Abstract
It is important that advanced driver assistance systems provide drivers with alarms at the time when the driver fails to detect vulnerable road users (VRUs). In this study, we employed the paradigm which was developed by Yang et al. to analyze the detectability of VRUs at the intersection. Nine VRUs were involved, including pedestrians, cyclists, and motorbikes in the same direction as the driver, in the opposite and orthogonal direction to the driver. As the results, the detectability of pedestrians and motorbikes were lower than the detectability of cyclists. On the other hand, the detectability of VRUs in the same direction as the driver was lower than the detectability of VRUs in other directions.
Introduction
Intersections are recognized as hotspots for accidents, with more than half of all accidents in Japan occurring at or near intersections (Japanese Police Department, 2022). Catin et al. (2009) highlighted that driver distraction at intersections poses a significant threat to vulnerable road users (hereafter, VRUs). Driver distraction often results in the failure to detect VRUs effectively, thereby increasing the likelihood of accidents. Consequently, Advanced Driver Assistance Systems (hereafter, ADAS) have been widely adopted to support drivers in detecting VRUs and reducing accidents. ADAS issues alarms to drivers in hazardous situations. However, according to Lee et al. (2002), drivers may experience discomfort when ADAS alarms are perceived as being issued too early. Therefore, it is crucial that ADAS functions effectively and provides drivers with necessary information regarding detectability of VRUs at appropriate times. Currently, numerous studies are being conducted on detectability of VRUs. For instance, Kircher and Ahlström (2020) investigated detectability of cyclists for truck drivers. However, most studies focus primarily on the detectability of a single type of VRUs.
To address this issue, the authors’ previous study (Yang et al., 2024) developed a framework capable of examining the detectability of various types of VRUs and investigated how detectability of VRUs is influenced by the temporal changes in collision risk. Utilizing the Fuzzy Signal Detection Theory (Parasuraman et al., 2000), the detectability index (hereafter, d') was analyzed as a function of the time to the point of closest approach (hereafter, t), exploring how changes in t affect d'. The higher the value of d', the more accurately the driver can detect VRUs from the environment. Furthermore, to clearly understand how drivers detect to different VRUs, four detectability patterns were defined:
1. Gradual Increasing Pattern: With the reducing of t, d’ gradually increases (Figure 1a). This pattern indicates that as t shortens, more drivers can detect VRUs with increasing accuracy, thus, it is considered a safe pattern.
2. High Detectability Pattern: Regardless of t, d’ remains consistently high (Figure 1b). This pattern signifies that many drivers can always detect VRUs accurately, irrespective of the length of t, thus, it is considered a safe pattern.
3. Immediate Increasing Pattern: d’ remains at a low level and rises just before reaching the point of closest approach to the VRUs (Figure 1c). This pattern indicates that the proportion of drivers noticing VRUs remains low for a long period and only increases just before a collision occurs, thus, it is considered a dangerous pattern.
4. Low Detectability Pattern: d’ consistently remains low regardless of t (Figure 1d). This suggests that many drivers cannot accurately detect VRUs until reaching the point of closest approach, leading to an inability to avoid collisions, thus, it is considered a dangerous pattern.
In the previous study (2024), four detectability patterns were defined, but no examination was conducted on the classification of these patterns. This study proposes a method to classify the four patterns through statistical analysis (Figure 2).
Four types of detectability patterns.
Typologies of detectability patterns.
Firstly, if the d’ at t = 5s/6s is already two or more, this pattern is classified as High Detectability Pattern, Secondly, if d’ at t = 5 /6 s is less than two, the next step is to compare the d’ at t = 5/6 s with the d’ at t = 3 s. If the d’ at t = 3s is statistically significantly higher than t = 5/6 s, this pattern is classified as Gradual Increasing Pattern, Thirdly, if there is no significant difference, the next step is to compare the d’ at t = 3 s with the d’ at t = 0 s. If the d’ at t = 0 s is statistically significantly higher than the d’ at t = 3 s, this pattern is classified as Immediate increasing Pattern. And lastly, if there is no significant difference between the values of d’ at t = 3 s and t = 0 s, this pattern is classified as Low Detectability Pattern.
In regard to the classification method, the focus was on the timing of t = 5/6 , 3, and 0 s. The rationale behind each of these is elucidated below. The reason for focusing on the d' at t = 5/6 s (Study 1: t = 5 s, Study 2: t = 6 s) is that we can assess when there was sufficient time to avoid the collision based on the d’ at this point. Wada (2022) indicated that if there is a grace period of approximately 5 s remaining before a collision occurs, the driver is able to decelerate independently and avoid the collision. The reason for focusing on the d' at t = 3 s is that we can assess whether the driver is able to avoid the collision without assistance based on the d’ at this point. When 3 s remain before a collision occurs, it becomes challenging to avoid it independently and it becomes necessary to present an alarm or to allow the car to brake automatically (Hirose & Kubota, 2017). And lastly, the reason for focusing on the d’ at t = 0 s is that we can assess whether the VRU was detected immediately before the collision based on the d’ at this point. The greater d', the more accurately VRUs can be detected. A value of two or above is considered sufficient for accurate detection. (Liang et al., 2016). In light of the preceding studies, a method to classify the four patterns was proposed.
This study concentrated on right-turn at intersections and the VRUs moving in the same direction as the driver. For right-turn at intersections, Yoshitake et al. (2020) pointed out that it poses a higher risk of accidents with VRUs compared to left-turns and straight-ahead movements, due to the involvement of multiple lanes and traffic directions. For the VRUs moving in the same direction as the driver, Hagiwara et al. (2015) argued that they are often within the driver’s blind spot, resulting in a higher risk of accidents compared to VRUs moving in the opposite direction. Therefore, this study will use statistical tests to examine in detail the more dangerous right-turn scenarios (Study 1) and VRUs moving in the same direction as the driver (Study 2).
Experiments
This study examined how d' changes with the shortening of t. Specifically, d' was used as the dependent variable, while the time margin t until the driver’s vehicle reaches the point of closest approach with the VRUs was used as the independent variable (Study 1: t = 5, 4, 3, 2, 1, 0 s; Study 2: t = 6, 3, 0 s; Figure 3a). As moderators, the distance of the closest point of approach (hereafter, DCPA) (1, 5 m), was used (Figure 3b). Driving videos for each condition were created using Unity, and experiments were conducted based on these videos. Additionally, videos without the target VRUs were also created for the calculation of d'. The videos were set so that the driver’s vehicle and the VRUs would reach the point of closest approach at 20 s into the video playback.
Descriptions of t/DCPA condition.
The evaluation targets differed between the two studies. Study 1 focused on five types of VRUs during right turns (motorbikes, cyclists, pedestrians in opposite direction, cyclists in orthogonal direction, and pedestrians in same direction). Study 2 focused on five types of VRUs moving in the same direction (motorbikes when driver turn left, cyclists and pedestrians when drivers turn left/right; Figure 4).
The left figure shows the evaluation targets for Study 1, and the right figure shows the evaluation targets for Study 2.
Participants for the Studies 1 and 2 were recruited through the cloud service Crowd works. 320 participants (166 males, 153 females, one other,
The procedures for the two studies were identical. At the beginning of the experiment, participants were randomly divided into multiple groups based on the t conditions (Study 1: six groups, Study 2: three groups) to ensure that a single participant dose not observe more than one t condition for any evaluate target. The reason for this is that if multiple t conditions are observed, the subsequent one may exhibit a higher d'.
The experiment consisted of an observation phase and a test phase. In the observation phase, participants were instructed to watch videos from the actual driver’s perspective. Text information indicating whether the video showed a right turn, or a left turn was presented at the bottom of the video. The video stopped and blacked out t seconds before reaching the closest point of approach (Figure 5). For example, in the case of t = 3 s, the video ended 3 s before reaching the closest point of approach, that is, at 17 s into the playback.
Observation phase’s screenshot.
Following the observation phase, the test phase commenced. The image of the evaluation target, and the location of the participant’s vehicle and evaluation target at the end of the video, were displayed on the screen. Participants were asked to answer the question, “When the video stopped, was the target shown on the screen present” They responded using a 101-point slider bar. This evaluation scale ranged from “Definitely not” (0) on the left end to “Definitely was” (100) on the right end (Figure 6).
Test phase’s screenshot.
Results and Discussions
The results of Study 1 (Figure 7) indicated that the detectability patterns of pedestrians in the DCPA 5 m condition, motorbikes, cyclists in opposite direction and orthogonal direction showed Gradual Increasing Pattern. High Detectability Pattern was showed in the detectability patterns of pedestrians in opposite direction in the DCPA 1 m condition. Additionally, the detectability patterns of pedestrians in the same direction in the DCPA 1 m condition showed Immediate Increasing Pattern. Furthermore, the detectability patterns of pedestrians in the same direction in the DCPA 5 m condition showed Low detectability Pattern. The detectability patterns of VRUs in the opposite and orthogonal direction showed safe patterns, while the detectability patterns of VRUs in the same direction showed dangerous patterns.
The results of Study 1.
The VRUs in the same direction approached from behind the driver’s vehicle, resulting in shorter visibility time and lower visibility compared to VRUs in the opposite and orthogonal directions, making them more challenging to detect. This indicates that the direction of the VRUs significantly influences the detectability of VRUs during right turns.
The results of Study 2 (Figure 8) showed that the detectability pattern of cyclists showed High Detectability pattern. The detectability patterns of motorbikes and pedestrians in the DCPA 1m condition showed Immediate Increasing Pattern. The detectability pattern of pedestrians in the DCPA 5 m condition showed Low detectability pattern. The detectability patterns of motorbikes and pedestrians in the same direction as the driver’s vehicle showed dangerous patterns, while the detectability patterns of cyclists showed safe patterns.
The results of Study 2.
According to the study by Kaya et al. (2021), drivers with bicycle experience are likely to enhance their visual attention related to cyclist detection while driving. Japan exhibits a relatively high rate of bicycle ownership, with an estimated one bicycle per every two individuals. Therefore, the detectability of cyclists is considered higher than pedestrians and motorbikes. This indicates that the type of the VRUs significantly influences the detectability of VRUs in the same direction as the drivers.
The findings of this study are expected to contribute to the improvement of ADAS alarm functions. ADAS issues alarms to drivers in hazardous situations. However, the deployment of alarms for targets that already receive sufficient attention may result in the inappropriate allocation of the driver's cognitive resources. This study indicated that when making right turns at intersections, ADAS does not need to issue strong alarms for VRUs in the opposite direction, but rather more prominent alarms are necessary for VRUs moving in the same direction due to the potential difficulty for drivers in detecting them. Furthermore, the detectability of VRUs moving in the same direction also varies depending on the type of VRUs. Since cyclists are detected more effectively, ADAS should issue weaker for cyclists.
Finally, the limitations and future directions of this study will be discussed. First, the control of the speed of VRUs and the time they remain within the driver’s field of vision was insufficient in this study. VRUs that remain within the driver’s field of vision for longer periods are likely to be detected more easily. Therefore, future research should rigorously consider these factors. Second, the study involved participants viewing driving videos, but did not require them to perform any driving operations. Future research should follow up these studies in more realistic environments such as VR or driving simulations.
Footnotes
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: DENSO CORPORATION supported this work. However, note that this article solely reflects the opinions and conclusions of its authors and not DENSO CORPORATION.
