Hidden Markov Model for Action Recognition Using Joint Angle Acceleration

Abstract

AbstractThis paper proposes a recognition method of human actions in video by adding new features, the joint angle acceleration to the feature space. In this method, human body is described as three-dimensional skeletons. The features consist of vectors of several important joint angles on the human body and the joint angle accelerations are also considered as a part of features. Hidden Markov Model (HMM) is used as classification scheme. The HMM models are trained by sequences extracted from the CMU graphics lab motion capture database. This method is invariant to scale, coordinate system and transition. A system is implemented to recognize 4 different types of actions (walk, run, jump and jumping jack) both on the dataset of CMU and Weizmann[7]. Each video clip contains a single action type. The experimental results show excellent performance of the proposed approach. A maximum 10.3% accuracy gain can be achieved by our method compared with the method without considering acceleration.KeywordsHuman action recognitionskeletonjoint angleHMMangle acceleration