Human Pose Estimation using Part-based Region Matching

Abstract

In this thesis, a part-based region matching algorithm is proposed for human pose estimation in 2D images. A new notion of part, named a semantic part is introduced. A semantic part is represented as a combination of classic rigid parts and contains partial semantic information of body pose. Region proposals are used to form a set of candidate bounding boxes for semantic parts. These regions are matched between target and source images and their confidences are evaluated by computing the matching score. The regions with high confidence is the final semantic parts which form a body pose. Based on a data-driven approach, the final pose is estimated by getting information of joint positions from the source correspondences of the final semantic parts. Using semantic information catches more meaningful pose information and the part-based region matching has simple algorithm and performs effectively for a large number of data.