Digital Signal Processing Reference
In-Depth Information
where
B
k
(
.
)
is the current background image,
η ∈
(0
,
1)
is a weight.
The variance
σ
2
that characterizes the noise level presents in the video sequence is
crucial for calculating the local sum
Δ
in Eq. (1). The algorithm described in [11] is
adopted in this paper. For reducing the computation cost, the estimation is integrated
in the process of calculating the local sum. That is both the
computation and the
convolution for noise estimation ( see [11] for details ) are completed in one scan.
Δ
2.2
Shadow Elimination Based on Intensity and Chromacity
After the variation detection, the pixels in the input image are classified into two cate-
gories: foreground and background. The former consists of moving objects, the moving
cast shadows, and the false positive pixels. The goal of the following step is to eliminate
the shadows and false positives by using simple intensity and chromacity features. In
our system, we propose the use of YCbCr color space. The reason is that this format is
typically used in video surveillance systems for video coding. Using the same, instead
another, format for object segmentation will avoid the extra computation required in
color conversion.
In real-world outdoor scenes, the shadows are formed when the light source is
blocked. So, the shadow pixels decrease their intensity in comparison with the refer-
ence background. Further observation suggests that blue component of a shadow pixel
may be enhanced by the reflection of the sky in blue spectrum [10]. Therefore, we use
the following simple assumptions to eliminate shadow pixels in this stage: (1) A fore-
ground pixel cannot be a shadow if it has higher Y component value in current frame
than in background; (2) A foreground pixel cannot be a shadow if it has higher Cr or
lower Cb value in current frame than in background.
In order to suppress the influence of noise, decision like we are faced with is usually
based on evaluating a set of pixels inside a small region instead of a single pixel. We
thus compute the local mean of Y, Cr and Cb components inside a
sliding win-
dow of current frame and the background. These mean values are then evaluated using
above mentioned two heuristic rules. For every pixel in the set of moving pixels in the
frame difference mask, if either of the rules is satisfied, it can be excluded from further
evaluating.
3
×
3
2.3
Shadow Elimination Based on HOG
We have observed that local textures can be described rather well by the local inten-
sity gradients. In the shadow detection scenarios, an important fact that shadows pixels
decrease their intensity in comparison with the reference background, however do not
significantly modify the gradient directions, can be exploited. In this paper, we make
use of histograms of oriented gradient (HOG), which is first introduced by N. Dalal and
B. Triggs [12] to solve pedestrian detection issues and has shown superior performance
to most other existing feature sets, as feature descriptor.
An overview of the HOG descriptors can be described as follows: First, the image
gradients are computed by convolving the original image with a horizontal gradient
kernel [
]
T
, respectively. At each pixel,
the orientation of the gradient, which covers a range of
−
1
,
0
,
1
] and a vertical gradient kernel [
−
1
,
0
,
1
0
◦
−
180
◦
, is quantized to a
