Drivers’ Individual Design Preferences of Takeover Requests in Highly Automated Driving

3.1 Participants

A total of 53 participants (26 females, 26 males, 1 no answer) participated in the study. Their mean age was 32 yrs (SD=9.86yrs). Forty-eight participants (90 %) possessed a valid drivers’ license, the other 10 % were student drivers. Forty-two participants (79 %) received a university degree; the other participants were regular employees with varying diplomas or no qualifying degree at all (9 %). Twenty-two of the participants (41 %) were students and 20 (37 %) were employed. The other participants were self-employed. The study was subject to the evaluation of the local ethics committee and all participants signed an informed consent.

3.2 Design of the Human-Machine Interface for Take-Over Requests

Table 1 shows screenshots of each of the take-over request interfaces. Take-over request interfaces differed regarding the following three dimensions: take-over request procedure (one-step or two-step procedure), visual take-over request modality (text or text and pictogram), and auditory take-over request modality (tone or speech). A one-step procedure issued auditory and visual take-over requests at one point in time whereas a two-step procedure issued an initial warning followed by an alarm. The warning was: “Warning – Roadworks ahead – Take-over vehicle control soon” and the alarm was “Alert – Take-over vehicle control now”. The visual take-over request modality differed with respect to the additional information that was provided by the pictogram. It visualised a standard road works sign for the warning and the vehicle’s pedals and steering wheel for the alarm if applicable (see take-over request interface no. 6 in Table 1). Finally, auditory information consisted of a non-specific beep or a mechanical voice reading the warning and the alarm that was displayed on the dashboard.

Take-over request interfaces were dynamic in the sense that they were presented as short film clips of 8 seconds length for a one-step take-over request and 16 seconds for a two-step take-over request. Each film clip followed the same plot. The first 3 seconds showed the dashboard of the vehicle followed by the take-over request which lasted for about 3 seconds. The take-over request faded and another 2 seconds of showing the dashboard were added to end the take-over request smoothly. In the two-step take-over procedure condition the plot was repeated a second time which doubled its length. The videos of the best and the worst rated interfaces can be accessed in the supplementary material.

Table 1

Screenshots of the 8 take-over request interfaces.

	One-step procedure (alarm)		Two-step procedure (warning and alarm)
	text	text and pictogram	text	text and pictogram
tone	no. 1	no. 2	no. 5	no. 6
speech	no. 3	no. 4	no. 7	no. 8

1. Note. The loudspeaker symbolises a simple beep tone; the talking head visualises speech output of the warning “Roadworks ahead! Take-over vehicle control soon” and/or the alert “Take-over vehicle control now” if applicable.

3.3 meCUE Questionnaire

The meCUE questionnaire is based on the Component of User Experience model of Thüring and Mahlke [35]. The instrument allows for the modular evaluation of key aspects of User Experience. It consists of five modules relating the perception of different product characteristics (usefulness, usability, visual aesthetics, status, commitment) to users’ emotions (both positive and negative emotions) and to consequences of product use (product loyalty and intention to use). The fifth module allows for a global assessment of the product. Participants are asked to answer the items on a 7-point Likert-type scale ranging from 1 (disagree) to 7 (agree), except for the general evaluation item that provides a scale ranging from 0 to 10. The questionnaire has been subject to an extensive validation procedure (Minge, Thüring, & Wagner [25]; Minge, Thüring, Wagner, & Kuhr [24]). We used the following dimensions of the meCUE: utility (3 items), perceived usability (3 items), visual aesthetics (3 items), positive emotions (6 items), negative emotions (6 items), intention to use (3 items), product loyalty (3 items), and general evaluation (1 item) for evaluating the participants’ personal highest ranked and their personal lowest ranked interface.

3.4 Procedure

Participants’ were invited to participate in the online-study via facebook™ and other social media platforms. When they clicked on the invitation link they were forwarded to SociSurvey, the online tool that was used for data assessment.

The online study consisted of four parts: (1) an introduction, (2) a general evaluation of each take-over request interface, (3) a ranking of take-over request interfaces, and (4) a detailed evaluation of the personal best and worst take-over request interfaces. The introduction explained the general procedure and obtained the participants data assessment agreement. Participants were asked some demographic questions that were followed by a brief description of the take-over request interface and the procedure of evaluation. It was highlighted that participants would need to switch on their loudspeakers.

In the general evaluation section each of the 8 take-over request interfaces were shown separately, in randomised order. The sequence of interfaces was randomised for each participant. Each evaluation showed a short introductory text on the top of the screen saying that participants should imagine sitting in a highly automated car and reading text messages on their smartphone. Participants were instructed to play a short film clip to see the take-over request that is going to be issued by the car asking them to resume control over their vehicle. They were then asked to evaluate the take-over request interface after playing the video that was located under the introductory text regarding the following four items: (a) This take-over request interface is intuitive, (b) I find this take-over request interface useful, (c) I find this take-over request interface attractive, and (d) I find the amount of information appropriate. Participants were asked to rate each interface on a scale ranging from 0 (totally disagree) to 100 (totally agree) by moving a slider.

In the third part of the questionnaire, participants were asked to bring the interfaces in descending order with the best interface on top (rank 1) and the worst interface on the bottom (rank 8). Take-over request interfaces were shown via the screenshots depicted in Table 1. Finally, in the last part of the questionnaire participants were asked to evaluate their personal best and their personal worst interface using the dimensions of the meCUE (Minge, Thüring, Wagner & Kuhr [25]) that were mention in section 3.3. This evaluation of best and worst interfaces was conducted to derive more detailed insights into the reasons for the preference structure regarding take-over request interfaces. The whole study lasted for about 25 minutes.

4.1 General Evaluation of Take-Over Request Interfaces

The evaluation of the take-over request interfaces revealed significant main effects for each of the independent variables: (s) take-over request procedure (one-step or two-step procedure), (p) visual take-over request modality (text or text and pictogram), and (t) auditory take-over request modality (tone or speech), and an interaction (p*t). Table 2 summarises the effects of the independent variables on the dependent variables intuitiveness, usefulness, attractiveness, and appropriateness of the information. Table 2 shows that participants found the two-step procedure to be more intuitive, useful, and attractive. The two-step procedure also showed the information more appropriately. Participants rated text and pictogram more favourably than mere text with respect to the intuitiveness and the attractiveness of the take-over request interface. They stated speech was more intuitive, more useful and provided information more appropriately than an unspecific beep. Finally, one interaction was observed on two outcome variables. It is depicted in Figure 1.

Figure 1 shows that adding a pictogram to text does not increase ratings of usefulness and appropriateness of information when speech is used as auditory output. However, when an unspecific tone is used as auditory output, the ratings of usefulness and appropriateness of information are improved by adding a pictogram to text.

4.2 Take-Over Request Interface Ranking and Conjoint Analysis

Analysing the ranking of the 8 different take-over request interfaces, it was found that each interface appeared on all possible ranks. The range for the rankings of each interface included the 1^st and the 8^th rank. However, clear preferences became evident when analysing the median rank of each take-over request interface. Friedman’s ANOVA for non-parametric data showed a significant difference indicating that the rank distributions of the interfaces are not similar, p<.001. Table 3 shows each interface, its median rank and the corresponding interquartile range as measure of spread.

Table 3 shows that take-over request interface no. 7 received the lowest median rank and therefore was the best interface. It combined text with speech in a two-step procedure. Interface no. 8 showed the second lowest median rank. The best and the second-best interfaces differed with respect to the presence of pictograms that were contained in the second best but not in the best interface. The take-over request interface with the highest rank (worst interface) presented a tone and the combination of text and pictograms in a single step (interface no. 2).

A conjoint analysis was computed using the three design aspects as factors and the participants’ ranking results as dependent variable. Nine of the 53 (16 %) participants did not complete the whole ranking procedure and were therefore excluded from analysis. The conjoint analysis revealed that factors differed with respect to their relative importance on participants’ rating. The number of steps that are used to announce a take-over request was the most important factor (relative importance = 38.63 %) followed by the utilisation of text or text and pictograms to visualise the take-over request message (relative importance = 32.18 %). The usage of a tone or speech had the lowest relative importance = 29.18 %. Table 4 shows the utility scores for each factor and its attributes. A higher utility score indicates a greater preference.

Table 4 shows the part-worth scores for each factor which can be used to compute the take-over request design with the highest utility. Maximum utility can be inferred by using the constant and by adding the single part-worth scores for each attribute. Table 4 indicates that a two-step take-over request procedure has by far the highest utility (4.40) and should be combined with speech auditory output (0.57) to design a take-over request with maximum total utility (7.28 = constant + two-step + speech). Visual output always decreases the total utility score.

4.3 Detailed User Experience Evaluation of the Best and the Worst Take-Over Request Interfaces

In order to examine the differences between the best and the worst rated take-over request interfaces, the data of the participants were used that put interface no. 7 on rank 1 and interface no. 2 on rank 8. A subsample of n=15 out of 53 participants fulfilled both criteria. Dependent group t-tests were computed for each of the meCUE dimensions. Table 5 shows the results of the detailed evaluation of the best and the worst interfaces.

Table 5 shows that interface no. 7 and interface no. 2 differed regarding all user experience dimensions that were measured with the meCUE questionnaire. Participants rated the winning interface no. 7 higher on all dimensions, except for negative emotions. These were higher for the interface no. 2. Therefore, the difference between the best and the worst rated interface did not just manifest itself in some dimensions of the meCUE but could be found on all dimensions. It might be noted here that both interfaces scored below the midpoint of the scale (3.50) for positive emotions and the intention to use showing the potential for improvement.

5.1 General Evaluation of Take-Over Request Interfaces

The general evaluation of take-over request interfaces revealed that participants perceived clear differences regarding the design aspects that were manipulated in the study. Participants indicated that human-machine interfaces (HMIs) issuing take-over requests should implement a two-step procedure because it is more intuitive, useful, and attractive. The two-step procedure shows the information more appropriately than the one-step procedure. Moreover, text and pictogram were rated to be more intuitive and more attractive than mere text. For the items “usefulness” and “appropriateness of information”, the results were less clear: there was no difference between text and text with an additional pictogram regarding usefulness and appropriateness of information when speech was used as auditory output. However, when the auditory output consisted of a mere tone, text with an additional pictogram was rated more favourably regarding these measures. This finding could be explained by the additional information that can be conveyed by speech in comparison to an unspecific tone. Using speech output can help to provide important information about the driving situation and necessary actions to the driver. Conveying this information via the auditory channel, can make additional visual information provided by a pictogram unnecessary. In line with this argumentation, the results on the auditory modality showed that speech was rated more intuitive, more useful and provided more information than an unspecific beep. These findings suggest that take-over request interface designers should consider other interface design elements than a simple beep and some text that are issued in a single step. This take-over request interface design is associated with problematic driver performance (Naujoks et al. [26]; Brandenburg & Skottke [5]; Gold, Damböck, Lorentz, & Bengler [14]), delayed reallocation of attention resources to the driving task (Merat, Jamson, Lai, Daly, & Carsten [21]; Louw, Kountouriotis, Carsten, & Merat [18]), low TTC (time to collision), strong braking behaviour and high criticality ratings of drivers (Radlmayr, Gold, Lorenz, Farid, & Bengler [33]).

Designers should rather provide more information in a step-by-step procedure (cf. Brandenburg and Chuang [4]) that can be beneficial in case of engagement in a secondary task. Jamson, Merat, Carsten and Lai [15] showed that high levels of automation lead to drivers’ engagement in highly distracting secondary tasks associated with in-vehicle entertainment (e. g. the use of radio or DVD). Petermann-Stock, Hackenberg, Muhr, and Mergl [28] varied drivers’ tasks that they had to accomplish during a phase of highly automated driving from a conversation (low load) to reading (intermediate load) and to reading and writing (high load). They found that drivers’ reaction time increased with increasing load. Drivers also distributed less attention to the road and more attention to the secondary task with increasing load. The take-over request should therefore reallocate drivers’ attention to the road and to the necessary driver actions (Borojeni, Chuang, Heuten, & Boll [3]). A two-step procedure provides a more gradual take-over process because the driver is disengaged from a secondary task and engaged in the driving task in a more stepwise manner. The second step of a two-step take-over request also provides a fallback level in case drivers missed the first warning due to engagement in a secondary task. However, a two-step procedure might not be possible in critical driving situations where immediate driver reactions are necessary. Future studies should investigate in which driving situations two-step procedures and one-step procedures could be better suited to trigger transitions of control.

In addition, participants’ preference ratings indicated the wish for more information than a beep and a short text stating why that beep occurred. This finding could indicate the participants’ want to update their situation awareness [9]. Following Radlmayr et al. [33] the drivers’ situation awareness decreases during phases of highly automated driving. Informative take-over requests have the potential to support drivers understanding of the driving situation and the preparation of necessary actions. Knowing why actions have to be performed might provide useful information about the future projection of the course of action. It might also deliver information helping to estimate the urgency of driver reactions (cf. Stahl, Donmez, & Jamieson [34]). This information could be provided by speech output. In line with the findings by Epple et al. [10], our results show that the use of speech output in a two-step procedure could be beneficial for take-over requests in highly automated driving.

5.2 Take-Over Request Interface Ranking and Conjoint Analysis

The ranking of the eight interfaces revealed that participants clearly favoured one of the eight designs. The winning interface combined a textual visualisation with speech output in a two-step procedure. Adding a pictogram to this layout decreased participants’ appraisal of the take-over request visualisation. This finding is interesting as it is in contrast with the results of the general evaluation of take-over request interfaces. In the general evaluation, an interface containing text and pictogram was rated to be more attractive and intuitive than mere text. This discrepancy between general evaluation and ranking of take-over request interfaces may be due to a discrepancy between the aspects that we assessed in the general evaluation (intuitiveness, usefulness, attractiveness, and appropriateness of information) and the criteria that participants used to rank the interfaces. Possibly, participants perceived other criteria to be more decisive for their ranking, than the ones we assessed. One possible criterion could have been visual clutter. The interface with text and pictogram may contain too much visual information and therefore degrade information processing. Typically, take-over request interfaces that are reported in literature use pictograms to visualise the status of the automated system [18], [27] or the drivers need to resume vehicle controls (Albert, Lange, Schmidt, Wimmer, & Bengler [1]; Flemisch et al. [12]; Radlmayr et al. [33]).

Moreover, the ranking of interfaces revealed that speech was preferred over a single tone. Recently, an increasing number of studies used speech output to issue take-over requests to drivers (e. g., Miller et al. [23]; Politis et al. [31]; Walch et al. [37]). Politis et al. [31] and Epple et al. [10] found that take-over requests using speech were associated with considerably shorter reaction times than written text output. The present study highlights the importance of speech output when issuing take-over requests for subjective preference ratings. Speech output may be superior to mere visual output due to the drivers’ tendency to engage in demanding and mostly visual secondary tasks during highly automated driving.

The worst ranked take-over request interface combined an unspecific tone with a text and pictogram visualisation in a one-step procedure. This is what some car manufacturers and researchers propose as visualisation for take-over requests [33], [26]. Based on the results of the current study it could be recommended to refrain from using this take-over request interface any longer.

The conjoint analysis revealed additional insights in participants’ preferences regarding take-over request design aspects. Again, it underlined that participants value a two-step take-over request procedure over a one-step procedure. In fact, take-over request procedure was the most important factor for the participants’ preference ratings. This finding is in line with previous research of Brandenburg and Chuang (in review) on the design of take-over request procedures that also found a two-step procedure to be superior to a one-step procedure. However, a two-step procedure will be applicable in regular take-over request situations where the vehicle passes on the driving task to the driver with much time available; for example at the end of a phase of highly automated driving on an autobahn. It might not be applicable in highly dynamic situations where driver reactions have to follow the take-over request promptly. Future studies should examine the applicability of two-step take-over request procedures in varying driving situations. Brandenburg and Chuang (in review) showed that take-over request characteristics differently affect participants take-over request related performance depending on the characteristics of the driving course. Additionally, Jamson et al. [15] and Naujoks et al. [26] showed that drivers post-take-over request performance is affected by road characteristics, which further stresses the need for future work on the interaction of take-over request design and road characteristics.

Finally, the conjoint analysis extended the simple analysis of the participants’ take-over request interface ranking by the fact that both attributes of the factor text or text and pictogram are negatively related to the participants’ utility of take-over request interfaces. In fact, the take-over request interface with the highest subjective utility consists of a speech output in a two-step procedure. Any type of visualisation would decrease participants’ utility ratings. This finding is interesting because it relates to the issue of an increasing number of visual displays in in-car HMI. Implementing more and more assistance systems currently goes along with adding more visual elements in the in-car HMI as well. Drivers might feel overwhelmed by the high number of icons and pictograms that might impose feelings of stress and cognitive strain on them. However, it should be emphasised that this is a study on drivers’ preferences without a driving task. Participants may rate the interfaces differently in a real driving scenario. Moreover, driver preference does not necessarily accord with safety. Visual output could provide a backup when drivers miss the warnings in other modalities due to inattentiveness or auditory masking. Therefore, visual output may still contribute to a safer transition of control.

Based on the current work we conclude that the industries current focus on additional non-visual driver-to-car and car-to-driver interaction is supported by our findings (Gellatly, Hansen, Highstrom, & Weiss [13]). Pictograms and beeps might no longer be the primary way of communicating information from the car to the driver. Instead, the interaction of drivers and vehicles could be based on more meaningful auditory information to reach a high level of immersion in a driver-vehicle system [2].

5.3 User Experience Evaluation of the Best and the Worst Interfaces

The final evaluation of the best and the worst interfaces revealed that both differed regarding all aspects of user experience that were assessed in this study. This result sheds some light on possible reasons for the ranking of the best and the worst interface. Interface no. 7 was not just rated more favourably than interface no. 2 on product characteristics, like usefulness and usability. It also received better ratings for users’ emotions (both positive and negative emotions) and consequences of product use (product loyalty and intention to use). Whereas this finding might not be very surprising, the examination of group means revealed that the best and the worst interfaces differed on a low level of absolute values for some important UX dimensions. Non-instrumental aspects like positive emotions and the consequences of user experience like intention to use the product were on a low level indicating the potential for improvement. In addition, these effects were of medium size only, indicating that study participants did not perceive large differences in user experience between the interfaces. These results might partially be due to the study setting. Participants were asked to imagine that they sit in a highly automated vehicle. They played short film clips instead of experience a TOR situation in a driving simulator or a real car. Their ratings might therefore change when experiencing a TOR instead of seeing it on a computer screen.