Courtesy Behavior for Highly Automated Vehicles on Highway Interchanges.pdf

海尔

202

6页

0次

2021-02-23

50墨值下载

Courtesy Behavior for Highly Automated Vehicles

on Highway Interchanges

Cristina Men

endez-Romero

, Mustafa Sezer

, Franz Winkler

, Christian Dornhege

and Wolfram Burgard

Abstract— On the highway, human drivers continuously

make decisions adapting their driving behavior. In some of

these, for example near highway ramp-entrances, they adjust

intuitively their own behavior in order to facilitate the merging

of the incoming vehicles. On highly automated vehicles, this

decision should be taken by the system. In such situations, not

only the own goal should be optimized but also the comfort of

the surrounding trafﬁc. The ability to plan adequate courtesy

behaviors improves public acceptance of autonomous systems

and the comfort of the surrounding vehicles without consider-

ably decreasing the own comfort. We present a novel method

that automatically adapts the driving behavior, integrating the

merging intention of other vehicles. In contrast to other systems,

robustness is achieved by considering not only the most likely

evolution, but also the expected value of the possible outcomes

in real time. The ﬂexibility of this method allows us to integrate

it within different planning systems. We are therefore able to

offer courtesy behaviors to other vehicles, thereby improving

the collective comfort and also safety of the situation. We

evaluate the method in simulation and in real world experiments

with our test vehicle. Results show an improvement of the

aggregate trafﬁc comfort and in addition a reduction of critical

situations, as a result of applying our courtesy behavior to

different planning strategies.

I. INTRODUCTION

Vehicles of different automated levels are already ope-

rating on the streets and interacting with purely manually

human-driven vehicles. One challenge lies in situations that

a human driver solves in an intuitive way and that are still

non-trivial for the machine.

Particularly interesting are those situations that arise from

the politeness of the trafﬁc participants. For many merging

situation as presented in Figure 1, human drivers antici-

pate the merging intention of other vehicles driving on the

neighbor lane and select a cooperative behavior. Drivers also

expect such cooperative behavior from automated vehicles.

In dense trafﬁc situations, the vehicle should be able to adapt

its strategy in such situations where a light decrease of the

own comfort considerably improves the collective one. For

example if a merging vehicle is reaching the end of lane or

approaching a slower vehicle, the ego vehicle could decide

to open a gap by decelerating or changing the lane.

This paper focuses on such cooperative behaviors. After

the identiﬁcation of a conﬂicting situation for a potential

merging vehicle, our vehicle assesses the options between

behaving cooperatively or following its own interest. The

C.Men

endez-Romero, M.Sezer and F.Winkler are with BMW Group,

Munich, Germany cristina.menendez@bmw.de

C.Dornhege and W.Burgard are with Autonomous Intelligent Systems,

University of Freiburg, D-79110 Freiburg, Germany

Fig. 1: On a merging scenario, incoming vehicles should

select the appropriate gap to merge and vehicles in the main

ﬂow can cooperate to facilitate the maneuver.

challenge is to optimize safety and comfort over the evo-

lution of several possible scenes without the explosion of

computational costs.

Many works address the prediction of other trafﬁc partici-

pants behavior. However, this behavior is only partially pre-

dictable and only accurate for a short time horizon. Classical

planning systems consider only the most likely evolution of

the situation, neglecting less likely actions that could result

in more dangerous situations. On the other extreme, some

approaches include all possible scene evolutions resulting in

too conservative systems or too high computational load to

be used for real-world applications.

The decision-making should deal with the uncertainties

derived from the behavior of other trafﬁc participants in

an efﬁcient way. This means that the trade-off between

computational load, comfort and safety should be optimized.

The focus of this work concerns the integration of the

intention of the merging vehicles within the decision-making

process.

In this paper we present a courtesy behavior method that

enhances several existing planning approaches. This method

allows to assess the adequacy for the ego vehicle to adapt

its own strategy in order to facilitate the merging maneuver

of a potential merging vehicle. As the main contribution the

improvement of two already existing planning approaches

using our method is presented. The limits between a good

2018 IEEE Intelligent Vehicles Symposium (IV)

Changshu, Suzhou, China, June 26-30, 2018

intention and planning algorithm and the re-planning ability

of the system are also analyzed.

II. RELATED WORK

Human drivers analyze and anticipate the trafﬁc situation.

In a similar manner, autonomous vehicles should integrate

a prediction of the behavior of other participants into their

driving task.

The problem of robots interacting on populated human

environments is not only focus of autonomous driving but

also point of interest of other robotic ﬁelds. Bennewitz et

al. [1] presented a method to predict the trajectories of per-

sons and improve the navigation behavior of a mobile robot.

Kuderer [2] and Kretzschmar [3] present a cooperative navi-

gation model for mobile robots interacting with pedestrians.

Nevertheless, when navigating on freeways and highways,

the topology is more structured and the velocities of the

trafﬁc participants are higher, which requires consideration

of speciﬁc solutions.

In a near future, as presented by Hobert [4], the intercom-

munication between vehicles and infrastructure (V2X) will

allow to acquire a precise information about the intentions of

the trafﬁc participants and the evolution of the situations. The

authors of [5] proposed an on-ramp merging system which

assessed the road trafﬁc conditions and transmit the instruc-

tions to the vehicles on the surrounding area. The system

performs well but relies on an advanced infrastructure. V2X

technology presents encouraging results but the technology

is still not mature enough.

Thourought the last years several prediction methods for

trafﬁc participants have been intensively studied in the lit-

erature. Different motion models like physic-based models,

maneuver-based models or interaction-aware based models

can be used for the prediction [6]. The integration of pre-

diction and intention information within the decision-making

process plays a crucial role in the system performance.

In [7] the authors compute the complete set of the

collision-free start and end points of trajectories for each

vehicle. This kind of intensive computations provide accurate

results in detriment of the online capability of the system.

Other strategies take the current most likely prediction and

rely on a continuous update of the available information and

fast re-planning system, as the multilevel planning system

presented by Men

endez-Romero et al. [8]. Carvalho et al. [9]

integrate the most likely cut-in prediction information to

improve the autonomous cruise control. The combination of

the most likely prediction with a fast re-planing works for

most of the situations quite well, but still does not consider

other possible interactions between the agents involved.

Iterative planning strategies combine the planning and pre-

diction tasks. Wei et al. [10] propose an intention prediction

based strategy generation. In [11] the authors suggest a game

theoretic approach which can model the re-planning capa-

bilities of the drivers. In [12] the decision-making is based

on Partially Observable Markov Decision Process (POMDP).

The multi-policy decision-making presented by Cunningham

et al. [13] also simulates the scene evolution using the most

likely evolution of the other agents involved in order to

reason about the policies. The problem with such iterative

planning approaches is that they only consider the most

likely evolution of the other trafﬁc participants. Especially in

longer prediction horizons the model predictions can become

inaccurate and overlook some critical situations.

One step further, our proposed system not only anticipates

the behavior of other trafﬁc participants to improve their own

safety but also plans a cooperative behavior to improve the

aggregate trafﬁc comfort.

III. APPROACH

The objective of this work is to provide our system with

a courtesy behavior, which identiﬁes the intention of other

trafﬁc participants and assesses the cost of adapting the own

ego strategy. We achieve this by integrating an intention

prediction algorithm into the decision-making. Thereby we

gain a better foresight of the scene evolution by including

all possible outcomes to provide our system with robustness

over false predictions. Decisions are made based on maxi-

mizing the expected utility of the involved trafﬁc participants.

A. Problem and Task Description

We semantically determine the possible actions for the

ego and the conﬂicting vehicle as shown in Tables I and II.

All possible ego actions are clustered as No Cooperative

(NC) and Cooperative (CO). Thus, the action space of

the ego vehicle is deﬁned as A

eg o

:= {NC, CO}. For the

conﬂicting vehicle normal, courtesy and forced merge actions

are combined into the No Yield actions (N Y ), corresponding

to merging in front of the ego vehicle, and the Yield actions

(Y ). Thus, the action space of the conﬂicting vehicle is

deﬁned as A

:= {Y, NY }. The Cartesian product A =

eg o

× A

deﬁnes the joint action space.

Our goal is to choose an action for the ego vehicle so that

the combined expected utility (U(a)) is maximized, i.e.,

∗

eg o

= argmax

a∈A

ego

E(U(a))

= argmax

a∈A

ego

∈A

p(a

| a

eg o

) · U(a

eg o

, a

(1)

This makes necessary to predict the intention of the merging

vehicle p(a

| a

ego

), which is presented in the next sec-

tion. The description of the expected utilities is given in

Section III-C.

B. Prediction Module

In the state-of-the-art methods, the costs of actions are

computed for the most likely actions of other trafﬁc partici-

pants. But these approaches lack the robustness against false

predictions. Our aim, on the other hand, is to predict the

probability of unlikely behavior of other trafﬁc participants,

speciﬁcally the misprediction probability for the conﬂicting

vehicle. Using this probability we can compute the expected

utility of an ego action considering two possible decisions

of the conﬂicting vehicle.

In order to predict the probability of unlikely outcomes, we

use an ensemble learning based Gentle Boost classiﬁer [14].

944

Courtesy Behavior for Highly Automated Vehicles

on Highway Interchanges

Cristina Men

endez-Romero

, Mustafa Sezer

, Franz Winkler

, Christian Dornhege

and Wolfram Burgard

Abstract— On the highway, human drivers continuously

make decisions adapting their driving behavior. In some of

these, for example near highway ramp-entrances, they adjust

intuitively their own behavior in order to facilitate the merging

of the incoming vehicles. On highly automated vehicles, this

decision should be taken by the system. In such situations, not

only the own goal should be optimized but also the comfort of

the surrounding trafﬁc. The ability to plan adequate courtesy

behaviors improves public acceptance of autonomous systems

and the comfort of the surrounding vehicles without consider-

ably decreasing the own comfort. We present a novel method

that automatically adapts the driving behavior, integrating the

merging intention of other vehicles. In contrast to other systems,

robustness is achieved by considering not only the most likely

evolution, but also the expected value of the possible outcomes

in real time. The ﬂexibility of this method allows us to integrate

it within different planning systems. We are therefore able to

offer courtesy behaviors to other vehicles, thereby improving

the collective comfort and also safety of the situation. We

evaluate the method in simulation and in real world experiments

with our test vehicle. Results show an improvement of the

aggregate trafﬁc comfort and in addition a reduction of critical

situations, as a result of applying our courtesy behavior to

different planning strategies.

I. INTRODUCTION

Vehicles of different automated levels are already ope-

rating on the streets and interacting with purely manually

human-driven vehicles. One challenge lies in situations that

a human driver solves in an intuitive way and that are still

non-trivial for the machine.

Particularly interesting are those situations that arise from

the politeness of the trafﬁc participants. For many merging

situation as presented in Figure 1, human drivers antici-

pate the merging intention of other vehicles driving on the

neighbor lane and select a cooperative behavior. Drivers also

expect such cooperative behavior from automated vehicles.

In dense trafﬁc situations, the vehicle should be able to adapt

its strategy in such situations where a light decrease of the

own comfort considerably improves the collective one. For

example if a merging vehicle is reaching the end of lane or

approaching a slower vehicle, the ego vehicle could decide

to open a gap by decelerating or changing the lane.

This paper focuses on such cooperative behaviors. After

the identiﬁcation of a conﬂicting situation for a potential

merging vehicle, our vehicle assesses the options between

behaving cooperatively or following its own interest. The

C.Men

endez-Romero, M.Sezer and F.Winkler are with BMW Group,

Munich, Germany cristina.menendez@bmw.de

C.Dornhege and W.Burgard are with Autonomous Intelligent Systems,

University of Freiburg, D-79110 Freiburg, Germany

Fig. 1: On a merging scenario, incoming vehicles should

select the appropriate gap to merge and vehicles in the main

ﬂow can cooperate to facilitate the maneuver.

challenge is to optimize safety and comfort over the evo-

lution of several possible scenes without the explosion of

computational costs.

Many works address the prediction of other trafﬁc partici-

pants behavior. However, this behavior is only partially pre-

dictable and only accurate for a short time horizon. Classical

planning systems consider only the most likely evolution of

the situation, neglecting less likely actions that could result

in more dangerous situations. On the other extreme, some

approaches include all possible scene evolutions resulting in

too conservative systems or too high computational load to

be used for real-world applications.

The decision-making should deal with the uncertainties

derived from the behavior of other trafﬁc participants in

an efﬁcient way. This means that the trade-off between

computational load, comfort and safety should be optimized.

The focus of this work concerns the integration of the

intention of the merging vehicles within the decision-making

process.

In this paper we present a courtesy behavior method that

enhances several existing planning approaches. This method

allows to assess the adequacy for the ego vehicle to adapt

its own strategy in order to facilitate the merging maneuver

of a potential merging vehicle. As the main contribution the

improvement of two already existing planning approaches

using our method is presented. The limits between a good

2018 IEEE Intelligent Vehicles Symposium (IV)

Changshu, Suzhou, China, June 26-30, 2018

intention and planning algorithm and the re-planning ability

of the system are also analyzed.

II. RELATED WORK

Human drivers analyze and anticipate the trafﬁc situation.

In a similar manner, autonomous vehicles should integrate

a prediction of the behavior of other participants into their

driving task.

The problem of robots interacting on populated human

environments is not only focus of autonomous driving but

also point of interest of other robotic ﬁelds. Bennewitz et

al. [1] presented a method to predict the trajectories of per-

sons and improve the navigation behavior of a mobile robot.

Kuderer [2] and Kretzschmar [3] present a cooperative navi-

gation model for mobile robots interacting with pedestrians.

Nevertheless, when navigating on freeways and highways,

the topology is more structured and the velocities of the

trafﬁc participants are higher, which requires consideration

of speciﬁc solutions.

In a near future, as presented by Hobert [4], the intercom-

munication between vehicles and infrastructure (V2X) will

allow to acquire a precise information about the intentions of

the trafﬁc participants and the evolution of the situations. The

authors of [5] proposed an on-ramp merging system which

assessed the road trafﬁc conditions and transmit the instruc-

tions to the vehicles on the surrounding area. The system

performs well but relies on an advanced infrastructure. V2X

technology presents encouraging results but the technology

is still not mature enough.

Thourought the last years several prediction methods for

trafﬁc participants have been intensively studied in the lit-

erature. Different motion models like physic-based models,

maneuver-based models or interaction-aware based models

can be used for the prediction [6]. The integration of pre-

diction and intention information within the decision-making

process plays a crucial role in the system performance.

In [7] the authors compute the complete set of the

collision-free start and end points of trajectories for each

vehicle. This kind of intensive computations provide accurate

results in detriment of the online capability of the system.

Other strategies take the current most likely prediction and

rely on a continuous update of the available information and

fast re-planning system, as the multilevel planning system

presented by Men

endez-Romero et al. [8]. Carvalho et al. [9]

integrate the most likely cut-in prediction information to

improve the autonomous cruise control. The combination of

the most likely prediction with a fast re-planing works for

most of the situations quite well, but still does not consider

other possible interactions between the agents involved.

Iterative planning strategies combine the planning and pre-

diction tasks. Wei et al. [10] propose an intention prediction

based strategy generation. In [11] the authors suggest a game

theoretic approach which can model the re-planning capa-

bilities of the drivers. In [12] the decision-making is based

on Partially Observable Markov Decision Process (POMDP).

The multi-policy decision-making presented by Cunningham

et al. [13] also simulates the scene evolution using the most

likely evolution of the other agents involved in order to

reason about the policies. The problem with such iterative

planning approaches is that they only consider the most

likely evolution of the other trafﬁc participants. Especially in

longer prediction horizons the model predictions can become

inaccurate and overlook some critical situations.

One step further, our proposed system not only anticipates

the behavior of other trafﬁc participants to improve their own

safety but also plans a cooperative behavior to improve the

aggregate trafﬁc comfort.

III. APPROACH

The objective of this work is to provide our system with

a courtesy behavior, which identiﬁes the intention of other

trafﬁc participants and assesses the cost of adapting the own

ego strategy. We achieve this by integrating an intention

prediction algorithm into the decision-making. Thereby we

gain a better foresight of the scene evolution by including

all possible outcomes to provide our system with robustness

over false predictions. Decisions are made based on maxi-

mizing the expected utility of the involved trafﬁc participants.

A. Problem and Task Description

We semantically determine the possible actions for the

ego and the conﬂicting vehicle as shown in Tables I and II.

All possible ego actions are clustered as No Cooperative

(NC) and Cooperative (CO). Thus, the action space of

the ego vehicle is deﬁned as A

eg o

:= {NC, CO}. For the

conﬂicting vehicle normal, courtesy and forced merge actions

are combined into the No Yield actions (N Y ), corresponding

to merging in front of the ego vehicle, and the Yield actions

(Y ). Thus, the action space of the conﬂicting vehicle is

deﬁned as A

:= {Y, NY }. The Cartesian product A =

eg o

× A

deﬁnes the joint action space.

Our goal is to choose an action for the ego vehicle so that

the combined expected utility (U(a)) is maximized, i.e.,

∗

eg o

= argmax

a∈A

ego

E(U(a))

= argmax

a∈A

ego

∈A

p(a

| a

eg o

) · U(a

eg o

, a

(1)

This makes necessary to predict the intention of the merging

vehicle p(a

| a

ego

), which is presented in the next sec-

tion. The description of the expected utilities is given in

Section III-C.

B. Prediction Module

In the state-of-the-art methods, the costs of actions are

computed for the most likely actions of other trafﬁc partici-

pants. But these approaches lack the robustness against false

predictions. Our aim, on the other hand, is to predict the

probability of unlikely behavior of other trafﬁc participants,

speciﬁcally the misprediction probability for the conﬂicting

vehicle. Using this probability we can compute the expected

utility of an ego action considering two possible decisions

of the conﬂicting vehicle.

In order to predict the probability of unlikely outcomes, we

use an ensemble learning based Gentle Boost classiﬁer [14].

944

of 6

50墨值下载

database

关注

评论