Pose estimation and tracking using multivariate regression

This paper presents an extension of the relevance vector machine (RVM) algorithm to multivariate regression. This allows the application to the task of estimating the pose of an articulated object from a single camera. RVMs are used to learn a oneto-many mapping from image features to state space, thereby being able to handle pose ambiguity. Key words: Regression, Relevance Vector Machines, Tracking, Articulated Motion This paper considers the problem of estimating the 3D pose of an articulated object such as the human body from a single view. This problem is difficult due to the large number of degrees of freedom and the inherent ambiguities that arise when projecting a 3D structure into the 2D image [5,10]. Once the pose estimation task is solved, temporal information can be used to smooth motion and resolve potential pose ambiguities. This divides continuous pose estimation into two distinct tasks: (1) estimate a distribution of possible configurations from a single frame, (2) comb...

• Type: Journal
• Year: 2008
• Where: PRL
• Authors: Arasanathan Thayananthan, Ramanan Navaratnam, Björn Stenger, Philip H. S. Torr, Roberto Cipolla
• Keywords: Articulated Object, Pose Estimation Task, PRL 2008, Relevance Vector Machines