Robust Registration and Tracking Using Kernel Density Correlation

Maneesh K. Singh
Himanshu Arora
Narendra Ahuja

Beckman Institute and ECE Department,
University of Illinois at Urbana Champaign
Urbana, IL 61801.

{msingh,harora1,nahuja}@vision.ai.uiuc.edu

- Abstract

Challenges to accurate registration come from three factors
---presence of background clutter, occlusion of the pattern being
registered and changes in feature values across images. To address
these concerns, we propose a robust probabilistic estimation
approach predicated on representations of the object model and the
target image using a kernel density estimate. These
representations are then matched in the space of density functions
using a correlation measure, termed the Kernel Density Correlation
(KDC) measure. A popular metric which has been widely used by
previous image registration approaches is the Mutual Information
(MI) metric. We compare the proposed KDC metric with the MI metric
to highlight its better robustness to occlusions and random
background clutter---this is a consequence of the fact that the
KDC measure forms a re-descending M-estimator. Another advantage
of the proposed metric is that the registration problem can be
efficiently solved using a variational optimization algorithm. We
show that this algorithm is an iteratively reweighted least
squares (IRLS) algorithm and prove its convergence properties. The
efficacy of the proposed algorithm is demonstrated by its
application on standard stereo registration data-sets and real
tracking sequences.