A Novel Pedestrian Orientation Estimation Method for Autonomous Driving

摘要

Pedestrian orientation estimation is a vital component of autonomous driving system. The challenging factors for pedestrian orientation estimation include pose variations, fast motions, background clutters and crowded people flow. In this paper, we explore a novel pedestrian orientation estimation unified framework, which is based on a monocular camera. Firstly, pedestrian images are normalized to the same size, then extract histogram oriented gradient feature (HOG) which is one of the most effective image descriptor. In addition, we utilize structured Support Vector Machines (SVM) to generate binary classification result. Moreover, Error Correcting Output Coding (ECOC) framework combines with structured SVM to deal with multi-class classification problem. Finally, we conduct our approach on public pedestrian datasets and achieve competitive performance.