Embedded Implementations of Real Time Video Stabilization Mechanisms A Comprehensive Review

International Journal of Trend in Scientific Research and Development (IJTSRD)
Volume 5 Issue 3, March-April
@ IJTSRD | Unique Paper ID – IJTSRD39
Embedded Implementations
Stabilization Mechanisms: A Comprehensive Review
Mohammed Ahmed
1,2Rabindranath Tagore University, Bhopal, Madhya Pradesh
ABSTRACT
Video Stabilization has been widely researched and still is under active
research considering the advancements that are being done in the field of
digital imagery. Despite all the works, Hardware based Real time
Stabilization systems; especially works dealing with implementation of
Technology on prototype Implementation boards have been very few. This
review works focuses on the specific aspect of the modern day Stabilization
systems implemented in prototyping boards and the algorithms that have
been found suitable for such implementation taking considerations of cost,
size and speed as the principal criterions.
KEYWORDS: Motion Estimation, Real-time Stabilization, FPGA, SIFT, SURF
1. INTRODUCTION
Digital Video stabilization has been the more favored
technique in recent times as compared to its more
accurate but expensive counterparts like mechanical and
optical stabilization. This is primarily because of the fact
that Software based implementations provide a viable
alternative to solutions having costly and bulky hardware
assemblies like sensors, gyroscopes, lenses, making DVS
based setups portable and cost effective. A generic DVS
system consists of three major blocks as shown in fig
1.The first block in Digital stabilizatio
Estimation where global motions in the frames of the
video are extracted. Then, intentional motions are
separated from the global motions in the second step
which leads to motion correction stage. The last block
deals with image correction to produces the final
stabilized video using the estimated unintentional
motions.
Motion estimation is the most time consuming and
difficult part in digital stabilization which also forms the
basis of the operational speeds of the stabilization system.
For real time stabilization mechanisms, computational
speeds are of critical importance which actually are a
function of the employed algorithm being computationally
less intensive. Motion estimation, being the most
complicated section, needs to be focused on
achieving the objectives of real time stabilization. Various
mechanisms have been used to implement the motion
April 2021 Available Online: www.ijtsrd.com e
39963 | Volume – 5 | Issue – 3 | March-April
Embedded Implementations of Real Time
Mohammed Ahmed1, Dr. Laxmi Singh2
1Research Scholar, 2Professor,
ath Tagore University, Bhopal, Madhya Pradesh, India
Video Stabilization has been widely researched and still is under active
research considering the advancements that are being done in the field of
digital imagery. Despite all the works, Hardware based Real time Video
Stabilization systems; especially works dealing with implementation of
Technology on prototype Implementation boards have been very few. This
review works focuses on the specific aspect of the modern day Stabilization
ng boards and the algorithms that have
been found suitable for such implementation taking considerations of cost,
time Stabilization, FPGA, SIFT, SURF
How to cite this paper
Ahmed | Dr. Laxmi Singh "Embedded
Implementations of Real Time Video
Stabilization Mechanisms: A
Comprehensive
Review" Published
in International
Journal of Trend in
Scientific Research
and Development
(ijtsrd), ISSN: 2456
6470, Volume
Issue-3, April 2021, pp.685
www.ijtsrd.com/papers/ijtsrd39963.pdf
Copyright © 20
International Journal of Trend in
Scientific Research and Development
Journal. This is an
distributed under
the terms of the
Creative Commons
Attribution License (CC BY 4
(http://creativecommons
has been the more favored
technique in recent times as compared to its more
accurate but expensive counterparts like mechanical and
optical stabilization. This is primarily because of the fact
that Software based implementations provide a viable
ve to solutions having costly and bulky hardware
assemblies like sensors, gyroscopes, lenses, making DVS
based setups portable and cost effective. A generic DVS
system consists of three major blocks as shown in fig
1.The first block in Digital stabilization is of Motion
Estimation where global motions in the frames of the
video are extracted. Then, intentional motions are
separated from the global motions in the second step
which leads to motion correction stage. The last block
o produces the final
stabilized video using the estimated unintentional
Motion estimation is the most time consuming and
difficult part in digital stabilization which also forms the
basis of the operational speeds of the stabilization system.
real time stabilization mechanisms, computational
speeds are of critical importance which actually are a
function of the employed algorithm being computationally
less intensive. Motion estimation, being the most
complicated section, needs to be focused on the most for
achieving the objectives of real time stabilization. Various
mechanisms have been used to implement the motion
estimation stage having differing computation speeds and
complexity levels, and also with varying degrees of
accuracy. The paper discusses the following main things;
the first part discusses the working details of Video
Stabilization Techniques, the second part discusses the
works and stabilization techniques that have been utilized
for realtime applications, especially prototyped
implementations. The last part discusses the realtime
performance of feature based descriptors like SIFT and
SURF and the associated advantages with them.
Fig 1 Generic Video Stabilization System
2. Motion Estimation Techniques
The motion Estimation is of five types in general:
1. Gradient based methods
2. Pel Recursive Techniques
3. Block based Matching methods
4. Feature Matching Methods
2.1. Gradient techniques
For image sequence analysis based applications, gradient
techniques can be employed. Gradient
use of the optical flow constrain equation. [1,2]
e-ISSN: 2456 – 6470
April 2021 Page 685
f Real Time Video
, India
How to cite this paper: Mohammed
Ahmed | Dr. Laxmi Singh "Embedded
Implementations of Real Time Video
Stabilization Mechanisms: A
Comprehensive
Review" Published
in International
Journal of Trend in
Scientific Research
and Development
(ijtsrd), ISSN: 2456-
6470, Volume-5 |
, April 2021, pp.685-693, URL:
www.ijtsrd.com/papers/ijtsrd39963.pdf
Copyright © 2021 by author(s) and
International Journal of Trend in
Scientific Research and Development
This is an Open Access article
distributed under
the terms of the
Creative Commons
Attribution License (CC BY 4.0)
//creativecommons.org/licenses/by/4.0)
estimation stage having differing computation speeds and
complexity levels, and also with varying degrees of
iscusses the following main things;
the first part discusses the working details of Video
Stabilization Techniques, the second part discusses the
works and stabilization techniques that have been utilized
for realtime applications, especially prototyped
plementations. The last part discusses the realtime
performance of feature based descriptors like SIFT and
SURF and the associated advantages with them.
Generic Video Stabilization System
Motion Estimation Techniques
mation is of five types in general:
Gradient based methods
Pel Recursive Techniques
Block based Matching methods
Feature Matching Methods
For image sequence analysis based applications, gradient
techniques can be employed. Gradient techniques make
use of the optical flow constrain equation. [1,2]
IJTSRD39963

International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com eISSN: 2456-6470
@ IJTSRD | Unique Paper ID – IJTSRD39963 | Volume – 5 | Issue – 3 | March-April 2021 Page 686
v. ∇ I r, t +
,
= 0 (1)
in order to find the motion vector ~v on the position ~r
with some additional constraints. For example, a Horn-
Schunck method [3] minimizes the square of the optical
flow
+ and + (2)
Some preliminaru works employing gradient mechanisms
have been done by Toshiaki Kondo, Pramuk Boonsieng
et.al [4,5] wherein thet discussed two conventional
gradient based motion estimation techniques. Since Unit
gradient vectors have the advantage of being insensitive to
constantly varying image intensity, the methods suggested
by [4,5 ] were further improved in [5] by using unit
gradient vectors rather than utilizing image intensities.
The works resulted in better motion estimation
techniques more robust to irregular lighting conditions as
compared to conventional systems.
In [6], information from consecutive frames was uses to
account for the trans-rotational motions. The work used
an optical flow based affine model for motion estimation
using the Horn Schunck algorithm. A model fitter was the
used to stabilize video sequences afterwards.
2.2. Pixel recursive Techniques
As a variation of gradient techniques, Pixel recursive
techniques were introduced. It is basically an iterative
gradient mechanism which minmizes predictive error
using Displaced Frame Difference (DFD). It owes its roots
to the Netravali Robbins method [7], which works on the
principle of recursively updating the DFD vector according
to the formula
d !
= d −∈ DFD&r, t, d '. ∇ I&r − d, t − ∆t' (3)
2.3. Block Matching Techniques
Block Matching Techniques [8] work on the minimization
of a differentiative measure. In other words, blocks in
current frame are matched with with those in previous
frame. The best prediction is done by matching between
the current block and all blocks in the search area, also
known as full search algorithm. If the algorithm uses MSE
as parameter for matching, then for every block size of
16×16, the algorithm will require 256 subtractions, 256
multiplications and 255 additions which is fairly resource
exhaustive [9]. This is the primary reason the algorithm is
rarely used, especially in real time applications.
Fig 2 Block matching a macro block of size 16 × 16
pixels and a search Parameter p of size 7 pixels.
Three step search (TSS) is a moderated version of the full
search algorithm, with a search location at the center of
the search area and searches in search window with sides
of 4. Three motion estimation algorithm have been used in
[10, 11]. The authors in [10] suggested a proposed a
modified adaptive rood path search (N-ARPS) algorithm
with small motion prejudgment (SMP), where it will
decide on the block search strategy taking into account
motion properties. Computational efficiency in [12] was
improved using the three step search (TSS) along with
GCBP matching that performed a competent search during
correlation measure calculation. Another variation of
block based searching uses Hexagon based search
algorithm (HEXBS) [14] which is based on the same search
pattern as in diamond search but instead uses a hexagon
in place of diamond search pattern. An adaptation of the
HEXBS was used by [14] wherein the direction of motion
was predicted by a rood shaped pattern incorporated with
Hexagonal based Search (HexBS) for refining search
process.
In context of realtime implementations, BBGDS (block
based gradient descent search) and diamond search (DS)
have been seen to outperform other algorithms in terms of
lesser number of computations. The DS algorithm consists
of a small diamond search pattern (SDSP) and large
diamond search pattern (LDSP). The LDSP pattern is
targeted at finding a match that will occur at the center of
the LDSP. The works done in the field have shown
simulation results using DS outperform TSS with a
matched NTSS in terms of compensation error. It also has
shown a much refined computational cost of the order of
20 to 25 percent. DS has shown better prediction quality
and lower complexity than TSS and 2D log search, making
it a more preferred choice for software implementations.
For hardware implementation, the two diamond sizes,
however, makes the control circuitry slightly more
complex. Another noteworthy work to address this issue
with an end application of stabilization has been done by
Song, Ma et al[16] wherein they have incorporated Hooke
Jeeves algorithm into the traditional diamond search
based fast block matching method, which as per their
results has resulted in the efficiency of motion estimation
being significantly improved. The testing has been done
with 36 blocks (8×8 pixels) is selected in each frame with
the proposed method consumed 24.79ms while 30.99ms
for DS, the average time processing the same frame of 10
consecutive frames; For a sample set of 49 blocks, it again
faired better with 25.39ms verses 31.79ms. When the
blocks are increased to 64, the proposed method
(35.47ms) rather than DS (41.35ms), still being a real time
video stabilization method.
2.4. Feature Matching Methods
Feature matching works on identifying scenes that are
easily recognizable. Here, motion estimation can be
performed by computing the displacements of these
points of interest in the entire video frame by frame and
by tracking the positions of these points of interest using
the properties of the selected features, forming
trajectories. One significant advantage is the fact that the
same point can be tracked and recognized across many
frames. Some of the more commonly used features
detection algorithms are the Kanade Lucas Tomasi (KLT)
feature tracker [52] that has been used to track features
across videos in several methods [8,16, 20,21,32,34].

Harris corners operator is also a features detector used for
this purpose [53]. These features are then tracked using
optical flow technique.
SIFT based mechanisms have been widely used
[54,38,40,41,43,45,33]. These features use descriptors
based on the image gradient to obtain very specific
descriptors with very reliable matching results. The
descriptors are rotation invariant, but the fact that, they
are slower than most alternatives, which makes them too
complicated for realtime applications. SURF points
although designed on similar principles [35], but are
optimized for speed, making them a good alternative
[46,8].
Other interesting features like Maximally Stable Extremal
Regions (MSER) [55] or FAST corners using BRIEF
descriptors [56,26] have been successfully used. Feature
matching provides accurate and fast results, and the
obtained trajectories allow for additional temporal
analysis in the remaining steps of the process, although
scenes with large uniform regions can sometimes yield
few features per frame. This is one of the limitations of this
kind of feature matching methods.
3. Real time Embedded Video Stabilization Systems
used for different applications
This section primarily focuses on the works that have been
done keeping in mind the realtime application of the end
product and have dealt with a hardware implementation
of the proposed algorithm. FPGA based prototyped
implementations can give a correct analysis of the real-
time usability of the VS algorithm as we can analyze the
algorithms for computational intensiveness, resource
utilization, speed, accuracy and output quality when
implemented in a prototyping platform. Most of these
works have been targeted for end applications like UAVS,
off road vehicles, Robots etc, requiring different
approaches, making use of varied computational resources
for them to be implementable on embedded systems in
real time. However, most of the referred methods have
involved finding the 2D motion model to estimate the
global motion path. Then the path is low pass filtered to
remove the high frequency jitter component. The low
frequency parameters are then posted onto frames. This
mechanism has been found to be useful for sequences with
minimal dynamic movement which is critical to for
stabilizing UAV aerial videos. However, it becomes
complicated when immediate stabilization, is required like
in realtime processing only presuming the global motion
path using a given window of frames.
In [26], Oriented FAST and Rotated BRIEF algorithm was
utilized for feature extraction followed by matching
between consecutive frames. Here, interframe motion was
computed using using Random Sample Consensus
(RANSAC). The authors proposed framework to
implement the whole feature based video stabilization
process on a single FPGA chip, in order to achieve real
time performance. A fully pipelined FPGA architecture was
proposed to substantially accelerate feature based video
stabilization in a highly parallel manner, which also
provides a reference to accelerating other feature-based
video processing tasks such as object tracking and video
fusion on FPGA. The authors in [27] worked on a design
specifically designed for real time embedded systems
where the present and previous video frames were used to
estimate the motion vector, followed by ﬁltering
operations. The implemention was done on a a Xilinx
Spartan 6 LX45 board that was tested on a 640x480 pixel
IR camera output. In Vazquez and Chang [28], a
smoothening operation was done using the Kanade Lucas
tracker to enable the detection of interest points. The
undesired motion was balanced out by adjusting for
additional rotation and displacements due to the
vibrations. The stabilizing speeds achieved through this
mechanism have been found to be in the range of 20fps to
28fps for images having resolutions roughly of the order
of 320x240 pixels. The proposed system was finally tested
on MAC using a 2.16GHz Processor clock and a three frame
delay.
Wang [29], suggested a video stabilization method
specifically designed for UAVs. In this literature, he
proposed a 3 step mechanism using a FAST corner
detector to locate the feature points existing in frames. In
the second step, after matching, the key points were then
used for estimation of affine transform to separate false
matches. [30] discussed the video stabilization that has
been implemented on an FPGA in the form of a Driven
Mobile Robot. Further works include [31] where a real
time video stabilization system has
been implemented using high frame rate jitter sensor and
a HS camera to retrieve feature points in gray level
512×496 image sequences at 1000 fps. In (32), the
stabilization work is based on local trajectories and robust
mesh transformation. Every frame is refined in a different
local trajectory matrix, helpful in modeling a nonlinear
video stream. The method has been proposed to be
performing better than conventional methods based on
feature trajectories. L. M. Abdullah, N. Md Tahir et. al[36]
use a system where Corner Detector System Object is used
to find corner values using Harris Corner Detection which
is one of the fastest algorithms to find corner values. After
the salient points from each frame are obtained the
correspondence between the points that are identified
previously need to be picked. Authors in [33] have
presented a feature-based approach for video stabilization
that produces stabilized videos, while preserving the
original resolution. . Lowe’s method is used to extract
extract the SIFT feature points for every frame in a video.
Fischler and Bolles’s RANSAC algorithm is used in the final
fine stage to screen out the feature pairs of mismatch.
Then, top eight pairs of feature points are chosen to form a
linear system for deriving the perspective transformation
between two consecutive frames. Another work that falls
in line with the real time implementation of Video
Stabilization in hardware has been represented in [34]
Here an embedded real time video stabilization system
targeted toward a Virtex-5 FPGA Xilinx board is given. The
horizontal and vertical global video frame movements are
calculated by making use of an integral projection
matching approach
4. Advantages of using Feature based extractors like
SIFT and SURF for real time applications
Apart from the other descriptors, the focus has been
primarily on the works using SIFT and SURF descriptors,
especially when it comes to real time stabilization works.
It is because of the fact that their accuracy has been pretty
good, but more so because they are computationally
efficient. Scale invariant feature transform (SIFT) extracts
and connects feature points in images which are invariant

to image scale, rotation and changes in illumination.
Moreover, it provides distinctive descriptors that can find
the correspondences between features in different images.
Because of all these advantages, it is very suitable for
estimating motion between images and hence can be very
effectively deployed for motion estimation in Video
Stabilization.
However, again as per the referred literatures it was
inferred that although SIFT has achieved remarkably
success in video stabilization, it suffers from
comparatively costly computation, especially for realtime
applications. This leads to need for viable replacements
with a much lower computational cost. Demonstrably, one
of the best of these methods that has been found in
comparison to other descriptors has been found to be
Speeded up robust features (SURF).
SURF deploys the Hessian matrix due to the very reason
that it has had a good history of performance on the
performing standards of computation time and accuracy.
In place of employing a different parameter for choosing
the location and the Hessian Laplace detector, Speeded up
Robust Features makes use of the determinant of the
Hessian matrix for both the operations. For a particular
pixel, the Hessian matrix of this pixel can be given as:
H&f x, y ' = -
./
0.
./
0 1
./
0 1
./
1.
2 (4)
For the purpose of adaptive scaling, the image is then
filtered making use of a Gaussian kernel, so that for a point
X = (x, y), the Hessian matrix becomes defined as:
H& x, σ ' = 3
L00 x, σ L01 x, σ
L00 x, σ L11 x, σ
5 (5)
where Lxx depicts the convolution of the Gaussian 2nd
order derivative of the image I at point x, and likewise Lxy
and Lyy. Gaussians derivatives are in general optimized
for scale space analysis but at the same time without
discretization and cropping. Due to this there is significant
observed loss in repeatability while image rotations
around in odd multiples. This is a major problem when it
comes to Hessian based detectors.
Despite this, the detectors have still been found to perform
well, and this performance glitch doesn’t outclass the
advantage of fast convolutions brought as a result of the
discretization and cropping. For determinant of the
Hessian matrix calculation, convolving is done with
Gaussian kernel second order derivative. Lowe used LoG
approximations with considerable success. SURF pushed
the boundaries of the approximation for both convolution
as well as second order derivative even more with the use
of box filters. The approximated second order Gaussian
derivatives are evaluated at a very low computational cost
making use of integral images being independent of size
hence enabling SURF algorithm the capability of fast
computation which makes it the perfect choice for the
applications of this work.
The 9 × 9 box filters as approximations for Gaussian
second order derivatives with σ= 1.2 and depicts the
lowest scale and the highest spatial resolution for
computing the blob response maps. The approximations
have been denoted by Dxx, Dyy, and Dxy. representing scale
for computing the blob response. The weights applicable
to rectangular regions are simple in nature to minimize
computation. This finally produces the approximated
determinant of the Hessian as:
det&H788 90' = D00D11 − &wD01'
w = 0.9 Bays Suggestion
Or
det&H788 90' = D00D11 − &0.9D01' (6)
Making use of box filters and integral images, SURF
doesn’t require to recursively apply similar filter to the
output of a prior filtered layer however can apply such
filters of various sizes at similar speeds directly on the
actual image, simultaneously. Therefore, the scale space is
verified by up scaling the filter size (9×9 to 15×15 to
21×21 to 27×27 etc) instead of recursively minimising the
image size. An octave is a series of filter maps achieved by
convolving input image with a filter of increasing size,
consists of a scaling factor of 2 and also subdivided into a
constant number of scale levels. Due to the discrete nature
of integral images, the minimum scale difference between
2 subsequent scales depends on the length l0 of the
positive or negative lobes of the partial second order
derivative in the direction of derivation (x or y), which has
been fixed to a third of the filter size length. For the 9×9
filter, this length is 3. For two successive levels, size is
incremented by a minimum of 2 pixels in order to keep the
size uneven and ensuring presence of the central pixel,
finally resulting in an increased mask size. For localizing
image interest points and over scales, a non maximum
suppression in a 3 × 3 × 3 neighborhood is applied. The
scale space is verified by upscaling the filter size in place
of recursively minimising the image size. The following
layers are then obtained by gradually filtering the image
with larger masks. The biggest advantage of this type of
sampling is its computational efficiency. Also, as we do not
have to downsample the image, there is no aliasing. On the
flipside, box filters tend to protect HF components are lost
in magnified scenes, thereby limiting scale invariance.
SURF descriptor creation is done normally as a two step
mechanism, where first step is to fix a orientation on the
basis of information from a circular region around the
keypoint and then a region is constructed and aligned to
the chosen orientation to extract the SURF descriptor. For
rotational invariance, SURF identifies an orientation for
the interest points. For achieving, SURF involves
estimation of the response of Haar wavelets in xy
directions. Besides this, the sampling step is totally scale
dependent and is finally chosen to be s, with the wavelet
responses being estimated at that current scale s.
Similarly, increased scaling of the size of the wavelets is
also substantial. Hence for this reason, the integral images
are again employed for fast filtering. Then finally, the sum
of vertical and horizontal wavelet responses is estimated
in a scanning area, leading to change in the scanning
orientation, and then again recalculating, till the time
orientation with largest sum value is found, which in turn
is the main orientation of feature descriptor. However, this
step employs extra load of computation on the resources
especially which are available on applications like offroad
vehicles and UAVs in certain cases. Hence in order to
further optimizing the SURF algorithm and in order to
reduce the complexity for realtime applications, a

curtailed version of SURF termed as USURF can be
employed. The authors in [35] proposed an upright
version of the descriptor that is image rotation invariant
resulting in faster computation and better suited for
applications where the camera remains more or less
horizontal Hence, the modified SURF also called U-SURF
(or Upright SURF) has been found to be suited better for
realtime applications for discussed applications.
5. Conclusions
Motion estimation and Digital Video Stabilization has been
achieved using multiple mechanisms. However, for
Realtime performance, considerable success has been
performance enhancement has been achieved with the use
of Feature based motion estimation mechanisms like
SURF. We have seen that systems, and more specifically
embedded systems, have shown better performance in
terms of better outputs, frame processing speeds and
utilization of resources.
SURF as a descriptor has shown a good tradeoff between
accuracy and speed. Most of the other descriptors
including SIFT had better accuracies than SURF , however
still we preferred SURF over the others because of the
very fact speed was more important for us to ensure
realtime functionality. An even faster variant of SURF i.e U-
SURF can be made use of to further enhance the speeds
taking into the account the fact that most of the
application discussed in this paper have more significant
translational motions than rotational motions. The
architecture of the real time embedded system can be
further optimized for even quicker performance by using
pipelining of different stabilization system blocks and can
facilitate the input frames to be fed and processed block by
block.
References
[1] Mostafa Attaran , Dawood Seyed Javan “Motion
Estimation Using the Gradient Method by Genetic
Algorithm” IEEE International Conference on
Computer Modelling and Simulation, 2010,
[2] Peter O’Donovan “Optical Flow: Techniques and
Applications” 20053. van Leeuwen, J. (ed.):
Computer Science Today. Recent Trends and
Developments. Lecture Notes in Computer Science,
Vol. 1000. Springer-Verlag, Berlin Heidelberg New
York, 1995
[3] Z.E Baarir, F. Charif “Fast modified Horn & Schunck
method for the estimation of optical flow fields”
IEEE Workshop on Signal Processing Systems (SiPS),
2011
[4] T. Kondo, P. Boonsieng and W. Kongprawechnon,
"Improved gradient-based methods for motion
estimation in image sequences," 2008 SICE Annual
Conference, Tokyo, 2008, pp. 1120-1123, doi:
10.1109/SICE.2008.4654826.
[5] Boonsieng, Pramuk & Kondo, Toshiaki &
Kongprawechnon, Waree. (2014). A robust optical
flow estimation technique using gradient
orientations. Science Asia. 40. 73.
10.2306/scienceasia1513-1874.2014.40.073, 2014
[6] W. Xu, X. Lai, D. Xu, N. A. Tsoligkas, "An Integrated
New Scheme for Digital Video Stabilization",
Advances in Multimedia, vol. 2013, Article ID
651650, 8 pages, 2013.
https://doi.org/10.1155/2013/651650
[7] A. N. Netravali and J. D. Robbins, "Motion
compensated television coding: Part I", Bell Syst.
Tech. J., vol. 58, pp. 631-670, Mar. 1979
[8] W. Hassen and H. Amiri, "Block Matching
Algorithms for motion estimation," 2013 7th IEEE
International Conference on e-Learning in Industrial
Electronics (ICELIE), Vienna, 2013, pp. 136-139, doi:
10.1109/ICELIE.2013.6701287.
[9] H. A. Choudhury and M. Saikia, "Survey on block
matching algorithms for motion estimation," 2014
International Conference on Communication and
Signal Processing, Melmaruvathur, 2014, pp. 036-
040, doi: 10.1109/ICCSP.2014.6949794
[10] H. Chen, J. Ding and Y. Lee, "Novel Adaptive Rood
Path Searches with Small Motion Prejudgments for
Fast Block Motion Estimation," 2018 IEEE Asia
Pacific Conference on Circuits and Systems (APCCAS),
Chengdu, 2018, pp. 560-563, doi:
10.1109/APCCAS.2018.8605600.
[11] W. Lei, M. Shi and G. Wang, "Research and Verilog
HDL implementation of technology for motion
estimation based on new three — step method,"
2017 IEEE 2nd Information Technology, Networking,
Electronic and Automation Control Conference
(ITNEC), Chengdu, 2017, pp. 1715-1719, doi:
10.1109/ITNEC.2017.8285089.
[12] L. Garlin Delphina and VPS Naidu, “Implementation
and Validation of Video Stabilization using
Simulink” In: NACES-11, 21-23 Sep. 2011, MSRIT,
Bangalore,2011
[13] Y. Wang, Q. Huang, D. Zhang and Y. Chen, "Digital
Video Stabilization Based on Block Motion
Estimation," in 2017 International Conference on
Computer Technology, Electronics and
Communication (ICCTEC), Dalian, China, 2017 pp.
894-897. doi: 10.1109/ICCTEC.2017.00198
[14] I. Ali, G. Raja, M. Muzammil and A. K. Khan,
"Adaptive Modified Hexagon Based Search Motion
Estimation algorithm," 2014 IEEE Fourth
International Conference on Consumer Electronics
Berlin (ICCE-Berlin), Berlin, 2014, pp. 147-148, doi:
10.1109/ICCE-Berlin.2014.7034206
[15] H. Chen, J. Ding and Y. Lee, "Novel Adaptive Rood
Path Searches with Small Motion Prejudgments for
Fast Block Motion Estimation," 2018 IEEE Asia
Pacific Conference on Circuits and Systems (APCCAS),
Chengdu, 2018, pp. 560-563, doi:
10.1109/APCCAS.2018.8605600.
[16] Jifei Song, Xiaohong Ma , "A novel real-time digital
video stabilization algorithm based on the
improved diamond search and modified Kalman
filter," 2015 IEEE 7th International Conference on
Awareness Science and Technology (iCAST),
Qinhuangdao, 2015, pp. 91-95, doi:
10.1109/ICAwST.2015.7314026.
[17] Mingzhou Song and Jing-ao Sun, “Motion Estimation
in DCT Domain”, IEEE Transactions on
Communication, Vol. 25, pp. 1004 - 1009, Aug. 2003

[18] Y. Keller, A. Averbuch, and M. Israeli,
“Pseudopolarbased estimation of large translations,
rotations and scalings in images,” IEEE Trans. Image
Processing, vol. 14, no. 1, pp. 12–22, 2005.
[19] Hanzhou Liu, Baolong Guo and Zongzhe Feng,
"Pseudo-log-polar Fourier transform for image
registration," in IEEE Signal Processing Letters, vol.
13, no. 1, pp. 17-20, Jan. 2006, doi:
10.1109/LSP.2005.860549
[20] A. Briassouli and N. Ahuja, "Integration of frequency
and space for multiple motion estimation and
shape-independent object segmentation," in IEEE
Transactions on Circuits and Systems for Video
Technology, vol. 18, no. 5, pp. 657-669, May 2008,
doi: 10.1109/TCSVT.2008.918799
[21] A. Briassouli and N. Ahuja, "Integration of frequency
and space for multiple motion estimation and
shape-independent object segmentation," in IEEE
Transactions on Circuits and Systems for Video
Technology, vol. 18, no. 5, pp. 657-669, May 2008,
doi: 10.1109/TCSVT.2008.918799.
[22] Y. Wang, Q. Huang, D. Zhang and Y. Chen, "Digital
Video Stabilization Based on Block Motion
Estimation," 2017 International Conference on
Computer Technology, Electronics and
Communication (ICCTEC), Dalian, China, 2017, pp.
894-897, doi: 10.1109/ICCTEC.2017.00198.
[23] 23.Babagholami-Mohamadabadi B., Jourabloo A.,
Manzuri-Shalmani M.T. (2012) A Robust Global
Motion Estimation for Digital Video Stabilization. In:
Thielscher M., Zhang D. (eds) AI 2012: Advances in
Artificial Intelligence. AI 2012. Lecture Notes in
Computer Science, vol 7691. Springer, Berlin,
Heidelberg. https://doi.org/10.1007/978-3-642-
35101-3_12
[24] Shen, Yao & Guturu, Parthasarathy & Damarla,
Thyagaraju & Buckles, Bill & Namuduri, Kamesh.
(2009). Video Stabilization Using Principal
Component Analysis and Scale Invariant Feature
Transform in Particle Filter Framework. IEEE
Transactions on Consumer Electronics, 55(3). 1714 -
1721. 10.1109/TCE.2009.5278047.
[25] Chenguang, Guo & Xianglong, Li & Linfeng, Zhong &
Xiang, Luo. (2009). A Fast and Accurate Corner
Detector Based on Harris Algorithm. 3rd
International Symposium on Intelligent Information
Technology Application, IITA 2009. 2.
10.1109/IITA.2009.311.
[26] J. Li, T. Xu and K. Zhang, "Real Time Feature-Based
Video Stabilization on FPGA," in IEEE Transactions
on Circuits and Systems for Video Technology, vol. 27,
no. 4, pp. 907-919, April 2017, doi:
10.1109/TCSVT.2016.2515238.
[27] L. Araneda and M. Figueroa, "Real-Time Digital
Video Stabilization on an FPGA," 2014 17th
Euromicro Conference on Digital System Design,
Verona, 2014, pp. 90-97, doi: 10.1109/DSD.2014.26
[28] Vazquez, M., Chang, C.: Real-time video smoothing
for small rc helicopters. In: Systems, Man and
Cybernetics, 2009. SMC 2009. IEEE International
Conference on, IEEE, pp. 4019–4024 (2009)
[29] Wang Y, Hou Z, Leman K, Chang R (2011) Real-time
video stabilization for unmanned aerial vehicles.
IAPR Conference on Machine Vision Applications,
June 13-15, 2011, Nara, JAPAN, 2011 pp 336–339
[30] J. Li, T. Xu and K. Zhang, "Real-Time Feature-Based
no. 4, pp. 907-919, April 2017, doi:
10.1109/TCSVT.2016.2515238
[31] Raut, S., Shimasaki, K., Singh, S. et al. Real-time high-
resolution video stabilization using high-frame-rate
jitter sensing. Robomech J 6, 16 (2019).
https://doi.org/10.1186/s40648-019-0144-z
[32] Z. Zhao and X. Ma, "Video stabilization based on
local trajectories and robust mesh transformation,"
2016 IEEE International Conference on Image
Processing (ICIP), Phoenix, AZ, 2016, pp. 4092-4096,
doi: 10.1109/ICIP.2016.7533129.
[33] Y. H. Chen, H. Y. S. Lin and C. W. Su, "Full-Frame
Video Stabilization via SIFT Feature Matching,"
2014 Tenth International Conference on Intelligent
Information Hiding and Multimedia Signal
Processing, Kitakyushu, 2014, pp. 361-364, doi:
10.1109/IIH- MSP.2014.96.
[34] J. Li, T. Xu and K. Zhang, "Real-Time Feature-Based
no. 4, pp. 907-919, April 2017, doi:
10.1109/TCSVT.2016.2515238
[35] Herbert Bay, Andreas Ess, Tinne Tuytelaars, Luc
Van Gool, “Speeded-Up Robust Features (SURF)”,
Computer Vision and Image Understanding, Volume
110, Issue 3, 2008, Pages 346-359, ISSN 1077-3142,
https://doi.org/10.1016/j.cviu.2007.09.014
[36] L. M. Abdullah, N. Md Tahir and M. Samad, "Video
stabilization based on point feature matching
technique," 2012 IEEE Control and System Graduate
Research Colloquium, Shah Alam, Selangor, 2012, pp.
303-307, doi: 10.1109/ICSGRC.2012.6287181.
[37] W. -x. Jin, X. -g. Di and S. -w. Fu, "A fast digital image
stabilization based on large-angle rotational motion
estimation," Proceeding of the 11th World Congress
on Intelligent Control and Automation, Shenyang,
2014, pp. 5615-5620, doi:
10.1109/WCICA.2014.7053676
[38] Chang Li and Yangke Liu, "Global motion estimation
based on SIFT feature match for digital image
stabilization," Proceedings of 2011 International
Conference on Computer Science and Network
Technology, Harbin, 2011, pp. 2264-2267, doi:
10.1109/ICCSNT.2011.6182425.
[39] Lim, A., Ramesh, B., Yang, Y. et al. Real-time optical
flow-based video stabilization for unmanned aerial
vehicles. J Real-Time Image Proc 16, 1975–1985
(2019). https://doi.org/10.1007/s11554-017-
0699-y

[40] M. Fang, H. Li and S. Si, "A video stabilization
algorithm based on affine SIFT," 2018 International
Conference on Computing, Mathematics and
Engineering Technologies (iCoMET), Sukkur, 2018,
pp. 1-4, doi: 10.1109/ICOMET.2018.8346332.
[41] M. Sharif, S. Khan, T. Saba, M. Raza and A. Rehman,
"Improved Video Stabilization using SIFT-Log Polar
Technique for Unmanned Aerial Vehicles," 2019
International Conference on Computer and
Information Sciences (ICCIS), Sakaka, Saudi Arabia,
2019, pp. 1-7, doi: 10.1109/ICCISci.2019.8716427.
[42] Huan Yu and Wenhui Zhang, "Moving camera video
stabilization based on Kalman filter and least
squares fitting," Proceeding of the 11th World
Congress on Intelligent Control and Automation,
Shenyang, 2014, pp. 5956-5961, doi:
10.1109/WCICA.2014.7053740.
[43] Shen, Yao & Guturu, Parthasarathy & Damarla,
Thyagaraju & Buckles, Bill & Namuduri, Kamesh.
(2009). Video Stabilization Using Principal
Component Analysis and Scale Invariant Feature
Transform in Particle Filter Framework. Consumer
Electronics, IEEE Transactions on. 55(3). 1714 -
1721. 10.1109/TCE.2009.5278047.
[44] Dong, Jing & Liu, HaiBo. (2016). Video Stabilization
for Strict Real-Time Applications. IEEE Transactions
on Circuits and Systems for Video Technology. 27. 1-
1. 10.1109/TCSVT.2016.2589860.
[45] Yang, Junlan & Schonfeld, Dan & Mohamed, Magdi.
(2009). Robust Video Stabilization Based on
Particle Filter Tracking of Projected Camera Motion.
IEEE Transactions on Circuits and Systems for Video
Technology. 19. 945 - 954.
10.1109/TCSVT.2009.2020252.
[46] Gao Chunxian, Zeng Zhe, Liu Hui, "Hybrid Video
Stabilization for Mobile Vehicle Detection on SURF
in Aerial Surveillance", Discrete Dynamics in Nature
and Society, vol. 2015, Article ID 357191, 12 pages,
2015. https://doi.org/10.1155/2015/357191
[47] N. Ejaz, W. Kim, S. I. Kwon and S. W. Baik, "Video
Stabilization by Detecting Intentional and
Unintentional Camera Motions," 2012 Third
International Conference on Intelligent Systems
Modelling and Simulation, Kota Kinabalu, 2012, pp.
312-316, doi: 10.1109/ISMS.2012.73.
[48] Chao Zhanga, Fugen Zhoua, Bindang Xuea, Wenfang
Xueb, “Stabilization of atmospheric turbulence-
distorted video containing moving objects using the
monogenic signal ”Journal of Signal Processing:
Image Communication, Elsevier, Pages 19-29,
Volume 63, April 2018.
[49] Xie, Q. & Chen, Xi & Zhang, L. & Jiang, A. & Cui, F.
(2016). A novel rapid and efficient video
stabilization algorithm for mobile platforms. 1-4.
10.1109/VCIP.2016.7805487.
[50] C. Fang, T. Tsai and C. Chang, "Video stabilization
with local rotational motion model," 2012 IEEE Asia
Pacific Conference on Circuits and Systems,
Kaohsiung, 2012, pp. 551-554, doi:
10.1109/APCCAS.2012.6419094.
[51] M. Wang et al, "Deep Online Video Stabilization
With Multi-Grid Warping Transformation
Learning," in IEEE Transactions on Image
Processing, vol. 28, no. 5, pp. 2283-2292, May 2019,
doi: 10.1109/TIP.2018.2884280.
[52] Carlo Tomasi, Takeo Kanade, Detection and
Tracking of Point Features, School of Computer
Science, Carnegie Mellon Univ. Pittsburgh, 1991.
[53] Javier Sánchez, Comparison of motion smoothing
strategies for video stabilization using parametric
models, Image Process. Online 7 (2017) 309–346.
[54] Bing-Yu Chen, Ken-Yi Lee, Wei-Ting Huang, Jong-
Shan Lin, Capturing intention based full-frame
video stabilization, Comput. Graph. Forum 27 (7)
(2008) 1805–1814.
[55] Manish Okade, Prabir Kumar Biswas, Video
stabilization using maximally stable extremal
region features, Multimedia Tools Appl. 68 (3)
(2014) 947–968.
[56] Semi Jeon, Inhye Yoon, Jinbeum Jang, Seungji Yang,
Jisung Kim, Joonki Paik, Robust video stabilization
using particle keypoint update and l1-optimize
camera path, in: Sensors, Sensors 17 (2) (2017)
337.
Table 1 Comparative analysis of Motion estimation Techniques and the relevant work in DVS based Stabilization
Author(s)
Motion Estimation
Method
Motion
Model
USP Limitations
Jifei Song ,X.
Ma[16]
Block
Based(Diamond
Search)
2D
Incorporating Hooke Jeeves into
traditional DS
Makes use of Bilinear
interpolation, thereby
making it time consuming
and slow
Y. Wang, Q.
Huang[13]
Block Motion
Estimation
2D
Use of SAD in place of MSE, SSD
reduces computational intensity.
Reduced picture resolution
due to Low pass filtering , No
performance metrics
evaluated
W. -x. Jin, X. -g.
Di[37]
Block Motion
Estimation
2D
Bit Plane Matching, Performs well
even for large angle rotational
motion, Speeds upto 29 fps
Estimation of translational
after rotational motion
makes it complicated
Chang Li and
Yangke Liu
[38]
Feature Based 2D
Block based key point extraction for
SIFT to reduce complexity
Decreased Accuracy due to
use of blocks, Below average
performance for scenes with
moving scenes

Lim, A.,
Ramesh, B.,
Yang[39]
Feature Based 2D
Processing Speed of 50 fps, Truly
realtime, Multithreaded Work Flow
Performance drops in scenes
with constant motion
J. Li, T. Xu and
K. Zhang[26]
Feature Based
(Oriented FAST and
Rotated BRIEF and
RANSAC integrated)
2D
FPGA based implementation,
Pipelined Architecture
Stringent hardware
requirements for embedded
Processors
M. Fang, H. Li
and S. Si[40]
Feature
Based(Affine SIFT)
2D
Complete Affine Variance as a result
of usage of ASIFT
Use of Affine SIFT and
Matsushita Gaussian Filter
makes it non real time.
M. Sharif, S.
Khan[41]
Feature Based (SIFT
with Log Transform
2D Rotation invariance
Requires Storing of Video
Sequence, Not realtime
Babagholami-
Mohamadabadi
B.,
Jourabloo[23]
Frequency
Transform
based(2D Radon
transform, 1D
Fourier transform
and 1D Scale
transform)
2D
Rotational Invariance upto +-15
degrees
Doesn’t deal with affine
models, Frequency domain
issues render it useless for
real time
Huan Yu and
Wenhui
Zhang[42]
Global Motion
estimation, Corner
matching by DS
Method
2D
Enhanced Kalman filtering for
motion correction stage
Diamond search for corner
matching, complicated
kalman filtering stage
Yao Shen,
Parthasarathy
Guturu[43]
Feature Based (PCA
-SIFT)
2D
Use of the SIFT-BMSE for particle
weight determination further
improves its speed and accuracy
Requires adaptively
changing particle size to
reduce the complexity
W. Xu,X. Lai[6]
Gradient
Descent(Optical
Flow motion
estimation
technique)
2D
Use of Savitzky-Golay filter in place
of Kalman filters for processing
estimated affine global motion
parameters,
Not realtime
L. Garlin
Delphina and
VPS Naidu[12]
Block Based Motion
Estimation (TSS
along with GCBP)
2D
Good PSNR and ITF values for
Bloack based mechanisms
GCBP Computation is slow
and also assumes within 4
chosen sub images as
uniform
Jing Dong,
Haibo Liu[44]
Homography
Estimation
2D Truly realtime performance
Homographies are employed
to represent the video
motion over a short time
interval only. Hence assumes
virtually no camera
translation within a short
timeframe
Yang, Junlan &
Schonfeld[45]
SIFT 2D
Particle Filtering for Global Motion
Estimation Between Successive
Frames
Use of SIFT reduces realtime
capabilities
Wang Y, Hou Z,
Leman K[29]
FAST Corner
Detection
2D Realtime
Max performance till 30 fps,
keypoint matching is based
on grey image only
Gao Chunxian,
Zeng Zhe[46]
SURF 2D
Tradeoff between accuracy and
Speed
Computational overhead due
to lack of pipelining and
multithreading in design
architecture
Raut, S.,
Shimasaki[31]
2D
Completely Realtime, Works on HR
Images
For Grayscale images
N. Ejaz, W. Kim
[47]
Optical Flow based
motion esti mation
2D
Use of Image morphing to
reconstruct jittered frames
Low PSNR values
Chao Zhanga,
Fugen
Zhoua[48]
Frequency Based 2D
Generating stable background
frames using the monogenic signal
Not realtime due to presence
of complex transforms
Z. Zhao and X.
Ma[32]
Feature points
extraction (KLT)
2D
Feature points by Kanade Lucas
Tracker and Outlier removal by
RANSAC
Very low processing speed of
8 fps

Xie, Q. &
Chen[49]
Global motion
estimation(K Means
Clustering)
2D
Applicable to handheld devices with
optimal resources
Coarse to fine global motion
estimation method based on
the tiny block matching
suffers from computation
cost and complex motion
vector classification
A. J. Amiri and
H. Moradi[8]
SURF 2D
Use of BRISK as detectors along
with RANSAC
15 fps only
C. Fang, T.
Tsai[50]
Block Matching
Estimation
2D
Use of rotation-based block
matching method for local motion
estimation
Not realtime
Y. H. Chen, H. Y.
S. Lin[33]
SIFT 2D Full frame Stabilization Not realtime
M. Wang[51]
Deep Neural
Networks
2D Development of StabNet Database
Frames need to be buffered
and stored
L. Araneda and
M.
Figueroa[27]
Global Motion
Estimation
2D Fast and realtime
Better MSE results can be
obtained

Embedded Implementations of Real Time Video Stabilization Mechanisms A Comprehensive Review

More Related Content

What's hot (18)

Similar to Embedded Implementations of Real Time Video Stabilization Mechanisms A Comprehensive Review (20)

More from YogeshIJTSRD (20)

Recently uploaded (20)

Embedded Implementations of Real Time Video Stabilization Mechanisms A Comprehensive Review