26 Acta Electrotechnica et Informatica Vol. 8, No. 3, 2008, 26–30
ISSN 1335-8243 © 2008 FEI TUKE
THE FAST SEARCH MOTION ESTIMATION ALGORITHMS AND THEIR ERRORS
Ján GAMEC, Mária GAMCOVÁ
Department of Electronics and Multimedia Communications
University of Technology Košice
Park Komenského 13, Košice, Slovak Republic
e-mail: jan.gamec@tuke.sk
ABSTRACT
This paper presents the analysis of errors produced by block matching motion estimation methods with lower computational cost
(fast search algorithms). We try to find an answer on the question what measure can be used to determine correctness of searched
motion vectors and which motion vectors are founded as fault (corrupted by noise). We consider the vector fields obtained by using
of full search algorithm (FSA) as lossless data. We take the vectors of movement obtained by using 2D logarithmic (2D log) search
procedure as data corrupted by noise. The results of analysis can help optimize post processing of motion vectors in endeavoring to
reduce computational cost or regularize vector fields.
Keywords: motion estimation, block matching, motion vector
1. INTRODUCTION
The purpose of the motion estimation (ME) and
compensation is reduction of redundancy caused by
interframe correlation of movement objects [1-6].
However, the estimation and coding of movement vectors
should be appropriated to computational costs and bit
rates at the perspective high compression systems. That’s
way is very important relationship between accuracy of
movement estimation and simplicity of the description
vector fields. Better motion estimation means higher space
decorrelation of prediction errors in time area.
The most popular approach is to reduce the number of
search locations by using the assumption of unimodal
error surface in which the matching error decreases
monotonically when the searching location approaches to
the global optimum. However, this assumption is not
usually satisfied, thus resulting in local optimal solution.
Instead of limiting the number of search locations, another
interesting tech-nique aims at reducing computation of
block matching with pixel subsampling, successive
elimination algorithm (SEA).
The above two techniques achieves computation
reduction with or without loss of search performance.
This paper is concentrated on analyze one of often
mentioned method of ME with reduced searching steps -
2Dlog method. The 2Dlog method is analyzed from point
of view of estimation and localization of potentially
possible errors of motion vectors in the searching
procedure.
2. THE REASONS OF ERRORS
The essential condition of the correct displacement
vector finding (for ME methods with reduced searching
steps) is flatness of matching criteria. It means that
function of matching criteria monotonically increases as
we move away from direction of minimum matching
criteria in each of the four quadrants [7].
The example of founded motion vectors (MV’s) from
the image sequence “Railway station” is in fig. 1. The
black arrows are MV’s founded by 2Dlog method for
subblock size BS = 16x16 pel and proposed maximum
displacement d
m
= 13 pel in all directions. The white
arrows are MV’s founded by full search (FS) method with
the same searching parameters and can be seen for the
subblocks in which vectors are different from 2Dlog
MV’s. A darker rectangle in fig. 1 highlights one of
subblocks with mismatch vector (2Dlog-black, FS-white).
Fig. 1 The frame from the image sequence “Railway station”
with motion vectors founded by FS method (white) and 2Dlog
method (black)
All computed values by FS method of matching
criteria (MAD - Mean Absolute Difference) of
highlighted subblock in fig. 1 are plotted in fig. 2. The
number of these values is (2d
m
+ 1)x(2d
m
+ 1), i.e. 27x27
values. It can bee seen that values of matching criteria do
not increase monotony. The position of global matching
criterion (MAD) minimum corresponds with actual MV.
The values of criterion MAD are represented as gray
levels in fig. 2b. The white point in very dark area
indicates the position of minimum. The white arrow is
searched MV by FS method. The positions of searching