Temporal key poses for human action recognition

In this paper, we present a simple yet effective approach to recognizing human activities from video sequences. Our approach integrates the advantages of human action recognition in static images using action key poses and motion based approaches using the variants of Motion History Images (MHI) and Motion Energy Images(MEI). We combine both methodologies to extract a new representation of temporal key poses. In an evaluation of this on well established benchmark data we achieve high recognition rates. For the task of action recognition using the MuHAVi data set, we achieve an accuracy of 98.5% in a leave-one-out cross validation procedure. For single-view action recognition using the popular Weizmann data sets, we achieve an accuracy of 100%. In more difficult evaluation setups where the number of training samples for certain individuals or views are restricted, the proposed method exceeds recently published results of other approaches. Moreover, the introduced approach is computationally efficient, robust with respect to parameter selection, and straight forward to implement as it builds on well established and understood concepts.