![](bg6.png)
Table 2: False negative (FN) and false positive (FP) hits of the reference template matching (TM) algorithm in 4 different
scales, and the proposed method on the five testsequences.
Seq1 Seq2 Seq3 Seq4 Seq5
Method Scale FN FP FN FP FN FP FN FP FN FP
TM 0.1 22 12 31 13 14 5 8 50 0 1
TM 0.25 20 6 28 6 0 0 25 15 0 1
TM 0.5 34 2 28 4 0 0 6 11 0 1
TM 1.0 0 0 0 0 0 0 0 0 0 0
Prop. Alg. 1.0 0 0 0 0 1 0 1 0 0 0
Table 3: Computational time (in sec) for one fixation point
for the reference template matching (TM) algorithm in 4
different scales, and the proposed method (PM) on the five
testsequences, using C++ implementation and a Pentium
laptop (Intel(R) Core(TM)2 CPU, 2GHz).
Processing time
Me. SF Seq1 Seq2 Seq3 Seq4 Seq5
TM 0.1 0.32 0.42 0.05 0.60 0.28
TM 0.25 2.55 4.42 0.05 4.56 2.29
TM 0.5 9.17 21.8 0.11 22.9 9.33
TM 1.0 > 45 > 45 0.27 >45 > 45
PM 1.0 0.14 0.15 0.08 0.23 0.19
with reasonable speed having only a few templates
(Seq. 3), but as the template number grows the com-
putational time dramatically increases (Seq. 4). Al-
though the reference algorithm can run in real time if
the input data is downscaled to one tenth of the origi-
nal resolution, in that case the number of false positive
and false negative hits is extremely high. With less
drastic down scale the method is more accurate but
the computational time highly increases so the pro-
cess is not real time any more. On the other hand, the
proposed algorithm runs in real time (which means
that the processing of a fixation point is finished be-
for the next data arrives) with high accuracy on each
sequence even with the highest template number.
7 CONCLUSIONS
In this paper we presented an on-line method for sta-
tistical evaluation of dynamic web advertisements by
tracking the users’ eye movements. For registering
the eye movement a special camera was used. The
events when the eye “visits” an advertisement were
automatically detected with a quick template match-
ing algorithm. The proposed method was tested on
real data and found to be very accurate (less then 1%
error) and fast (real time) even in case of a high tem-
plate number.
ACKNOWLEDGEMENTS
The authors are grateful to Zolt´an Vidny´anszky, Vik-
tor G´al, and M´arton Fernezelyi for providing the test
data. This work was supported by the C3 (Center for
Culture and Communication) Foundation, Hungary.
REFERENCES
Comaniciu, D., Ramesh, V., and Meer, P. (2003). Kernel-
based object tracking. IEEE Trans. Pattern Anal.
Mach. Intell., 25(5):564–575.
Intel (2005). Documentation of the Intel Open
Source Computer Vision Library (OpenCV),
http://www.intel.com/technology/computing/opencv/.
Kikuchi, H., Kato, H., and Akahori, K. (2002). Analysis
of children’s web browsing process: Ict education in
elementary schools. In Proc. ICCE, page 253, Wash-
ington, DC, USA. IEEE Computer Society.
Lucas, B. and Kanade, T. (1981). An iterative image regis-
tration technique with an application to stereo vision.
In Proc. of International Joint Conference on Artificial
Intelligence, pages 674–679, Vancouver, BC, Canada.
Reddy, B. and Chatterji, B. (1996). An FFT-based tech-
nique for translation, rotation and scale-invariant im-
age registration. IEEE Trans. on Image Processing,
5(8):1266–1271.
Schweitzer, H., Bell, J. W., and Wu, F. (2002). Very fast
template matching. In Proc. ECCV, pages 358–372,
London, UK. Springer-Verlag.
SensoMotoric (2005). Documentation of the iView X Sys-
tem, http://www.smi.de/iv/.
Velayathan, G. and Yamada, S. (2006). Behavior-based web
page evaluation. In Proc. WWW, pages 841–842, New
York, NY, USA. ACM Press.
Viola, P. and Jones, M. (2001). Rapid object detection using
a boosted cascade of simple features. In Proc. IEEE
CVPR, volume 1, pages 511–518, Hawaii, USA.
Yoshimura, S. and Kanade, T. (1994). Fast template match-
ing based on the normalized correlation by using mul-
tiresolution eigenimages. In Proc. IROS, volume 3,
pages 2086 – 2093.
FAST TEMPLATE MATCHING FOR MEASURING VISIT FREQUENCIES OF DYNAMICWEB ADVERTISEMENTS
233