Action Recognition using Temporal Bag-of-Words from Depth Maps

In this paper, we present a methodology for human action recognition from a sequence of depth maps obtained using Microsoft Kinect. Specifically, we use a Temporal Bag-of-Words model as representation scheme to capture the variation of features across the temporal domain. Our methodology builds the Temporal Bag-of-Words model on top of the spatiotemporal features extracted from interest points. The local spatio-temporal features provide some invariance to scale, viewpoint changes by capturing the local information. In order to make the representation insensitive to temporal sequence misalignment, we propose using the Temporal Bag-of-Words model in a hierarchical manner by recursively partitioning the depth maps sequence into sub-sequences in temporal domain. Classification is done using SVM. We test our algorithm on our own dataset consisting of eight different actions.