Action recognition by extracting pyramidal motion features from skeleton sequences

Abstract

Human action recognition has been a long-standing problem in computer vision. Computational efficiency is an important aspect in the design of an action-recognition based practical system. This paper presents a framework for efficient human action recognition. The novel pyramidal motion features are proposed to represent skeleton sequences via computing position offsets in 3D skeletal body joints. In the recognition phase, a Naive-Bayes-Nearest-Neighbors (NBNN) classifier is used to take into account the spatial independence of body joints.We conducted experiments to systematically test our framework on the public UCF dataset. Experimental results show that, compared with the state-ofthe- art approaches, the presented framework is more effective and more accurate for action recognition, and meanwhile it has a high potential to be more efficient in computation.