Ego-motion classification for body-worn videos

Abstract

Portable cameras record dynamic first-person video footage and these videos contain information on the motion of the individual to whom the camera is mounted, defined as ego. We address the task of discovering ego-motion from the video itself, without other external calibration information. We investigate the use of similarity transformations between successive video frames to extract signals reflecting ego-motions and their frequencies. We use novel graph-based unsupervised and semi-supervised learning algorithms to segment the video frames into different ego-motion categories. Our results show very accurate results on both choreographed test videos and ego-motion videos provided by the Los Angeles Police Department.