Posture Estimation of Human Body Based on Connection Relations of 3D Ellipsoidal Models

摘要

We propose new method of estimating human body posture from connection relations of three-dimensional (3D) ellipsoidal models. First, 3D ellipsoidal models with enlargement and reduction transformations are constructed. Next, two-dimensional (2D) appearance models are constructed from 2D projected images of the 3D model. The appearance models are related to each other by employing a network data structure. They are then matched with an image of the body made from actual thermal images. By using the connection relations between the head and body, the head and the body can be recognized. Differences between individuals can be simply treated by using different sized parts. Moreover, this method can be applied also to recognition of arms and legs.