Qualitative Recognition of Ongoing Human Action Sequences

This paper deals with the qualitative visual recognition of continuous human action sequences. Based on an analysis of the structure of physical actions, a framework for temporal segmentation and qualitative classi cation of physical actions is proposed. In order to achieve correctness as well as e ciency in real time action recognition, a hierarchical spatio-temporal attention control method is developed. As the integration of these ideas, a cognitive architecture for action recognition is proposed. An experimental system is built for recognizing block assembly tasks performed by human workers. This shows the possibility of generating symbolic descriptions of the shown tasks in real time.