which is now more complex than the known model, are learned numerically from labeled data as part of an application process.

With the proposed method, the driving task B of the observed moving target vehicle TV can be deduced more clearly and thus more quickly from the available measured values. Inconsistencies or errors in measured values can be better filtered out so that they affect the result of the inference much less. This is essentially based in particular on the fact that the proposed method depicts real relationships more stringently.

A core idea of the proposed method is to provide a model and algorithms suitable for the computational execution of the model in order to implement a specific function for an automated driving system or driver assistance system, taking into account boundary conditions for execution on an electronic computing system (embedded system). will. This can also be applied in a similar way to other functions of the ego vehicle 1, such as, for example, when entering a roundabout, recognizing whether a target vehicle TV approaching from the left in the roundabout will exit the roundabout or in the direction of the ego vehicle 1 continues.

For this purpose, an estimation of the driving task B is modeled or formulated or broken down in such a way that existing algorithms can be used and good results are achieved, and an implementation on embedded hardware is possible.

A new model structure and a method for carrying out the inference in a driver assistance system are proposed, with the necessary steps for determining the sub-models of the model structure being specified.

The proposed model structure is in 3 where each node of the model represents a sub-model, each sub-model representing a probability distribution for the random variable specified in the node, conditional on the random variables connected by edges incoming to the node. Hatched nodes of the model represent measured variables determined by sensors.

Essential elements of the proposed model structure are:

(i) Regarding the sub-model for the transition between different driving tasks:

The probability distribution for the driving task B in the current time step t depends not only on the situation and the driving task B _-1 in the previous time t-1, but also on the transverse position y _-1 of the target vehicle TV in the previous time t-1 and on the position y _L of the lane markings 10a...10n of the lane used by the target vehicle TV. In this way, it is modeled that switching to another driving task B, eg from a “lane change” driving task to a “keep lane” driving task, can depend on the progress of the previous driving task B.

The addition of this partial model means that a more precise hypothesis about the driving task B at the current point in time t is generated in the prediction step.

(ii) Regarding the partial model for the execution of driving task B:

The lateral speed v _y of the target vehicle TV depends on the driving task B and the position of the lane markings y _L , which together represent a target value, and on the previous lateral position y _-1 of the observed target vehicle TV, which is the actual value for executing the driving task Are you. The transverse speed v _y of the observed target vehicle TV is to be interpreted as an action that is carried out to achieve the driving task B. This model structure thus depicts the behavior of a driver who carries out a specific driving task and takes into account the essential influencing variables.

This structure has the advantage that dependencies between the transverse position y and the transverse speed v _y of the observed target vehicle TV can be modeled. This allows the real driver behavior to be modeled more realistically than in the conventional model, which supports a more stringent conclusion from the measurement data in the inference.

(iii) Concerning the partial model for the kinematics:

The transverse position y at the current point in time t depends on the transverse speed v _y and the transverse position y _-1 in the previous point in time t-1, as a result of which a simple physical relationship (kinematics) is depicted. The introduction of this partial model in the inference means that when new information (measurement data) is merged with the previous estimation of the driving task probabilities, a plausible physical behavior is required. Information parts that contradict this can be filtered out. In other words, the inference takes into account the requirement that driving task B should be consistent with the development of the measured values over time.

In order to use the proposed model for probabilistic inference, some preliminary steps are required.

In a first step, the model structure is simplified by assuming an uninformative prior both for the situation (velocities and positions of neighboring surrounding vehicles SV) and for the positions y _L of lane markings. In this way, the variable S̃ takes the place of S in the graph, and ỹ _L takes the place of y _L . In a further step, the continuous physical states y and v _y are discretized.

A decision about the fineness of the discretization follows a trade-off between accuracy and computational effort. The transverse position y of the observed target vehicle TV can, for example, be discretized with five values per lane given, for example, three lanes, which are taken into account in the model. The Querge speed v _y can be discretized into, for example, seven values, ie zero as well as three positive and three negative values.

mit:

P

Wahrscheinlichkeit

S

Situation (Umgebungszustand, z.B. Position und Geschwindigkeit benachbarter Umgebungsfahrzeuge, inklusive Ego-Fahrzeug), wahrer Wert

S̃

Situation, Messwert

B

aktuell ausgeführte Fahraufgabe (z.B. Spurwechsel, Spur halten für spezifische Spuren, usw., kodiert als Integer) des beobachteten Zielfahrzeugs, wahrer Wert

B-1

Fahraufgabe des Ego-Fahrzeugs im vorhergehenden Zeitschritt

yL

Querposition der Fahrspurmarkierungen, wahrer Wert

ỹL

Querposition der Fahrspurmarkierungen, Messwert (aus Messung bzw. Perzeption)

y

Querposition des Zielfahrzeugs, wahrer Wert

ỹ

Querposition des Zielfahrzeugs, Messwert

vy

Quergeschwindigkeit des Zielfahrzeugs, wahrer Wert

ṽy

Querposition des Zielfahrzeugs, Messwert

The proposed overall model can thus be represented in the following mathematical form as a combination of sub-models (sub-model for the transition probabilities between driving tasks, sub-model for the lateral speed, kinematic model): $$\begin{array}{l}p\left(B_{-1},B,v_y,y,y_{-1},\tilde{S},{\tilde{y}}_{\text{L}},{\tilde{v}}_y,\tilde{y}\right)=P\left(B_{-1}\right)\cdot P\left(B|B_{-1},y_{-1},\tilde{S},{\tilde{y}}_{\text{L}}\right)\cdot P\left(v_y|B,y_{-1},y_{\text{L}}\right)\cdot P\left(y|v_y,y_{-1}\right)\cdot \\ P\left(y_{-1}\right)\cdot p\left({\tilde{v}}_y|v_y\right)\cdot p\left(\tilde{y}|y\right)\end{array}$$

With:

P

probability

S

Situation (environmental state, e.g. position and speed of neighboring surrounding vehicles, including ego vehicle), true value

S̃

situation, reading

B

Currently executing driving task (e.g. changing lanes, staying in lane for specific lanes, etc., encoded as an integer) of the observed target vehicle, true value

B-1

Driving task of the ego vehicle in the previous time step

yL

Lateral position of lane markings, true value

ỹL

Lateral position of the lane markings, measured value (from measurement or perception)

y

Lateral position of target vehicle, true value

ỹ

Lateral position of the target vehicle, measured value

vy

Lateral speed of the target vehicle, true value

ṽy

Lateral position of the target vehicle, measured value

Calculation formulas for determining the probabilities for the current driving task B and for the current transverse position y of the observed target vehicle TV can be derived analytically from this, the formulas comprising nested summations of the parameters of the partial models. This derivation corresponds to the use of the "variable elimination" method in the model graph. This now specifies what needs to be calculated when using the function for probabilistic inference.

In the following, a specific computational implementation of the sub-models mentioned and their parameterization are explained in more detail.

a1) The partial model for the transition probabilities between driving tasks, P(B|B _-1 , y _-1 , S̃, ỹ _L ), is parameterized by a neural network, with the exact execution of the neural network resulting from an existing computing time budget and accuracy requirements . For the training of the neural network, training data generated during road trips of the ego vehicle 1 are made available, which contain the transverse position at the previous point in time y _-1 and the measured values for the driving task B to be inferred later, as well as the parameters S̃, which contain the observations of the SV represent and the measured values of the positions ỹ _L of the lane markings 10a...10n.

The requirement that training data with assigned driving task B are available already applies to the data input of the conventional model 2 to.

Now there is a requirement that training data with substantially error-free transverse positions y _-1 are provided.

An alternative to avoid the effort for this is to apply a more complex training procedure for the sub-model for the transition probability between driving tasks. In concrete terms, this can be a so-called expectation-maximization procedure over the entire model, in which values are inferred for the lateral positions y _-1 and are then available for training the partial model for the transition probability between driving tasks B.

a2) The partial model for the lateral speed P(v _y |B, y _-1 ,ỹ _L ) of the target vehicle TV and the observed driving road users 2a...2n describes the implementation of the driving task by a typical driver. This partial model can be set up manually for continuous variables in an application process, as is already practiced for the models for lateral velocity and position in the conventional model. Previous knowledge about driver behavior is introduced.

The partial model for the lateral speed P(v _y |B, y _-1 ,ỹ _L ) of the observed moving target vehicle TV is then discretized and inserted into the overall model. Alternatively, it is also possible to learn the partial model for the lateral speed P(v _y |B, y _-1 ,ỹ _L ) of the target vehicle TV as a neural network from data, since data sets with all the required input and output data from measurements on the street can be provided. When dealing with the requirement for values of the lateral positions y _-1 from previous points in time, the same considerations apply here as for the partial model for the transition probabilities between driving tasks B.

a3) The kinematic model P(y|v _y ,y _-1 ) describes the physical relationship that applying a velocity to a transverse position for a given time interval causes a specific change in the position of the observed target vehicle TV, which is a simple deterministic relationship represents. From this, a model for the discretized variables is derived analytically, resulting in a probabilistic model that takes into account an uncertainty as to where exactly the target vehicle TV is actually located in the respective discretization interval.

• - Diskretisierung der kontinuierlichen Variablen des Modells von 3, um anschließend auf an sich bekannte Weise Summationsformeln herleiten zu können zu können
• - Herleitung der für die Inferenz zu berechnenden mathematischen Formeln
• - Bereitstellung von Trainingsdaten, wobei Messungen durchgeführt werden, bei denen viele fahrende Umgebungsfahrzeuge SV beobachtet und die entsprechenden Messungen aufgezeichnet werden. Anhand der derart durchgeführten Messungen (z.B. in Form von Videos, Radardaten, usw.) werden Daten bewertet, indem z.B. bewertet wird, dass ein Umgebungsfahrzeug gerade die Fahrspur wechselt oder die Fahrspur hält (Messdaten mit zugeordneter Fahraufgabe). Die Fahraufgabe B wird auf diese Weise gewissermaßen gemessen, indem jemand „von Hand“ die Daten mit Werten für die Fahraufgabe B versehen hat. Hier können auch automatische Verfahren verwendet werden, z.B. indem in den aufgezeichneten Daten in die Zukunft (z.B. wenige Sekunden) geschaut wird, und daraus auf die aktuelle Fahraufgabe geschlossen wird
• - Bestimmung der Modellparameter bzw. Modelltraining: Bei den erläuterten Teilmodellen ist beschrieben, welche Daten man braucht, um die Teilmodelle mathematisch zu konkretisieren.

In summary, the following steps, which are known per se, are to be carried out in preparation for a computational calculation of the proposed model:

• - Discretization of the continuous variables of the model of 3 , in order to then be able to derive summation formulas in a manner known per se
• - Derivation of the mathematical formulas to be calculated for the inference
• - Provision of training data, in which measurements are carried out, in which many moving surrounding vehicles SV are observed and the corresponding measurements are recorded. Based on the measurements carried out in this way (eg in the form of videos, radar data, etc.), data is evaluated by evaluating, for example, that a surrounding vehicle is changing lanes or is staying in lane (measurement data with assigned driving task). In this way, driving task B is measured, so to speak, by someone “manually” adding values for driving task B to the data. Automatic methods can also be used here, for example by looking into the future (for example a few seconds) in the recorded data and inferring the current driving task from this
• - Determination of the model parameters or model training: In the case of the explained sub-models, it is described which data is required in order to mathematically concretize the sub-models.

• - In jedem definierten Zeitpunkt t, wobei die Zeitpunkte t in definierten Zeitintervallen beabstandet sind, werden Messungen zur Ermittlung der Messwerte S̃, ỹ_L, ṽ_y und ỹ durchgeführt und in die Teilmodelle eingegeben, genau wie die Schätzung aus dem vorherigen Zeitpunkt
• - Ferner werden in jedem definierten Zeitpunkt t die resultierenden Werte in die vorab berechneten und in einem Programmcode des Steuergeräts implementierten Inferenzformeln eingesetzt, und damit die neuen Wahrscheinlichkeitsverteilungen (Posterior-Verteilungen, die alle bekannten Daten vollständig berücksichtigen) berechnet: P(B|S̃,ỹ_L,ṽ_y,ỹ) und P(y|S̃,ỹ_L, ṽ_y,ỹ), wobei beide Wahrscheinlichkeitsverteilungen in die nächste Schätzung der Wahrscheinlichkeitsverteilung als Wert des vorangegangenen definierten Zeitpunkts t-1 eingehen. Im Ergebnis wird auf diese Weise die Position des Zielfahrzeugs TV nachverfolgt.

When using the proposed method in an embedded system, e.g. an electronic control unit (radar control unit or the like), the following steps result from this:

• - At each defined point in time t, where the points in time t are spaced at defined time intervals, measurements to determine the measured values S̃, ỹ _L , ṽ _y and ỹ are performed and entered into the sub-models, just like the estimate from the previous point in time
• - Furthermore, at each defined point in time t, the resulting values are inserted into the inference formulas calculated in advance and implemented in a program code of the control unit, and the new probability distributions (posterior distributions that take all known data completely into account) are calculated: P(B|S̃, ỹ _L ,ṽ _y ,ỹ) and P(y|S̃,ỹ _L , ṽ _y ,ỹ), where both probability distributions are included in the next estimate of the probability distribution as the value of the previously defined point in time t-1. As a result, the position of the target vehicle TV is tracked in this way.

• - In einem ersten Schritt (Prädiktionsschritt) wird das Vorwärtsmodell ausgeführt zur Berechnung eines joint prior P(B, v_y, y|S̃, ỹ_L)
• - In einem zweiten Schritt (Inferenz-Schritt) werden die Messwerte ṽ_y und ỹ in das Modell eingebracht und damit der joint posterior P(B, y|S̃, ỹ_L, ṽ_y, ỹ) berechnet
• - Daraus werden dann durch Marginalisierung die posterior P(B|S̃, ỹ_L, ṽ_y, ỹ) und P(ỹ|S̃, ỹ_L, ṽ_y, ỹ) bestimmt

In a further alternative of the proposed method, it is also conceivable to divide the implementation of the inference into several steps in order to present it more intuitively:

• - In a first step (prediction step) the forward model is executed to calculate a joint prior P(B, v _y , y|S̃, ỹ _L )
• - In a second step (inference step), the measured values ṽ _y and ỹ are introduced into the model and the joint posterior P(B, y|S̃, ỹ _L , ṽ _y , ỹ) is calculated with them
• - From this, the posterior P(B|S̃, ỹ _L , ṽ _y , ỹ) and P(ỹ|S̃, ỹ _L , ṽ _y , ỹ) are then determined by marginalization

Although this may not be the most resource-efficient way of calculating the inference, it advantageously allows an insight into intermediate variables and can thus help in a development process when troubleshooting a computer program product on which the proposed model is based.

Instead of deriving inference formulas in advance and then using them at runtime, the known variable-elimination algorithm can be executed directly on the model graph at runtime, which processes the model in a specific order. As a result, this provides an alternative way of determining the posterior distributions.

In an advantageous variant of the proposed method, there is no discretization of the variables and no analytical inference, but instead an approximate inference based on continuous variables. In order to use the model for probabilistic inference, the model structure is simplified as a preparatory step by assuming an uninformative prior for both the situation S (e.g. represented by speeds and positions of neighboring observed surrounding vehicles SV) and for the lane marking positions y _L . In this way, the quantity S̃ takes the place of S in the graph, and the quantity ỹ _L takes the place of ỹ.

• b1) Das Teilmodell für die Übergangswahrscheinlichkeiten zwischen Fahraufgaben, P(B|B_-1, y_-1,S̃,ỹ_L) wird durch ein neuronales Netz parametriert, wobei sich die genaue Ausführung aus einem vorhandenen Rechenzeitbudget und Genauigkeitsanforderungen ergibt. Für das Training des neuronalen Netzes werden in Straßenfahrten erzeugte Trainingsdaten zur Verfügung gestellt, die die Querposition im vorhergehenden Zeitschritt y_-1 und die Messwerte für die später zu inferierende Fahraufgabe B enthalten, sowie die Parameter S̃, die die Fahrsituation der Umgebungsfahrzeuge SV repräsentieren und die Messwerte ỹ_L der Fahrspurmarkierungen.

The following is a concrete implementation of the partial models of 3 and their parameterization explained:

• b1) The partial model for the transition probabilities between driving tasks, P(B|B _-1 , y _-1 ,S̃,ỹ _L ) is parameterized by a neural network, with the exact execution resulting from an existing computing time budget and accuracy requirements. For the training of the neural network, training data generated during road trips are made available, which contain the lateral position in the previous time step y _-1 and the measured values for the driving task B to be inferred later, as well as the parameters S̃, which represent the driving situation of the surrounding vehicles SV and the Measured values ỹ _L of the lane markings.

The requirement that training data with assigned driving task B be available already applies to the data input of the conventional model. The requirement that training data be provided with almost error-free transverse positions y _-1 is a new requirement compared to the conventional model. An alternative to avoid the effort involved is to use a more complex training method for the model, e.g. in the form of an expectation-maximization method over the entire model, in which the values of the positions of the transverse positions y _-1 at the previously defined point in time t- 1 are inferred and are then available for training the partial model for the transition probabilities between driving tasks P(B|B _-1 ,y _-1 ,S̃,ỹ _L ).

b2) The partial model for the lateral speed, P(v _y |B, y _-1 ,ỹ _L ) of the target vehicle TV describes the implementation of the driving task B by a typical driver and could be referred to as a "driving strategy". This partial model can be set up manually for continuous variables in an application process, as is already practiced for the models for transverse speed and position in the known model. Previous knowledge about driver behavior is introduced.

Alternatively, it is possible to learn the named partial model or the driving strategy as a neural network from data, since data sets with all the required input and output data from measurements on the road can be made available. The same considerations as for the partial model for the transition probabilities apply here for dealing with the requirement for values of the transverse positions y ₋₁ in the previously defined point in time t−1.

b3) The kinematics model P(y|v _y , y ₋₁ ) describes the physical relationship that the application of a speed to a transverse position for a given time interval causes a specific change in the transverse position of the target vehicle TV observed moving surrounding vehicle SV gently. This is a simple deterministic relation y = y _-1 + v _y x Δt. The proposed model has the special structure that a model part in continuous variables, here the kinematics model with the driving strategy, is controlled by a model part in discrete variables (driving task), with the continuous model part having an influence on discrete variables over time.

For this type of model, methods for approximate inference are known from the literature, e.g. from David S. Gonzales (2019), Chapter 6 from https://www.theses.fr/2019GREAM012.pdf or from Gabriel Agamennoni et al. (2012), https://ieeexplore. ieee.org/iel5/8860/4359257/06196233. pdf.

Applying these methods to the proposed model results in an update of an estimate for P(B|S̃, ỹ _L , ṽ _y , ỹ) and P(y|S̃, ỹ _L , ṽ _y , ỹ) at each time step.

These methods for approximate inference are well adapted to the model structure, so that, in contrast to using a particle filter, a good quality of inference for determining driving task B is possible with little computational effort.

4 shows a block diagram of an embodiment of the proposed device 100.

A detection device 20 that is functionally connected to a determination device 30 can be seen. In this case, the detection device 20 has, for example, at least one camera and/or at least one radar sensor, which provide the sensor-detected measured values explained above and transmit them to the determination device 30. The determination device 30 carries out the determination of probabilities of a driving task B in the manner explained in detail above. The device 100 can be embodied as an embedded system, e.g. as an electronic control unit (radar control unit or the like) of an at least partially automated vehicle.

5 shows a basic sequence of a proposed method in a highly schematic manner.

In a step 200, a transverse position of the at least one target vehicle TV is detected by sensors in relation to lanes 10a...10n used by the target vehicle TV.

In a step 210, a lateral speed of the at least one target vehicle TV is detected by sensors in relation to the ego vehicle 1.

In a step 220, distances in the longitudinal and transverse directions between the target vehicle TV and at least one surrounding vehicle SV are detected by sensors, and longitudinal and transverse speeds of the at least one surrounding vehicle SV are detected by sensors.

In a step 230, a position y _L of at least one lane marking 10a...10n of a lane in which the target vehicle TV is driving is determined by sensors, with a driving task B and a transverse position of the target vehicle TV being determined when the driving state is determined at a current time t the data recorded by sensors are determined, with an estimated driving task B and an estimated transverse position y being taken into account at a previous point in time t-1 in order to determine the driving task B and the transverse position of the target vehicle TV.

The methods according to the invention and the explained devices 20, 30 of the device 100 can advantageously be implemented as software that runs on the proposed device 100, for example. A simple adaptability of the method is supported in this way. It is conceivable that the software of the device 100 is at least partially cloud-based.

The person skilled in the art will modify and/or combine the features of the invention in a suitable manner without departing from the essence of the invention.

Claims (13)

Method for determining a driving status of at least one moving target vehicle (TV) in the vicinity of an ego vehicle (1), comprising the steps: - sensory detection of a transverse position of the at least one target vehicle (TV) in relation to lanes (10a. ..10n); - Sensory detection of a transverse speed of the at least one target vehicle (TV) in relation to the ego vehicle (1); - Sensory detection of distances in the longitudinal and transverse direction between the target vehicle (TV) and at least one surrounding vehicle (SV), and sensory detection of longitudinal and transverse speeds of the at least one surrounding vehicle (SV); - Sensory determination of a position (ỹL) of at least one lane marking (10a...10n) of a lane on which the target vehicle (TV) travels; - when determining the driving state at a current point in time (t), a driving task (B) and a transverse position of the target vehicle (TV) are determined from the data recorded by sensors, with the determination of the driving task (B) and the transverse position of the target vehicle (TV ) an estimated driving task (B) and an estimated lateral position (y) at a previous point in time (t-1) are taken into account. procedure after claim 1 , wherein a partial model for a transition probability between the driving tasks (B) of the target vehicle (TV) is used. procedure after claim 1 or 2 , using a partial model for the lateral velocity (v _y ) of the target vehicle (TV), which is a hand-built model or a parameterized neural network. Method according to one of the preceding claims, wherein a partial model for kinematics of the target vehicle (TV) is used. procedure after claim 3 or 4 , wherein a neural network trained with estimated transverse positions (y) at the previously defined point in time (t-1) is used for the partial model for the transition probability between the driving tasks (B). Method according to one of the preceding claims, wherein discretized parameters are used for the partial models. procedure after claim 6 , where an analytical inference is performed with discretized parameters. Procedure according to one of Claims 1 until 6 , where continuous parameters are used for the partial models. procedure after claim 8 , whereby information is extracted from the sensor-acquired measurement data, with a method for approximate inference being used with the continuous parameters. Method according to one of the preceding claims, wherein only measured values recorded by sensors are used for a model-based recording of a situation surrounding the ego vehicle (1). Method according to one of the preceding claims, with the additional step of providing a warning signal and/or control signal for controlling the ego vehicle (1), the warning signal and/or the control signal being determined on the basis of the determined driving state of the target vehicle (ZF). Device (100) for determining a driving state of at least one moving target vehicle (TV) in the vicinity of an ego vehicle (1), having: - a detection device (20); and - a determination device (30) functionally connected to the detection device (20); - the detection device (20) is designed to detect a transverse position of the at least one target vehicle (TV) by sensor in relation to lanes (10a...10n) used by the target vehicle (TV); a - sensory detection of a lateral speed of the at least one target vehicle (TV) in relation to the ego vehicle (1); a sensory detection of distances in the longitudinal direction between the target vehicle (TV) and at least one surrounding vehicle (SV), and sensory detection of a longitudinal speed of the at least one surrounding vehicle (SV); a - sensory determination of a position (ỹL) of at least one lane marking (10a...10n) of a lane on which the target vehicle (TV) travels; - and where - the determination device (30) is designed to determine a driving task (B) and a transverse position of the target vehicle (TV) from the data recorded by sensors when determining the driving state at a current point in time (t) and to determine the driving task (B) and the lateral position of the target vehicle (TV) to take into account an estimated driving task (B) and an estimated lateral position (y) at a previous time (t-1). Computer program product with program code means for carrying out the method according to one of Claims 1 until 10 when it runs on an electronic device (100) for determining a driving condition of at least one target vehicle (TV) in the vicinity of an ego vehicle (1) or is stored on a computer-readable data medium.

Images