Real-Time Human Action Recognition Using DMMs-Based LBP and EOH Features

This paper proposes a new feature extraction scheme for the real-time human action recognition from depth video sequences. First, three Depth Motion Maps (DMMs) are formed from the depth video. Then, on top of these DMMs, the Local Binary Patterns (LBPs) are calculated within overlapping blocks to capture the local texture information, and the Edge Oriented Histograms (EOHs) are computed within non-overlapping blocks to extract dense shape features. Finally, to increase the discriminatory power, the DMMs-based LBP and EOH features are fused in a systematic way to get the so-called DLE features. The proposed DLE features are then fed into the l ₂ -regularized Collaborative Representation Classifier (l ₂ -CRC) to learn the model of human action. Experimental results on the publicly available Microsoft Research Action3D dataset demonstrate that the proposed approach achieves the state-of-the-art recognition performance without compromising the processing speed for all the key steps, and thus shows the suitability for real-time implementation.