Hagai Tapiro, Anat Meir, Yisrael Parmet & Tal Oron-Gilad

Presentation at HFES-EU Annual meeting, Torino 2013

Abstract

Children are over-represented in road accidents, often due to their limited ability to perform well in road crossing tasks. The present study examined children’s visual search strategies in hazardous road-crossing situations. A sample of 33 young participants (ages 7-13) and 21 adults observed 18 different road-crossing scenarios in a 180° dome shaped mixed reality simulator. Gaze data was collected while participants made the crossing decisions. It was used to characterize their visual scanning strategies. Results showed that age group, limited field of view, and the presence of moving vehicles affect the way pedestrians allocate their attention in the scene. Adults tend to spend relatively more time in further peripheral areas of interest than younger pedestrians do. It was also found that the oldest child age group (11-13) demonstrated more resemblance to the adults in their visual scanning strategy, which can indicate on a learning process that originates from gaining experience and maturation. Characterization of child pedestrian eye movements can be used to determine readiness for independence as pedestrians. The results of this study, emphasize the differences among age groups in terms of visual scanning. This information can contribute to promote awareness and training directions.

Dirichlet regression model and analysis

For each scenario, five areas of interest were defined (as shown in the Figure). The close range central area was defined as the 10 meters of road in each side from the pedestrian’s point of view (AOI 3). Then symmetrically areas to the right of the center and to the left were defined. The medium right/left range (AOIs 2/4) was the part of the road distant at least 10 meter to the right/left of the point of view but less than 100 meters away. The far right/left range (AOIs 1/5) was the part of the road at least 100 meter or more to the right/left of the pedestrian point of view.

Open this link to see a sample video of a scenario as seen by a young pedestrian

Why Dirichlet?

• For each participant and scenario, the total Gaze distribution over the five AOI’s sums up to one.
• Therefore Gaze distribution is compositional data i.e., non-negative proportions with unit-sum.
• These types of data arise whenever we classify objects into disjoint categories and record their resulting relative frequencies, or partition a whole measurement into percentage contributions from its various parts.
• Attempts to apply statistical methods for unconstrained data often lead to inappropriate inference.
• Dirichlet regression suggested by Hijazi and Jernigan (2009) is more suitable for such cases.

How to use?

• The Dirichlet regression model was fitted using DirichletReg package, in R Language. Applying a backward elimination procedure found the best fitting model has three significant main effects.

What did we find?

• The dependent variable was the vector of AOIs and the independent variables were Age-group, POV and FOV; all of them were statistically significant (p <0.05). Predicted means for the percentage of time spent in each AOI for each age group based on the Dirichlet regression model are shown in the following figure and reveal differences among age groups. Note how children aged 9-10 spend more time gazing at the central area, note also the differences between mid-left and mid-right.

Predicted means (in each AOI) using Dirichlet model across all scenarios per age group