18-07-2012, 11:58 AM
Visual Tracking with Histograms and Articulating Blocks
Visual Tracking with Histograms.pdf (Size: 894.02 KB / Downloads: 28)
Introduction
Developing an accurate, efficient and robust visual
tracker is always challenging, and the task becomes even
more difficult when the target is expected to undergo significant
and rapid variation in shape as well as appearance.
While the audience is delighted and awed by the virtuoso
performance of the world-renown skaters (Figure 1), their
graceful movements and dazzling poses offer multiple challenges
for any visual tracker. In this example (and many
others), the appearance variation is mainly due to change
in shape while the foreground intensity distribution remains
roughly stationary. An important problem is then to efficiently
exploit this weak appearance constancy assumption
for accurate visual tracking amidst substantial shape variation.
PreviousWork
There is a rich literature on shape and appearance modeling
for visual tracking, and a comprehensive review is of
course beyond the scope of this paper. In this section, we
discuss the most relevant works within the context of object
tracking.
Active contours using parametric models [5, 17] typically
require offline training, and expressiveness of these
models (e.g., splines) is somewhat restrictive. Furthermore,
with all the offline training, it is still difficult to predict the
tracker’s behavior when hitherto unseen target is encountered.
For example, a number of exemplars have to be
learned from training data prior to tracking in [24], and the
tracker does not provide any mechanism to handle shapes
that are drastically different from the templates. Likewise,
there is also an offline learning process involved in the active
shape and appearance models [8].
Tracking Algorithm
We present the details of the proposed tracking algorithm
in this section. The output of the proposed tracker consists
of a rectangular window enclosing the target in each frame.
Furthermore, an approximated boundary contour of the target
is also estimated, and the region it encloses defines the
estimated target region. Our objective is to achieve a balance
among the three somewhat conflicting goals of efficiency,
accuracy and robustness. Specifically, we treat the
tracking problem as a sequence of detection problems, and
the main feature that we use to detect the target is the intensity
histogram. The detection process is carried out by
matching foreground intensity histogram and we employ integral
histograms for efficient computation. In the following
discussion, we will use the terms histogram and density interchangeably.
Conclusion and Future Work
In this paper, we have introduced an algorithm for accurate
tracking of objects undergoing significant shape variation
(e.g., articulated objects). Under the general assumption
that the foreground intensity distribution is approximately
stationary, we show that it is possible to rapidly and
efficiently estimate it amidst substantial shape changes using
a collection of adaptively positioned rectangular blocks.
The proposed algorithm first locates the target by scanning
the entire image using the estimated foreground intensity
distribution.