Joint-oriented Features for Skeleton-based Action Recognition

Abstract

In this paper, we propose joint-oriented features for skeleton-based action recognition, which aims to decrease the influence of ambiguous joints in the skeleton sequences. When doing skeleton-based action recognition, we noticed that if the skeleton data contains noisy joints, the result would be influenced by the noise. Since the selections of disambiguous joints might be impossible, it would be a hard task for humans to distinguish whether the joint in the frame is correct. To deal with this situation, we propose joint-oriented features to train joint-oriented models. If some joints are noise in the frame, the corresponding joint-oriented models would not perform well on the case. As we could not preFigure the noise in the data, we apply ensemble modeling with our joint-oriented models to let the disambiguous joints correct the ambiguous ones. To demonstrate the effectiveness of our proposed method, we conducted the experiments on three benchmark skeleton-based action datasets, including the large-scale challenging NTU-RGBD, and our approach achieves competitive performance over the state-of-the-art.