Toward the Design of Personalised Adaptive Driver Assistance for Truck

Docking

Marc-Andr

e B

uchner

1 a

, Pedro Ribeiro

1 b

, Andr

e Frank Krause

1 c

, Jan Benders

, Christian

Ressel

, and Kai Essig

Rhine-Waal University of Applied Sciences, 47475 Kamp-Lintfort, Germany

HAN University of Applied Sciences

{marc-andre.buechner2, pedro.ribeiro, andrefrank.krause, christian.ressel,

Keywords:

Advanced Driver Assistance, Personalization, Adaptive User Interface, Expertise Estimation, Machine

Learning, Support Vector Machine, Truck Docking, Virtual Reality Simulator

Abstract:

Backwards manoeuvring of articulated vehicles towards the loading bay at a distribution centre has been ac-

knowledged as a signiﬁcant source of accidents. Individualized driver assistance aimed to support professional

truck drivers in this challenging scenario is still very much in its early days. This paper presents the ongo-

ing design and development of a user interface and expertise estimation for adaptive and personalised driver

assistance to support truck drivers through the scenario of parking a semi-trailer truck in a loading bay. The

proposed system aims to adapt in real-time based on the current driving performance and adaptation rules.

1 INTRODUCTION

Performing the manoeuvre of docking a truck combi-

nation towards the loading bay of a distribution cen-

tre can be a very challenging task. This compli-

cated manoeuvre can potentially result in accidents

and costs due to damage caused, for example, by

driving errors or information overload (Kural et al.,

2016; Kusumakar et al., 2017). Generally, the main

causes of truck accidents are related to human fac-

tors, technical failures or environmental conditions

(Union, 2007) - with the ﬁrst one being the most im-

pactful cause. There is a plethora of aspects that in-

ﬂuence the complexity of this speciﬁc task of docking

a truck combination, some are driver related such as

the experience, physical, emotional or perceptual fac-

tors, and others are associated with the environment

and vehicle conditions. In that context, the platform

VISTA-Sim (Ribeiro et al., 2021) was being devel-

oped in the project VISTA (VIsion Supported Truck

docking Assistant).

VISTA-Sim uses a virtual reality (VR) environment

to simulate and investigate driver performance, train

drivers and develop and evaluate different forms

https://orcid.org/0000-0003-0530-2975

https://orcid.org/0000-0002-9829-8268

https://orcid.org/0000-0002-9685-278X

of context-sensitive and personalised driver assis-

tance provided by the In-Vehicle Information System

(IVIS) that is running on a virtual tablet installed in

the truck cabin (see Figure 1). One of the pivotal

advantages of the VR simulator is, that it allows the

tracking of driver and vehicle behaviour data, which

can be interpreted to discover driver problems and to

adapt personalised assistance.

This paper aims to further develop the VISTA-Sim

platform with a particular focus on a mechanism

that will allow the personalised adaptation of driver

assistance. For that, this mechanism shall be ca-

pable of learning from the driver’s behaviours, es-

timating driver performance and deciding upon the

most appropriate driver assistance. The develop-

ment of such a mechanism is a challenge: Humans,

particularly drivers, might be hesitant to follow AI-

generated instructions that are not easily justiﬁed, in-

terpretable and ultimately trustworthy (Arrieta et al.,

2020). Therefore, this paper aims to apply the pro-

cess of personalisation in which the driver assistance

is tailored to match the driver’s needs and prefer-

ences (Hasenj

ager et al., 2019). Accordingly, it is

intended to explore a hybrid approach that combines

the potential of machine learning (ML) with back-

ground knowledge in the form of logical statements.

This paper presents the current progress of a person-

alised adaptive user interface for an assistive applica-

BÃijchner, M., Ribeiro, P., Krause, A., Benders, J., Ressel, C. and Essig, K.

Toward the Design of Personalised Adaptive Dr iver Assistance for Truck Docking.

DOI: 10.5220/0011907100003612

In Proceedings of the 3rd Inter national Symposium on Automation, Information and Computing (ISAIC 2022), pages 131-137

ISBN: 978-989-758-622-4; ISSN: 2975-9463

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

131

tion that supports drivers performing the backwards

manoeuvring of articulated vehicles. The proposed

mechanism aims to automatically adjust the driver as-

sistance based on a performance estimation module

that uses the VISTA-Sim capacities to track the driv-

ing behaviour. For example, the steering wheel an-

gle, trailer’s position and rotation, driver’s head po-

sition and orientation and eye-tracking data such as

the direction of the gaze ray and ﬁxated objects can

be fed into the performance estimation module. The

performance-dependent assistance is adjusted by ap-

plying ML-based expertise estimation in combination

with a rule-based adaptation mechanism.

2 RELATED WORK

Fan and Poole (Fan and Poole, 2006) deﬁned the per-

sonalisation of digital technologies as “a process that

changes the functionality, interface, information ac-

cess and content, or distinctiveness of a system to in-

crease its personal relevance to an individual or a cate-

gory of individuals”. They also proposed a high-level

framework to classify approaches to personalisation

along three dimensions: (1) what to personalise (e.g.

content and functionality of the driver assistance pro-

vided by the IVIS); (2) to whom to personalise (e.g.

drivers); and (3) the explicitness of the personalisa-

tion or how the personalisation is achieved (Fan and

Poole, 2006). Regarding the third dimension, there

are two possibilities: (1) the implicit personalisation

where the information needed for personalisation is

collected automatically; and (2) the explicit person-

alisation where the driver has direct control over the

parameterization needed for the personalisation.

One of the current trends in the automotive sector

is the exploration of the personalisation of advanced

driver assistance systems (ADAS) and the individ-

ualised adaptation of the driver-vehicle interaction

(Hasenj

ager et al., 2019; Lilis et al., 2017). Prop-

erly adapting the IVIS functionalities according to

the individual characteristics and needs of a driver

requires that the driver remains in the loop (Riener

et al., 2016). Personalisation is seen as an impor-

tant safety aspect as it has the potential to prevent

distractions and cognitive overload, and ultimately re-

duce the risks of accidents (Lilis et al., 2017; Ulahan-

nan et al., 2021). Additionally, according to a recent

study, the personalisation of autonomy increases the

familiarity and willingness to trust Autonomous vehi-

cles (AVs) or self-driving cars (Sun et al., 2020).

As mentioned previously, there are two approaches to

personalise IVIS functionalities: the explicit and the

implicit approach (Fan and Poole, 2006; Stuerzlinger

et al., 2006). Although the explicit approach has the

potential to promote a personalisation that is better

aligned with the driver’s expectation it also has some

disadvantages. First, it requires drivers to direct at-

tention and effort toward the process of personalisa-

tion, which might eventually distract them from the

docking process. This direct control may also result

in a limitation of the number of parameters that can be

subjected to personalisation. Additionally, this limi-

tation can be further aggravated by the “inaccessibil-

ity” of parameters which are difﬁcult to understand

(Hasenj

ager et al., 2019). Finally, the driver may not

be completely aware of what information is neces-

sary to maximise the driving performance. In fact,

even the safety might be compromised by a decision

grounded on preferences that can prevent the presen-

tation of critical information (Ulahannan et al., 2020;

Ulahannan et al., 2021).

With the implicit approach, the information provided

by the driving assistance is automatically presented

or hidden. With this approach, the driver might ex-

perience that the information provided by the IVIS is

not under his/her control and might not even under-

stand the rationale of the driving assistance provided

(Hasenj

ager et al., 2019). However, the implicit ap-

proach has the beneﬁt of reducing the effort, the at-

tention and the cognitive load. Additionally, it has the

potential to offer a more complex, precise and safety-

oriented adaptation (Hasenj

ager et al., 2019).

To implicitly (automatically) adapt the information

being presented in the driving assistance, it is nec-

essary to determine which processes will inﬂuence

the adaptation. In that sense, a data-driven process

is commonly employed that implies the creation of

a driver model which is based on the observation of

driving data (Wang et al., 2014; Lin et al., 2014). Us-

ing such a driver model, it is then possible to recog-

nize driving behaviour or expertise in certain tasks,

e.g. parking an articulated vehicle.

Besides this driver model, other information needs to

be considered as an input for the adaptation (Khan and

Khusro, 2020). For instance, information about the

environment, the vehicle, the IVIS, as well as the pref-

erences and proﬁles of the drivers. Based on such in-

put, the application of a rule-based approach has been

successfully employed to adapt the information being

presented to the user (Hussain et al., 2018; Khan and

Khusro, 2020). For instance, Hussain et al. (2018)

followed a model-based approach to adapt the user

interface. This approach is based on three different

models: (1) the context model (environmental factors

such as light level); (2) the user model (human fac-

tors such as experience or abilities) and (3) the device

model (such as the characteristics of the hardware and

ISAIC 2022 - International Symposium on Automation, Information and Computing

132

software). The resulting system is then capable to ap-

ply, at runtime, adaptation rules devised to properly

adapt the user interface according to the current con-

text. For example, whenever the light level is too low

the user interface is adapted to a night mode.

3 VR-SIMULATOR PLATFORM

In our VR-Simulator (VISTA-Sim, see Figure 2), the

driver wears a Virtual Reality headset that immerses

the driver inside a truck cabin located at a virtual dis-

tribution centre. Inside the VR truck cabin, there is

also an IVIS running on a virtual tablet computer. It

is responsible to provide driving assistive information

showing for example a bird-eye view of the distri-

bution centre, the distance to the target loading bay

and steering recommendations. Finally, VISTA-Sim

also allows the driver to operate the truck through

a steering wheel and pedals. VISTA-Sim is com-

posed of four main components: (1) a path plan-

ner combined with the path tracking controller which

runs in Simulink (MathWorks); (2) the VR-simulator

that uses the Unity 3D cross-platform game engine

(Unity); (3) the driver assistance Human Machine In-

terface (HMI); and (4) the Behaviour Analysis Mod-

ule (Ribeiro et al., 2021).

The focus of this paper is to present the current work

progress of the component “Driver Assistance HMI”.

The focus is on the design and implementation of an

adaptation mechanism that aims to provide individual

driver assistance based on expertise-estimation and

driving performance.

Figure 1: The high-level view of the VISTA-Sim architec-

ture.

4 ADAPTATION MECHANISM

ADAS should provide individualised and anticipative

feedback that minimizes driver distraction and pro-

vides only situation-speciﬁc, highly relevant informa-

tion. The proposed adaptation mechanism aims to

adaptively control the several IVIS subsystems to pro-

vide the best ﬁtting support for drivers given the cur-

rent context. The IVIS adaptation mechanism relies

on three components: (1) the Rules; (2) the Adap-

tor; and (3) the Models. The machine learning model

that estimates driver expertise (see section 4.2) is part

of the Models-component. As input, the adaptation

mechanism receives a real-time expertise estimation

from a driver model that is responsible to determine

how good a driver is at performing the parking ma-

neuver (Ribeiro et al., 2021). The estimation is based

on real-time driving data recorded by the simulator

(e.g. steering wheel angle, head movements and other

data) and is used to determine when the Adaptive User

Interface(AUI) should be active. For example, if the

estimation predicts expert behavior, then the AUI is

currently not needed. If novice driver behavior is de-

tected, the system will adjust the IVIS in ﬁxed inter-

vals to match the current needs of the user.

The Rule-based component describes when and under

which conditions changes to the IVIS User Interface

(UI) are applied, utilizing the commonly used Event,

Condition, Action (ECA) Approach (Hussain et al.,

2018). The event determines when a Rule is applied,

e.g. when a novice driver is detected by the perfor-

mance estimator. The Action states the change that

has to happen when the Event is triggered, like en-

abling more detailed assistance on the UI. The Condi-

tion holds information about what prerequisites need

to be met so that a Rule can be applied, e.g. a disabled

UI would prevent the application of any Rule with the

condition “UI is on”. The ECA template was mod-

iﬁed to have ﬁve components. First, a unique id to

identify and distinguish it from other rules. Second, a

description of what this rule does when and which UI

elements will be affected in a human-readable way.

Third, a collection of events that state when the rule

is triggered. Fourth, a collection of conditions that

prevent or enable it. Finally, the ﬁfth component is

a collection of actions that specify what happens to

which UI element.

On a conceptual level, the Adaptor component (see

Figure 3) is composed of: (1) the Evaluator; (2) the

Conﬂict resolver; and (3) the UI-Adaptor.

Whenever the Evaluator (see Figure 4) detects that

a value from the Performance estimator matches the

event of one or more rules, it checks if there is any

condition that prevents their application. When an

event matches multiple rules, it is important to verify

if the rules can be applied together. For example, the

rules “In the ﬁrst phase of the manoeuvre the steering

assistance should be disabled” and “If the ﬁrst phase

of the manoeuvre duration takes longer than two min-

Toward the Design of Personalised Adaptive Driver Assistance for Truck Docking

133

utes the steering assistance should be enabled” may

have a potential conﬂict.

Figure 2: Detailed view of the AUI Evaluator module of

the Adaptor-component. It uses four boolean decisions to

select a ﬁtting set of rules for the current context.

In that sense, the role of the Conﬂict resolver compo-

nent is to detect and solve conﬂicts between the two

rules. The conﬂicts will be detected by checking the

actions that are associated with the respective rules.

For example, by analysing if the rules affect the same

UI object or enable the same UI modality. After de-

tection, conﬂict resolution strategies must be applied.

Although this component is still under development,

it is envisioned to extend rules with priorities and/or

the expected utility of an action (e.g. a rule with the

“safety” property will have a higher priority than one

without). Both approaches allow to resolve conﬂicts

between rules by excluding rules that ﬁt less the cur-

rent requirements of docking safely. If there are con-

ﬂicting rules remaining after the conﬂict resolution,

then they are all equally ﬁt for the current context and

the resolver will return the ﬁrst. Ultimately, even the

active involvement of the driver in the process of con-

ﬂict resolution is under consideration. For example, a

user’s preference for a certain rule can be used to se-

lect this rule among the remaining rules after conﬂict

resolving. As the last step, a UI-Adaptor component

changes speciﬁc UI- elements according to the speci-

ﬁed action of the prioritised rule.

The Models component contains information that fa-

cilitates the selection of the rules according to their

applicability. Examples are task phases, user prefer-

ences or user abilities. Since this component is still

at an early stage of conceptualization it will not be

further discussed in this paper.

4.1 Implementation in Unity

The functional prototype is being implemented in

Unity 3D according to the conceptualization men-

tioned in Section 4. The adaptation process starts

in the Performance estimator where a generic Unity-

Event is triggered whenever the driver’s expertise hits

a value range that corresponds to a level of expertise

(e.g. novice). Each level of expertise has an event at-

tached and when triggered. It invokes a list of Rule

IDs (strings keys) that are supposed to be executed.

For example, for the event “driver expertise is equal

to Novice” we could have one or more rules associ-

ated, such as “enable HMI”.

The Rule IDs are then processed in the Evaluator-

Class where they are used as keys to select the cor-

rect Rule-Object (Rule) from a map. The implemen-

tation of a Rule contains its Actions, which are com-

posed of a reference to all the Adaptable-UI-Objects

that are affected by it as well as which methods are to

be called and their parameters. Afterwards, the Ac-

tions reference is used to fetch its objects from the

Object-Map to adapt the UI.

The implementation of the adaptation does not have

a singular UI Adaptor. Instead, the adaptation is han-

dled by each Adaptable-UI-Object by a script that in-

herits from a virtual base class which provides them

with the Execute-Method that is called whenever a

rule is applied. The individual version of Execute-

Method handles how the adaptation is done. The cur-

rent implementation of the Adaptation mechanism is

not yet capable of taking system states or the user into

account or resolving conﬂicts.

4.2 Driver Expertise Estimation Model

Using machine learning, the system can adapt to the

driver based on previous behaviour. If the trained

system detects unexpected driver behaviour in a fa-

miliar situation, it can trigger warnings or supportive

feedback. Here, a ML-model was trained to predict

the driver’s expertise on-the-ﬂy within a certain time-

frame. If a situation of low expertise is detected, the

system will provide context sensitive feedback to ease

the truck docking process.

Assuming that support can be helpful especially

during initial training of the docking process (e.g.

for novices), the ML-classiﬁer should provide well-

balanced classiﬁcation accuracy for both expert- and

novice behaviour. That way, novices will be provided

with extensive support, while expert drivers will see

only very speciﬁc, non-disturbing support.

To train a ML-model, a small dataset was collected

using the VISTA Sim: Four participants with no prior

truck driving experience performed a familiarization

session at the ﬁrst day of docking training. Three

days later, the ﬁrst training session started, followed

by another session the next day and a ﬁnal session

ISAIC 2022 - International Symposium on Automation, Information and Computing

134

at the third day. Each day, ﬁve to six trials were

recorded. Recording of a trial was stopped if dock-

ing failed after seven forward-backward reposition-

attempts (turns). The maximum turn amount was

based on a previous pilot-dataset, where an already-

trained driver completed docking with no more than

ﬁve turns. Two extra turns were given to ease initial

learning progress and to aquire more trials with a rat-

ing above zero (see below).

The ﬁnal dataset consists of 74 trials. Each trial in the

dataset was rated by the ﬁrst-author on a scale from

0 to 5 using the docking quality as an objective cri-

terion (rating: 0 = did not park; 1,2: unloading not

possible; 3,4: unloading possible; 5: perfect dock-

ing). This docking-quality rating was used to separate

trials into novice- and expert trials. The dataset con-

tains a large number of trials rated with 0, underlining

that truck-docking is indeed a quite complicated task.

Most machine learning methods work best given a

balanced dataset with equal number of trials per class.

Because of the comparatively small number of trials

and the strong imbalance, the dataset was split into

only two classes: expert- and novice trials. All trials

with a zero-rating (’did not park’) were assigned to the

novice class (n=41) and all other trials were assigned

to the expert class (n=33). With these two classes, the

dataset is still moderately imbalanced.

Figure 3: Features for ML-learning, extracted from the

dataset. Some features correlate with expertise, e.g. head

horizontal movements (e) and the cabin-trailer angle (f).

The recorded dataset was analysed with respect to

several key performance indicators suggested by a

professional truck-driving instructor (for details see

(Ribeiro et al., 2021)). Here, we focus on perfor-

mance indicators that are locally observable inside

the truck cabin: docking duration, gear shifts, vehicle

acceleration, steering wheel movements, head move-

ments and the cabin-trailer angle. Figure 5 shows that

some of the analysed performance indicators clearly

correlate with the expertise level, visible by non- or

slightly overlapping box-plots.

For comparison with ML-trained based classiﬁers, a

simple baseline-classiﬁer (BC) was implemented us-

ing the three features with strong predictive power

(vehicle acceleration, head movements and cabin

trailer angle, see Figure 5). The BC performs

simple thresholding per feature and combines the

feature-speciﬁc classiﬁcation results using majority

voting. Optimal threshold values were selected by

grid-search. With two classes, we have a binary clas-

siﬁcation task: Expert trials were assigned to the pos-

itive class, and novice trials to the negative class. Ap-

plied to the complete dataset and full trial-length, the

BC yields within-class accuracies of 82% for the ex-

pert class (i.e. true-positive rate (TPR), also known as

recall / sensitivity) and 90% for the novice-class (true

negative rate (TNR), also known as speciﬁcity) and a

balanced accuracy (the average of TPR and TNR) of

86%. This is a promising but probably over-optimistic

result, because the BC-thresholds were adjusted given

the full dataset.

A classiﬁer trained on full trials can only provide

an expertise estimation at the end of a trial. This

estimation could be used to provide adequate feed-

back for the next docking attempt, but a better so-

lution would be a classiﬁer that can estimate driver

performance within a shorter timeframe (sliding win-

dow approach). Therefore, support vector machines

(SVM) were trained on three different, bootstrapped

datasets consisting of sub-trials with a predeﬁned du-

ration (25 seconds, 50s and 75s). From each trial

in the dataset, 50 sub-trials were randomly sampled.

The SVMs (RBF kernel, C=0.001) were trained using

leave-one-out cross-validation, with generalisation

evaluated at the level of individual trials. Hence, each

trial was left out once, resulting in 74 folds. The train-

ing set of each fold was balanced using the synthetic

minority over-sampling technique (SMOTE, (Chawla

et al., 2002)), because SVM training is very sensi-

tive to class-imbalance (Akbani et al., 2004). Cross-

validation was repeated ten times, because SMOTE is

stochastic.

SVM-Training Results: The balanced accuracy de-

pends - as expected - on the window size. It drops

only slightly from a window size of 75s to a window

size of 50s (see Figure 6). Hence, a duration of 50

seconds appears to be a suitable compromise between

“classiﬁer response delay” (time until the ﬁrst exper-

tise estimation is available) and accuracy. Accuracy –

especially for the expert class – drops noticeably com-

pared to observing full trials (see baseline classiﬁer).

This means that the proposed hybrid approach, mix-

ing ML and the rule-based approach might help to re-

duce uncertainty regarding observed driver expertise

within shorter timeframes.

Toward the Design of Personalised Adaptive Driver Assistance for Truck Docking

135

Figure 4: The performance of trained SVMs depends on the

observation timeframe (window size in seconds).

5 CONCLUSIONS & FUTURE

WORK

In this paper, we have presented ongoing work that

targets the design and development of an adaptation

mechanism for driver assistance. This adaptation

mechanism will provide individualised driver assis-

tance based on estimated driver expertise and aims to

support professional truck drivers in the challenging

task of backwards manoeuvring articulated vehicles

towards a loading bay at a distribution centre.

Rule-based adaptation: As a next step, the design

of the adaptation rules will be guided by a detailed

analysis of the thinking aloud protocols and question-

naires collected during recording of the dataset, e.g.

problems experienced while docking. The observa-

tions and comments will inform a post-interview on

which the adaptation rules will be based. The not

yet fully implemented AUI components will be fur-

ther developed. The development of a task model

will require a task analysis (a breakdown into phases,

tasks, and subtasks) of the docking manoeuvre. Fur-

ther, a detection mechanism is required to determine

in which phase the driver currently is. The rule adap-

tion needs such a detector to know what the driver is

currently supposed to do to apply rules that are made

for this phase/task or sub-task. The Driver model (i.e.

an individual proﬁle) will require further research on

user proﬁles for automotive applications. Individ-

ual driver proﬁles will allow the system to also take

the current driver requirements (preferences, impair-

ments, skills, etc.) into account when applying rules.

For example, such an individual user model could be

initialized once by the driver and then continuously

updated with a running-average of previous docking-

performances to provide a robust and stable adapta-

tion of the UI. For the Resolver, the concept described

in section 4 needs to be implemented, and the priority

and the expected utility properties need to be deﬁned

further.

Machine learning: The current dataset is compara-

tively small. Larger or more complex models like

recurrent neural networks (RNN) or attention-based

RNN (Vaswani et al., 2017) might show overﬁtting.

Therefore, it is planned to evaluate and compare spe-

cial machine learning methods well suited for small

datasets, e.g. few-shot learning (Wang et al., 2020).

Further, methods with a low energy consumption and

fast training times (e.g. random projection methods,

(Krause et al., 2013)) will be explored. This is impor-

tant for the future goal of continuous, “lifelong” ma-

chine learning (Chen and Liu, 2018) on a low-power,

embedded system inside a vehicle to continuously up-

date an individual expertise estimator model for each

driver.

Future work: In the future, it is also planned to do

a study regarding implicit and explicit adaptation be-

cause - as shown in previous studies (Hussain et al.,

2018; Bongartz et al., 2012; Todi et al., 2021) - fre-

quent, unnotiﬁed changes can frustrate and confuse

the user. To answer, if implicit or explicit adaptation

is better for the AUI of the VISTA-Sim, it is planned

to do an A/B test where one group performs docking

with an implicit and the other with an explicit version

of the AUI. Explicit here would likely mean by voice

command e.g. ”Can you help me” which would trig-

ger the adaptation of the AUI. To not waste too many

resources by implementing both versions a Wizard of

Oz approach is likely.

ACKNOWLEDGEMENTS

This Interreg V-A project was co-ﬁnanced by the

European Union via the INTERREG Deutschland-

Nederland programme, the Ministerium f

Wirtschaft, Innovation, Digitalisierung und En-

ergie of Nordrhein-Westfalen and the Province of

Gelderland.

ISAIC 2022 - International Symposium on Automation, Information and Computing

136

REFERENCES

Akbani, R., Kwek, S., and Japkowicz, N. (2004). Applying

support vector machines to imbalanced datasets. In

European conference on machine learning, pages 39–

50. Springer.

Arrieta, A. B., D

ıaz-Rodr

ıguez, N., Del Ser, J., Bennetot,

A., Tabik, S., Barbado, A., Garc

ıa, S., Gil-L

opez, S.,

Molina, D., Benjamins, R., et al. (2020). Explainable

artiﬁcial intelligence (xai): Concepts, taxonomies, op-

portunities and challenges toward responsible ai. In-

formation fusion, 58:82–115.

Bongartz, S., Jin, Y., Patern

o, F., Rett, J., Santoro, C., and

Spano, L. D. (2012). Adaptive user interfaces for

smart environments with the support of model-based

languages. In International Joint Conference on Am-

bient Intelligence, pages 33–48. Springer.

Chawla, N. V., Bowyer, K. W., Hall, L. O., and Kegelmeyer,

W. P. (2002). Smote: synthetic minority over-

sampling technique. Journal of artiﬁcial intelligence

research, 16:321–357.

Chen, Z. and Liu, B. (2018). Lifelong machine learning.

Synthesis Lectures on Artiﬁcial Intelligence and Ma-

chine Learning, 12(3):1–207.

Fan, H. and Poole, M. S. (2006). What is personaliza-

tion? perspectives on the design and implementation

of personalization in information systems. Journal of

Organizational Computing and Electronic Commerce,

16(3-4):179–202.

Hasenj

ager, M., Heckmann, M., and Wersing, H. (2019).

A survey of personalization for advanced driver assis-

tance systems. IEEE Transactions on Intelligent Vehi-

cles, 5(2):335–344.

Hussain, J., Ul Hassan, A., Muhammad Bilal, H. S., Ali, R.,

Afzal, M., Hussain, S., Bang, J., Banos, O., and Lee,

S. (2018). Model-based adaptive user interface based

on context and user experience evaluation. Journal on

Multimodal User Interfaces, 12(1):1–16.

Khan, I. and Khusro, S. (2020). Towards the design of

context-aware adaptive user interfaces to minimize

drivers’ distractions. Mobile Information Systems,

2020.

Krause, A. F., Essig, K., Piefke, M., and Schack, T.

(2013). No-prop-fast-a high-speed multilayer neu-

ral network learning algorithm: mnist benchmark and

eye-tracking data classiﬁcation. In International Con-

ference on Engineering Applications of Neural Net-

works, pages 446–455. Springer.

Kural, K., Besselink, I., Xu, Y., Tomar, A., and Nijmeijer,

H. (2016). A driver support system for improved ma-

neuvering of articulated vehicles using an unmanned

aerial vehicle. In HVTT14: International Sympo-

sium on Heavy Vehicle Transport Technology, Ro-

torua, New Zealand.

Kusumakar, R., Kural, K., Tomar, A. S., and Pyman, B.

(2017). Autonomous parking for articulated vehicles.

HAN University of Applied Science, Netherlands.

Lilis, Y., Zidianakis, E., Partarakis, N., Antona, M., and

Stephanidis, C. (2017). Personalizing hmi elements

in adas using ontology meta-models and rule based

reasoning. In International Conference on Universal

Access in Human-Computer Interaction, pages 383–

401. Springer.

Lin, N., Zong, C., Tomizuka, M., Song, P., Zhang, Z., and

Li, G. (2014). An overview on study of identiﬁcation

of driver behavior characteristics for automotive con-

trol. Mathematical Problems in Engineering, 2014.

Ribeiro, P., Krause, A. F., Meesters, P., Kural, K., van Kolf-

schoten, J., B

uchner, M.-A., Ohlmann, J., Ressel, C.,

Benders, J., and Essig, K. (2021). A vr truck docking

simulator platform for developing personalized driver

assistance. Applied Sciences, 11(19):8911.

Riener, A., Boll, S., and Kun, A. L. (2016). Automotive

user interfaces in the age of automation (dagstuhl sem-

inar 16262). In Dagstuhl reports, volume 6. Schloss

Dagstuhl-Leibniz-Zentrum fuer Informatik.

Stuerzlinger, W., Chapuis, O., Phillips, D., and Roussel, N.

(2006). User interface fac¸ades: towards fully adapt-

able user interfaces. In Proceedings of the 19th an-

nual ACM symposium on User interface software and

technology, pages 309–318.

Sun, X., Li, J., Tang, P., Zhou, S., Peng, X., Li, H. N., and

Wang, Q. (2020). Exploring personalised autonomous

vehicles to inﬂuence user trust. Cognitive Computa-

tion, 12(6):1170–1186.

Todi, K., Bailly, G., Leiva, L., and Oulasvirta, A. (2021).

Adapting user interfaces with model-based reinforce-

ment learning. In Proceedings of the 2021 CHI Con-

ference on Human Factors in Computing Systems,

pages 1–13.

Ulahannan, A., Cain, R., Thompson, S., Skrypchuk, L.,

Mouzakitis, A., Jennings, P., and Birrell, S. (2020).

User expectations of partial driving automation capa-

bilities and their effect on information design prefer-

ences in the vehicle. Applied Ergonomics, 82:102969.

Ulahannan, A., Thompson, S., Jennings, P., and Birrell,

S. (2021). Using glance behaviour to inform the de-

sign of adaptive hmi for partially automated vehicles.

IEEE Transactions on Intelligent Transportation Sys-

tems, 23(5):4877–4892.

Union, I. R. T. (2007). A scientiﬁc study “etac” european

truck accident causation: Executive summary and rec-

ommendations.

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones,

L., Gomez, A. N., Kaiser, Ł., and Polosukhin, I.

(2017). Attention is all you need. Advances in neural

information processing systems, 30.

Wang, W., Xi, J., and Chen, H. (2014). Modeling and rec-

ognizing driver behavior based on driving data: A sur-

vey. Mathematical Problems in Engineering, 2014.

Wang, Y., Yao, Q., Kwok, J. T., and Ni, L. M. (2020). Gen-

eralizing from a few examples: A survey on few-shot

learning. ACM computing surveys (csur), 53(3):1–34.

Toward the Design of Personalised Adaptive Driver Assistance for Truck Docking

137