Informative shape representations for human action recognition

Abstract

Shape and kinematics are two important cues in human movement analysis. Due to real difficulties in extracting kinematics from videos accurately, this paper proposes to address the problem of human action recognition by spatiotemporal shape analysis. Without explicit feature tracking and complex probabilistic modeling of human movements, we directly convert an associated sequence of human silhouettes derived from videos into two types of computationally efficient representations, i.e., average motion energy and mean motion shape, to characterize actions. Supervised pattern classification techniques using various distance measures are used for recognition. The encouraging experimental results are obtained on a recent dataset including 10 different actions from 9 subjects.