Toward a unified framework of motion understanding

Computer vision researchers have made a great deal of progress in several domains of motion understanding including human activity recognition, vehicle trajectory analysis, and facial expression recognition, motivated by applications. In order to construct a unified theory for recognition of motion, feature-level understanding of motion must be performed first. Johansson's pioneering experiments suggest that human joint locations contain sufficient information for humans to identify what humans are doing. Context often serves as unified knowledge to support motion understanding. Researchers are required to explore this direction further to consider joint recognition of motion, objects and scene contexts including social roles as well as more implicit contexts such as spatio-temporal inconsistency in observation.