![](bg6.png)
It can be seen from Fig.7 that the proposed
approach also shows good ability to adapting to
hand appearance and pose changes.
4.2.3 Tracking in Low Frame Rate
The fourth experiment is implemented on a down-
sampled outdoor pedestrain video sequence, which
contains typical fast object motion due to low
sample rate as in (Faith, 2005). Fig 8 indicates with
some tracking frames of our approach.
Figure 8: Tracking results of successive frames under low
sample rate (frame # 94, 95, 142, and 143).
In Fig.8, the outdoor pedestrian video is down-
sampled like (Faith, 2005), which contains typical
rapid motion. We can see that the objects in
consecutive frames 94 and 95 have no overlap, and
so do frame 142 and 143. The proposed algorithm
works well on this dramatically rapid motion video.
Notice that the distance of consecutive object is
comparatively bigger in the scenario, which is
caused by the heavy down-sampling for testing.
5 DISCUSSION
Rapid movement of object using normal camera
always accompanies with motion blur of image,
which will result in the color drift and content
degeneracy of the image due to long exposal time.
That is, object appearance will be deformed or
blurred under fast motion. Here, it should be pointed
out that sample images in Fig.5 b) show that the
proposed algorithm works well in above situations.
Based on the description of amending strategy in
part 2 and 3, the extra computation cost of
discriminative linear color feature and pre-
refinement strategy before kernel tracking is
moderately low, and hence the proposed tracking
algorithm can be used in real-time tracking tasks
unlike the works of (Faith, 2005) and (Arulampalam,
2002).
Proposed method uses a pre-refinement method
to modify the ill-conditioned initialization position
before kernel tracker, which is somewhat a fake
multiple-kernel strategy. It is similar but different
with (Faith, 2005) for we resolve using pre-
processing with low computation cost and (Faith,
2005) use real multiple-kernel and post fusion of
multiple position with high computation cost.
6 CONCLUSIONS
In this paper, we propose an efficiently
discriminative combination of color feature for
tracking problem, which introduces foreground /
background classification idea into object
representation. Also we propose a low-cost pre-
refinement method to better resolve the ill-
initialization problem of kernel tracker, which could
enhance the performance of kernel tracker under
object rapid motion. With respect to experiment
results, our proposed representation and multiple
kernel strategy works better than popularly used
Camshift and BKT under quick motion situation. It
also partly diminishes the effect of background
cluster and illumination change’s influence on
tracking result.
REFERENCES
Dorin Comaniciu, Visvanathan Ramesh, Peter Meer, 2003.
Kernel-Based Object Tracking, IEEE transaction on
Pattern Analysis and Machine Intelligence, Vol 25,
Issue 5, pp. 564-577.
Klaus Robert Muller, Sebastian Mka, Gunnar Ratsch,
2001. An Introduction to Kernel Based Learning
Algorithms, IEEE Transactions on Neural Networks,
Vol 12, No 2, pp. 181-201.
Faith Porikli, Oncel Tuzel. 2005. Multi-Kernel Object
Tracking, ICME, pp. 1234-1237.
Hanger G D, Dewan M. 2004. Multiple Kernel Tracking
with SSD, Proc. CVPR. Vol.1, pp. 790-797.
Ahmed Elgammal, Ramani Duraiswami, David Harwood.
2002. Background and Foreground Modeling using
nonparametric kernel density estimation for visual
surveillance. Proceedings of the IEEE , Vol 90, Issue
7, pp. 1151-1163.
Arulampalam, M. S. Maskell, S. Gordon. 2002. A tutorial
on particle filters for online nonlinear/non-Gaussian
Bayesian tracking. IEEE Transaction on Signal
Processing. Vol 50, No 2, pp. 174-188.
VISAPP 2010 - International Conference on Computer Vision Theory and Applications
266