A brief pc-based hazard prediction training program improves young novice drivers’ hazard perception skills compared to a control group over time

Over time, the brief PC-based hazard prediction training program improved young novice drivers’ hazard perception skills compared to a control group.

Authors:van der Kint, S. T., van Schagen, I., Vlakveld, W., Mons, C., de Zwart, R., & Hoekstra, T.

1.Introduction

Young novice drivers have a high accident rate, partly due to poor hazard perception skills. This study developed and tested a training method to investigate whether a brief PC-based hazard prediction training could enhance these skills. Previous research developed an online training program that did not consist solely of “what happens next” exercises and lasted six sessions instead of one. This study aims to analyze the effects of a hazard prediction program that only includes “what happens next” exercises, conducted in a single phase. The study helps participants predict potential hazards and explains how experienced drivers avoid these hazards.

2.Research Content

This study evaluates a brief PC-based hazard prediction training, specifically the “what happens next” exercises. Both groups completed simulator driving before and after the intervention, during which gaze directions to explicit and implicit potential hazards were recorded. Approximately five months later, participants from both groups and another group of experienced drivers completed an online hazard perception test. Therefore, four hypotheses were proposed:

(1) In the second simulator driving (post-test), the hazard prediction training group gazes at more potential hazards on average than the control group (including explicit and implicit potential hazards), while controlling for possible differences between the two groups during the first simulator driving (pre-test);

(2) Five months later, participants in the hazard prediction training group score higher on the hazard perception test than participants in the control group;

(3) Experienced drivers score higher than participants in the control group on the online hazard perception test, but not necessarily better than participants in the hazard prediction training group;

(4) There is a correlation between the number of potential hazards participants noticed during simulator driving and their scores on the online hazard perception test.

3.Research Method

3.1 Experimental Participants

The study recruited participants for the hazard perception training group and control group from the TeamAlert participant database and social media. Inclusion criteria were: aged between 18 and 24 years, holding a valid driver’s license, no known tendency to motion sickness (as participants had to drive in a simulator), and not wearing glasses or hard contact lenses. The inclusion criteria for experienced drivers were: holding a driver’s license for more than 10 years, aged between 34 and 60 years, and driving more than 5000 kilometers annually. Young novice drivers were randomly assigned to the hazard prediction training group or the control group. One participant from the hazard prediction training group and one from the control group did not complete the online hazard perception test approximately five months later. Sample characteristics are shown in Table 1.

Table 1 Sample Characteristics

3.2 Experimental Setup and Stimuli

(1) Hazard Prediction Training

The hazard prediction training is a PC-based training program consisting of eight training modules. The training includes eight short videos filmed from the driver’s perspective, culminating in a collision or emergency braking incident. The training includes four dashcam videos showing implicit potential hazards and four animated videos showing explicit hazards. In this study, a workshop with eight driving examiners was organized, who were informed about what constitutes explicit and implicit hazards. In the following 1 minute, they predicted as many “what happens next” scenarios as possible. After that, the videos were replayed, including the remaining parts of the collision or emergency braking events. This was accompanied by a countdown timer on the screen. After the countdown ended, the entire video, including the collision, was played. Then, using highlighted areas and arrows in the video, a voice-over explained why the collision occurred and other potential hazards that experienced drivers should have noticed to avoid a collision or emergency braking incident. The time to complete the hazard perception training was approximately 20 minutes. The collision scenes from each video are shown in Table 2.

Table 2 Description of Training Videos

(2) Control Group Training

The control group training was as similar as possible to the hazard prediction training, without training in hazard prediction. It consisted of six dashcam videos, which were also played twice, just like the hazard perception training. After the second presentation of the videos, participants had to answer a question regarding the content they had seen in the videos, such as: “How many bus stops did you see in the video?” Or after another video: “How many oncoming cars did you see?” They then had 15 seconds to write down their answers before replaying the video and providing the answers. These videos were clips taken from online dashcam videos to enhance engagement and participants’ attention to the videos.

(3) Experimental Scenario

Participants took approximately 15 minutes to complete the simulator driving. The scenario included urban roads, rural roads, and a two-lane highway. During this driving, there were eight potential hazards: four explicit potential hazards and four implicit potential hazards. These potential hazards did not materialize as the simulated car approached them.Potential hazards embedded in the driving simulator are shown in Table 3.Due to only two near-transfer scenarios in the driving simulator, the effect of training on near-transfer scenarios and far-transfer could not be compared.

Table 3 Potential Hazards Embedded in the Driving Simulator

(4) Questionnaire

During the simulator study, participants were asked to complete two short questionnaires. The first questionnaire was completed before the first simulator driving and asked about demographics and driving experience, such as years holding a driver’s license and annual mileage. The second questionnaire was completed after the simulator driving, asking how realistic they thought driving in the simulator was and whether they experienced symptoms of simulator sickness.

(5) Online Hazard Perception Test

In the online hazard perception test, clips were not included because they minimally distinguished between professional drivers and learning drivers in this study. All ten animated video clips were filmed from the driver’s perspective and included a high-priority potential hazard (cover or obvious potential hazard), but did not materialize. All video clips also included several minor potential hazards, which were much less likely than the high-priority hazards. The scenes of the animated video clips were created in a workshop with eight driving examiners. After explaining the conceptual distinction between explicit and implicit potential hazards, they were asked if they could mention events where candidates failed to notice and identify potential hazards. The examiners also mentioned potential hazards that were unlikely to materialize in each video clip. For each item, participants first watched the entire video clip. Before the video clip, a simple overhead view was presented showing the movement of the video car (e.g., a pictogram of a car turning left with the text “You are turning left”). Participants were told to imagine themselves as the driver of the car and pay attention to situations that could develop into likely collisions. They were told that there would be no urgent situations requiring immediate action to avoid a collision. At the end of the video, four screenshots of the video clip appeared on the screen, one of which displayed the primary potential hazard. Participants were asked to select the screenshot that made them think, “Oh, I hope this doesn’t happen” while watching. After selecting the screenshot, participants had to enter their reasons for choosing that specific screenshot. When the correct screenshot was selected and a correct description of the potential hazard was provided, the potential hazard was considered correctly identified. Participants were instructed to complete the hazard perception test on their laptop or PC, not on their smartphones. To mitigate possible learning effects, the videos were presented in random order. Notably, the applied online hazard perception test differs from many other hazard perception tests because the high-priority potential hazards remained latent and did not fully materialize in the video clips.

(6) Driving Simulator

The driving simulator is a fixed-base research simulator consisting of a car seat in front of a fixed device with a steering wheel, pedal components, and gear shift. This model is placed in front of a configuration of three 50-inch LCD screens, each with a resolution of 1280 × 800 pixels. These screens are positioned approximately 84 centimeters from the “driver”’s head in the car seat, providing the “driver” with about 180° horizontal and about 40° vertical field of view. The refresh rate is 60 Hz. The driving simulator is shown in Figure 1.

Figure 1 Driving Simulator

(7) Head-Mounted Eye Tracker

During driving, participants’ gaze direction was recorded using the Pupil Lab Pro head-mounted eye tracker. This eye tracker uses an infrared light-based “dark pupil” detection method. The head-mounted device consists of a lightweight frame with a scene camera and two infrared eye cameras (one for each eye). The frame rate of the scene camera is 60 Hz, and the frame rate of the dual-eye cameras is 120 Hz. The system’s output includes video clips from the scene camera with a crosshair overlay indicating the gaze direction. When the eye tracker is correctly calibrated, the gaze accuracy is 0.60°.

3.3 Experimental Procedure

The experimental flow diagram of the study is shown in Figure 2. The simulator study was conducted at SWOV in The Hague. After being informed about the study, participants signed an informed consent form. They then completed a pre-questionnaire. After that, participants sat in the driving simulator and wore the eye tracker. First, each participant drove a short route to practice driving and get accustomed to the eye tracker. After that, the eye tracker was calibrated, and the first driving session was completed. After the training group participants completed the hazard perception training, the control group participants completed their training. Subsequently, all participants drove the same driving simulator, starting from different positions to mitigate learning effects. After the second driving session, participants completed a post-questionnaire, and the simulator experiment lasted 60 – 90 minutes.

Figure 2 Experimental Flow Diagram

About five months after the simulator experiment, all participants received a WhatsApp message containing a link to the survey questionnaire and were asked to complete the online hazard perception test on their PC or laptop. Experienced drivers were also asked to complete the hazard perception test on their personal computer or laptop. The online hazard perception test began with information about the test, and participants could only proceed after giving their consent. The version for experienced drivers included a section with questions about their demographics and driving behavior. Upon completion, participants from the experimental and control groups received a €20 gift card, but experienced drivers did not. Completing the online hazard perception test took approximately 20 minutes.

3.4 Experimental Design and Analysis

For the simulator study, a 2 × 2 mixed-factor design was used. Generalized linear mixed modeling (GLMM) with a logistic link function was applied to the binomial data. The between-subject factor was “group” (hazard prediction training group vs. control group), and the within-subject factor was “time” (pre-test vs. post-test). The dependent variable was the proportion of correctly gazed potential hazards (i.e., the number of “successes” in the seven “trials”). Both “group” and “time” were fixed factors, while participant ID numbers and the number of phase potential hazards were random factors. The same statistical analysis was applied to three phase explicit potential hazards and four phase implicit potential hazards.

For the online hazard perception test conducted approximately five months after the intervention, a 1 × 3 between-subject factor design was used. Generalized linear modeling (GLM) with a logit link function was applied to the binomial data, with participant ID and the number of video clips as random factors. The dependent variable was the proportion of correct responses to ten questions. The fixed factor was “group”: hazard perception training group, control group, and experienced driver group. Additionally, four explicit potential hazards and six implicit potential hazards were analyzed separately.

Ideally, the measurement of hazard perception skills should be the same before the intervention, after the intervention, and five months post-intervention. Then, preliminary analyses could cover all three measurements across time and group variations. In the hazard detection tests conducted using the eye-tracking driving simulator and the online hazard perception test, the stimuli were both implicit and explicit potential hazards that did not materialize. Although the stimuli were very similar, the measurement methods were different. Therefore, caution is needed when interpreting the results of the two testing methods. Due to the different types of measurements, Spearman correlation was applied to examine whether there was a correlation between scores on the driving simulator and scores on the online hazard perception test.

4.Research Results

4.1 Driving Simulator Study

The marginal estimated means and standard errors (SD) of the proportion of gazed potential hazards before (pre-test) and after (post-test) the intervention for all hazards (n = 7), only explicit potential hazards (n = 3), and only implicit potential hazards (n = 4) are shown in Table 4.

Table 4 Average Proportion of Gazed Potential Hazards in the First and Second Driving Simulator

For binomial data, GLMM with a logit link function was used, where the scores for all potential hazards after the test were the dependent variable, and the scores for all potential hazards before the test were covariates, with participant ID numbers and the number of seven phase hazards as random factors. From Table 4, it can be seen that the hazard prediction group scored significantly higher than the control group in the post-test, controlling for pre-test scores, with a medium effect size. This indicates that, prior to the intervention, the hazard prediction training group gazed at significantly more potential hazards during the second simulator driving compared to the control group. The same GLMM was applied for explicit potential hazards, showing that the hazard prediction group scored significantly higher in the post-test than the control group, with a small to medium effect size; under the pre-test scores of the control group, the hazard prediction group scored significantly higher than the control group, with a medium effect size. These results support Hypothesis 1, indicating that hazard prediction training overall improved scanning for potential hazards, and improved scanning for both explicit and implicit potential hazards separately.

4.2 Online Hazard Perception Test

For the online hazard perception test with a logit link function for binomial data, the scores for all ten items were the dependent variable, with “group” (hazard prediction training group, control group, experienced driver group) as the fixed factor, and participant ID numbers and the item number of the hazard perception test as random factors. The marginal estimated means and standard errors (SD) of the proportion of correctly identified potential hazards in the online hazard prediction test for the hazard prediction training group, control group, and experienced drivers are shown in Table 5.

Table 5 Mean Proportion of Correctly Identified Potential Hazards

From Table 5, it can be seen that the proportion of correctly identified potential hazards significantly differed among the three groups. Post hoc pairwise comparisons with Bonferroni correction indicated that the hazard prediction group scored significantly higher than the control group, with an effect size between small and medium. Furthermore, experienced drivers scored significantly higher than the control group, with a small effect size. However, there was no significant difference between the scores of the hazard prediction training group and the experienced drivers. The combined results of explicit potential hazards and implicit potential hazards in the online hazard perception test support Hypothesis 2, indicating that approximately five months later, the hazard prediction training group had better hazard perception skills than the control group. Experienced drivers scored significantly better than the control group in the hazard perception test, but there was no significant difference in the overall scores between the hazard prediction training group and the experienced drivers, supporting Hypothesis 3.

The same GLMM was applied to the four items of the hazard perception test with explicit potential hazards, showing significant differences in the average proportion of correctly identified explicit potential hazards among the three groups. Post hoc pairwise comparisons with Bonferroni correction indicated that the hazard prediction group did not score significantly better than the control group on the four explicit potential hazard items. However, experienced drivers scored significantly higher than the control group on these four explicit potential hazard items, with a medium effect size. Finally, the scores of the hazard prediction training group participants on the four explicit potential hazard items were not significantly lower than those of the experienced driver group.

The same GLMM was applied to the six items of the hazard perception test with implicit potential hazards, showing significant differences in the average proportion of correctly identified implicit potential hazards among the three groups. Post hoc pairwise comparisons with Bonferroni correction indicated that the hazard prediction group scored significantly higher than the control group on the six implicit potential hazard items, with an effect size from medium to small. However, although experienced drivers scored higher than the control group on the six implicit potential hazard items, their scores did not significantly improve. Finally, the participants in the hazard prediction training group scored only slightly higher than the experienced drivers on the six implicit potential hazard items. The results for implicit potential hazards support Hypothesis 2, but the results for explicit potential hazards do not support it. The results for explicit and implicit potential hazards only partially support Hypothesis 3, as experienced drivers outperformed the control group on explicit potential hazards but did not perform well on implicit potential hazards.

4.3 Gaze at Potential Hazards and Scores on Hazard Perception Tests

During the second simulator driving, all participants’ gazes at potential hazards were relatively weak, but significantly correlated with the scores of the hazard perception test approximately five months later. This partially confirms the correlation between gaze at potential hazards and scores on the online hazard perception test, supporting Hypothesis 4.

5.Conclusion

(1)The video-based hazard prediction (HP) training method improved young drivers’ attention to potential hazards.

(2)Participants who completed the hazard prediction training correctly identified significantly more potential hazards (both explicit and implicit) in the test than the control group.

(3)Compared to young novice drivers who completed the training, experienced drivers scored higher in the hazard perception test, but not necessarily better than those who completed the hazard prediction training.

6.References

van der Kint, S. T., van Schagen, I., Vlakveld, W., Mons, C., de Zwart, R., & Hoekstra, T. (2024). A brief pc-based hazard prediction training program improves young novice drivers’ hazard perception skills compared to a control group over time. Transportation Research Part F: Traffic Psychology and Behaviour, 102, 64-76.

7.Review

This training method improved young drivers’ attention to potential hazards, and five months post-training, young drivers retained their hazard prediction skills; additionally, the hazard prediction skills of trained young drivers were comparable to those of experienced drivers. However, the study still has the following limitations: the effects of hazard prediction training were not measured in real traffic. Therefore, it is currently unclear whether hazard prediction training can improve driving skills in the real world.

Note:Due to limited proficiency, inaccuracies or errors may exist in the translation; please refer to the original text and feel free to provide corrections.

Knowledge Transporter (Translator) | Zhang Siqing

Reviewed by | Gao Ya

Produced by | Hefei University of Technology, Transportation and Safety Research Institute

Copyright Statement

This article is exclusively written by the “Transportation and Safety Research Institute of Hefei University of Technology”. For reprints,

please indicate the source, thank you!