Stacked Filter Bank based descriptor for Human Action Recognition from Depth Sequences

Registering the motion cues from a video to produce a compact representation is a crucial stage in video based Human Action Recognition (HAR). Exploiting the most prominent features using an efficient descriptor from such a representation also plays equally significant role in the performance of recognition models. In this work, we present a concise Depth Motion Map with striding which registers the motion cues from depth sequences on a video and a novel Filter Bank based descriptor, wherein a Taylor Series Expansion (TSE) filter, a Riesz filter and a gradient filter are stacked together to extract the prominent features. We empirically evaluate the feasibility of our method on MSR Action 3D dataset under standard protocols, achieving state-of-the-art results.