Self-Report Data

We performed a dependent t-test for paired samples to analyse the perceived levels of distraction, helpfulness, liking, and comfort of the visual and olfactory modalities experienced in the process of driving (see Figure 11.3).

0

Figure 11.3: Mean ratings of distraction, helpfulness, liking, and comfort of the visual and olfactory modalities (1= "Not at all", 7= "Very much"). Error bars, ± s.e.m., ∗ ∗ ∗p < .001

The olfactory modality (M= 2.29, SD= 1.45) has been reported significantly less distracting (t(20)= 5.510, P<.001) than the visual modality (M= 4.57, SD= 1.78).

Par-ticipants also found the olfactory modality (M= 5.00, SD= 1.14) more helpful (t(20)=

-5.477, P<.001) in understanding the notifications than the visual (M= 3.00, SD= 1.87).

Moreover, the olfactory modality (M= 5.38, SD= 2.19) was liked significantly more (t(20)=

-7.345, P<.001) than the visual modality (M= 2.19, SD= 1.40). Finally, the olfactory modality (M= 5.19, SD= 1.25) was perceived significantly more comfortable (t(20)=

-4.298, P<.001) than the visual (M= 3.10, SD= 1.64).

11.5 D ^ISCUSSION

This paper is the first to demonstrate the use of olfactory conditioning to instruct drivers which olfactory notification is assigned to which driving-relevant message (based on the conditioning procedure of [115]). The results show that all participants were able to correctly assign at least six scents out of nine (in line with prior work achieving >60%

success rate [104]).

Moreover, the driving behaviour data shows that participants made significantly less mistakes when receiving visual-olfactory notifications. This is in line with previous findings on the positive effect of ambient scents on performing the driving task (e.g.

lemon scent promoted better braking performance in [130]). Such results suggest that scents are a promising notification modality in the car.

Furthermore, in our study, olfactory feedback was perceived less distracting, more helpful, and more comfortable than the visual notifications alone. The olfactory modal-ity was also liked more than the visual. This might be because visual notifications were not salient enough. However, the findings on scents match the current automotive trends regarding the wellbeing in the car (e.g. like in a new Mercedes-Benz [35], Bentley [18], or BMW [20].

The driving simulator software we have used does not allow customisation of visual notifications. Results might change if different visual stimuli (e.g. icons) and in a different location on the screen are used. However, there is evidence that the olfactory feedback in combination with the visual information can help us become more aware of notifications without the need to shift our attention. Understanding crossmodal integrations can enable better design of in-vehicle notification systems. More studies on multimodal in-car interaction (e.g. like [164]) should be carried out.

Although olfaction is related to many constraints, including interpersonal and cul-tural differences, health issues (e.g. adverse reactions to certain scents), and ventilation, our findings underline the benefits of the olfactory modality. When the olfactory

stimu-lation is applied, controlling the delivery parameters [45], we can achieve both a better interaction performance and a better user experience. Such a finding is very important for the design of in-car user interfaces.

11.6 C ONCLUSION AND F UTURE W ORK

The findings of our user study suggest that (when scent-delivery parameters are con-trolled) olfactory notifications can, not only increase our hedonic experience, but also act as a non-distracting, helpful, and comfortable interaction modality. Moreover, ol-factory feedback has the potential to improve our driving behaviour, giving us hints we could have missed when relying on visual stimuli only (e.g. in cases of exceeding the speed limit, short inter-vehicle distance, and lane departure).

Future research can now start investigating scents for more complex driving tasks (e.g. overtaking slower vehicles [121]) and other notifications (e.g. "Traffic jam ahead").

It is also worth exploring olfactory notifications in the presence of a secondary task (e.g. using a radio or a touchscreen [147]) and in combination with other visual (e.g.

ambient lights [123]) and auditory [27] stimuli. Finally, studies in a real car interior would demonstrate the effectiveness of olfactory notifications in the presence of other ambient scents and scent absorbing materials.

12 | S(C)ENTINEL - Monitoring Automated Vehicles With

Olfactory Reliability Displays

Wintersberger, P.^2,3,∗, Dmitrenko, D.^1,∗, Schartmüller, C.^2,3, Frison, A.K.^2,3, Maggioni, E.¹, Obrist, M.¹, Riener, A.²

1SCHI Lab, CTRG, School of Engineering and Informatics, University of Sussex, Brighton, United Kingdom

2Technische Hochschule Ingolstadt, Germany

3Johannes Kepler University, Linz, Austria

∗Both first and second author contributed equally.

In IUI ’19: Proceedings of the 24th International SIGCHI Conference on Intelligent User Interfaces, ACM, 2019.

DOI:https://doi.org/10.1145/3301275.3302332 Abstract

Overreliance in technology is safety-critical and it is assumed that it could have been a main cause of severe accidents with automated vehicles. To ease the complex task of monitoring vehicle behaviour in the driving environment, researchers have proposed to implement uncertainty displays. They allow the driver to estimate whether or not an upcoming intervention is likely. However, presenting uncertainty just visually adds more workload on drivers, who might also be engaged in secondary tasks. We suggest to use olfactory displays as a potential solution to communicate uncertainty and conducted a user study (N=25) in a driving simulator. Results of the experiment comparing both objective (task performance) and subjective (technology acceptance model, trust scales, semi-structured interviews) measures suggest that olfactory notifications could become a valuable extension for calibrating trust in automated vehicles.

12.1 I NTRODUCTION

Trust in automation is an important topic for a safe use of automated driving systems (ADSs) [216]. According to the classification of autonomy levels as proposed by SAE [33], ADSs currently available on the market mainly operate on level 2. Here, the driver is fully responsible for monitoring the vehicle’s actions and thus the overall safety. However, recent events (such as the fatal crash of a Tesla driver in May 2016, but also less critical situations) indicate that many drivers utilising such systems tend to overtrust them, and do not properly monitor ADSs even in scenarios they were not designed for [206].

This is especially dangerous when systems seem to work flawlessly for a long time and in varying situations [58]. Since monitoring is a challenging task, even for “highly motivated human beings” (irony of automation [12]), researchers have proposed to use so-called “reliability/uncertainty displays”, that have shown to provide benefits in both level 2 [17] and level 3/4 automated driving (AD) [78]. Such displays are able to reduce the chance of mode awareness failures while increasing situation awareness as well as system transparency, and thereby ultimately lead to better calibrated trust [149]. They present the actual system reliability (or uncertainty, what is the inversion of reliability, but still follows the same concept - which kind of information works better is still an ongoing research [149]) to the user to adjust his/her monitoring behaviour.

However, especially when drivers are visually engaged in secondary tasks, such dis-plays can merely act as “proxy” for the system state – instead of monitoring the vehicle and the environment itself, the driver has to frequently inspect the display, what still de-mands his/her visual attention. Since future intelligent and multimodal user interfaces should adapt to different types of users [159] while using the full range of human inter-action and communication capabilities [198], we claim that there is a need to evaluate other modalities for communicating reliability information. A potential modality in this regard could be the sense of smell, that, in contrast to other typical approaches (such as haptics [116] or auditory cues [114]), is still widely unused, but provides some unique advantages: The sense of smell is a very powerful interaction medium [106] enabling humans to extract meaningful information [184]. For example, it has been shown that odours trigger automatic and implicit retrieval of mental representations of information related to the object the scent is coming from [29], and enable automatic access to terms semantically related to odours [88]. Moreover, scents can be very efficient in activating the central neural system [110, 203, 10], which is essential to keep the driver alert and more attentive on the road [170]. Scents can also act as an arousing (e.g., when the

driver is tired or inattentive [220, 66]) or as a calming (e.g., when the driver is stressed [137, 89]) stimulus. In future automated vehicles (AVs), classical perception channels (i.e., visual and auditory) will often be occupied by secondary tasks (such as watching a video, what demands both visual and auditory attention), while olfactory notifications have proven to be a valid way to gain user’s attention [19]. Consequently, to the best of our knowledge, our study (see Figure 12.1) is the first experiment including olfactory notifications for trust calibration in AD.

Figure 12.1: Study setup: Participants had to frequently intervene by actuating the brake pedal in case the automated longitudinal system fails (low reliability indicated on the central in-vehicle display), while performing a detection-response task on a smartphone (left). To hide the activation sound of the olfactory device (located outside the vehicle, right), we used noise-cancelling headphones for the sound output of the driving simulation.

12.2 R ^{EL ATED} W ^ORK

Trust in automation can be defined as “the attitude that an agent will achieve an individ-ual’s goals in a situation characterised by uncertainty and vulnerability” [122], and is a complex construct built by analytic, analogical, and affective processes before (disposi-tional trust), during (situa(disposi-tional trust), and after (learned trust) direct system interaction [87]. To foster safe use of automated systems (and thereby prevent both disuse and misuse [160]), users should adjust their subjective trust levels to fit “an objective mea-sure of trustworthiness” (“calibration of trust” [139]). Reliability/uncertainty displays should assist in the process of trust calibration (especially to account for the problem of overtrust) by providing decision aids that allow users to estimate an automated system’s performance in a given situation [61].

Important groundwork in the domain of AD is the study conducted by Beller, Heesen and Vollrath [17], who demonstrated the potential of a binary reliability display for AVs in a dual-task experiment. Since then, various papers have addressed reliabil-ity/uncertainty displays in the driving domain. Helldin et al. could show that such displays can also improve performance and comfort in Take-Over scenarios [78]. Recent studies have addressed potential metrics and design approaches for in-vehicle displays [149], but also augmented reality [117], or less obtrusive modalities such as haptics [116].

The presentation of different levels of reliability/uncertainty became more and more fine grained in these experiments, aiming to provide drivers more detailed information about the system state. However, a problem reliability displays share with any warning information is that, if (due to an offensive warning strategy) users face too many false alarms, they might simply ignore them (“cry wolf effect” [24]). Considering vehicle safety, drivers already seem to often ignore warning lights in vehicles [92]. As in the near future more and more potentially safety critical systems will be operated by everyday consumers [216], overtrust/overreliance is widely debated in the field of robotics [206]

and AD [204]. For example, in a recent series of simulator studies conducted by Volvo, nearly 30% of drivers crashed in a provoked accident scenario, despite hands on the wheel and eyes on the road, and the authors conclude that more research is necessary to find out how system limitations can be communicated to drivers more effectively [204]. We claim that olfactory notifications could benefit the driver in such situations, as smell is a sense with a strong emotional component [56, 81, 4].

For example, Baron and Kalsher [15] proved that the scent of lemon increases alert-ness and the mood of the driver. The emotion-eliciting effect of scents is particularly useful in inducing mood changes because they are almost always experienced clearly as either pleasant or unpleasant [56]. For instance, Alaoui-Ismaïli et al. [2] used scents of vanillin and menthol to trigger positive emotions in their subjects (mainly happiness and surprise), as well as methyl methacrylate and propionic acid to trigger negative emotions (mainly disgust and anger). The scents of lemon, peppermint, rose, and lavender have been shown as efficient in improving the hedonic experiences of the user [48, 46, 168], whereas lemon and lavender have also demonstrated to be a good medium of conveying useful information in the context of driving and beyond [49, 46, 129, 23, 50].

On this aspect, it is essential to keep olfactory stimuli synchronised with other modali-ties [140]. Further, scents have already been proven to have a positive impact on driving performance/behaviour. Martin and Cooper [130] showed that the scent of lemon can improve drivers’ braking performance, while Dmitrenko et al. [46] demonstrated

that the scents of lemon, peppermint, and lavender could help to reduce the number of errors. Further, scents of peppermint, rosemary, eucalyptus and lemon have been proven to be useful for keeping drowsy drivers awake [220, 66, 157, 83]. Scents could also help to remind drivers on certain driving-relevant activities, as the sense of smell is known to have a strong link with memories [179, 80, 23].