Human Action Recognition in Video by Fusion of Structural and Spatio-temporal Features

The problem of human action recognition has received increasing attention in recent years for its importance in many applications. Local representations and in particular STIP descriptors have gained increasing popularity for action recognition. Yet, the main limitation of those approaches is that they do not capture the spatial relationships in the subject performing the action. This paper proposes a novel method based on the fusion of global spatial relationships provided by graph embedding and the local spatio-temporal information of STIP descriptors. Experiments on an action recognition dataset reported in the paper show that recognition accuracy can be significantly improved by combining the structural information with the spatio-temporal features.