Augmented reality HUD vs. conventional HUD to perform a navigation task in a complex driving situation

This study aims at investigating the added value of an augmented reality head-up display (AR-HUD) in relation to a conventional head-up display (C-HUD) to perform navigation tasks in a complex road situation. The notion of complexity was defined according to two main factors: infrastructure and traffic. It was used to identify and select real road situations presenting different sources of complexity. This study focuses on one of these situations, which was reproduced on a simulator and broken down into three use cases. A total of 32 participants performed three navigation tasks, using the AR-HUD or the C-HUD. Both objective and subjective data were collected. Data analyses, using linear mixed model analyses of variance and multilevel logistic regression, indicate a slight advantage of the AR-HUD. Participants using the AR-HUD make fewer errors and drive faster on average. Moreover, the AR-HUD is assessed to be more useful and easier to understand than the C-HUD. However, this interface shows limitations, in particular, because it does not enable drivers to anticipate the manoeuvre to be made. The study raises questions about the design of an instrument system that would help drivers not only identify, but also build a representation of a forthcoming manoeuvre to be performed.


Introduction
Navigation in unknown road environments represents a costly cognitive activity for drivers. Numerous studies have demonstrated the difficulties drivers encounter to plan and follow routes (Ross and Burnett 2001). These difficulties may induce dangerous behaviours such as late lane changes. Michon (1985) breaks down the driving task into three levels: strategic, tactical, and operational. Ross and Burnett (2001) point out that the navigation task also involves these three levels. Planning the route falls within the strategic level. It impacts upon the manoeuvres performed at the tactical level (e.g., overtaking, controlling the inter-vehicle distance) as drivers need to follow the planned route while dealing with interactions with the other road users. On the operational level, manoeuvres involve lateral and longitudinal control (e.g., slowing down at a highway off-ramp). Burnett (1998) proposes to break down the navigation task into several phases that precede and follow a navigation task manoeuvre: -Preview: drivers assess the time/distance remaining before the next manoeuvre and envision the manoeuvre to be performed. -Identify: they identify the location of the next manoeuvre and the direction to take, adapt their speed, and correctly position their vehicle. -Confirm: they check that they have performed the right manoeuvre.
In parallel with these phases, two processes are executed. Drivers make sure that the route taken is the correct one. They also build a representation of their environment and of their vehicle location. Navigation systems may provide assistance to the drivers at each of the three phases. They take the form of a vehicleindependent personal assistant, a head-down display (HDD), a head-up display (HUD), or even a head-mounted display (HMD). Harkin et al. (2005) indicate that using an HUD that projects information onto the vehicle windshield is expected to reduce the time spent not looking at the road environment (i.e., eye-off-road time); in turn, this would improve the drivers' situation awareness and safety in contrast to what happens with a personal assistant or an HDD. Gabbard et al. (2014) distinguish between conventional HUDs (C-HUDs) providing information in symbolic form and augmented reality-based HUDs (AR-HUDs) that superimpose information onto outside-world objects or draw attention to these objects. The authors explain that the latter HUDs (termed world-fixed head-up optical seethrough displays) are more relevant than the first type of HUDs, because they highlight information relating to the primary task of driving (e.g., wayfinding, hazard warnings). Augmented reality-based HUDs facilitate drivers' perception of useful information. Drivers do not have to change focus as their attention remains attached to the driving scene, and it is directed towards the most relevant objects in their environment. Utsuboya et al. (2013) argue that AR-HUDs provide intuitive wayfinding by superimposing the planned route onto the actual route followed by the vehicle and direction arrows onto intersections, or identifying landmarks on the buildings appearing in the visual scene.
Regarding the performance of navigation tasks, studies have demonstrated the advantage of using different types of HUD-based presentations of information over a personal assistant (Nwakacha et al. 2013) and the advantage of AR-HUDs over HDD (Kim and Dey 2009;Medenica et al. 2011;Smith et al. 2017) or HMD systems (Jose et al. 2016). Some studies also point out the advantages of AR-HUDs over C-HUDs (Bauernfeind et al. 2019, 2021Bolton et al. 2015;Chauvin et al. 2019). Bolton et al. (2015) examine different uses of augmented reality on an HUD. Using a simulator-based town-driving scenario, the authors compare four driving conditions: one shows the direction to follow using a C-HUD, and the other three use augmented reality with an arrow indicating the direction to follow, or an arrow or a box highlighting a landmark in the environment. The experiment shows the merit of using AR to draw attention to points of interest for a navigational task (e.g., petrol stations, schools, railway stations, and places of worship). Using AR to put these points of interest inside boxes has a significant effect upon participants' response time and successful completion of the task. Out of the four conditions being tested, the C-HUD is associated with the lowest performance and the highest mental load; it is also the participants' least preferred condition.
In two other studies, Bauerfeind et al. (2019Bauerfeind et al. ( , 2021 compare an AR-HUD and a C-HUD to perform a navigation task in an ambiguous urban area. The authors hypothesize that augmented reality displays are particularly beneficial in ambiguous situations. They use scenarios including many possible turns that are very close to each other. The results confirm their hypothesis. As they spend more time observing the road scene, participants using the AR-HUD make earlier decisions, drive faster, and make fewer errors than with a C-HUD (Bauernfeind et al. 2019). Furthermore, augmented reality information requires less of their mental load in understanding and interpreting and leads to more confident driving behaviours (Bauerfeind et al. 2021). More than the ambiguity of the situation, it is its complexity that is central in the preliminary study of Chauvin et al. (2019). A real driving environment is reproduced. It is a suburban road with an X-shaped interchange where participants have to change lanes. The results Chauvin et al. (2019) obtain were less conclusive than those pointed out by Bauerfeind et al. (2019Bauerfeind et al. ( , 2021. The AR-HUD supports drivers, since it contributes to avoiding poor performance (characterised by abrupt actions on the steering wheel and low time-tocollision with the front vehicle), but there is no positive effect of the use of an AR-HUD on the vehicle speed. One explanation is that the situation is not complex enough to observe a clear added value of the AR-HUD.
The present study aims at investigating the following issue more thoroughly: what are the benefits of an AR-HUD, compared with a C-HUD, in a complex and real driving situation? The proposal of Boelhouwer et al. (2020) was used to define complex driving situations according to two main factors: infrastructure and traffic. In line with the literature, two general hypotheses guided the study: -H1-In complex situations, the AR-HUD facilitates the identification of the manoeuvre to be performed and therefore is expected to be associated with better driving performance than the C-HUD: participants make fewer errors when using the AR-HUD (H1.1), they adopt a higher speed (H1.2), and their actions are less abrupt (H1.3). -H2-In complex situations, the AR-HUD is associated with a better user experience than the C-HUD: participants have a better comprehension of the information provided by the AR-HUD (H2.1), they find the AR-HUD more useful than the C-HUD (H2.2), they have higher confidence in the manoeuvre when using the AR-HUD (H2.3), and they feel more comfortable with the AR-HUD than with the C-HUD (2.4).
Because Boelhouwer et al. (2020) also underline a possible effect of familiarity on the perceived complexity of driving situation, this factor was controlled.

Definition and design of real and complex driving situations
Several real and complex situations were identified in a pre-study during which 10 individuals who were used to driving in the Paris area were asked to list places where it is difficult to find one's way. In this pre-study, 31 problem areas were listed, and 14 were retained to make open observations with eight drivers in real-life driving situations. The drivers were asked to think aloud along the route defined by the experimenter. They could use their favourite navigation app on their smartphone. When reaching their destination, drivers had to indicate the reasons for their discomfort, the influence of traffic, the information used to complete the task, and they had to assess the difficulties they experienced to find their way.
In line with the proposal of Boelhouwer et al. (2020), the overall complexity of these road situations was determined by the characteristics of the infrastructure (the type of infrastructure and the presence of road elements that may increase the complexity for the drivers) and the traffic features (traffic density, traffic predictability, and type of priority). Concerning the type of infrastructure, the selected situations were made up of intersections or interchanges. They were characterised by one or several of the following road elements: absence of road lines, hidden road signs, unfamiliar road arrangement (e.g., entering a freeway from the left lane), and short distance between two route changes. In some use cases (UCs), those factors were supplemented by dense traffic or unclear driving priority rules.
Three geographic areas (Le Plessy, Le Gallo, Bercy), representative of the 31 situations listed at the beginning, were reproduced in the simulator-based experiment.

Participants
For the simulator-based experiment, 32 participants were recruited. The panel consisted of 15 women and 17 men aged between 25 and 60 years (M = 41.8, SD = 9.8). They had at least 2 year driving experience, and drove at least 5000 km a year. They used a navigation system regularly (once a week), whether on-board or not.

Experimental design
In the experiment, participants were required to drive on a suburban road, in three selected geographical areas. Each area involved different UCs, defined by the manoeuvre to be executed to change route (e.g., turning left, taking a particular direction, or taking a junction on the right).
Driving in the first geographical area (Le Plessy) was the learning phase; hence, the data from this area were not analysed. The area called Le Gallo comprised two UCs. The data collected as part of these two UCs are not dealt with in this paper, since their analysis has already been published and its main results outlined in the introduction (Chauvin et al. 2019). The Bercy area comprised four UCs (UC1 turning left, UC2 taking the Metz-Nancy direction, UC3 taking the middle lane at the first junction on the right, UC4 taking the right-hand lane at another junction). As only 10 s separate UC3 and UC4, they were examined together. The present study deals with these three UCs (UC1, UC2, and UC3-4). In all UCs, the speed limit was 50 km/h (31 mph).
UC1 takes place at an intersection. It involves changing lane and then taking the second road on the left (Fig. 1). Participants were required to choose between three directions (driving straight ahead, turning left immediately, or taking the second road on the left). In this UC, speed is limited by infrastructure and traffic. Complexity is determined by the fact that there are two close intersections on the left-hand side and, therefore, two possible turns. This UC is close to the situation investigated by Bauerfeind et al. (2021). It was expected that the AR-HUD would help drivers identify the correct turn.
UC2 presents an intersection where the participants had to go through a vehicle flow to take the Metz-Nancy direction (Fig. 2). In this case, speed is limited by traffic. Complexity is determined by the fact that the priority rule is violated as the flow of vehicles on the larger lane does not give way to vehicles approaching from the right. Because this UC does not involve any manoeuvre, the benefits of the AR-HUD were expected to be limited.
UC3-4 is made up of two interchanges (Fig. 3). Drivers needed: -to change lane to go to the middle lane in an X-shaped junction (two lanes become two lanes heading in different directions). When their vehicle reached the start point, it was followed by another vehicle and they also needed to take a motorcycle into account to adapt their trajectory, because the motorcycle undertook late and unpredictable actions. Two other vehicles were already present in the right-hand lane; -to change lane to go to the right-hand lane in a Y-shaped interchange (one lane becomes two lanes heading in different directions). This manoeuvre took place 10 s after the previous one. The ego vehicle was followed by a vehicle travelling at the same speed with a time headway of 2 s. This UC displays several factors of complexity due to the infrastructure (consecutive X-shaped and Y-shaped interchanges, hidden road signs, and absence of road lines) and to the traffic. There are several other vehicles of different types behind and in front of the vehicle, in the same lane and in the adjacent one. One of those vehicles, a motorcycle, shows unexpected behaviour when overtaking the vehicle on the right at high speed (70 km/h vs. the speed limit of 50 km/h). The speed differential between the motorcycle and the vehicle is high enough to ensure that all participants will face this event, even if they drive at a speed that is a little above 50 km/h. Because of these features, and as Boelhouwer et al. (2020) point out that traffic issues (traffic density, varying road user types, predictability of other road users) have a larger influence on the overall complexity than infrastructure considerations, it was thought that this UC is more complex than the other ones. Hence, the benefits of an AR-HUD were expected to appear clearly.
The UCs were presented to the participants in fixed order (UC1, UC2, and UC3-4). In all UCs, half the participants

Hardware and the human-machine interface (HMI)
The experiment was conducted with the Driving Simulator for Human-Machine Interaction studies (Dr SIHMI) of the Institute for Technological Research System X (IRT Sys-temX). This static simulator, shown in Fig. 4, consists of a driving station containing a steering wheel, acceleration, and brake pedals, an HUD functioning as AR-HUD or C-HUD depending upon the experiment condition, and a GPS. This basic GPS system shows a map of the entire area, oriented north and clearly indicating the vehicle location. It is displayed on a screen located in the middle of the dashboard. A curved projection screen with an aperture of 170° was used to show the simulation of the different driving situations with the SCANeR simulation software. The HUD is a binocular mirrored one with a surface of 9*15°. As defined by Gish (1995), it is similar to a fully functional HUD, except that there are no optics between the image source and the combiner (a large piece of plate glass). Thus, the virtual image distance is equivalent to the sourceto-eye distance. In this simulator, the distance between the eye of the driver and the screen where the driving scene is projected is equal to 1.3 m. The HUD was developed, so that the virtual image distance is also equal to 1.3 m.  The AR-HUD unit has a dedicated surface of 8*15°. It projects virtual information onto the windshield in the form of arrows that indicate the direction to follow. In Fig. 5, the AR information overlaid corresponds to the blue arrows with a triangular form. The arrows appear 400 m before the manoeuvre end-of-point. The system takes the traffic into account and indicates the manoeuvre to be performed only if this action is possible given the vehicles present in the vicinity. Other information appears on a fixed section at the bottom of the HUD: the speed limit, the vehicle speed, the turn-by-turn icon indicating the next change of direction, the remaining distance before reaching it, and an indication of the direction to follow, shown as a road sign similar to the one that appears on the road.
The C-HUD shows a map indicating the route to follow (Fig. 6). It reproduces the system available on some highend vehicles. The same items of information as those listed above also appear on the HUD (speed limit, vehicle speed, etc.).

Experiment sessions
The experiment sessions lasted 1h30. Participants were invited to read an information document and sign consent forms. They were then introduced to the simulator, interfaces, and the test (5 min duration). A learning phase entailed driving in the Le Plessis area, first with, then without AR or conversely; the sequence was similar to that selected for the experimental phases.
The first experimental phase entailed driving in the Le Gallo or Bercy area with the AR-HUD or C-HUD interface. In case of failure, the driving phase was interrupted, and participants had to start over again. As participants could replay this phase up to three times, should they fail the first and second driving test, the duration of this phase varied between 20 and 30 min. A debriefing interview was then held. The second experimental phase was held in similar fashion in the second area and with the other interface. Finally, participants were asked to complete a questionnaire relating to their preferences and their profile (about 10 min duration). Ross and Burnett (2001) explain that the assessment of navigation systems is based on objective and/or subjective dimensions. The present study uses both assessment categories. The objective dimensions, relating mainly to driving performance, were used to test the first general hypothesis: in complex situations, the AR-HUD facilitates the identification of the manoeuvre to be performed and therefore is expected to be associated with better driving performance than with the C-HUD. The subjective dimensions were used to test the second general hypothesis related to the added value of the AR-HUD on the user experience.  Table 1 shows the metrics that were calculated for each UC from the data collected by the simulator. In addition to their average values, the maximum values of the vehicle speed, pedal deflection speed, steering wheel rotation speed, and acceleration were examined as they may be indicative of the roughness (vs. smoothness) of the driving actions. Following Eriksson and Stanton (2017), the variability of actions on the steering wheel was also considered, since the authors view this metric as a good indicator of the drivers' mental workload.

The data collected
Failures were recorded; they correspond to navigation errors. The experimenter also noted all gazing at the manoeuvre-adapted mirror.

Subjective dimensions
After each experimental phase, a debriefing interview was held with participants to determine, for each UC: -the sources of information used (GPS, the AR-HUD or C-HUD interface, the turn-by-turn icon, the distance, and the road sign reproduced at the bottom of the HUD); -the participants' impressions (understanding/usefulness of the information from each information source).
In addition, participants were asked to assess their feelings using the following four 6-point Likert scales: -the level of comprehension of each piece of information provided (GPS, turn-by-turn icon, turn-by-turn distance, direction sign, C-HUD, and AR-HUD); -the perceived usefulness of each piece of information provided (GPS, turn-by-turn icon, turn-by-turn distance, direction sign, C-HUD, and -HUD); -the degree of confidence placed in the manoeuvre (can you assess the level of confidence you felt in the route to be followed: I felt almost no confidence/very little confidence/little confidence/average confidence/high confidence/very high confidence); -a comfort score (can you assess the level of comfort you experienced: I experienced almost no comfort/very little comfort/little comfort/average comfort/high comfort/very high comfort).
At the end of the experiment, participants were asked to indicate their preference for one or the other of the two interfaces. Additionally, participants were asked about their knowledge of the area and the route.

Data processing
The data analysis was conducted in three phases: an objective data analysis followed by an analysis of the subjective data, and then an evaluation of the relationship between the two types of data.
The first phase aimed to evaluate the effect of AR-HUD on driving performance to address the first hypothesis, confirm the specific status of UC3-4, and control for a possible learning or experience effect. To this end, the analysis focused on the effects of HMI, UC, and area knowledge on the objective variables: compliance with instructions, navigation errors, and vehicle parameters.
The second phase aimed to test the hypothesis related to the user experience by analysing the subjective information collected from the participants. The investigation focused on the effects of HMI, UC, and area knowledge on comprehension, utility, confidence, and comfort scores.
The analyses were conducted in R (R Core Team 2022). The lme4 package (Bates et al. 2015) was used to perform linear mixed model analyses of variance and multilevel logistic regression. The models were run with HMI, UC, area knowledge, and their interactions as fixed effects and participant intercepts as random effects. The P values were obtained by likelihood ratio tests of the full model with the effect under study vs. the model without the effect under study, and the assumptions of homogeneity and normality were assessed by visual inspection of the residual plots. Besides, repeated-measures ANOVAs were performed with Mauchly's test for assessing sphericity; in case of violation, the Greenhouse-Geisser correction was used to report the results. Finally, the Chi-square test of independence was used to measure the association between categorical variables.

Results
Among the 32 participants, 10 already knew the Bercy area in which the experiment took place, and 3 of them also had some knowledge of the UCs. In addition, 8 participants made navigation errors that led to re-running the driving sessions; 2 of them knew the area, and 1 knew the route. To take into account a possible learning effect, a binary variable "knowledge of the area" was introduced. It indicates whether the participants knew the area, either from experience or because they failed in a first run and had to replay a UC. In all, 17 participants knew the area (9 AR-HUD, 8 C-HUD), 14 did not (8 AR-HUD, 6 C-HUD), and this information was missing for one participant.
Results are presented by type of variable. Analysis of the objective metrics provides answers to the first general hypothesis related to the effect of the HMI on the driving performance (H1), whereas the analysis of the subjective metrics provides answers to the second hypothesis related to the effect of the HMI on the user experience (H2).

Objective metrics analysis
This first part presents the analysis of navigation errors and compliance with instructions on the one hand and the analysis of vehicle metrics on the other.

Navigation errors and compliance with instructions
The analyses reported below aim at testing Hypothesis H1.1 stating that participants make fewer errors when using the AR-HUD.
A total of 14 navigation errors (i.e., failures) were recorded during the driving sessions: 2 participants failed both UC1 and UC3-4, 2 failed only UC1, and 8 failed only UC3-4. The failures are presented in Table 2 by HMI, UC, and prior knowledge of the area. A multilevel logistic regression was performed to determine the effects of these factors and their interactions on the likelihood that participants would fail to complete the driving task. It revealed a significant main effect of the UC (χ 2 (2) = 14.57, p < 0.001). There was no significant effect of HMI (χ 2 (1) = 0.07, ns), prior knowledge of the area (χ 2 (1) < 0.001, ns) nor their interactions. UC2 was associated with a zero risk of failure and probability of failure raised to 9.3% and 27.7% for UC1 and UC3-4, respectively.
In addition to the failure analysis, participants' compliance with procedures was examined, which proved conclusive for UC1. In this UC, participants had to perform two successive lane changes (Fig. 7): shifting to the left to go from lane 3 to lane 4 (1), and then turning at the traffic lights (2).
Using the AR-HUD significantly promotes compliance with the procedure ( 2 (32) = 18.15, p < 0.001, Cramer's V = 0.75). It is worth noting that a large majority of the participants using the C-HUD continued driving on lane 3, and then, they cut across lane 4 to turn at the traffic light (Table 3).
Among the participants who complied with the procedure, lane changing occurs later under the C-HUD. The median distance between the lane change point and the end-of-manoeuvre point was greater under the C-HUD (Mdn = 87.4 m) than under the AR-HUD (Mdn = 56.6 m). A Mann-Whitney test indicated that this difference was statistically significant [U(3,16) = 4; p = 0.023; rank biserial correlation = − 0.83].

Analysis of the driving performance (mean and maximum speed)
This analysis tests Hypothesis H1.2 stating that participants adopt a higher speed with the AR-HUD than with the C-HUD.  Mixed linear model analyses of variance were used to investigate the effects of HMI, UC, and area knowledge on the speed metrics. HMI, UC, area knowledge, and all their possible interactions were entered as fixed factors, and participants intercepts as random factors.

Analysis of the driving actions
The same methodology was used to investigate the effects of HMI, UC, and area knowledge on the metrics related to the driving actions (acceleration and actions on the brake pedal, the acceleration pedal, and the steering wheel). The analyses carried out aim at testing Hypothesis H1.3 stating that actions are less abrupt with the AR-HUD than with the C-HUD. Two significant effects of the UC were found: one concerned the average acceleration, and the other concerned the steering wheel rotation speed. No effect of the HMI was found.
Mean acceleration. There was only one significant effect of the UC on the average vehicle acceleration ( 2 (2) = 10.31, p < 0.01). Pairwise comparisons indicated that the differences of UC1 vs. UC2 (d = 0.496, t(30) = 6.08, p < 0.001), UC1 vs. UC3-4 (d = 0.73, t(30) = 8.98, p < 0.001), and UC2 vs. UC3-4 (d = 0.08, t(30) = 2.91, p < 0.01) were statistically significant. Figure 10 shows the main effect that was revealed: the average vehicle acceleration is at its lowest in UC1 and at its highest in UC2.  Steering wheel rotation speed. The analysis revealed a significant main effect of the UC on both the maximum steering wheel rotation speed ( 2 (2) = 74.05, p < 0.001) and on the steering wheel rotation speed variability ( 2 (2) = 69.24, p < 0.001). There was also a significant interaction effect of UC and area knowledge on the maximum steering wheel rotation speed ( 2 (2) = 9.68, p < 0.01) and on its variability ( 2 (2) = 9.54, p < 0.01). Pairwise comparisons indicated that when the area is not known, the smoothest actions on the wheel and the lowest variability of the steering wheel rotation speed are significantly associated with UC2, followed by UC1 and then UC3-4. When the area is known, the smoothest actions on the steering wheel and the lowest variability of the steering wheel rotation speed are still associated with UC2. However, there is no significant difference between UC1 and UC3-4 (Fig. 11a). Furthermore, the variability of the steering wheel rotation speed and the roughness of the actions on the wheel in UC3-4 tend to decrease when the area is known (Fig. 11b).
Thus, the objective metrics analysis shows that -H1.1 is partially supported. In UC 1, the AR-HUD interface promotes compliance with instructions. However, and regardless of the UCs, there was no effect of the HMI on navigation errors. -H1.2 related to the vehicle speed is supported. There is a significant effect of the HMI on the average speed and a marginal effect of the HMI on the maximum speed. As expected, speed is higher with an AR-HUD than with a C-HUD. -H1.3 related to the driving actions is not supported.
Actions carried out when using an AR-HUD were not less abrupt than with those carried out using a C-HUD.
Besides those results concerning the effects of the AR-HUD, significant effects of the UCs on driving performance were observed. Some of them (e.g., differences in mean and maximum speeds) may have been induced by the specific characteristics of the UCs. The speed of the vehicle is necessarily limited in UC1, since the driver must stop at the intersection. It is limited by the traffic in UC2. Conversely, the speed limit (50 km/h) is the only constraint in UC3-4. However, and even if the situation features are the most critical factors, a learning effect may have impacted the average and maximum speed in the latter UC.
Other effects reflect differences between UCs attributable to their level of complexity. The fact that navigation odds ratios are higher under UC3-4, while the risk of failure is null in UC2 confirms the specific status of UC3-4 as the most complex case and of UC2 as the easiest one.  Similarly, the steering wheel rotation speed variability, a variable that is an indicator of mental workload, takes its lowest value in UC2 and its highest ones in UC3-4.

Subjective metrics analysis
This analysis aims at testing the second hypothesis related to the effect of the AR-HUD on user experience. It focuses on two dimensions: (i) the information used by the participants and its evaluation in terms of understandability and usefulness; (ii) the confidence and comfort scores associated with the manoeuvres.

Sources of information
Participants had several sources of information at their disposal: GPS, three items at the bottom of the HUD indicating the next direction to take (turn-by-turn icon), the remaining distance before the direction change (turn-by-turn distance), a road sign showing the direction to follow (road sign), a 3D arrow under AR-HUD, and a map under C-HUD. Table 4 illustrates the information sources that the participants stated they had used in their debriefing interviews. Results show that -In UC1, the participants used more information sources than in other UCs. -The GPS was hardly used.
-Under the AR-HUD, the participants mainly used the 3D arrow and the road sign at the bottom of the HUD screen.
In the most complex UC, UC3-4, participants relied even more on the 3D arrow to the detriment of other information sources. -Under the C-HUD, participants mainly relied on the road sign. The C-HUD map was barely used in UC3-4 and not much at all in UC1 and UC2.
At the end of the driving sessions, the participants were asked to assess the usefulness and understandability of the available information, on a scale ranging from 1 to 6. With a mean score of 1.55 (SD = 1.06), the GPS is the source of information viewed as the least useful, although it is easy to understand (M = 4.41; SD = 0.72). The three items of information shown at the bottom of the HUD (i.e., the turn-by-turn icon, turn-by-turn distance, and road sign) were viewed as useful or even very useful, since their mean scores are 4.86 (SD = 1.11), 5.08 (SD = 1.46), and 4.92 (SD = 1.71), respectively. These items are also easy to understand as participants assessed this information at over 5 for the three items: 5.41 (SD = 0.72), 5.61 (SD = 0.94), and 5.44 (SD = 0.79), respectively.
The mean score for usefulness of the AR-HUD is 4.33 (SD = 1.89), and that of the C-HUD is 2.95 (SD = 1.87). The mean score given to understandability is 4.3 (SD = 1.67) for the AR-HUD and 2.85 (SD = 1.75) for the C-HUD. Repeated-measures ANOVAs revealed a significant effect of the sources of information on predicting the usefulness scores (F(5,155) = 27.58, p < 0.001, 2 =0.47) and the understandability scores (Greenhouse-Geisser correction of sphericity, F(3.25,71.51) = 16.22, p < 0.001, 2 =0.42). Pairwise comparison tests, with Holm correction to adjust for p, established that the map (C-HUD) is the least understandable source of information and that the information at the bottom of the HUD screen (turn-by-turn icon, turn-by-turn distance, road sign) is easier to understand than the C-HUD map, the GPS, and the AR-HUD. The GPS is the least useful information source, followed by the C-HUD map (Table 5).
Thus, 3D arrows (AR-HUD) are more useful and easier to understand than the map (C-HUD). However, they are considered less easy to understand than the bottom HUD information, without any evidence of lesser usefulness. Various problems were reported by AR-HUD users during the debriefing interviews, including a lack of anticipation that reduces the perceived usefulness of this interface. It was also sometimes difficult to interpret the direction of the arrows; some participants indicated that they were not accurate enough, not on track, or they appeared too suddenly. At the same time, the participants who used the C-HUD mentioned the difficulty of finding their location on the map and the disruptive changes of representation. Several also thought that this artefact carried too much information ("too many criss-crossing lines") in a complex situation; hence, the map added confusion.  Table 6 shows the mean and standard deviations of confidence and comfort scores associated with the manoeuvres, by UC, area knowledge, and HMI. A linear mixed model analysis of variance revealed a significant main effect of UC ( 2 (2) = 18.09, p < 0.001) and prior knowledge of the area ( 2 (1) = 5.52, p < 0.05) in predicting confidence scores, and a significant main effect of UC ( 2 (2) = 8.33, p < 0.05) in predicting comfort scores. There was no significant effect of the HMI on confidence or comfort levels. Pairwise comparison tests, using Holm correction to adjust p, indicated that on average, confidence and comfort scores were significantly lower in UC3-4 than in UC1 and UC2. There was no evidence of a difference between UC1 and UC2 (Table 7).

Correlations between the objective and subjective dimensions
In each UC, the correlations between the vehicle metrics on the one hand and the confidence and comfort scores on the other were assessed. In UC1, lower average acceleration and smoother actions on the steering wheel were associated with higher confidence and comfort scores (Table 8).
In UC3-4, there was a direct correlation between the minimum vehicle speed and the confidence scores: confidence scores tended to be lower when vehicles were driven more slowly and vice versa. In addition, confidence and comfort scores tended to be lower for higher magnitudes of action on the steering wheel. The average rotation speed of the steering wheel was inversely correlated with comfort and confidence scores, and the maximum rotation speed was inversely correlated with the confidence score (Table 9).
In UC2, there was no significant correlation between objective and subjective metrics.
Thus, the analysis of the subjective metrics shows that -Hypotheses H2.1 and H2.2 are supported. Participants have better comprehension of the information provided with the AR-HUD than with the C-HUD (H2.1), and they find the AR-HUD more useful (H2.2). However,  the AR-HUD is viewed as less easy to understand than the bottom HUD information (the turn-by-turn icon, turnby-turn distance, and road sign). -Hypotheses H2.3 and H2.4 are not supported. There was no effect of the HMI on the confidence and comfort scores.
Those results indicate, once again, the special status of UC3-4, which is associated with the lowest confidence and comfort scores. They also show the learning or experience effect on these scores and uncover significant correlations between several objective metrics on the one hand and confidence and comfort scores on the other.

Discussion
This study shows the higher performance of participants using the AR-HUD in comparison with those who used the C-HUD, and at the same time, it brings to light the problems encountered with the AR-HUD.
In addition, it confirms the singular status of UC3-4. Associated with a high risk of failure and with the lowest confidence and comfort scores, it appears to be the most difficult situation.

Superior performance with the AR-HUD
The first hypotheses (H1.1 and H1.2) are validated. The AR-HUD has a positive effect on rule compliance in UC1, and it is associated with an increased mean speed. Speed adjustments are known to enable drivers to deal with the difficulties they encounter (Fuller 2005). In a complex situation, slowing down provides extra time to obtain useful information and to understand the situation. The study results show that in the most complex UC (UC3-4), speed, and specifically minimum speed, correlates with the confidence score. Drivers adopt a low speed when they are unsure of the route to take. Nwakacha et al. (2013) also note the influence of the interface (arrow C-HUD vs. a sat-nav personal assistant) upon the speed adopted by the participants during a navigation task; the authors argue that "The lowest speed values recorded with the sat nav suggested the possibility that the participants may have found the tasks more difficult to carry out with the sat nav" (p. 265). Bauerfeind et al. (2021) also point out that participants drive more slowly with a conventional HUD than with an AR display.
These findings, based on the analysis of the objective data, are confirmed by the analysis of the subjective data. The map displayed in the C-HUD condition remained a secondary, even minor source of information in the most complex UC. Furthermore, the information provided by the AR-HUD is easier to understand and considered as more useful than the information provided by the C-HUD. These results confirm some of the hypotheses related to the user experience (i.e., H2.1 and H2.2).
However, and against H1.3, the HMI has no effect on the smoothness (vs. roughness) of the actions performed, nor on their variability. Furthermore, and against H2.3 and H2.4, it has no effect on confidence and comfort.

The AR-HUD: room for improvement
Those mitigated results suggest that the AR-HUD does not provide optimal support. Some participants explained that it precludes anticipation. It is necessary to specify that this navigation system is based on the SCANeR simulation software package and takes the surrounding traffic into consideration; hence, it indicates a manoeuvre only when the latter is possible. The AR-HUD supports drivers in the guidance phase, but it does not enable them to build an overall representation of the route. In addition, participants do not use the GPS system, certainly because it is somewhat basic. GPS versions that are zoomed and self-centring would surely help drivers construct a comprehensive vision of the route.
The AR-HUD is assessed as being more useful and easier to understand than the C-HUD, but less so than the indication of direction (road sign) or the turn-by-turn information. Some participants reported that the arrow in the augmented reality view "oscillates", which makes it difficult to interpret the information. Pfannmüller et al. (2015) also point out that lack of precision in the position of the information displayed in augmented reality has a negative impact upon drivers' performance and the user experience.

Limitations of the study
One limitation of this study is reflected in the participants' interviews. It is related to the fact that the version of the AR-HUD that was used may be considered as not mature enough to play the role of "paradigm".
The other main limitation of this study is that eye-tracking data were not collected. The analysis of the sources of information used by the participants was based only on verbatim statements and the number of glances in the rear-view mirrors. A more detailed investigation of the visual scanning processes would have been useful, since it has been shown that the AR-HUD could have a negative impact on the perception of a dangerous or unexpected event ) that may frequently occur in an urban context. As explained by Prinzel and Rissel (2004), using an HUD in complex situations may lead to some dangers because of focused attention; "the eyes become fixated on the symbology and, without the benefit of scanning, may reduce the effectiveness of the HUD to perceive peripheral information" (Prinzel and Rissel 2004, p.8).
Even if no difference between AR-HUD and C-HUD was noticed concerning the glances in the rear-view mirrors, a detailed investigation of the visual scanning processes would be necessary to verify the following statement: "Optimal AR graphics would require a few glances of short duration in the direction of the graphics and would increase the amount of visual attention available for drivers to allocate to other areas with potential hazards and other driving-relevant information" (de Oliveira Faria et al. 2020, p. 764).

Conclusion
This study provides methodological and practical contributions to the research field of AR navigation interface in complex driving environments.
From a methodological perspective, its main originality is to reproduce real driving situations, experienced as complex by drivers and assessed as such as regards the factors defined by Boelhouwer et al. (2020). The study uses objective and subjective data and shows their complementarity. The mean speed variable discriminates between AR-HUD and C-HUD, which confirms the value of examining the effect of an interface on driving performance in terms of speed control. The study also draws attention to the variables relating to the steering wheel rotation speed (i.e., mean and maximum) whose value correlates negatively with the confidence score and the comfort score when the situation is very complex. This result complements the existing findings, indicating that driving task complexity has an impact on steering wheel corrections (see, for example, Faure et al. 2016).
From a practical perspective, the study demonstrates the value of using an AR-HUD to complete a navigation task in a driving situation characterised by several factors of complexity: a complex infrastructure, little guidance information, and the presence of other vehicles. However, the limitations pointed out by the participants, particularly those relating to the difficulties of anticipation, call for further investigations. On the one hand, this study paves the way for the investigation of a finer-grained incorporation of the traffic situation (i.e., contextualizing the presentation of information rather than delaying it). On the other hand, it leads to further consideration of ways of helping drivers not only in the phase preceding the manoeuvre (the "identify" stage in Burnett's model) but also in the "preview" phase that enables drivers to construct a representation of the manoeuvre. Topliss et al. (2019), who have identified the same limitation, suggest incorporating support using augmented reality in a system comprising other sources of information (e.g., a map or spoken instructions). Hence, it is necessary to consider integrating different sources of information in an instrument system (Forzy 2002), so that they may be used as complements or alternatives depending upon the phases of the navigation activity and the characteristics of the situation.