Integrating depth-HOG and spatio-temporal joints data for action recognition

Abstract

In this paper, we propose an approach for human activity recognition using gradient orientation of depth maps and spatio-temporal features from body-joints data. Our approach is based on an amalgamation of key local and global feature descriptors such as spatial pose, temporal variation in ‘joints’ position and spatio-temporal gradient orientation of depth maps. Additionally, we obtain a motion-induced global shape feature describing the motion dynamics during an action. Feature selection is carried out to select a relevant subset of features for action recognition. The resultant features are evaluated using SVM classifier. We validate our proposed method on our own dataset consisting of 11 classes and a total of 287 videos. We also compare the effectiveness of our method on the MSR-Action3D dataset.