Keywords

1 Introduction

Pedestrians are the most vulnerable road users that accounting for 22% of the fatalities in traffic accidents (World Health Organization 2013). The conflicts between vehicles and pedestrians often occur as a result of unsafe interaction (Kaparias et al. 2015; Ni et al. 2016), especially when drivers do not yield on time or pedestrians inappropriately accept a small gap to cross. To solve the former issue, the desire for deploying automated vehicles to increase crossing safety has surged up these years (Köhler et al. 2013; Schneemann and Heinemann 2016). And for the later issue, researchers suppose pedestrians should be partially responsible for their safety without any help from the drivers, thus previous studies attempted to identify pedestrian perceptual and cognitive failures that may lead to unsafe road-crossing decision-making (Koh and Wong 2014; Pawar and Patil 2015).

Automated vehicles were expected to adopt a conservative strategy, where once the automated system recognizes pedestrians’ intention to cross, they should give way to pedestrians (Schneemann and Heinemann 2016). In order to help pedestrians to recognize and accept vehicles’ yielding, automated vehicles were suggested to equip the external human-machine interface (eHMI) to convey vehicle intentions. The eHMI was expected to establish a new interaction mode with other road users, to compensate for the absence of common human-human interactions (Rasouli and Tsotsos 2019; Schieben et al. 2019). Following this line of research, many studies have evaluated the effect of different types of messages (texts, symbol, lights, animations, etc.) on pedestrian stated feeling of safety and crossing behavior, especially when automated vehicles would decelerate or yield to pedestrians (Holländer et al. 2019; Lee et al. 2019a, b; Nuñez Velasco et al. 2019). For example, pedestrians were found to cross more often when they encountered a yielding vehicle with the eHMI, which was thought to increase the efficiency of interaction between pedestrians and automated vehicles (de Clercq et al. 2019; Song et al. 2018). In another study, researchers added eyes on the car to establish an eye-contact communication between automated vehicles and pedestrians (Chang et al. 2017). When the automated vehicles equipped with this “eyes on a car” interface, pedestrians could make road-crossing decisions more quickly and feel safer.

However, in contrast, some studies implied the limited effect of such communication information. In a video-based road-crossing task, Nuñez Velasco et al. (2019) suggested that automated vehicles equipped with interfaces to communicate with pedestrians could only increase their perceived risk, but pedestrians’ crossing intentions were not affected. Some researchers reckoned that pedestrian may rely on their legacy strategy to make decisions (e.g., estimating time-to-arrival, vehicle speeds or distance to make decisions), instead of the communication information (Clamann et al. 2017). Another possible explanation is that vehicle deceleration rate could also serve as informal communication information to help pedestrians understand the intentions of automated vehicles, which then could support pedestrians to detect yield behavior and cross immediately (Ackermann et al. 2019; Fuest et al. 2018). Besides, pedestrians’ understanding of the automated vehicles’ intentions may vary for different designs of the message (Lee et al. 2019a) or the level of familiarity (i.e., learning effect) (Lee et al. 2019b). Judging from the above studies, pedestrians could recognize intentions of automated vehicles by communication information, but it is still not determined that to what extent such information affects pedestrians’ road-crossing decision-making.

Compared with the conservative strategy, the current study suggests a strategy to assist pedestrians’ gap selection decision-making rather than consistent yielding. In this strategy, automated vehicles would not yield to pedestrians when pedestrians do not have road rights. Instead, automated vehicles just send assistance information to pedestrians to support the existing gap selection strategy to increase safety. This strategy is necessary to ensure efficient vehicle flow, especially at places where pedestrians do not have road rights. Besides, considering pedestrians’ decision-making is easy to be affected by several factors (e.g., vehicle size, vehicle speeds, traffic environment) (Kaparias et al. 2015; Pawar and Patil 2015; Petzoldt 2016), this strategy is expected to reduce errors in pedestrians’ gap selection decision-making. By this strategy, automated vehicles behave similarly to human-driven vehicles except for the assistance information. Will pedestrians make gap acceptance decisions based on assistance information or will they treat the automated vehicles just as human-driven vehicles? It is necessary to compare pedestrians’ road-crossing behavior when interacting with these two types of vehicles and identify pedestrian strategies. Moreover, if the pedestrians had different crossing strategy towards the two types of vehicles, then whether automated vehicles driven in traffic that were surrounded by human-driven vehicles may be another concern to address. We expect that pedestrians may need to transit strategies constantly if the two types of vehicles driven in mixed traffic conditions, which causing additional cognitive load.

2 Methods

2.1 Participants

Participants were recruited with rewards from Shaanxi Normal University. A sample of 48 students (mean age = 18.78 years, SD = 1.08 years, 17 males and 31 females) took part in the experiment. All participants had normal or corrected to normal visions.

2.2 Apparatus

A custom program was used to carry out the experiment. The program was developed based on the jMonkeyEngine, which is a 3-D game engine developed by Java language. The program ran on a high-performance computer and render the virtual environment to an HTC VIVE Pro headset. The headset refreshed at 90 Hz with a resolution of 2880 * 1600 pixels (1440 × 1600 pixels per eye) and with a field of view of 110°.

The model of road environment (as shown in Fig. 1.) was developed by Esri CityEngine and Blender software. A four-lane, two-way road was used and the lane width was 3.5 m. The sidewalks, street trees, and road lamps were also included in order to provide a realistic experience of the traffic environment.

Fig. 1.
figure 1

The traffic scene from pedestrians’ perspective (human-driven vehicles were presented)

2.3 Stimulus and Design

Two types of vehicles were designed to simulate human-driven vehicles and automated vehicles. Compared to the appearances of human-driven vehicles, the simulated automated vehicles were equipped with a signal on top to distinguish itself with human-driven vehicles and a frontal display conveying the assistance information to pedestrians. Vehicles could either drive at a mixed traffic environment (human-driven and automated vehicles shared a lane and were presented randomly) or not (only human-driven vehicles or automated vehicles were present). Therefore, there are four kinds of interaction situations between pedestrians and vehicles as follows:

  1. (1)

    pedestrians interact with human-driven vehicles only (HUM-only);

  2. (2)

    pedestrians interact with automated vehicles only (AV-only);

  3. (3)

    pedestrians interact with human-driven vehicles in a mixed traffic environment (HUM-mixed);

  4. (4)

    pedestrians interact with automated vehicles in a mixed traffic environment (AV-mixed).

Vehicles approached pedestrians at a constant speed of 30 km/h. When the vehicle interacted with pedestrians, the time gap ranged from 1.4 s to 6.4 s (at 1.0 s intervals). For automated vehicles, the assistance information was designed to reflect the risk of crossing, according to the time-to-arrival (TTA) between vehicles and pedestrians. The signal simulated the common traffic light with three states in red (TTA < 3.4 s), yellow (3.4 s < TTA < 4.8 s) and green color (TTA > 4.8 s). At the beginning of each trial, participants always required to stand 0.5 m away to the curb. Therefore, they had to walk 4 meters to complete the crossing task during the experiment.

2.4 Procedure

At first, participants were explained about the task and procedure of the experiment. They were told that the vehicle would not yield or decelerate, and they had to accept an appropriate time gap to cross safely. In addition, they were explained about the meaning of assistance information: the red state means that it may be dangerous if they chose to cross, even they crossed in a hurry; the green state means it would be safe to cross, even they crossed at leisure; and the yellow state is a situation between the red state and green state, whereas it is not recommended to cross for the reason of safety.

Then they had to cross the lane for at least three times, in order to get familiar with the environment and adapt themselves to walk with the headset in the virtual reality environment. After that, they had to complete eight blocks of crossing task. The order of blocks was balanced. There were two blocks for only human-driven vehicles (HUM-only) or only automated vehicles (AV-only) respectively and four blocks for mixed traffic environment (HUM-mixed and AV-mixed). In each block, participants would cross the road for 15 times (i.e., 15 trials).

For each trial, a sequence of vehicles approached the pedestrian from left to right. The first vehicle always appeared very close to the pedestrian with the time-to-arrival of 1 s (excluded for the analysis), in order to present an ecological perception for the following vehicles. And the following vehicles provide a randomized gap from 1.4 to 6.4 s (at 1.0 s intervals). Participants could walk to cross the road at any time at any walking pace and then go back to the starting position to prepare for the next trial. The experiment lasted about 45 to 60 min.

3 Results

For each trial, pedestrians could either accept or reject the current time gap. Therefore, trials were labeled as “accepted” or “rejected” respectively according to the movement of the VR headset. Because of the limitation of the program, if pedestrians started to cross when a vehicle was passing by, it would record a response with an incorrect timestamp (13 trials were excluded from all analyses). To sum up, a total of 12218 decisions (6471 waiting decision and 5747 crossing decision) were used for the following analyses.

Considering the binary nature of crossing decisions, a Generalized Linear Mixed Model (GLMM, using glmer function from lme4 package 1.1-21 for R 3.6.1) with a logit link function was used for modeling pedestrians’ decisions. Specifically, the independent variables (interaction conditions and gap size) and their interactions were modeled as the fixed effects, with a random effect of individual difference (i.e., the difference of individual from the overall participants), by maximum likelihood method. The linear mixed model (LMM) was also used to estimate the reaction time of pedestrians. Continuous variables were scaled (M = 0, SD = 1) to fit the model.

The summary of the model is shown in Table 1. The predicted probability was calculated according to the model and was plotted in Fig. 2. Both the main effects of situations and gap size are significant (ps < 0.02), as well as the interaction effect (p < 0.001).

Table 1. Result of the generalized linear mixed model for pedestrians’ road-crossing decisions.
Fig. 2.
figure 2

Predicted probability of pedestrians’ gap acceptance

The results indicate that, in general, the likelihood of pedestrians accepting a specific gap increased with gap size (p < 0.001). More importantly, pedestrians crossed more often when interacting with automated vehicles, regardless of whether in a mixed traffic condition (ps < 0.05). The significant effect of interaction between gap size and interaction situation means that the transition from rejecting to accepting was more abruptly when pedestrians interacted with automated vehicles, regardless of the traffic environment (ps < 0.02). No significant difference was found between HUM-only and HUM-mixed conditions, or between AV-only and AV-mixed (all ps > 0.17). As shown in Fig. 2, for larger gaps, pedestrians crossed more often in front of automated vehicles than human-driven vehicles.

4 Discussion

The current study compared pedestrians’ gap acceptance behavior when interacting with human-driven vehicles and automated vehicles in the different traffic environments. For both two types of vehicles, pedestrians showed a general tendency to cross at larger gaps. In addition, pedestrians’ road-crossing decision-making was proved to be affected by assistance information when interacting with automated vehicles. Pedestrians accepted more large gaps in front of automated vehicles. And when automated vehicles and human-driven vehicles shared a lane, the effect of assistance information remained the same as they drive separately.

When pedestrians interact with human-driven vehicles, especially when human-driven vehicles drive on a separate lane, pedestrians’ gap acceptance behavior reflected the legacy strategy. For automated vehicles, the assistance information showed the objective and unbiased estimated risk of crossing, which should be more reliable than pedestrians’ estimation of the crossing risk. Compared with human-driven vehicles, pedestrians accepted more larger gaps when interacting with automated vehicles. And pedestrians’ gap acceptance behavior was more sensitive to object time-to-arrival with the help of assistance information. This result suggested that the proposed strategy which just conveys information to support existing gap acceptance behavior was feasible.

Pedestrians’ gap acceptance behavior also indicated that pedestrians inclined to make decisions based on their legacy strategy. For example, when gap size smaller than 3.4 s, the automated vehicles reminded pedestrians a higher level of risk to cross for smaller gaps. If pedestrians take a strategy to follow the signals conveyed by automated vehicles, they would not cross for small gaps. However, pedestrians still crossed in such cases (TTA <= 3.4 s) from time to time as they do when interacting with human-driven vehicles.

Studies have proved that the accuracy of time-to-arrival estimation is affected by several factors in traffic, such as vehicle size (Petzoldt 2016), approaching speed (Petzoldt 2016, 2014), driving experience (DeLucia and Mather 2006), age of perceivers (Rusch et al. 2016). Estimation of time-to-arrival was reckoned as a main source of information for pedestrians’ gap acceptance decisions (Petzoldt 2014). In the current study, the vehicle size, vehicle speeds remained unchanged, which then helped pedestrian estimate time-to-arrival more accurately. Even so, assistance information still played a role in pedestrians’ decision-making. It is expected that in a more complex traffic environment where different kinds of vehicles present with changed vehicle speeds, automated vehicles with assistance information would further increase crossing safety.

As to the traffic environment, no difference was found for human-driven vehicles in mixed or non-mixed traffic, nor the automated vehicles. One reason may be that in the current study, assistance information only had a minor effect on pedestrians’ decision. Considering unchanged vehicle speeds, reduced uncertainty of time-to-arrival estimation may serve as the main source of information, even when assistance information was conveyed by automated vehicles. Therefore, similar strategies were employed for all four kinds of interaction situations. A further step is to introduce the variability of vehicle speeds and compare pedestrians’ strategies of decision-making when their estimation of time-to-arrival became more inaccurate.

5 Conclusions

In the current study, automated vehicles just send messages to assist pedestrians’ gap acceptance decisions, whereas vehicles did not yield as usually do. Assistance information conveyed by automated vehicles was found to help pedestrians make safer road-crossing decisions. No difference for the traffic environment reflected pedestrians rely on their legacy gap acceptance strategy to cross and assistance information had a minor effect.