Colloquium: Robust Collaborative Trackers: Its Application for Medical and Natural Object Tracking

Collaborative Tracking and Robust Dictionary Learning Using Sparse
Representation

Dr. Lin Yang, Biomedical Informatics, University of Kentucky

4 PM Wednesday, February 1st, Davis Marksbury Building Theater

Cookies at 3:45

 

In this talk, I will first present a robust, fast and accurate 3D tracking algorithm: Prediction Based Collaborative Trackers (PCT). In PCT, a novel one-step forward prediction is introduced to generate the motion prior using motion manifold learning. Marginal space learning, which searches the global optimum in the marginal space instead of the original parameter space, is used to speed up both the training and testing procedures. PCT is completely automatic and computationally efficient. It requires less than 1.5 seconds to process a 3D volume which contains millions of voxels.

Although PCT can be used to track the 2D/3D object accurately and rapidly, it is not online updated. Online learning is widely used for its adaptive ability to handle dynamic changes. However, it introduces potential drifting problems due to the accumulation of errors during the self-updating. We have developed a robust tracking algorithm using a local sparse appearance model (SPT). A novel sparse representation-based voting map and sparse constraint regularized mean-shift are fused for adaptive object tracking through online learning.

In this talk, I will also introduce a novel dictionary learning algorithm with a locally constrained sparse representation, called K-Selection. The most popular existing dictionary learning algorithms, such as K-SVD, are some generalized variants of K-means, which compute the dictionary columns to minimize the overall reconstruction errors. K-Selection (KS) constructs the dictionary by directly choosing its basis from the training data which proves to be superior in many machine learning and computer vision applications, such as object tracking, recognition and dynamic background subtraction. I will also briefly cover part of my ongoing research in developing a novel and robust K-selection (RKS) algorithm which is not only sparse but robust to outliers.
                 
                                                                                   
Brief Bio: Lin Yang is an assistant professor with the Division of Biomedical Informatics, Dept. of Biostatistics in University of Kentucky. He received his B. E. and M. S. from Xian Jiaotong University in 1999 and 2002, and his Ph. D. in Dept. of Electrical and Computer Engineering from Rutgers, the State University of New Jersey in 2009. He did part of his research in Siemens Corporate Research and IBM T. J. Watson Research Center in 2007 and 2008. He was an assistant professor in the Dept. of Radiology in University of Medicine and Dentistry of New Jersey, Cancer Institute of New Jersey, and the Dept. of Biomedical Engineering in Rutgers University from 2009-2011.

His major research interests are focused on medical image analysis, imaging informatics, computer vision and machine learning. He is also working on high performance computing and computer aided diagnostics. He is the winner of NIH young investigator paper award on 2008 IEEE international symposium on biomedical imaging.