A Part Semantic Based Action Recognition Method

Aiming at the deficiency that model high level semantic motion pattern and semantic movement evolution in the research of traditional video action recognition, this paper proposes a part semantic based video action recognition method. First, it extracts high level semantic features based on the position information of the body parts. Second, we propose a feature descriptor for action recognition based on human part semantic and its movement evolution. The method is implemented on Caffe. The experiment results display good performance.