Skeleton-based Tai Chi action segmentation using trajectory primitives and content

Recognizing and analyzing human action is an important problem in many applications. Most studies focus on single motions, but human activity usually appears as a complex action sequence. The attendant problem is that segmenting and labeling action data manually is expensive and time-consuming, especially motions in professional fields. In this paper, we introduce Tai Chi as the background of action segmentation and propose a supervised method for Tai Chi action sequence segmentation based on trajectory primitives and geometric features. The concept of trajectory primitives is inspired by how humans recognize actions based on action fragments. They can be learned by unsupervised clustering through the self-organizing feature map. Also, we extract geometric features based on the content of motion. The work contains an experimental analysis of the proposed method on the Tai Chi dataset. In the experiment, we argued various parameters and considered the abnormal sequences. Experimental results demonstrate that our method achieves state-of-the-art performance. To allow future use by interested researchers, we release the Tai Chi dataset used in this paper.