Adaptive temporal sampling for real-time human action recognition

Abstract

A new temporal sampling method for human action recognition using depth maps is proposed. The depth map frames are projected onto three orthogonal Cartesian planes respectively and the absolute difference between two successive projected ones is stacked to form the motion energy. Three sub-action frame sequences are selected accordingly, which constructs Adaptive Temporal Sampling (ATS) descriptor, capturing the dynamic features of one action. In order to represent the static features, the Depth Static Map (DSM) is adopted to constitute ATS-DSM descriptor. An l <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> -regularized collaborative representation classifier is applied. The developed approach is proven to be computationally efficient and the experimental results on the testing dataset outperform the state-of-the-art performance.