Enhancing the performance of 3D auto-correlation gradient features in depth action classification

The 3D auto-correlation gradient features have demonstrated only limited success on depth action data, whereas the 2D auto-correlation gradient features have been successful in the domain. In this paper, we propose to calculate three depth motion map sequences from each depth action video by accumulating only the motion information of the action. We then obtain the three vectors of 3D auto-correlation gradient features by applying the space-time auto-correlation of gradients (STACOG) descriptor on the depth motion map sequences. The three vectors are then concatenated and passed to an unsupervised classifier to recognize the action. The experimental evaluation on four public datasets (MSR-Action3D, DHA, UTD-MHAD, and MSR-Gesture3D dataset) demonstrates the superiority of our proposed method over state-of-the-art methods.