International Core Journal of Engineering 2020-26 | Page 101

IV. E XPERIMENTS
Experiments are conducted on a computer with Intel Core
i7 CPU(3.60GHz). The code is implemented by MATLAB
2018. The threshold parameter in the de-noising process is
set to 16 in the experiments.
3) In the de-noising processing, a threshold parameter is
used to process the Difference_Gabor, to get the
Difference_Gabor_Threshold. The Difference_Gabor_T-
hreshold is converted into a binary graph, by setting non-
zero pixel values to 1, and others to 0. Then the method
described in the postprocessing part is used to process the
binary graph to obtain, Bitmap_Gabor.
Fig. 5. The Gabor function act on the highway dataset. (a):Total difference
image; (b): Total difference after binary processing; (c): Total difference
after de-noising processing; (d): detection result: Bitmap_Gabor. A. The Cdnet dataset
The Change Detection.net dataset ( Cdnet ) is the first
large-scale benchmark dataset established by the IEEE
Change Detection Workshop organizing committee in 2012.
It is a video dataset devoted to the evaluation of change and
motion detection approaches. In 2014, the IEEE Change
Detection Workshop organizing committee expanded the
dataset by adding 22 additional videos in 5 new categories.
The Cdnet dataset covers most application scenarios of
moving target detection and has become a main platform for
fair comparison of algorithms in the research of moving
target detection.
C. Postprocessing
The binary image is processed by the morphological
closing [19], filled and smoothed by median filter [20] to
obtain the detection result. The functions in MATLAB 2018
are used in the experiments. The proposed method is evaluated on the Cdnet 2014
dataset [21] which totally includes 53 scenarios, categorized
into 11 classes. The baseline category is the best test data to
evaluate the generality of exercise objective testing
algorithms. It contains four catalogs that are highway, office,
pedestrian, and PETS2006.
An example of detection using the texture features is
shown in Fig. 5.
(a)
(b)
(c)
(d)
First, the original image is processed by the
morphological closing which performs closing with the
structuring element. To complete this step, the
IMCLOSE(IM,SE) function is used . IM can be any numeric
or logical class and any dimension, and must be non-sparse.
SE is a single structuring element object. In our experiment,
IM is the input binary image, SE is the shape of a rectangle
of 10 by10 for doing closing operations.
B. Results
The results of the proposed algorithm on the baseline
dataset are shown in Fig. 7, 8, 9, and 10. In all these figures,
(a) is the original video frame of the baseline dataset, (b) is
the final result of the moving target extracted with the
proposed method, (c) is the binary graph, Bitmap_Gabor,
obtained only by using the texture based inter-frame
difference detection, (d) is a binary graph, Bitmap_Lab,
obtained only based on color based inter-frame difference
detection, and (e) is the binary graph, Binary_image, based
on both texture features and color features.
Then the region and holes in the image are filled. The
IMFILL(BW1, LOCATIONS) function is used, which is
based on the principle of a flood-fill operation on
background pixels of the input binary image BW1. It starts
from the points specified in LOCATIONS, and the parameter
of LOCATIONS is chosen to be holes.
Fig. 7 shows the experimental results conducted on the
Highway dataset. It can be seen that both texture features
(Bitmap_Gabor) and color features (Bitmap_Lab) can detect
the moving cars well, while texture features work better on
the little cars and color features work better on the closest car.
And it can be seen that combination of the two features
(Binary_image) produces the best result for the car detection.
Finally, the binary image is operated by the median
filtering operation to smoothing the noise in the image. The
MEDFILT2 (A, [M N]) function performs median filtering
of the matrix A in two dimensions. Each output pixel
contains the median value in the M-by-N neighborhood
around the corresponding pixel in the input image.
Fig. 8 shows the experiment results conducted on the
Office dataset. Fig. 9 shows the experiment results conducted
on the Pedestrian dataset. It can be seen that for the two
datasets, texture features (Bitmap_Gabor) can detect the full
body of the person successfully, as shown in Fig. 8(c) and
Fig. 9(c), while the color features (Bitmap_Lab) only detects
few parts of the body, as shown in Fig. 8(d) and Fig. 9(d).
The combination results shown in Fig. 8(e) and Fig. 9(e) are
mainly the same as the results from the texture features.
D. Combination
1) Binary_image is obtained by combining the
difference binary image representing color features,
Bitmap_Lab, and the different binary image representing
texture features, Bitmap_Gabor. The combination is done
by the logical AND operation.
2) Apply the Binary_image as a mask on the original
frame F2 to get the moving target.
Fig. 10 shows the experiment results conducted on the
PETS2006 dataset. It can be seen that texture features
(Bitmap_Gabor) perform much better than color features
(Bitmap_Lab), while there still exits some holes in the result
of the texture features and the combination result. The person
closest to the center of the image is detected successfully,
while the two people on the bottom right corner are only
partially detected. This may be due to that the two people on
the bottom right corner have just entered into the view
shortly.
An example for the combination is shown in Fig. 6.
Fig. 6. The combination result
79