Abstract
The analysis of non-rigid motion has been an important research topic for the applications in many areas, such as virtual reality, computer animation, video conferencing, image coding, medical diagnosis and intelligent human-computer interface(HCI). We propose the methods to analyze the human body motion from two approaches: 2-D approach for posture recognition and analysis, and 3-D approach for posture identification.We apply two methods in the 2-D approach; the first method extracts the body signatures from the image sequence, and then generate the feature vectors to construct the HMM and ASM models. The processes of the method consist of three main processing phases: pre-processing phase, model construction phase and motion analysis phase. The other method is to recognize the body posture by training the hidden Markov models of the image sequence, which uses the relative distance and motion information of the markers as the feature vector. The method is composed of training phase and recognition phase.The method used in the 3-D approach is based on 3-D data-model-fitting processes by tracking the markers. The model fitting processes are (1) the directional fitting which determines the moving direction, and (2) the extremity fitting which locates the angle parameters of the body joints and extremity joints. In the experiments, we will illustrate that our processing system can provide effective and accurate results.