Articulated video object extraction by the combination of spatial and temporal segmentation

Generally moving object extraction schemes rely on either optical flow or frame difference. Though optical flow based methods can deal with moving camera situation with ease, the inconsistency at object boundaries causes extraction with inaccurate object boundaries. While frame difference approach, such as three frame difference, yields accurate object boundaries, it usually cannot deal with moving camera. In this paper we present a novel technique for automatic object extraction captured by a moving camera. Our approach blends motion information from both optical flow and frame difference, and combines spatial region segementation for smooth object boundary detection. As a reprentative example of articulated object, experimental results of human extraction are presented.