Temporal Activity Segmentation for Depth Cameras Using Joint Angle Variance Features

Abstract

In this work, we look to address the problem of temporal segmentation of human activity for videos captured with depth cameras. We propose a novel feature definition (namely, joint angle variance (JAV)) which works with skeleton joint data from depth videos. We define the angle between nearby joints as joint angles and compute the variance of each of these joint angles over a set of frames. The variance values for all the joint angles are concatenated to form the feature vector for that particular window of frames. We then use these feature vectors to train a support vector machine (SVM). Given a new test sequence, we can classify each frame belonging to a particular class by computing the class probabilities of the JAV feature computed from its corresponding window. We demonstrate the efficacy of our approach on a standard dataset consisting of ten different actions. Our work highlights that a rather simplistic and efficient formulation can also yield a good temporal segmentation performance on RGBD data.