![](bg3.png)
Table 3: The two columns on the left show true posi-
tives and negatives percentages of algorithm MBAC for the
Wallflower benchmark. On the right, results for the Stauffer
algorithm.
MBAC Stauffer BAC
TP TN TP TN TP TN
bootstrap 0.52 0.94 0.44 0.97 0.60 0.91
camouflage 0.73 0.90 0.73 0.92 0.75 0.76
foregroundAperture 0.48 0.92 0.50 0.85 0.48 0.90
lightSwitch 0.24 0.97 0.73 0.07 0.28 0.98
movedObject - 1.00 - 1.00 - 1.00
timeOfDay 0.36 0.97 0.41 0.98 0.36 0.98
wavingTree 0.75 0.75 0.86 0.90 0.78 0.67
where α ∈ [0,1] is a learning rate factor. For every
pixel p ∈ fSet, its m background models are ordered
in descending order according to their confidences.
We use a parameter gamma to control the speed at
which models are changed or updated in the back-
ground model. The closer gamma is to 1, the quicker
models will be added or updated. In the case it is
verified that s = Σc
m
p
(i),1 ≤ m ≤ K(p) < γ , a new
model will be added or the worst model will be re-
placed. For the new model m, the algorithm sets
B
m
p
(i) = F(i),c
m
p
(i) = 0.01.
3 EXPERIMENTS AND RESULTS
We used the Wallflower benchmark (Toyama et al.,
1999) in order to compare our approach to Stauffer’s
algorithm (Stauffer and Grimson, 1999) and BAC
(Rosell-Ortega et al., 2008). We compared the num-
ber of pixels classified as foreground and labelled as
foreground in the control image (true positives) and
those pixels classified as background and also classi-
fied as background in the control image (true nega-
tives).
We used K = 5 and T = 0.8 as parameters for the
Stauffer algorithm. Parameters for MBAC were set
after a previous study of their impact in the execution.
Tables 2 and 1 show the results obtained by varying
the values of κ in equation 1 and its impact depending
on the value of γ used. Results seem to be better with
a low γ. Tables 4 and 5 show the results of using each
definition of pFore in equation 4. As a conclusion
of these experiments, the segmentation seems to im-
prove slightly when a strict value for τ is chosen. The
remaining parameters of MBAC are κ = 20, µ = 0.85,
γ = 0.4 and τ = 0.8.
Table 3 illustrates the results with the Wallflower
benchmark. Qualitative results are shown in figure
1. In sequence lightSwitch, MBAC manages prop-
erly the sudden light change restarting the model,
while Stauffer’s algorithm fails to deal with the sit-
uation. The most significant improvement of MBAC
over BAC, is achieved in sequences wavingTrees and
camouflage. In all cases MBAC achieved over 80%
of success in the classification of background pixels.
4 CONCLUSIONS AND FUTURE
WORKS
We introduced an approach in which similarity and
motion features are used to classify pixels as fore-
ground or background. Considering motion at the
same level as background subtraction with several
models produces accurate background models but at
the expense or reducing the amount of regions of in-
terest detected if thresholds are not accurate enough.
This issue remains as an open line for further re-
search.
REFERENCES
Elgammal, A., Harwood, D., and Davis, L. (2000).
Non-parametric model for background subtraction.
ECCV00, pages 751 767.
Jabri, S., Duric, Z.,Wechsler, H., and Rosenfeld, A. (2000).
Detection and location of people in video images un-
sing adaptive fusion of color and edge information.
IEEE Proc. ICPR00, pages 627 630.
M. Heikkila, M. and Pietikainen, M. (2006). A texture-
based method for modeling the background and
detecting moving 0bjects. IEEE Trans. PAMI,
28(4):657662.
Mason, M. and Duric, Z. (2001). Using histograms to detect
and track objects in color video. Proc. Applied Imag-
inery pattern Recognition Workshop, pages 154159.
Rosell-Ortega, J., Andreu-Garcia, G., Rodas-Jorda, A.,
and Atienza-Vanacloig, V. (2008). Background mod-
elling in demanding situations with confidence mea-
sure. IEEE Proc. ICPR08.
Stauffer, C. and Grimson, W. E. L. (1999). Adaptive back-
ground mixture models for real-time tracking. Proc.
IEEE CVPR99, pages 246 252.
Toyama, K., Krumm, J., Brumitt, B., and Meyers, B.
(1999). Wallflower: Principles and practice of back-
ground maintenance. IEEE ICPR99, Kerkyra, Greece,
pages 255261.
Wixson, L. (2000). Detecting salient motion by accumulat-
ing directionally-consistent flow. IEEE Trans. PAMI,
8(22):774 780.
Wren, C. R., A. Azarbayenjani, T. D., and Pentland, A. P.
(1997). Pfinder: rel-time tracking of the human body.
IEEE Trans. PAMI, 10(7):780 785.
Zang, Q. and Klette, R. (2004). Robust background subtrac-
tion and maintenance. IEEE ICPR04, pages 90 93.
BACKGROUND MODELING WITH MOTION CRITERION AND MULTI-MODAL SUPPORT
421