View-invariant action recognition based on local linear dynamical system

Abstract

To address recognition of human actions under view changes, this paper proposes a view-invariant human action recognition approach based on local linear dynamical system and sparse coding. We utilize the bag-of-words (BoW) approach, local patches are modeled as linear dynamical systems and the model parameters are used as the descriptors of local patches. The model parameters capture the dynamics in human actions which is insensitive to view changes. The sparse coding algorithm is then applied to learn discriminative codebook and to avoid the initialization problem in the k-means algorithm. The proposed approach is tested on the IXMAS dataset. The experimental results demonstrate that this approach can recognize the view-invariant actions, obtain high recognition rates, and achieve comparable results in cross-views action recognition.