동적 베이스 망 기반의 계층적 보행 동작 분석

Abstract

Human gait is perhaps the most common activity which is a highly stereotyped, structural and inherently cyclical behavior. This paper proposes a new method for a detailed analysis of the dynamic characteristics of human gait into different poses and postures. The proposed model is called the ‘Gait Motion Decoder’(GMD) which is a sophisticated state space design using of Dynamic Bayesian Network(DBN). Inference and learning procedures are also developed for the proposed model. Unlike previous studies which focused only on human identification, this work makes an explicit and separate modeling of pedestrian pose and posture in the hierarchical structure of the model using factored state variables in the GMD. This allows the model to recognize gait direction and detect orientation change elegantly and intuitively in the probabilistic framework. Experimental results showed 96.5% in pose identification which is well over that of other models. Another important feature is that the model proposed makes a detailed analysis of gait motions and can thus infer 3D structural information about the pedestrian given a 2D video sequence.