-
PPT(1)
大家上午好！今天我汇报的主题是：基于改进型
LBP
算法的运动目标检测系统。运动
目标检测
技术能降低视频监控的人力成本，
提高监控效率，
同时也是运动
目标提取、
跟踪及
识别算法的基础。
图
像信号具有数据量大，
实时性要求高等特征。
随着算法的复杂度
和图像
清晰度的提高，
需要的处理速度也越来越高。
幸运的是，
图像处理的固有特性是并行的，尤
其是
低层和中间层算法。这一特性使这些算法，比较容易在
FPGA
等并行运算器件上实现，
今天汇报的主题就是关于改进型
LBP
算法在硬件上的实现。
good
morning everyone.
My report is about a
Motion Detection System Based on Improved LBP
Operator.
Automatic
motion
detection
can
reduce
the
human
cost
of
video
surveillance
and
improve
efficiency
[
?
'f
??
(
?
)ns
?
]
，
it
is
also
the
fundament
of
object
extraction,
tracking
and
recognition
[rek
?
g'n
??
(
?
)n]
.
In
this
work,
efforts
['ef
?
ts]
were
made
to
establish
the
background
model
which
is
resistance to the variation of
illumination. And our video surveillance system
was realized on a
FPGA based platform.
PPT(2)
目前，
常用的运动目标检测算法有背景差分法、
帧间差分法等。
帧间差分法的基本原理
是将相邻两帧图像的对应像素点的灰度值进行减
法运算，若得到的差值的绝对值大于阈值，
则将该点判定为运动点。
但是帧间差分检测的结果往往是运动物体的轮廓，
无法获得目标的
< br>完整形态。
Currently, Optic
Flow, Background Subtraction and
Inter-
frame difference
are regard as the
three
mainstream algorithms to detect
moving object.
Inter-frame
difference
based method need not model
['m
?
dl]
the background. It detects moving
objects based on the frame difference
between two continuous frames.
The
method is easy to be
implemented
and
can
realize
real-
time
detection,
but
it
cannot
extract
the
full
shape
of
the
moving
objects
[6]
.
PPT(3)
在摄像头固定的情况下，背景差分
法较为简单，且易于实现。
若背景已知，并能提供完
整的特征数
据，
该方法能较准确地检测出运动目标。
但在实际的应用中，<
/p>
准确的背景模型很
难建立。
如果背景模型
如果没有很好地适应场景的变化，
将大大影响目标检测结果的准确性。
< br>像这副图中，背景模型没有及时更新，导致了检测的错误。
The basic principle of background
removal method is building a background model and
providing
a classification of the
pixels into either foreground or background
[3-5]
. In a complex and
dynamic
environment, it is difficult to
build a robust
[r
?
(
?
)'b
?
st]
background model.
PPT(4)
上述的帧间差分法和背景差分法都
是基于灰度的。
基于灰度的算法在光照条件改变的情
况下，性能
会大大地降低，甚至失去作用。
The
algorithms
we
have
discussed
above
are
all
based
on
grayscale.
In
practical
applications
especially outdoor environment, the
grayscales of each pixel are unpredictably shifty
because of
the variations in the
intensity and angle of illumination.
PPT(5)
为了解决光照改变带来的基于灰度的算法失效的
问题，我们考虑用纹理特征来检测运
动目标。而
LBP
算法是目前最常用的表征纹理特征的算法之一。首先在图像中提取相邻
9
个像素点的灰度值。
然后对
9
个像素中除中心像素以外的其他
8
个像素做二值
化处理。
大于
等于中心点像素的，标记为
1
，小于的则标记为
0
。最后将中心
像素点周围的标记值按统一
的顺序排列，得到
LBP
值，图中计算出的
LBP
值为
10001111
。当某区域内所有像素的灰度
都同时增大
或减小一定的数值时，
该区域内的
LBP
值是不会改变的，
这就是
LBP
对灰
度的平
移不变特性。它能够很好地解决灰度受光照影响的问题。
In
order
to
solve
the
above
problems,
we
proposed
an
improved
LBP
algorithm
which
is
resistance to the
variations of illumination.
Local binary pattern (LBP) is widely
used in machine vision applications such as face
detection,
face
recognition
and
moving
object
detection
[9-11]
.
LBP
represents
a
relatively
simple
yet
powerful
texture
descriptor
which
can
describe
the
relationship
of
a
pixel
with
its
immediate
neighborhood.
The
fundamental
of
LBP
operator
is
showed
in
Fig
1.
The
basic
version
of
LBP
produces 256 texture
patterns based on a 9 pixels neighborhood. The
neighboring pixel is set to 1
or 0
according to the grayscale value of the pixel is
larger than the value of centric pixel or not.
For example, in Fig1 7 is larger than
6, so the pixel in first row first column is set
to 1. Arranging
the 8 binary numbers in
certain order, we get an 8 bits binary number,
which is the LBP pattern
we need. For
example in Fig.1, the LBP is 10001111. LBP is
tolerant
['t
?
l(
?
)r(
?
)nt
]
against illumination
changing.
When
the
grayscales
of
pixels
in
a
9
pixels
window
are
shifted
due
to
illumination
changing, the LBP value will keep
unchanged.
PPT(6)
图中的一些常见的纹理，都能用一些简单的
< br>LBP
向量表示，对于每个像素快，只需要
用一个
8
比特的
LBP
值来
表示。
There are some
textures , and they can be represent by some
simple 8bit LBP patterns.
PPT(7)
从这幅图也可以看出，虽然灰度发生了很大的变化，但是纹理特征并没有改变，
LBP
值
也没有变化。
You can see, in these picture ,
although the grayscale change alot, but the LBP
patterns keep
it value.
PPT(8)
上述的算法是
LBP<
/p>
算法的基本形式，但是这种基本算法不适合直接应用在视频监控系
统中。主要有两个原因：第一，在常用的视频监控系统中，特别是在高清视频监控系统中，
9
个像素点覆盖的区域很小，在如此小的区域内，各个像素点的灰度值十分接近，甚至
是相
同的，纹理特征不明显，无法在
LBP
值上体现。第二，由于以像素为单位计算
LBP
值，像素<
/p>
噪声会造成
LBP
值的噪声。
这两个原因导致计算出的
LBP
值存在较大的随机
性，
甚至在静止
的图像中，相邻两帧对应位置的
LBP
值也可能存在差异，从而引起的误检测。
为了得到更好的检测性能，
我们采用基于块均值的
< br>LBP
算法。
这种方法的基本原理是先
< br>计算出
3
×
3
< br>个像素组成的的像素块的灰度均值，以灰度均值作为该像素块的灰度值。然后
以<
/p>
3
×
3
个像素块
（即
9
×
9
个
像素）为单位，计算
LBP
值。
The typical LBP cannot meet the need of
practical application of video surveillance for
two
reasons:
Firstly, a
“window” which only contains 9 p
ixels
is a small area in which the grayscales of
pixels are similar or same to each
other, and the texture feature in such a small
area is too weak
to be reflected by a
LBP
. Secondly, pixel noise will
immediately cause the noise of LBP
,
which may
lead to a large number of
wrong detection. In order to obtain a better
performance, we proposed
an
improved
LBP
based
on
the
mean
value
of
“block”.
In
o
ur
algorithm,
one
block
contains
9
pixels. Compared with original LBP
pattern calculated in a local 9 neighborhood
between pixels,
the
improved
LBP
operator is
defined
by
comparing
the
mean
grayscale value
of central
block
-
-
-
-
-
-
-
-
-
上一篇：2017江西会考信息技术考试知识点大全(必修)
下一篇：售后服务承诺翻译11