Dynamic Scene Understanding and Recognition from Properties of Visual Data Structures

Wei Wang

Abstract: In this thesis, we present multiple contributions to scene understanding, object recognition, and pose estimation, deformable motion extraction, articulated object modeling, and human motion estimation. We propose a viewpoint oriented color-shape histogram for object recognition. A hierarchical MRF optimization method is designed for matching and a depth-based skeletonization is realized to extract the skeleton from visual observations. We validate our approach in daily-life complex scenarios.