Local salient motion analysis for action recognition

Abstract

Recently, local space-time feature based action elements were shown to be an efficient video representation for action recognition and achieved state-of-art results. However, these features are easily corrupted by camera motions and background noisy. To take account global camera motion and overcome the irrelevant features, this paper presents a novel motion filter to detect salient motion parts in which a contrast of local dynamic information is used to enhance local motions. After detection of salient motions, a histogram oriented descriptor on these motion maps (HOM) is calculate to describe action elements. The experiment shows that with local salient motion detection, the local space-time motion descriptor can achieve a significant improvement compare to the existing action descriptors.