In this paper, we propose a novel method to recognize human actions using 3D human skeleton joint points. First, we represent a skeleton pose by a feature vector with three descriptors: limb orientation, joint motion orientation and body part relation. Then, we mine discriminative local basic motions based on the sequences of feature vectors. These local basic motions contain the discriminative motions of key joints and can well represent human actions. Experiments conducted on MSR Action3D Dataset and MSR Daily Activity3D Dataset demonstrate the effectiveness of the proposed algorithm and a superior performance over the state-of-the-art techniques.

Date Published: 4 November 2014
PDF: 7 pages
Proc. SPIE 9273, Optoelectronic Imaging and Multimedia Technology III, 92731R (4 November 2014); doi: 10.1117/12.2073573

Published in SPIE Proceedings Vol. 9273:
Optoelectronic Imaging and Multimedia Technology III
Qionghai Dai; Tsutomu Shimura, Editor(s)