Deep Learning Techniques for Skeleton-Based Action Recognition: A Survey

Interpreting human behavior from entirely performed actions is called human action recognition (HAR). HAR applications rapidly expand into robotics, CCTV surveillance, self-driving vehicles, gaming, and video retrieval. Among different data modalities, skeleton data offers compact representation and computational efficiency. In recent years, much work has gone into developing a robust and accurate deep-learning framework for skeleton-based HAR. The paper reviews state-of-the-art methods for skeleton-based HAR. The survey also summarizes evaluation results on a large-scale benchmark dataset. Trends in action recognition research are discussed.