Learning universal multiview dictionary for human action recognition

Abstract

Recently, many sparse coding based approaches have been proposed for human action recognition. However, most of them focus on learning a discriminative dictionary without explicitly taking into account the common patterns shared among different action classes. In this paper, we propose a novel discriminative dictionary learning framework by formulating a universal dictionary which consists of a shared sub-dictionary and a set of class-specific sub-dictionaries. As a result, inter-class differences can be better characterized with sparse codes obtained from the class-specific dictionaries. In addition, group sparsity and locality constraints are utilized to preserve the relationship and structure among features. In order to leverage the benefits of multiple descriptors, a dictionary is learned for each view, and the corresponding sparse representations of those descriptors are fused in a low dimensional feature space together with temporal information. The experimental results on three challenging datasets demonstrate that our method is able to achieve better performance than a number of state-of-the-art ones.