Visual Clutter Study for Pedestrian Using Large Scale Naturalistic Driving Data

Some of the pedestrian crashes are due to driver’s late or difficult perception of pedestrian’s appearance. Recognition of pedestrians during driving is a complex cognitive activity. Visual clutter analysis can be used to study the factors that affect human visual search efficiency and help design advanced driver assistant system for better decision making and user experience. In this thesis, we propose the pedestrian perception evaluation model which can quantitatively analyze the pedestrian perception difficulty using naturalistic driving data. An efficient detection framework was developed to locate pedestrians within large scale naturalistic driving data. Visual clutter analysis was used to study the factors that may affect the driver’s ability to perceive pedestrian appearance. The candidate factors were explored by the designed exploratory study using naturalistic driving data and a bottom-up image-based pedestrian clutter metric was proposed to quantify the pedestrian perception difficulty in naturalistic driving data. Based on the proposed bottom-up clutter metrics and top-down pedestrian appearance based estimator, a Bayesian probabilistic pedestrian perception evaluation model was further constructed to simulate the pedestrian perception process