Characteristic-based descriptors for motion sequence recognition

Abstract

This paper proposes an approach based on characteristic descriptors for recognition of articulated and deformable human motions from image sequences. After extracting human movement silhouettes from motion videos, we apply Tensor Subspace Analysis to embed normalized dynamic silhouette sequences into low-dimensional forms of multivariate time series. Structure-based statistical features are then extracted from such multivariate time series to summarize motion patterns (as descriptors) in a compact manner. A multi-class Support Vector Machine classifier is used to learn and predict the motion sequence categories. The proposed method is evaluated on two real-world state-of-the-art video data sets, and the results have shown the power of our method for recognizing human motion sequences with intra- and inter-person variations on both temporal and spatial scales. © 2008 Springer-Verlag Berlin Heidelberg.