Abstract
In this thesis, we introduce a real time multiple-views human motion analysis system. This system analyzes the motion of the human object in a real scene and synthesize virtual actor in a corresponding virtual world by computer graphic OpenGL. To analyze the motion of the human object, we use the 3-D human model, which consists of 10 cylindrical primitives, representing torso, head, arms, and legs, and 10 joints connecting these cylinders. First, we introduce a simple camera calibration and 2-D human position finding method that enable us to find the 2-D human position and simulate the perspective projection by applying a perspective-scaling factor on the orthographic projection. Then, we use the 3-D human model-based method to find the BAPs based on the searching for the best matching between the 2-D projection model and the foreground object. Here, we also introduce a new overlapped tri-tree search algorithm with less running time and wider search range, which enables us to track some fast human motions, such as hands waving, in a lower frame rate. In our system, there are two cameras controlled by two viewers and the viewing direction of one camera is orthogonal to that of the other one. To integrate and arbitrate the information from multiple-views, we introduce a new integration and arbitration method, which integrates high-level information from multiple-views.