Predicting Vehicles Appearing from Blind Spots Based on Pedestrian Behaviors

Kensho Hara, Hirokatsu Kataoka, Masaki Inaba, Kenichi Narioka, Ryusuke Hotta, Yutaka Satoh

Abstract: This study aims to enable the prediction of objects appearing from blind spots, such as outside a field-of-view and regions occluded by other objects. Conventional prediction approaches for traffic scenes primarily predict the objects based on their current observations (i.e., not in blind spots). Unlike such approaches, we propose a novel problem definition for traffic scenes that predicts the objects in blind spots based on current observations of other visible objects. To this end, we provides a prediction problem for vehicles appearing from blind spots. We propose a method that predicts vehicles appearing from a blind spot based on the behaviors of visible pedestrians who observe vehicles in the blind spot of a camera. We build a dataset that includes videos capturing real traffic scenes and experimentally confirmed that our proposed method effectively predicted vehicles appearing from blind spots in traffic scenes, which include pedestrians, at a level similar to that afforded by humans. Our dataset except RGB sequences will be made publicly available.