Conference Proceedings

Characteristic-based descriptors for motion sequence recognition

Liang Wang, Xiaozhe Wang, Christopher Leckie, Kotagiri Ramamohanarao, T Washio (ed.), E Suzuki (ed.), KM Ting (ed.), A Inokuchi (ed.)

ADVANCES IN KNOWLEDGE DISCOVERY AND DATA MINING, PROCEEDINGS | SPRINGER-VERLAG BERLIN | Published : 2008

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 s..

View full abstract