detection, because smoke has characteristics similar to steam.
Gottuk et al. [11] evaluated the effectiveness of several commer-
cial video-based fire detection systems for small spaces on navy
ships. Their experiments showed that the video-based fire detec-
tion systems detected more fires faster than the traditional
systems. Yuan [12] adopted the integral image technique to
quickly estimate the motion of moving objects and proposed an
accumulative motion model for the video smoke detection. To
reduce false alarms, the orientation is accumulated over time to
compensate the results for the inaccuracy of orientation. Han and
Lee [13] presented a flame and smoke detection method to be
used in tunnels by analyzing color and motion information.
1.2. Texture analysis
Texture analysis is an effective method for the smoke detec-
tion. Most of the existing methods are sensitive to rotation and
illumination. Although some methods of texture analysis, such as
co-occurrence matrix methods [14], are insensitive to rotation, it
depends greatly on illumination conditions. Histogram equaliza-
tion is often performed to reduce the adverse effects of varying
illumination. However, histogram equalization may decrease
other characteristics of images. Ojala et al. [15] presented local
binary patterns (LBP) for rotation and illumination invariant
texture classification. Huang et al. [16] improved this method
by computing the derivative based LBPs for face alignment. To
prevent loss of global information caused by an LBP, Jafari-
Khouzani and Soltanian-Zadeh [17] used the radon transform to
estimate the principal orientation of the texture image and then
computed the wavelet energy features along the principal orien-
tation. The radon transform though is time consuming. Generally,
texture can be precisely characterized by a spatial statistical
measure and image contrast. So Ojala et al. [15] used the LBP
and variance joint histogram for rotation invariant texture classi-
fication. Guo et al. [18] observed that a quantization step is
required due to the continuity attribute of the variance value.
The quantization step and the computation of the joint histogram
were completely avoided by proposing a new operator, which is
called the local binary pattern variance (LBPV). In fact, this
operator can be regarded as the integral projection along the
variance axis. Liao et al. [19] proposed dominant local binary
patterns for texture classification by regarding the first 80% most
frequent patterns as dominant features.
In this paper, we make use of the histogram sequence of LBP
and LBPV pyramids to propose a new approach to the video-based
smoke detection. The LBP and LBPV features, which were pre-
sented by Ojala et al. [15] and Guo et al. [18], respectively, are
computed at each level of the 3-level image pyramid with
uniform, rotation-invariance and rotation-invariance-uniform
patterns. Then, all the histograms of the LBP and LBPV pyramids
are concatenated into a feature vector for smoke detection. Hence,
the feature vector contains both local and global informations.
This paper presents at least two innovative ideas on the video-
based smoke detection. First, LBP-related texture analysis meth-
ods are used to detect smoke. Second, different LBP patterns are
used at different levels of the image pyramid, in order to collect
local and global informations.
2. Local binary pattern
2.1. LBP
Local binary pattern (LBP) is a gray-scale texture operator,
which can capture spatial characteristics of images. At a center
pixel, a pattern number is computed by comparing its value with
values of its neighboring pixels. The LBP initial label for that
center pixel is given by
LBP
P, R
¼
X
P1
i ¼ 0
sðg
i
g
c
Þ2
i
ð1Þ
sðxÞ¼
1ifxZ 0
0 else
ð2Þ
where g
c
is the gray scale value of the center pixel, g
i
is the value
of its ith neighbor, P is the number of neighbors and R is the
radius of the neighborhood which is the Euclidean distance
between the center point and its neighbors.
Fig. 1 shows examples of circularly symmetric neighbor sets
for different P and R. LBP
8,1
, LBP
8,2
and LBP
16,2
denote a point set
with P¼8 and R¼1, one with P¼8 and R¼ 2, and one with P¼16
and R ¼2, respectively.
As the radius increases, an LBP contains more global informa-
tion of texture patterns and also needs more neighbors, thus
increasing the computational cost. If a large radius R is mixed
with small P, it will result in severe artifacts in the re-sampling. So
in our smoke detection, for the sake of balance between informa-
tion and computation, the number of neighbors is set to 8, and the
radius is set to 1.
Circularly symmetric neighbor sets require bi-linear interpola-
tion of pixel values. It may take a lot of time to interpolate values
of all the neighbors. In order to further improve the computa-
tional performance, the Euclidean distance is replaced with the
block distance, as shown in Fig. 2. Thus, re-sampling of the values
of the neighbors is completely avoided.
Ojala et al. [15] defined three different types of patterns, which
are uniform, rotation-invariant and rotation-invariant-uniform.
Fig. 1. LBP texture operator.
Fig. 2. Modified LBP texture operator with block distance (LBP
8,1
).
F. Yuan / Fire Safety Journal 46 (2011) 132–139 133