3D action recognition based on limb angle model

Abstract

Human action recognition technology has been applied to intelligent security surveillance, content-based image and video retrieval and natural user interface. How to make use of the new type of data, 3D skeleton joint position extracted by 3D depth camera, has been a highly active research topic. A posture representation model is proposed, which is invariant to limb length, length ratio between body parts and body orientation. This model contains polar angle and azimuthal angle of each limb in the spherical coordinate system which is established by the features of body joints. Hidden Markov Model (HMM) is exploited for recognition. Skeleton sequences of different body orientation are collected as experimental data. Experimental results demonstrate the effectiveness of our approach.