See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/270895993

An eXtended Center-Symmetric Local Binary

Pattern for Background Modeling and
Subtraction in Videos

Conference Paper · March 2015

DOI: 10.5220/0005266303950402

CITATIONS READS

3 372

3 authors:

Caroline Silva Thierry Bouwmans

Université de La Rochelle Université de La Rochelle
7 PUBLICATIONS 5 CITATIONS 57 PUBLICATIONS 775 CITATIONS

SEE PROFILE SEE PROFILE

C. Frelicot
Université de La Rochelle
66 PUBLICATIONS 587 CITATIONS

SEE PROFILE

Available from: Thierry Bouwmans

Retrieved on: 16 April 2016
 An eXtended Center-Symmetric Local Binary Pattern
for Background Modeling and Subtraction in Videos

Caroline Silva, Thierry Bouwmans, Carl Frélicot

Lab. Mathématiques, Images et Applications - Université de La Rochelle, 17000 La Rochelle, France
{caroline.pacheco do espirito silva, thierry.bouwmans,carl.frelicot}@univ-lr.fr

Keywords: Local binary patterns, Background modeling, Background subtraction.

Abstract: In this paper, we propose an eXtended Center-Symmetric Local Binary Pattern (XCS-LBP) descriptor for
background modeling and subtraction in videos. By combining the strengths of the original LBP and the
similar CS ones, it appears to be robust to illumination changes and noise, and produces short histograms,
too. The experiments conducted on both synthetic and real videos (from the Background Models Challenge)
of outdoor urban scenes under various conditions show that the proposed XCS-LBP outperforms its direct
competitors for the background subtraction task.

1 INTRODUCTION 57 60 58 1 1 1 code:
positive
52 56 51 0 0 11100110
The background subtraction (BS) is one of the main differences
24 72 68 0 1 1
steps in many computer vision applications, such as
object tracking, behavior understanding and activity
recognition (Pietikäinen et al., 2011). The BS process 57 60 58 0 1 0 code:
positive
consists basically of: a) background model initializa- 52 59 51 0 0 01000110
differences
tion, b) background model maintenance and c) fore- 24 72 68 0 1 1
ground detection. Many BS methods have been de-
Figure 1: Examples of LBP encoding
veloped during the last few years (Bouwmans, 2014;
Sobral and Vacavant, 2014; Shah et al., 2013), and the
main resources can be found at the Background Sub- The LBP feature of an image consists in building a
traction Web Site1 . histogram based on the codes of all the pixels within
The BS needs to face several challenging situa- the image. As it only adopts first-order gradient in-
tions such as illumination changes, dynamic back- formation between the center pixel and its neighbors,
grounds, bad weather, camera jitter, noise and shad- see (Xue et al., 2011), the produced histogram can be
ows. Several feature extraction methods have been rather long. A large number of local texture descrip-
developed to deal with these situations. Color features tors based on LBP (Richards and Jia, 2014) have been
are the most widely used, but they present several proposed so far for background modeling. In order to
limitations when illumination changes, shadows and be more robust to noise or illumination changes, most
camouflage occurences are present. A variety of local of them are unfortunately either very time-consuming
texture descriptors recently have attracted great atten- or produce a long feature histogram.
tion for background modeling, especially the Local In this paper, we propose to extend the variant by
Binary Pattern (LBP) because it is simple and fast to Heikkilä et al. (2009) by introducing a new neighbor-
compute. Figure 1 (top) shows how a (center) pixel is ing pixels comparison strategy that allows the descrip-
encoded by a series of bits, accordingly to the relative tor to be less sensitive to noisy pixels and to produce
gray levels of its circular neighboring pixels. It shows a short histogram, while preserving robustness to il-
great invariance to monotonic illumination changes, lumination changes and slightly gaining in time con-
do not require many parameters to be set, and have sumption when compared to its direct competitors.
a high discriminative power. However, the original The rest of this paper is organized as follows. Sec-
LBP descriptor in (Ojala et al., 2002) is not efficient tion 2 provides quite an exhaustive overview of LBP-
for background modeling because of its sensitivity to based descriptors. The new descriptor that we pro-
noise, see Figure 1 (bottom) where a little change of pose is described in Section 3. Comparative results
the central value greatly affects the resulting code. obtained on both synthetic and real videos are given
in Section 4. Finally, concluding remarks and some
1 https://sites.google.com/site/backgroundsubtraction/Home perspectives are drawn in Section 5.
 Table 1: Comparison of LBP and variants.

Robust to Uses Uses Histogram

Robust to
Descriptor illumination color temporal size with 8
noise
changes information information neighbors
Original LBP (Ojala et al., 2002) • 256
Modified LBP (Heikkilä and Pietikäinen, 2006) • • 256
CS-LBP (Heikkilä et al., 2009) • 16
STLBP (Shimada and Taniguchi, 2009) • • 256
εLBP Wang and Pan (2010) • 256
Adaptive εLBP (Wang et al., 2010) • 256
SCS-LBP (Xue et al., 2010) • • 16
SILTP (Liao et al., 2010) • 256
CS-LDP (Xue et al., 2011) • 16
SCBP (Xue et al., 2011) • 64
OCLBP (Lee et al., 2011) • 1536
Uniform LBP (Yuan et al., 2012) • 59
SALBP (Noh and Jeon, 2012) • 128
SLBP-AM (Yin et al., 2013) • • 256
LBSP (Bilodeau et al., 2013) • • 256
iLBP (Vishnyakov et al., 2014) • 256
CS-SILTP (Wu et al., 2014) • • 16
XCS-LBP (in this paper) • • 16

2 RELATED WORK lengths than the CS-LBP, but with a slightly lower
precision than the original LBP. Zhou et al. (2011)
One of the first descriptors based on the LBP for develop a Spatial-Color Binary Pattern (SCBP) that
background modeling can be found in (Heikkilä and fuse color and texture information. The SCBP out-
Pietikäinen, 2006). It improves the original LBP in performs LBP and SCS-LBP for background subtrac-
image areas where the gray values of the neighboring tion tasks. In (Lee et al., 2011), the authors propose
pixels are very close to the center pixel one, e.g. sky, an Opponent Color Local Binary Pattern (OCLBP)
grass, etc. that uses color and texture information. The OCLBP
Shimada and Taniguchi (2009) propose a Spatial- extracts several pixel’s pieces of information, but the
Temporal Local Binary Pattern (STLBP) which is ro- length of the produced histogram makes it useless for
bust to short-term illumination changes by using some some applications. An Uniform LBP Patterns with a
temporal information. Two variants of LBP, called new thresholding method can be found in (Yuan et al.,
εLBP and Adaptive εLBP, are developed in (Wang 2012). It appears to be tolerant to the interference
and Pan, 2010; Wang et al., 2010). They are fast to of the sampling noise. Yin et al. (2013) propose a
compute and less sensitive to the illumination varia- Stereo LBP on Appearance and Motion (SLBP-AM)
tion or some color similarity between foreground and which uses information from a set of frames of three
background. Heikkilä et al. (2009) propose the Center different planes. This texture descriptor is not only
Symmetric Local Binary Pattern (CS-LBP) descrip- robust to slight noise, but it also adapts quickly to
tor which generates more compact binary patterns by the large-scale and sudden light changes. A Local
working only with the center-symmetric pairs of pix- Binary Similarity Patterns (LBSP) descriptor is de-
els. In (Xue et al., 2010), a Spatial Extended Center- veloped in (Bilodeau et al., 2013). Based on abso-
Symmetric (SCS-LBP) is presented. It improves the lute absolute differences, it applies on small areas and
CS-LBP by better capturing the gradient information is calculated inside one image and between two im-
and hence, making it more discriminative. The au- ages. This allows LBSP to capture both texture and
thors explain that their SCS-LBP produces a relatively intensity changes. Noh and Jeon (2012) propose to
short feature histogram with low computationally improve the SILTP (Liao et al., 2010) thanks to a
complexity. Liao et al. (2010) propose the Scale In- codebook method. The derived descriptor gain in ro-
variant Local Ternary Pattern (SILTP) which is more bustness when segmenting moving objects from dy-
efficient for noisy images. The Center-Symmetric namic and complex backgrounds. Wu et al. (2014)
Local Derivative Pattern descriptor (CS-LDP) is de- extend SILTP by introducing a novel Center Symmet-
scribed in (Xue et al., 2011). It extracts more detailed ric Scale Invariant Local Ternary Patterns (CS-SILTP)
local information while preserving the same feature descriptor which explores spatial and temporal rela-
tionships within the neighborhood. The LBP descrip- 
tors present a significant drawback as it ignores the in- 1 if x > T
s(x) = (4)
tensity information. Because of this, there could be a 0 otherwise
wrong pixel comparison result when intensity values where T is a user-defined threshold. Since the gray
of pixels differ drastically, but their LBP values are levels are normalized in [0,1], the authors recommend
identical. To overcome this drawback, Vishnyakov to use of a small value. We will set it to 0.01 in the ex-
et al. (2014) propose an intensity LBP (iLBP) to build periments presented in Section 4. By construction, the
a fast background model is proposed in (Vishnyakov length of the histogram resulting from the CS-LBP
et al., 2014). It is defined as a collection of LBP de- P/2−1
scriptor values and intensity values of the image. The descriptor falls down to 1 + ∑i=0 2i = 2P/2 . For BS,
main characteristics of all the above reviewed LBP the CS-LBP encodes the two images to be compared
variants, including those we will compare our new de- as texture-based images with a lower quantization that
scriptor to, are summarized in Table 1. slightly favors robustness.
We propose to extend the CS-LBP operator
by comparing the gray values of pairs of center-
symmetric pixels so that the produced histogram are
3 THE XCS-LBP DESCRIPTOR short as well, but considering the central pixel also.
This combination makes the resulting descriptor less
The original LBP descriptor introduced by Ojala et al. sensitive to noise for the BS application. The new
(2002) has proved to be a powerful local image de- LBP variant, called XCS-LBP (eXtended CS-LBP),
scriptor. It labels the pixels of an image block by expresses as:
thresholding the neighbourhood of each pixel with the
center value and considering the result as a binary (P/2)−1
number. The LBP encodes local primitives such as XCS − LBPP,R (c) = ∑ s (g1 (i, c) + g2 (i, c)) 2i
curved edges, spots, flat areas, etc. In the context of i=0
BS, both the current image and the image represent- (5)
ing the background model are encoded such that they where the threshold function s, which is used to deter-
become a texture-based representation of the scene. mine the types of local pattern transition, is defined as
Let a pixel at a certain location, considered as the a characteristic function:
center pixel c = (xc , yc ) of a local neighborhood com-
1 if (x1 + x2 ) ≥ 0

posed of P equally spaced pixels on a circle of radius s(x1 + x2 ) = (6)
R. The LBP operator applied to c can be expressed as: 0 otherwise.
P−1
and where g1 (i, c) and g2 (i, c) are defined by:
LBPP,R (c) = ∑ s (gi − gc ) 2i (1)
i=0

g1 (i, c) = (gi − gi+(P/2) ) + gc
where gc is the gray value of the center pixel c and gi (7)
g2 (i, c) = (gi − gc ) (gi+(P/2) − gc )
is the gray value of each neighboring pixel, and s is a
thresholding function defined as: with the same notation conventions than in equations

1 if x ≥ 0 (1) and (3). It is worth noting that the threshold
s(x) = (2) function does not need a user-defined threshold value,
0 otherwise.
contrary to CS-LBP.
From (1), it is easy to show that the number of binary
terms to be summed is ∑P−1 i P
i=0 2 = 2 − 1, so that the
length of the resulting histogram (including the bin-
0 location) is 2P . The underlying idea of CS-LBP in
(Heikkilä et al., 2009) is to compare the gray levels
of pairs of pixels in centered symmetric directions in-
stead of comparing the central pixel to its neighbors.
Assuming an even number P of neighboring pixels,
the CS-LBP operator is given by:
(P/2)−1
CS − LBPP,R (c) = ∑ s(gi − gi+(P/2) ) 2i (3)
i=0 Figure 2: The LBP descriptor.
where gi and gi+(P/2) are the gray values of center-
symmetric pairs of pixels, and s is the thresholding The computation of the original LBP for a neigh-
function defined as: borhood of size P = 8 is illustrated in Figure 2 and
 Figure 3: The XCS-LBP descriptor.

Rotary (frame #1140) – scenes 122, 222, 322, 422 and 522 Street (frame #301) – scenes 112, 212, 312, 412 and 512

(a)

(b)

(c)

(d)

(e)

(f)

Figure 4: Background subtraction results using the ABL method on synthetic scenes – (a) original frame, (b) ground truth, (c)
LBP, (d) CS-LBP, (e) CS-LDP and (f) proposed XCS-LBP.

the computation of the proposed XCS-LBP is shown from the BMC (Background Models Challenge)
in Figure 3 in order to make the comparison more un- which comprises synthetic and real videos of outdoor
derstandable for the reader. Note the respective code situations (urban scenes) acquired with a static cam-
lengths of 8 and 4 that lead to respective image com- era, under different weather variations such as: wind,
pressions. sun or rain (Vacavant et al., 2012). We compare XCS-
The proposed XCS-LBP produces a shorter his- LBP with three other texture descriptors among the
togram than LBP, as short as CS-LBP, but it extracts reviewed ones, namely:
more image details than CS-LBP because (i) it takes
• original LBP (Ojala et al., 2002),
into account the gray value of the central pixel, and
(ii) it relies on a new strategy for neighboring pix- • CS-LBP (Heikkilä et al., 2009) and
els comparison. Since it is also more robust to noisy • CS-LDP(Xue et al., 2011).
images than both LBP and CS-LBP, the proposed
descriptor appears to more efficient for background We choose these two last descriptors on fair com-
modeling and subtraction. parison purpose. Indeed, among those who rely on the
same construction principle, i.e. Center Symmetric
(CS), they are the only ones that use neither color nor
4 EXPERIMENTAL RESULTS temporal information, see Table 1. For all descriptors,
the neighborhood size is empirically selected so that
Several experiments were conducted to illustrate P = 8 and R = 1, and we evaluate the performance
both the qualitative and quantitative performances of with two popular background subtraction methods,
the proposed descriptor XCS-LBP. We use datasets see (Bouwmans, 2014):
 Rotary (frame #1140) – scenes 122, 222, 322, 422 and 522 Street (frame #301) – scenes 112, 212, 312, 412 and 512

(a)

(b)

(c)

(d)

(e)

(f)

Figure 5: Background subtraction results using the GMM method on synthetic scenes – (a) original frame, (b) ground truth,
(c) LBP, (d) CS-LBP, (e) CS-LDP and (f) proposed XCS-LBP.

• Adaptive Background Learning (ABL) and achieved the best Recall using ABL, and scene Street-
• Gaussian Mixture Models (GMM). 112 for which LBP gives the best Recall using GMM.
Note that both CS-LBP and CS-LDP gives lower
First, we present results of background subtraction
scores (Precision and F-score) than LBP for some
on individual frames of five different scenes from two
scenes, while our XCS-LBP descriptor takes always
video sequences: Rotary (frame #1140) and Street
the advantage on the others, as shown by the average
(frame #301). Figures 4 and 5 show the foreground
scores reported at the bottom of each Table.
detection results using the ABL and the GMM meth-
ods, respectively. Our descriptor clearly appears to be Finally, we evaluate the proposed descriptor on
less sensitive to the background subtraction method, nine long duration (about one hour) real outdoor
whereas the three others are very useless in detecting video scenes from BMC. Each video sequence shows
the moving objects when using the ABL method, un- different challenging situations of real world: mov-
less a strong post-processing procedure. ing trees, casted shadows, the presence of a contin-
Next, we give quantitative results on the same uous car flow near to the surveillance zone, general
data. We use three classical measures based on the climatic conditions (sunny, rainy and snowy condi-
numbers of true positive T P pixels (correctly detected tions), fast light changes and the presence of big ob-
foreground pixels), false positive FP pixels (back- jects. The scores obtained using the ABL and the
ground pixels detected as foreground ones), false neg- GMM methods are given in Table 4 and 5, respec-
ative pixels FN (foreground pixels detected as back- tively. Once again, our descriptor achieved the best
ground ones), and true negative pixels (correctly de- scores on almost always scenes, even when using the
tected background pixels): simple ABL method whereas it dramatically compro-
mises the other descriptors. The average scores re-
TP
• Recall = , ported at the bottom of each Table show that our XCS-
T P + FN LBP outperforms the original LBP and both the simi-
TP lar construction-based CS-LBP and CS-LDP descrip-
• Precision = , and
T P + FP tors, the latter one being less performant than the LBP
Recall × Precision using GMM method. We use Matlab R2013a on a
• F − score = 2 × .
Recall + Precision MacBook Pro (OS X 10.9.4) equipped with 2.2 GHz
Tables 2 and 3 shows the scores of the different de- Intel Core i7 and 8 GB - 1333 MHz DDR3.
scriptors obtained on the Rotary and Street entire We collected the elapsed CPU times needed to
scenes when using the ABL and the GMM method, segment the foregrounds using the ABL and the
respectively. Best scores are in bold. The pro- GMM methods, averaged over the nine real videos
posed XCS-LBP gives the highest value for each of BMC. Since the reference is the (fastest) LBP de-
score on almost all scenes, except for scene Street- scriptor, the times are divided by LBP ones. Table 6
[112, 312,412], for which CS-LBP and CS-LDP has reports the resulting ratios for the compared CS de-
 Table 3: Performance of the different descriptors on syn-
scriptors. Our XCS-LBP shows slightly better time thetic videos of the BMC using the GMM method.
performance than both CS-LBP and CS-LDP.
Scenes Descriptor Recall Precision F-score
Table 2: Performance of the different descriptors on syn-
thetic videos of the BMC using the ABL method. LBP 0.817 0.701 0.755
Rotary CS-LBP 0.830 0.705 0.763
Scenes Descriptor Recall Precision F-score 122 CS-LDP 0.819 0.677 0.741
XCS-LBP 0.831 0.800 0.815
LBP 0.682 0.564 0.618
LBP 0.636 0.653 0.644
Rotary CS-LBP 0.832 0.520 0.640
Rotary CS-LBP 0.741 0.687 0.713
122 CS-LDP 0.809 0.523 0.635
222 CS-LDP 0.651 0.616 0.633
XCS-LBP 0.850 0.784 0.816
XCS-LBP 0.825 0.794 0.809
LBP 0.611 0.505 0.553
LBP 0.661 0.646 0.653
Rotary CS-LBP 0.673 0.504 0.577
Rotary CS-LBP 0.741 0.656 0.696
222 CS-LDP 0.753 0.510 0.608
322 CS-LDP 0.674 0.613 0.642
XCS-LBP 0.852 0.782 0.815
XCS-LBP 0.821 0.767 0.793
LBP 0.603 0.505 0.550
LBP 0.611 0.585 0.598
Rotary CS-LBP 0.647 0.504 0.566
Rotary CS-LBP 0.673 0.575 0.620
322 CS-LDP 0.733 0.507 0.600
422 CS-LDP 0.611 0.548 0.578
XCS-LBP 0.829 0.793 0.810
XCS-LBP 0.748 0.702 0.724
LBP 0.573 0.502 0.535
LBP 0.636 0.627 0.631
Rotary CS-LBP 0.609 0.503 0.550
Rotary CS-LBP 0.743 0.672 0.706
422 CS-LDP 0.733 0.508 0.600
522 CS-LDP 0.605 0.650 0.627
XCS-LBP 0.751 0.780 0.765
XCS-LBP 0.825 0.760 0.791
LBP 0.610 0.505 0.553
Rotary CS-LBP 0.663 0.504 0.573 LBP 0.940 0.674 0.785
522 CS-LDP 0.745 0.509 0.605 Street CS-LBP 0.924 0.675 0.780
XCS-LBP 0.852 0.732 0.787 112 CS-LDP 0.938 0.656 0.772
XCS-LBP 0.844 0.755 0.808
LBP 0.702 0.530 0.604
LBP 0.676 0.642 0.659
Street CS-LBP 0.839 0.512 0.636
Street CS-LBP 0.752 0.658 0.702
112 CS-LDP 0.826 0.525 0.642
212 CS-LDP 0.694 0.577 0.630
XCS-LBP 0.803 0.793 0.798
XCS-LBP 0.833 0.760 0.795
LBP 0.636 0.504 0.562
LBP 0.684 0.633 0.657
Street CS-LBP 0.716 0.503 0.591
Street CS-LBP 0.742 0.627 0.680
212 CS-LDP 0.798 0.513 0.624
312 CS-LDP 0.729 0.581 0.647
XCS-LBP 0.808 0.790 0.799
XCS-LBP 0.821 0.713 0.763
LBP 0.627 0.504 0.558
LBP 0.619 0.566 0.591
Street CS-LBP 0.699 0.503 0.585
Street CS-LBP 0.705 0.567 0.628
312 CS-LDP 0.801 0.511 0.624
412 CS-LDP 0.659 0.539 0.593
XCS-LBP 0.800 0.796 0.798
XCS-LBP 0.751 0.619 0.679
LBP 0.580 0.501 0.558
LBP 0.662 0.566 0.610
Street CS-LBP 0.599 0.501 0.546
Street CS-LBP 0.727 0.568 0.638
412 CS-LDP 0.754 0.507 0.607
512 CS-LDP 0.689 0.551 0.612
XCS-LBP 0.748 0.781 0.764
XCS-LBP 0.828 0.629 0.715
LBP 0.628 0.503 0.559
Street CS-LBP 0.677 0.503 0.577 LBP 0.694 0.629 0.658
512 CS-LDP 0.771 0.508 0.612 Average CS-LBP 0.758 0.639 0.693
XCS-LBP 0.800 0.575 0.669 scores CS-LDP 0.707 0.601 0.648
XCS-LBP 0.813 0.730 0.769
LBP 0.625 0.512 0.565
Average CS-LBP 0.695 0.506 0.584
scores CS-LDP 0.772 0.512 0.616
XCS-LBP 0.809 0.761 0.782
new variant XCS-LBP (eXtended CS-LBP) produces
a shorter histogram than LBP, by its CS-construction.
It is also tolerant to illumination changes as LBP and
CS-LBP are whereas CS-LDP is not, and robust to
5 CONCLUSION noise as CS-LDP is whereas LBP and CS-LBP are
not. We compared the XCS-LBP to the original LBP
In this paper, a new texture descriptor for back- and to its two direct competitors on both synthetic
ground modeling is proposed. It combines the and real videos of the Background Modeling Chal-
strengths of the original Local Binary Pattern (LBP) lenge (BMC) using two popular background subtrac-
and the Center-Symmetric (CS) LBPs. Thus, the tion methods. The experimental results show that the
 Table 5: Performance of the different descriptors on real-
proposed descriptor qualitatively and quantitatively world videos of the BMC using the GMM method
outperforms the mentioned descriptors, making it a
serious candidate for the background substation task Videos Descriptor Recall Precision F-score
in computer vision applications. Boring LBP 0.684 0.587 0.632
Future works will explore how to extend the proposed parking, CS-LBP 0.716 0.593 0.649
descriptor to include temporal relationships between active CS-LDP 0.674 0.579 0/623
neighboring pixels for dynamic texture classification bkbg XCS-LBP 0.680 0.607 0.641
LBP 0.695 0.778 0.734
or human action recognition.
CS-LBP 0.698 0.773 0.733
Big trucks
Table 4: Performance of the different descriptors on real- CS-LDP 0.649 0.758 0.699
world videos of the BMC using the ABL method XCS-LBP 0.630 0.792 0.702
LBP 0.704 0.667 0.685
Videos Descriptor Recall Precision F-score Wandering CS-LBP 0.700 0.640 0.668
Boring LBP 0.555 0.512 0.533 students CS-LDP 0.654 0.634 0.643
parking, CS-LBP 0.663 0.539 0.595 XCS-LBP 0.826 0.742 0.782
active CS-LDP 0.712 0.556 0.624 LBP 0.767 0.659 0.709
bkbg XCS-LBP 0.673 0.628 0.650 Rabbit in CS-LBP 0.826 0.626 0.712
LBP 0.456 0.490 0.473 the night CS-LDP 0.706 0.619 0.659
CS-LBP 0.664 0.583 0.621 XCS-LBP 0.805 0.684 0.740
Big trucks LBP 0.750 0.519 0.614
CS-LDP 0.675 0.673 0.674
XCS-LBP 0.623 0.788 0.696 Snowy CS-LBP 0.734 0.516 0.606
LBP 0.500 0.500 0.500 christmas CS-LDP 0.625 0.510 0.562
Wandering CS-LBP 0.632 0.525 0.573 XCS-LBP 0.726 0.538 0.618
students CS-LDP 0.691 0.566 0.622 LBP 0.657 0.685 0.671
XCS-LBP 0.854 0.714 0.778 Beware of CS-LBP 0.699 0.664 0.681
LBP 0.562 0.515 0.537 the trains CS-LDP 0.641 0.642 0.642
Rabbit in CS-LBP 0.657 0.515 0.577 XCS-LBP 0.759 0.731 0.744
the night CS-LDP 0.742 0.561 0.639 LBP 0.724 0.711 0.717
XCS-LBP 0.818 0.706 0.758 Train in CS-LBP 0.710 0.675 0.692
LBP 0.568 0.516 0.541 the tunnel CS-LDP 0.679 0.697 0.688
Snowy CS-LBP 0.640 0.508 0.567 XCS-LBP 0.695 0.680 0.687
christmas CS-LDP 0.684 0.513 0.586 LBP 0.523 0.509 0.516
Traffic
XCS-LBP 0.719 0.557 0.628 CS-LBP 0.553 0.520 0.536
during
LBP 0.542 0.511 0.526 CS-LDP 0.527 0.510 0.518
windy day
Beware of CS-LBP 0.608 0.556 0.581 XCS-LBP 0.532 0.518 0.525
the trains CS-LDP 0.711 0.618 0.662 LBP 0.867 0.574 0.691
XCS-LBP 0.780 0.674 0.723 One rainy CS-LBP 0.774 0.589 0.669
LBP 0.524 0.505 0.514 hour CS-LDP 0.797 0.556 0.655
Train in CS-LBP 0.636 0.640 0.638 XCS-LBP 0.761 0.628 0.688
the tunnel CS-LDP 0.668 0.659 0.663 LBP 0.708 0.632 0.663
XCS-LBP 0.655 0.688 0.672 Average CS-LBP 0.712 0.622 0.661
LBP 0.491 0.497 0.494 scores CS-LDP 0.661 0.612 0.632
Traffic
CS-LBP 0.597 0.528 0.560 XCS-LBP 0.713 0.658 0.681
during
CS-LDP 0.589 0.515 0.550
windy day Table 6: Elapsed CPU times (averaged on the nine real-
XCS-LBP 0.572 0.529 0.550
LBP 0.536 0.508 0.521 world videos of the BMC) over LBP times
One rainy CS-LBP 0.563 0.504 0.532
hour CS-LDP 0.658 0.520 0.581
Descriptor CS-LBP CS-LDP XCS-LBP
XCS-LBP 0.694 0.649 0.671
ABL 1.10 1.12 1.09
LBP 0.526 0.506 0.515
GMM 1.06 1.07 1.05
Average CS-LBP 0.629 0.544 0.583
scores CS-LDP 0.681 0.576 0.558
XCS-LBP 0.710 0.659 0.681
nary similarity patterns. In Int. Conf. on Com-
puter and Robot Vision, pages 106–112.

REFERENCES Bouwmans, T. (2014). Traditional and recent ap-

proaches in background modeling for fore-
Bilodeau, G.-A., Jodoin, J.-P., and Saunier, N. (2013). ground detection: An overview. In Computer
Change detection in feature space using local bi- Science Review, pages 31–66.
Heikkilä, M. and Pietikäinen, M. (2006). A texture- Vacavant, A., Chateau, T., Wilhelm, A., and
based method for modeling the background and Lequievre, L. (2012). A benchmark dataset for
detecting moving objects. IEEE Trans. on Pat- outdoor foreground/background extraction. In
tern Analysis and Machine Intelligence, 28:657– Asian Conf. on Computer Vision, pages 291–
662. 300.
Heikkilä, M., Pietikäinen, M., and Schmid, C. (2009). Vishnyakov, B., Gorbatsevich, V., Sidyakin, S., Vizil-
Description of interest regions with local binary ter, Y., Malin, I., and Egorov, A. (2014). Fast
patterns. Pattern Recognition, 42:425–436. moving objects detection using ilbp background
model. International Archives of the Photogram-
Lee, Y., Jung, J., and Kweon, I.-S. (2011). Hierarchi- metry, Remote Sensing and Spatial Information
cal on-line boosting based background subtrac- Sciences, XL-3:347–350.
tion. In Korea-Japan Joint Workshop on Fron-
tiers of Computer Vision (FCV), pages 1–5. Wang, L. and Pan, C. (2010). Fast and effective back-
ground subtraction based on εLBP. In IEEE Int.
Liao, S., Zhao, G., Kellokumpu, V., Pietikainen, M., Conf. on Acoustics, Speech, and Signal Process-
and Li, S. (2010). Modeling pixel process with ing, pages 1394–1397.
scale invariant local patterns for background
subtraction in complex scenes. In IEEE Int. Wang, L., Wu, H.-Y., and Pan, C. (2010). Adaptive
Conf. on Computer Vision and Pattern Recogni- εLBP for background subtraction. In Kimmel,
tion, pages 1301–1306. R., Klette, R., and Sugimoto, A., editors, Asian
Conf. on Computer Vision, LNCS 6494, pages
Noh, S. and Jeon, M. (2012). A new framework for 560–571. Springer.
background subtraction using multiple cues. In
Asian Conf. on Computer Vision, LNCS 7726, Wu, H., Liu, N., Luo, X., Su, J., and Chen, L. (2014).
pages 493–506. Springer. Real-time background subtraction-based video
surveillance of people by integrating local tex-
Ojala, T., Pietikäinen, M., and Mäenpää, T. (2002). ture patterns. Signal, Image and Video Process-
Multiresolution gray-scale and rotation invariant ing, 8(4):665–676.
texture classification with local binary patterns.
IEEE Trans. on Pattern Analysis and Machine Xue, G., Song, L., Sun, J., and Wu, M. (2011). Hy-
Intelligence, pages 971–987. brid center-symmetric local pattern for dynamic
background subtraction. In IEEE Int. Conf. on
Pietikäinen, M., Hadid, A., Zhao, G., and Ahonen, Multimedia and Expo, pages 1–6.
T. (2011). Computer vision using local binary
patterns, volume 40 of Computational Imaging Xue, G., Sun, J., and Song, L. (2010). Dynamic back-
and Vision. Springer-Verlag. ground subtraction based on spatial extended
center-symmetric local binary pattern. In IEEE
Richards, J. and Jia, X. (2014). Local binary patterns: Int. Conf. on Multimedia and Expo, pages 1050–
New variants and applications, volume 506 of 1054.
Studies in Computational Intelligence. Springer-
Verlag. Yin, H., Yang, H., Su, H., and Zhang, C. (2013). Dy-
namic background subtraction based on appear-
Shah, M., Deng, J., and Woodford, B. (2013). Video ance and motion pattern. In IEEE Int. Conf. on
background modeling: Recent approaches, is- Multimedia and Expo Workshop, pages 1–6.
sues and our solutions. In Machine Vision and
Yuan, G.-W., Gao, Y., Xu, D., and Jiang, M.-
Applications, pages 1–15.
R. (2012). A new background subtraction
Shimada, A. and Taniguchi, R.-I. (2009). Hybrid method using texture and color information.
background model using spatial-temporal lbp. In In Advanced Intelligent Computing Theories
IEEE Int. Conf. on Advanced Video and Signal and Applications, LNAI 6839, pages 541–548.
Based Surveillance, pages 19–24. Springer.

Sobral, A. and Vacavant, A. (2014). A comprehensive Zhou, W., Liu, Y., Zhang, W., Zhuang, L., and Yu,
review of background subtraction algorithms N. (2011). Dynamic background subtraction us-
evaluated with synthetic and real videos. Com- ing spatial-color binary patterns. In Int. Conf. on
puter Vision and Image Understanding, 122:4– Graphic and Image Processing, pages 314–319.
21. IEEE Computer Society.