Mean shift based nonparametric motion characterization

Motion content is a very powerful cue for organizing video data. Efficient and robust identification of the camera motion nature and the dominant object motion is important for the generation of useful motion annotations. Most of existing methods focus on the estimation of a parametric motion model from dense optical flow fields or block-based MPEG motion vector fields (MVF). However, it is hard to achieve a reliable model estimation in large amounts of video data. This is due to the violation of parametric assumption in the presence of large object motion and bad estimation of the optical flow in low-textured regions. In this paper, we employ the mean shift procedure and the histogram to propose a novel nonparametric motion representation. With this motion representation, we transform the motion analysis to the classification problem of camera motion patterns in the presence of dominant object motion and non-dominant object motion. The unique features include three main aspects: 1) instead of computationally expensive and vulnerable parametric regression, we base the motion characterization on the classification of motion patterns, 2) we employ machine learning to capture the knowledge of recognizing camera motion patterns from bad motion fields, and 3) with the mean shift filtering the proposed motion representation elegantly considers the spatial-range cues so as to remove noise and implement discontinuity preserving smoothing of motion fields. Promising results are achieved on 1096 motion vector fields extracted from compressed broadcast soccer video.