Joint Angle-Based Skeleton Action Recognition

Abstract

Human action recognition is a difficult task and topic of discussion in the computer vision community. There are consistent breakthroughs by large research groups using sophisticated state-of-the-art computation hardware. These advancements allow for very accurate predictions on complex action recognition datasets that can be generalized to tasks outside of these pre-defined datasets. However, a key weakness is that in order to run these models, the computation requirements are significant. We explore these models, as well as some models that utilize intermediate representations. These representations focus on the movement of the person throughout the frame, specifically the 'pose' data and translate this data into an intermediate representation that can be processed by compact models. We propose a novel intermediate representation utilizing only the angles of the joints of a person from one frame to another. This allows for the representation constructed to be completely invariant to the global position of the person in the frame, providing insights into how these small and very efficient representations and models can be used with good effectiveness.