Professional Documents
Culture Documents
END
303
2012 IEEE Control and System Graduate Research Colloquium (ICSGRC 2012)
A.
Figure 2: The detected strong corners from both frames where they marked with
green dots.
B. Corresponding points
Next, the initial correspondences between the points that is
identified from the previous step will be invoked.
Correspondences between the invoked points have to be picked
for each point, for that purpose a matrix of 9 x 9 blocks will be
extracted around each point from its consecutive image frames.
The most important here is matching the cost between points by
finding the Sum of Squared Differences (SSD) between the
consecutive image regions of frames. Thus we have to find the
lowest costs to consider them in the solution [8]. Figure 3
showed the same positions for the green color points of the
initial corresponding points existed in both frames.
304
2012 IEEE Control and System Graduate Research Colloquium (ICSGRC 2012)
C. Accurate correspondence
As mentioned above, there are several incorrect point
correspondences but strong estimation of geometric transform
between the two image frames can be determined using the
random sample consensus algorithm (RANSAC) [5][6]. This
algorithm searched through the given set of point
correspondences specifically valid linear correspondence as in
Figure 4.
D. Frames Correction
Further, the raw mean video frames and the mean of
corrected frame are computed as in Figure 5.
Figure 6:
Cod Figure 5:
Corrected frames
The left image showed the mean of the raw input frames that
resembled the distorted original video frame due to extreme
jittery. On the right side is the mean of the corrected frames with
less distortion. This proven that the stabilization algorithm
worked well. Several more samples of corrected video frames
are as depicted in Figure 6.
E. Quality
The output video quality is also measured based on the
proposed methods. This is evaluated based on SVD based
grayscale Image value and graphical measurement.
i. SVD Based Grayscale Image Quality
Singular value decomposition (SVD) is developed as a new
measurement that can express the quality of distorted images
either graphically that is in 2D measurement or numerically as a
scalar measurement, both near and above the visual threshold.
The experiments here utilized SVD based measurement that
outperformed the normally used PSNR [10]. Equation 1
represented the computed value for this purpose:
305
2012 IEEE Control and System Graduate Research Colloquium (ICSGRC 2012)
the graphical results, Vid1 has the most size of blocks among all
frames blocks to indicate that it is the worst video experienced
distortion followed by Vid2 and Vid3. This result agreed with
the SVD value calculated earlier.
where:
Dmid represents the midpoint of the sorted DiS
k is the image size
n is the block size
M SVD is the measurement of Singular value
decomposition
An example for the output quality for Vid1 based on Equation 1
with k=8, n=1, Di and Dmid represented by 256256 matrix
attained M-SVD of 22.20. Hence, the numerical quality
obtained from the three sample videos are tabulated in Table 1.
As visualize in Figure 6, it can be seen that Vid3 obtained the
best quality based on the calculated value that is 40.50%
followed by Vid2 with 39.21% and Vid1 22.20%. This
resembled that Vid1 has great distortion whilst Vid3 is least
distorted.
IV. CONCLUSION
In conclusion, the video stabilization technique based on
proposed method showed remarkable results in term of
stabilizing high jittery videos suffered from distortion. Initial
results also proven that due fusions of RANSAC algorithm,
Gaussian filter, Harris, Stephenss and SAD
efficiency
stabilization process succeeded based on the output quality
attained. Future work includes finding better feature detector and
overcome the consequences of extreme shaking of handheld
camera in feasible real time implementation for video
stabilization.
Acknowledgment
Funding for presenting this study was supported by Faculty of
Electrical Engineeering, UiTM Shah Alam, Selangor.
i. Graphical measurement
The criteria of measuring graphical quality in any image or
frame can be done as shown in Figure 7. Graphical measurement
will indicate the condition of video due to distortion. As seen in
306
2012 IEEE Control and System Graduate Research Colloquium (ICSGRC 2012)
Table I: Criteria of sample videos & Results (Type: RGB & Extension: AVI)
st
Sample 1 Frame
Sample
Inputs
Size
Bytes
No
Vid1
2x139
Vid2
Vid3
1112
Frames #
&
Length
34 & 2
8.85e5
8.81
Quality
Value
(M -SVD)
22.20 %
2x66
528
73 & 3
0.0824
9.18
39.21%
2x128
1024
132 & 4
3.553e3
9.28
40.50%
REFERENCES
[1] M. Gleicher and F. Liu., Re-cinematography: Improving the camerawork
of casual video, ACM Transactions on Multimedia
Computing,
Communications, and Applications, 5(1), pp 1- 28, 2008.
[2] C. Harris and M.J. Stephens, A combined corner and edge detector,
Proc of Alvey Vision Conference, pp 147152, 1988.
[3] Anu Suneja and Gaurav Kumar . An Experimental Study of Edge
Detection Methods in Digital Image, Global Journal of Computer
Science and Technology, 10(2), 2010.
[4] http://www.mathworks.com/products/computervision/demos.html?file=/products/demos/shipping/vision/videostabilize_
pm.html.
[5] Fischler, MA; Bolles, RC. "Random Sample Consensus: A Paradigm for
Model Fitting with Applications to Image Analysis and Automated
Cartography." Comm. of the ACM 24, 1981.
[6] Tordoff, B; Murray, DW. "Guided sampling and consensus for motion
estimation." 7th European Conference on Computer Vision, 2002.
[7] J. Jin, Z. Zhu, and G. Xu. Digital video sequence stabilization based on
2.5D motion estimation and inertial motion filtering, Real-Time
Imaging, 7(4):357365, 2001.
[8] http://siddhantahuja.wordpress.com/tag/sum-of-squared-differences/
[9] M. Pilu. Video stabilization as a variation problem and numerical
solution with the Viterbi method. In Proceedings of Computer Vision
and Pattern Recognition, pp 625630, 2004.
[10] Aleksandra Shnayderman, Alexander Gusev, and Ahmet M. Eskicioglu
An SVD-Based Grayscale Image Quality Measure for Local and Global
Assessment ,IEEE 15(2), 2006.
307
SAD
Value
Computational
Time (s)