Modeling Shape, Appearance and Self-Occlusions for Articulated Object Tracking
Yanchao Yang, Ganesh Sundaramoorthi · arXiv (Cornell University) · 2012
We present a method for object tracking so that precise object shape can be obtained. Unlike previous tracking methods that build on image segmentation (by separating foreground and background statistics), which are limited to uncluttered background and simple appearance of the object, and tracking by detection approaches that have complex appearance models but simplistic models of shape, we track by identifying stationary statistics of both appearance and shape over time, and therefore obtain accurate shape in complex changing backgrounds. Our method is an adaptive template matching scheme that applies to objects with large deformations and articulation as well as a camera that moves and changes viewpoint relative to the object. The significant aspect that must be addressed in such a template matching scheme is that the shape of the projected object into the imaging plane is quickly changing from the complex image induced transformation due to 3D articulation and deformation, viewpoint change, occlusions and disocclusions of the object due to viewpoint change, and self-occlusions and self-disocclusions due to articulation. We provide the fundamentals to understand occlusions and disocclusions of the object, a model of the shape and appearance of the projected object in time taking into account occlusion phenomena, and a computational algorithm to obtain the precise shape of the object. We illustrate the ideas on challenging video sequences and obtain state-of-the-art results. Index Terms Object tracking, occlusion detection, disocclusion detection, shape models, appearance models, non-rigid shape registration, large deformation registration, optical flow I.