Graph regularized implicit pose for 3D human action recognition

Abstract

We present a novel feature descriptor for 3D human action recognition using graph signal processing techniques. A linear subspace is learned using graph total variation and graph Tikhonov regularizers, transforming 3D time derivative information into a representation that is robust against noisy skeleton measurements. The graph total variation regularizer learns an action representation that encourages piece-wise constantness, which helps discriminating between different action classes. Graph Tikhonov regularization ensures the searched lowrank subspace is similar to the original feature. Experiments show that our approach learns a good representation of an action due to the explicit graph structure, and achieves a statistically significant improvement over the baseline moving pose method, resulting in a 93.5% accuracy on the challenging MSRAction3D dataset.

A Fine-to-Coarse Convolutional Neural Network for 3D Human Action Recognition

“…Different from the other studies using BP features [8,14,24], we extract a velocity together with a joint position from each joint of the raw skeleton. The velocity represents the variations over the time and has been widely employed in many previous studies, mostly in the handcrafted-feature-based approaches [10,32,34]. It is robust against the speed changes; and accordingly, is effective to discriminate actions with similar distance variations but with different speeds, such as punching and pushing.…”