Pedestrian Orientation Estimation using On-board Monocular Camera with Semi-Supervised Learning

摘要

We propose a method for estimating the head and body orientation of a pedestrian with considerable accuracy to assist advanced pedestrian safety systems. To train a classifier, a dataset with precise annotations is required. However, it is difficult to annotate manually the exact orientation of the head and body from an image because the boundary of adjacent orientations is not clear. In our approach, we use a semi-supervised learning method that employs both labeled and unlabeled data to annotate the head and body orientation automatically. Experiments using real pedestrian images show the effectiveness of our approach.