Keywords

1 Introduction

Contemporary Driving Automation (DA) is quickly approaching a level where partial autonomy will be available, leaving nothing but the highest level, strategic decisions, to the driver. According to the NHTSA and BASt [1, 2] DA in its initial stages will likely be constricted to highway conditions where complexity is low, transferring control back to the driver when the operational limits of DA are reached. At this level of automation it is envisaged that the driving task is both hands and feet-free [3], reliably and predictably, over extended periods of time. In such a scenario, the driver should be able to attend to other in-car activities such as reading, working or sleeping without worrying about potential automation failure [4]. This necessitates a controlled and predictable transfer of control between the driver and the automation. Failure induced transfer of control has been extensively studied [see 5–10] whereas transfer of control in a controlled and timely manner has not been explored to the same extent [11].

It is arguable that DA where there is a risk for failure induced, immediate transfer of control, is unlikely to be allowed on public roads as the feasibility of DA rest on the systems’ ability to cope with all but the most severe errors. This assumption rests on fact that the predicted time to collision headway may be set to less than one second and the anticipated response time of the driver, in a state of mental underload [12, 13], or engaged in non-driving tasks which may result in poor situation awareness, may be well above one second [14–16]. These factors may also make it difficult for the driver to produce a timely and suitable response to a sudden change in operational conditions [17]. Therefore, transfer of control between automation and driver must be designed as to allow ample time to complete the transfer. To ensure successful transfer of control, the coordination between driver and automation must be of an adequate level [18]. This is a cause for some concern as inappropriate transfers of control between operator and automation due to poor coordination often result in incidents, sometimes with severe consequences [19]. Therefore, situation dependent information must be provided based on both temporal and contextual factors and must be in line with the current system goal (for example, transferring control in time to go on an upcoming off-ramp). The type of information required for a safe transition is expected to be dependent on the immediate goal and whether that goal is on the tactical, strategic or operational level [20].

1.1 Test Design

As an initial step to investigate what information drivers require, an online test was constructed. The test was designed to assess the information priorities of drivers with regard to both temporal and contextual factors. The online test was inspired by the work of De Craen, Twisk, et al. [21] who used a similar platform to assess speed adaptation in relation to traffic complexity. The results from De Craen, Twisk et al. [21] show that drivers adapt their speed estimate to the complexity of the situation by reporting a lower speed estimate in the complex situation.

The results from De Craen, Twisk, et al. [21] show that it is possible to gather data about driving behaviour using relatively simple and cost-effective methods which was a motivator for the choice of this method. An additional motivator was that data could be gathered relatively quickly from a wide array of demographic groups by using an online platform. It was also theorised that the participants would be more open minded towards automation by presenting a still picture of a traffic situation as there were no cues of how a vehicle cockpit might look in the future. The stills also made it easy to simulate the feeling of being out of the driving loop as the sudden presentation of a situation necessitates orientation and information scanning.

The information preferences test interface consisted of three parts (as shown in Fig. 1); the traffic situation area (1) where a picture of a traffic situation was displayed to the participant, the information selection area (2) where the participant selected what information he/she prefers and the information display area (3) where the selected information was displayed to the participant.

Fig. 1.
figure 1

The interface of the online test. (1) is the stimuli presentation area, (2) is the information selection area and (3) is the information presentation are.

Full size image

The information accessible for the participant covered the tactical, operational and strategic levels of driving [20], including gas level, engine temperature, speed, turn-by-turn navigation, automation status, road conditions, temperature, eco-driving, tachometer and GPS (own position on a map) information. The traffic situations were split into three categories; low complexity, medium complexity and high complexity. The low complexity category displayed an empty road in good visibility conditions with minimum signage and other items that could attract attention. The medium complexity condition had moderate traffic and signage whereas the high complexity condition had rush-hour traffic as well as plenty of signage to be attended to (Fig. 2).

Fig. 2.
figure 2

An example of a complex traffic situation, pictures used in actual tests had no snow or other deviant conditions to ensure this was not a factor. Credit: Katja Kircher Photography AB.

Full size image

1.2 Hypotheses

To assess whether the online testing method was able to produce any differences in response to a set of stimuli two hypotheses were formulated:

  1. 1.

    Decision making time is influenced by

    1. A:

      traffic complexity in the sense that more complex conditions lead to longer response times.

    2. B:

      time constraints in the sense that shorter time constraints lead to shorter decision making times.

  2. 2.

    Information requirements are influenced by

    1. A:

      time constraints in the sense that shorter time constraints lead to the choice of immediately necessary information.

2 Method

2.1 Participants

A total of 81 participants holding a valid driver’s license in their country of residence were recruited through mailing lists and social media. Demographic information is presented in (Table 1).

Table 1. Age range and average age of participants, average annual mileage and range driven per year.
Full size table

2.2 Test Design

The testing platform was constructed using standard web-based programming- languages (PHP, Java Script, HTML, CSS) and was designed to work on most browsers (not including Internet Explorer as certain JavaScripts had compatibility issues). Compatibility of both browser and resolution was ensured using server- and client-side scripts evaluating whether the participant was complying with the limitations set (minimum resolution of 1024 × 760 pixels, no use of Internet Explorer). The use of mobile devices was also restricted using client-side scripts as it was suspected that the use of touch enabled devices could produce incomparable results both between different touch technologies and traditional PC-input methods.

2.3 Experimental Design and Procedure

The experiment had a 3 × 3 design with the factors “time constraint” and “traffic complexity”. Each time constraint level was paired with a picture of one of three traffic complexity levels. These pairings resulted in a set of 9 different stimuli. The order of presentation was randomised for each participant to avoid learning effects. The traffic complexity picture and time constraint pairs were kept constant for all participants. The time constraints were set to 15, 30 and 120 s and the participant was informed of the current time limit before each picture to simulate the time pressure one might experience as a driver in a handover situation.

Upon visiting the website where the test was hosted the participant was faced with a screen containing information about the test and its purpose. He or she was informed about what data were collected and that the data were anonymised and would be stored in an aggregated form. Lastly, the participant was informed that by continuing past the information screen consent was given to the data collection of the previously stated nature.

The participant then proceeded to fill out a demographical questionnaire about driving experience (annual mileage and types of licence), age, education and annual income. Following the demographical questionnaire, the participant had an opportunity to familiarize him- or herself with the test during a training phase in which four traffic situations were presented, the first two without any time pressure. This was to enable the exploration of the interface without rushing the participant. The following two traffic situations were coupled with a time constraint of 15 and 30 s to give the participant an idea of what was to come in the main test. The training phase was also intended to reduce the influence of learning and familiarity effects that could have interfered with the data collection in the main test.

After the completion of the training phase the participant proceeded to the main testing phase. Firstly, the participant received more detailed information driving automation and its characteristics. On the next screen the participant was shown a short text (i.e.: “You have been sleeping for several hours and are suddenly woken up by your car, it is time to take control, what do you want to know?”) as well as a statement of the time constraint for the coming picture. The participant then proceeded to press a button to display one of the nine stimuli that were set to appear in a randomised order. Upon presentation of each picture the participant’s task was to choose whichever information he/she found appropriate given the context. The first and second choice of the participants were recorded along with the time it took to decide on which information was selected.

3 Analysis

A bifactorial analysis of variance (2 × 2 ANOVA) using Bonferroni post-hoc corrections was carried out to assess the effects of external time pressure and the complexity of traffic on decision making time in the first choice made by participants. Additional Chi2 tests were conducted to assess whether the above factors influence the decisions made by the participant and whether the selected information was of a tactical, strategic or operational character.

4 Results

4.1 Effects of Time Constraints on Choice

The time constraints had a significant effect on the choice made by participants χ2(18, N = 729) = 50,743, p < .01 V = .187. The immediate priorities of the participants in the 15 s constraint were GPS and turn-by-turn navigation which were classified as strategic and tactical information. In the 30 s as well as the 120 s constraint the participants favoured GPS and speed, which were classified as strategic and operational information. In the 15 s constraint the participants also disregarded information such as eco-driving, engine temperature and RPM in comparison with the 30 and 120 s constraints (Fig. 3). Participants were also more likely not to select any information at all in the 15 s constraint (11.1 %) as compared to the 30- and 120-s constraints (4.9–1.6 %).

Fig. 3.
figure 3

The distribution of first information item chosen in the different time constraints

Full size image

4.2 Effects of Traffic Complexity on Choice

For all conditions the participants varied in their selection of the first item of information (see Table 2 and Fig. 4). Traffic complexity had an effect on the first choice made by the participant χ2 (18, N = 729) = 69.147, p < .01 V = .218. In the high and medium complexity situations participants favour tactical information such as turn-by-turn navigation. In the low complexity situations participants chose to look more at strategic information (such as gas level, and GPS). The results also indicate that not all participants were ready to make a choice within the given time constraints in the high and medium complexity levels (more than 7 % made no choice) as compared to the low complexity situation (3.3 % made no choice).

Table 2. The distribution of information chosen first by participants within the different time constraints and traffic complexities.The colour code indicates the frequency of which the information was chosen, i.e. green is low frequency and red is high frequency.
Full size table
Fig. 4.
figure 4

The distribution of first information item chosen in the different traffic complexities

Full size image

4.3 Effects of Time Constraints on Decision Making Time

There was a significant difference (\( {\text{F}}\left( {2,720} \right) = 13.91\;{\text{p}}\, < \,.05\;\upeta_{p}^{2} = 0.037 \)) of about 2 s in decision making time between the 15 s constraint and both the 30 s and 120 s constraints (Table 3).

Table 3. Mean and standard deviation values for the different time constraints. Bonferroni corrected post-hoc results of the effect oft he time constraints.
Full size table

4.4 Effects of Traffic Complexity on Decision Making Time

The results show a significant effect of traffic complexity on decision making time (\( {\text{F}}\left( { 2, 7 20} \right) = 7. 9 7,p\, < \,.0 5,\upeta_{p}^{2} = .022 \)). Post-hoc analysis shows a significant (p < .05) increase in decision making time of 1.1–1.8 s in higher complexity situations (high and medium complexity) as compared to lower complexity situations. There are no significant differences in decision making time between the medium and high complexity conditions (Table 4).

Table 4. Mean and standard deviation values for the different traffic complexities. Bonferroni corrected post-hoc results of the effect of traffic complexity.
Full size table

5 Discussion

Overall, the results follow the hypotheses. Response times were shorter for a short time constraint as well as for situations of low complexity. As there were no differences between the time constraints of medium and long duration, and neither between the medium and high complexity levels, it is possible that both factors only have an influence up to a certain ceiling.

The participants varied in their choice of the first item of information, but for situations of high and medium complexity the items navigation, GPS and speed were most popular. They are all highly relevant in this type of situation, as a tactical decision about where to go may be imminent. The current speed could not be gleaned from the still picture shown and is therefore also of relevance. For situations of low complexity the turn-by-turn navigation was not as popular, instead the gas level was requested much more frequently. As the low complexity pictures did not include any upcoming turns, this information was not immediately relevant. Instead, being faced with a long straight road it may feel more relevant to check the fuel level.

The method has very low face validity, so requires validation. It is promising, however, that the factors investigated produced differences in behaviour in line with the experimental hypotheses. Given the current state of automation development, there are no drivers that are experienced with the level of automation implied in the study. This is likely to influence the results, as drivers experienced with ACC have been shown to behave differently on a tactical level than drivers without any previous experience with ACC [22].

6 Conclusion

According to the results obtained in this study it appears that driver behaviour is affected by stimuli presented in an online environment, which support the findings of De Craen, Twisk [21]. Furthermore, the method used in this experiment was able to identify changes in behaviour caused by varying the traffic situation and external time pressure. If the results can be validated against a more realistic setting, this particular method may be a cost effective, easily disseminated tool which has potential to gather valuable insights about what information drivers prioritize when confronted with different situations. These insights can be used to guide designers in building interfaces that adapt on a situation dependent basis with might be of use during transfer of control in automated driving.