Pose Tracking
As opposed to the existing approaches that can only predict the location of an object, our goal is to track both pose and deformations in real-time. We aim to accurately estimate the scale changes, in-plane and out-plane rotations, affine and perspective motions, and any parametric deformations in addition to object's translational movements. To our advantage, this method can also be trained to track only specific object types, such as faces, vehicles, etc.
Background & Objective: This technology is an essential component of many vision tasks in addition mainstream surveillance applications. It improves object identification performance by enabling adaptive selection of the most suitable representation. For example, the accuracy of the face recognition significantly increases as the correct pose of the face is provided as a prior. It is also required in robot cameras that are commonly used for factory automation to monitor production lines, in computer animation and human interfaces to obtain more realistic rendering and feedback, in intelligent vehicle systems to achieve more salient situation awareness, etc. Even more importantly, pose tracking has potential to speed up the human and face detectors.
Technical Discussion: Conventional pose tracking approaches, e.g. active appearance models, either require computationally prohibitive algorithms or make restrictive assumptions on the object appearance and quantize the space of possible motions. Instead, we learn how pose changes reflect in the object features, then use this information in a non-linear regression framework that uses Lie algebra as it is modeled on a manifold. Our pose/deformation tracking method is very fast and runs in real-time since it does not need a search operation or testing of multiple hypotheses. We iteratively estimate the most likely pose by employing simple matrix multiplications.
Contact: Fatih Porikli
Technology Area: Imaging
Modification Date: September 17, 2007
