optimal projection of 2-d displacements for 3-d ...tziritas/papers/motionivc.pdf · optimal...
TRANSCRIPT
UNCORRECTED PROOF
Optimal projection of 2-D displacements for 3-D translational
motion estimation
Christophe Garcia, Georgios Tziritas*
Department of Computer Science, University of Crete, P.O. Box 2208, Heraklion, Greece
Abstract
Recovering 3-D motion parameters from 2-D displacements is a difficult task, given the influence of noise contained in these data, which
correspond at best to a crude approximation of the real motion field. Stability for the system of equations to solve is therefore essential. In this
paper, we present a novel method based on an unbiased estimator that aims at enhancing this stability and strongly reduces the influence of
noise contamination. Experimental results using synthetic and real optical flows are presented to demonstrate the effectiveness of our method
in comparison to a set of selected methods. q 2002 Published by Elsevier Science B.V.
Keywords: Motion estimation; Optical flow; Unbiased estimator
1. Introduction
The estimation of 3-D motion parameters from a
sequence of images is a fundamental task in computer
vision research with numerous applications, such as
egomotion and time-to-contact estimation for mobile robots
[12], video segmentation [4], depth layering [11], or more
generally 3-D scene reconstruction. Most methods for 3-D
motion analysis begin by extracting two-dimensional
motion information. Many algorithms have been proposed
for extracting 3-D motion parameters from optical flow. A
detailed review is proposed by Heeger and Jepson in Ref.
[9]. The pioneering work of Prazdny [15] assumes that
surfaces in the viewed scene are smooth and solves for
rotation, at high computational cost, using a set of nonlinear
equations that are independent of translation. Bruss and
Horn [5] propose a global approach that combines
information in the entire visual field to choose the 3-D
motion and structure that fits the flow field best in the least
squares sense. Adiv [1] minimizes the same residual
function as Horn and Bruss but locally in patches under
the assumption of planarity. Heeger and Jepson [9]
minimize also the same residual function but depth and
rotation parameters are eliminated in order to obtain a
measure of error as a function of translation which is then
analyzed to select the correct translation. Lobo and Tsotsos
[12] propose a voting scheme based on triplets of points
using the Collinear Point Constraint for cancelling rotation
and finding the focus of expansion. Daniilidis [6] makes use
of fixation on a scene point and projection of the spherical
motion field on two latitudinal directions to decouple the
motion parameter space, searching then along meridians of
the image sphere.
One main problem in correctly estimating the camera
motion parameters is the fact that the 2-D motion field
usually contains a set of noisy and partially incorrect data
(outliers), making most of the above mentioned methods
unstable. The set of incorrect data can be even larger if
independent motions exist throughout the image sequence.
The negative effects of this set of outliers on motion
estimation increase with the complexity of the motion
model which is used to describe the camera motion.
Komodakis and Tziritas [11] proposed a robust estimation
method to cope with the set of outliers and the use of a
hierarchy of motion models, where simplest models were
first tested, and then more complex models were considered.
In this paper, we focus on improving the motion parameter
estimation in the case of translational motion. Section 2
describes the equations linking the projected 2-D motions
and 3-D motions inside the image sequence, which yields an
overdetermined system of linear equations in the translation
case. In Section 3, we propose a survey of the methods
devised to solve these overdetermined systems, and select
some of them according to criteria of processing times, for
comparison to the method proposed in this paper. Section 4
presents our approach, based on the projection of the
equations’ coefficients into a different space, chosen
appropriately in order to reduce the influence of noise
0262-8856/02/$ - see front matter q 2002 Published by Elsevier Science B.V.
PII: S0 26 2 -8 85 6 (0 2) 00 0 88 -4
Image and Vision Computing xx (0000) xxx–xxx
www.elsevier.com/locate/imavis
* Corresponding author. Tel.: þ30-810-39-3136; fax: þ30-810-39-3501.
E-mail addresses: [email protected] (G. Tziritas), [email protected].
gr (C. Garcia).
IMAVIS 1900—22/7/2002—17:43—DMESSENGER—51394— MODEL 5
ARTICLE IN PRESS
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
UNCORRECTED PROOF
contamination. To do so, a noise model of optical flow is
proposed according to previous works. In Section 5,
experimental results using synthetic noisy optical flows
are analyzed in order to compare the selected methods and
to show the superiority of our approach. Experimental
results using real optical flow are also presented. Finally,
conclusions are drawn.
2. 3-D motion parameters from 2-D displacements
2.1. Optical flow
Consider a 3-D coordinate system OðX; Y ; ZÞ at the
optical center of a pinhole camera of focal length f, such that
the axis OZ coincides with the optical axis, as shown in Fig.
1.
Suppose that the camera is moving rigidly with respect to
its 3-D static environment with simultaneous 3-D transla-
tional motion ðTx; Ty; TzÞ and 3-D rotational motion
ðVx;Vy;VzÞ: For a point PðX; Y ; ZÞ; the velocity com-
ponents are given by:
X0 ¼dX
dt¼ 2Tx 2 ZVy þ YVz ð1Þ
Y 0 ¼dY
dt¼ 2Ty þ ZVx 2 XVz ð2Þ
Z 0 ¼dZ
dt¼ 2Tz 2 YVx þ XVy ð3Þ
Under perspective projection, a point PðX; Y ; ZÞ is projected
at pðx; yÞ onto the camera retina with:
x ¼Xf
Zand y ¼
Yf
Zð4Þ
Therefore, the 2-D retinal velocity field or optical flow ðu; vÞ
is:
u ¼dx
dt¼
X0f
Z2
XfZ 0
Z2¼
Y 0f
Z2 x
Z
Z 0ð5Þ
v ¼dy
dt¼
Y 0f
Z2
YfZ 0
Z2¼
Y 0f
Z2 y
Z 0
Zð6Þ
It yields:
u ¼2Txf þ xTz
ZþVx
xy
f2Vy
x2
fþ f
( )þVzx ð7Þ
v ¼2Tyf þ yTz
ZþVx
y2
fþ f
( )2Vy
xy
f2Vzx ð8Þ
Eqs. (7) and (8) describe a 2-D velocity field, which relates
the 3-D motion of points to their projected 2-D motion on
the image plane. By observing these two equations, one may
notice that (i) the effect of translational and rotational
components are separable, (ii) the vectors defined by the
translational components lies on lines going through the
point ðTxf=Tz; Tyf=TzÞ; which is called focus of expansion
(FOE), (iii) the rotational component of motion is
independent of scene structure, since the depth Z influences
the translational component only.
By eliminating Z from the motion field Eqs. (7) and (8)
and after some algebra, we obtain:
ðTyVy þ TzVzÞx2 þ ðTxVx þ TzVzÞy
2 2 ðTxVz þ TzVxÞxf
2 ðTyVz þ TzVyÞyf 2 ðTxVy þ TyVxÞxy þ ðTxVx
þ TyVyÞf2 þ Tyuf 2 Txvf þ Tzðxv 2 yuÞ
¼ 0 ð9Þ
Eq. (9) is difficult to solve in the general case, given the
products of terms from ðTx; Ty; TzÞ by terms from
ðVx;Vy;VzÞ: Some authors, for instance Gupta et al. [8],
try to solve the problem by differentiating this equation with
respect to x and y and solve for subsets of the basic motion
parameters using Least Squares methods. Flow derivatives
are involved which make the method even more sensitive to
the original noise contained in the optical flow. If the camera
motion is considered to be only translational, i.e.
ðVx;Vy;VzÞ ¼ ð0; 0; 0Þ; Eq. (9) may be rewritten as:
2Txvf þ Tyuf þ Tzðxv 2 yuÞ ¼ 0 ð10Þ
By writing Eq. (10) in matrix form, and considering n points
ðn q 3Þ where the optical flow is defined, we obtain a
homogeneous system of n linear equations in 3 variables.
Obviously, it is not possible to estimate the 3-D translation
vector, but only the ratios of the 3-D translation com-
ponents. Ratios over Tz are considered, if Tz is nonzero. The
particular case of Tz ¼ 0 will be considered later on.
In the case of Tz – 0; introducing the notation ða ¼
Txf=Tz; b ¼ Tyf=TzÞ; Eq. (10) becomes:
2v1 u1 x1v12y1u1...
2vi ui xivi2yiui...
2vn un xnvn2ynun
24
35
|fflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflffl}M1
ab1
� |{z}
B1
¼ 0 ð11Þ
The point ða;bÞ is called FOE and corresponds to the point
of intersection of the lines supporting the motion vectors
defined by the translational components. This may be
observed in the optical flow shown in the first line of Fig. 3.
This case of 3-D translation will be referred as full
translation.
In the case of Tz ¼ 0; the FOE is at infinity. Only the
direction of translation may be recovered. This direction is
defined by the ratio g ¼ Ty=Tx (or Tx=Ty). This case of 3-D
motion will be referred as fronto-parallel translation. In that
case, Eq. (10) gives rise to the following system of
IMAVIS 1900—22/7/2002—17:43—DMESSENGER—51394— MODEL 5
C. Garcia, G. Tziritas / Image and Vision Computing xx (0000) xxx–xxx2
ARTICLE IN PRESS
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
UNCORRECTED PROOF
equations:
u1 2v1...
ui 2vi...
un 2vn
24
35
|fflfflffl{zfflfflffl}M1
g1
h i|{z}B1 ¼ 0 ð12Þ
In both cases, the estimation of solution B1 consists in
solving the corresponding overdetermined system, i.e. Eq.
(11) or (12), where all coefficients of M1 are noisy, given
that they depend on u and v. Indeed, the observed optical
flow field is a very crude approximation of the motion field,
whatever method for computing it is used. An interesting
review of optical flow techniques including performance
analysis is presented in Ref. [3].
2.2. Point correspondences
Considering the discrete case where point correspon-
dences have been obtained, let ðx0; y0Þ be, at time t0; the 2-D
point corresponding to ðx; yÞ at time t.
Given that, in 3-D space,
X0
Y 0
Z 0
2664
3775 ¼
X
Y
Z
2664
37752
TX
TY
TZ
2664
3775; ð13Þ
we obtain the relations of image point coordinates
x0 ¼xZ 2 fTX
Z 2 TZ
; y0 ¼yZ 2 fTY
Z 2 TZ
: ð14Þ
By eliminating Z from the above two correspondence
equations, if TZ – 0; we obtain:
x0 2 x
y0 2 y¼
x0 2 a
y0 2 b; ð15Þ
where ða;bÞ can again be interpreted as the FOE of the 2-D
displacement vector field. By symmetry we can also write,
x 2 x0
y 2 y0¼
a2 x
b2 y: ð16Þ
Finally, we obtain one linear equation for each point
correspondence, which is quite similar to Eq. (10) obtained
with the optical flow vector:
2ðy0 2 yÞaþ ðx0 2 xÞb ¼ x0y 2 xy0 ¼ ðx0 2 xÞy 2 ðy0 2 yÞx:
ð17Þ
If we denote u ¼ x0 2 x and v ¼ y0 2 y; Eqs. (10) and (17)
are identical.
When the 3-D translation is parallel to the image plane,
we obtain
ðx0 2 xÞg ¼ y0 2 y; ð18Þ
where g is again the ratio of the two translation components.
3. Existing methods for solving overdetermined systems
Several main techniques have been proposed for solving
overdetermined linear systems. In the following paragraphs,
we will give an overview of these methods and select some
of them according to processing-time criteria as a basis of
comparison with the proposed approach.
3.1. Least squares
The most popular methods are the error minimizing
techniques which formulate a quadratic error function to be
minimized. The simplest and therefore most often used error
minimizing technique is Least Squares (LS). The goal is to
find B1 which minimizes the norm kM1B1k2; and the problem
is reduced to solve the linear system M2B2 ¼ A for B2 such
that kM2B2 2 Ak2 is minimized, with M1 ¼ ½M2l½2A�� and
Bt1 ¼ ½B21�t: The classical least square solution is given by
B2 ¼ ðMT2 M2Þ
21MT2 A: It may be noticed that if the noise is
Gaussian and affects only A, this solution is also the
maximum likelihood estimate of B2:
3.2. Total least squares
Although it offers a simple technique for solving the
Fig. 1. The camera coordinate system.
IMAVIS 1900—22/7/2002—17:43—DMESSENGER—51394— MODEL 5
C. Garcia, G. Tziritas / Image and Vision Computing xx (0000) xxx–xxx 3
ARTICLE IN PRESS
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
UNCORRECTED PROOF
problem, LS provides an unbiased estimate only if M2 is
noise free and all the errors are in A. In our case, it can be
easily observed that measurement errors affect both M2 and
A. In the case of equally distributed errors through the entire
measurement matrix M1; the total least square (TLS)
algorithm aims at solving the overdetermined system of
equations by finding ðDM2;DA) such that ðM2 2 DM2ÞB2 ¼
A 2 DA has an exact solution and kDM2;DAk2 is minimized.
This is performed via classical eigenanalysis on singular
value decomposition (SVD). From the theory of SVD,
solution B1 is known to be identical to the eigenvector of
matrix Mt1M1; corresponding to its smallest eigenvalue [8,
13].
As this method involves the solution of an eigensystem
problem, the stability in the computation of the eigenvalues
and the eigenvectors must be considered. The smallest
eigenvalue is selected and the solution depends on the
corresponding eigenvector. In the case where there are
multiple small eigenvalues, instability appears in the
solution of TLS, given the difficulty to select the correct
small eigenvalue associated with the expected eigenvector.
To solve this problem of instability, in some cases, a so-
called equilibration technique may be performed, which
consists in equilibrating the errors in different terms of the
data matrix M1 [8].
3.3. High positive-breakdown methods
Least-squares-based estimators may be completely
perturbed by a few bad leverage points or vertical outliers
as defined in Ref. [17]. The goal of positive-breakdown
methods is robustness against the possibility of several
unannounced outliers that may have occurred anywhere in
the data. There are several types of high-breakdown robust
methods, in particular the least median of squares (LMedS)
and the M-estimators. An interesting review is given in Ref.
[20].
The least-median-of-squares (LMedS) method of Rous-
seeuw [16] estimates the parameters by solving the
nonlinear minimization problem: min medir2i ; where ri is
the residual error of data i. That is, the estimator must yield
the smallest value for the median of squared residuals
computed for the entire data set. The LMedS method attains
the highest possible break-down value b ! 50%: For least
squares, the break-down value is 1=n ! 0% which means
that a single outlier may contaminate the solution. This
algorithm considers a trial subset of a selected number of
observations and computes the linear fit passing through
them. This procedure is repeated many times, and the fit
with the lowest median of squared residuals is retained. For
small data sets, it is possible to consider all subsets, whereas
for larger data sets many subsets are to be drawn at random.
As we have to deal with very large data sets, this procedure
would be very time-consuming and is not selected in our set
of methods.
Another popular robust technique is the so-called M-
estimators [10], but unlike the LMedS method, it can be
reduced to a weighted least-squares problem. It is used by
Komodakis and Tziritas in Ref. [11]. The M-estimation
problem could be expressed as follows: given a set of data
samples Yi and Xi; where Yi ¼ f ðXi; uÞ þ ri; estimate the
vector of parameters u; ri being the residual error of datum i.
The only underlying assumption is that the noise obeys a
symmetric, independent, identical distribution. The M-
estimators try to reduce the effect of outliers by replacing
the squared residuals r2i ; used in LS, by another function of
the residuals. The M-estimate u is defined as the minimum
of a global error function:
u ¼ arg minX
i
rðriÞ ð19Þ
where r is a symmetric, positive-definite function with a
unique minimum at zero, and is chosen to be subquadratic in
r. Instead of directly solving this problem, we can
implement it as the following iterated reweighted least-
squares problem:
minX
i
vðrðk21Þi Þr2
i ð20Þ
where the superscript k indicates the iteration number. The
weight vðrðk21Þi Þ should be recomputed after each iteration
in order to be used in the next iteration.
In the least squares regression, all data points are
weighted equally with vðriÞ ¼ 1: In robust M-estimation,
the function vðriÞ ¼ CðriÞ=ri provides adaptive weighting,
where CðxÞ ¼ drðxÞ=dx is called the influence function,
measuring the influence of a datum on the value of the
parameter estimate. There are several commonly used
influence functions defining M-estimators which provide
solutions for reducing the influence of ‘gross errors’, like the
Huber, the Cauchy, the Geman–McClure, the Welsch and
the Tukey M-estimators.
In our approach, we selected the Tukey (or biweight)
estimator which has the advantage of even suppressing the
outliers [11]. The Tukey’s M-estimator has the following
weighting function:
vcðrÞ ¼1 2
r
c
� �2( )2
lrl # c
0 lrl . c
8>><>>: ð21Þ
The parameter c in the above function is a scale parameter,
which plays a crucial role in the success of the M-estimator.
A 95% asymptotic efficiency on the standard normal
distribution of Tukey’s biweight estimator function is
obtained with the tuning c ¼ 4:6851 [20]. In order to
handle gross errors with respect to the data, we chose the
parameter as c ¼ c0 medianðlrilÞ; where c0 is a normalizing
constant in the range between 3 and 6 [11].
Among these high positive breakdown methods, we
decided to retain the M-estimators method described above,
that will be referred as RLS.
IMAVIS 1900—22/7/2002—17:43—DMESSENGER—51394— MODEL 5
C. Garcia, G. Tziritas / Image and Vision Computing xx (0000) xxx–xxx4
ARTICLE IN PRESS
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
UNCORRECTED PROOF
4. The proposed approach
All the previous methods, i.e. LS, TLS, RLS, try to solve
for the motion parameters using a large set of equations
where the coefficients are very unstable, given the noise
affecting the optical flow vectors u and v. This is the key
observation that has given rise to our method. Therefore, we
aim at building a minimal set of equations where the
parameters are required to be much more stable. The
coefficients of these equations are searched as optimal
according to criteria derived from the supposed noise model
of the optical flow. They are basically obtained by first
projecting the vector of coefficients of each original
equations into a low dimension space with a chosen basis
of vectors. This scheme greatly reduces the influence of
noise contamination in the new set of equation coefficients.
Moreover, unlike all the above-mentioned methods, our
estimator is designed to be unbiased.
4.1. Noise in optical flow observations
The proposed method is based on the model of the noise
affecting the optical flow data. Two noise models are
considered, both with zero-mean distribution. The case of
mean deviation, or biased motion vector estimates, is
considered separately in a subsequent paragraph, where a
robust technique is introduced. We suppose that the two
components of the motion field u and v are perturbed by
additive zero-mean Gaussian noise. The two noise processes
are assumed to be independent, and each of them is assumed
to be spatially uncorrelated. This last property is not
necessary for obtaining an unbiased estimator, but it is
included for simplifying the variance expressions.
The variance of the noise is supposed to be either
constant or proportional to the squared value of the
corresponding component. This model seems compatible
with the probability distribution of optical flow proposed in
Ref. [18] and the observations made in the review of optical
flow techniques by Barron et al. [3]. Similar noise models
are used in Refs. [7,8,12]. However, it should be noticed that
typical optical flow techniques on real sequences produce
results with an error distribution which has a substantial
number of outliers. We will focus on dealing with this issue.
Considering the proposed noise model, we have:
uðiÞ ¼ mðiÞ þ N1ðiÞ ð22Þ
vðiÞ ¼ nðiÞ þ N2ðiÞ ð23Þ
where i indexes the image points where an optical flow
vector is defined and mðiÞ and nðiÞ are the ideal optical flow
components at point i. When the ‘proportional’ model is
used the noise processes N1 and N2 are such that:
E{N1ðiÞ} ¼ E{N2ðiÞ} ¼ 0
;i – i0; E{NkðiÞNkði0Þ} ¼ 0; for k ¼ 1; 2
;i; ;i0; E{N1ðiÞN2ði0Þ} ¼ 0
E{N21 ðiÞ} ¼ s2m2ðiÞ
E{N22 ðiÞ} ¼ s2n2ðiÞ
We will describe our method first in the case of a 3-D
translation parallel to the image plane (fronto-parallel
translation) and then in the general case of full 3-D
translation.
4.2. Translation parallel to the image plane
We consider the case where the translational motion
along the optical axis is null, i.e. Tz ¼ 0: According to Eqs.
(7) and (8), we can write:
mðiÞ ¼ 2Txf
ZðiÞand nðiÞ ¼ 2
Tyf
ZðiÞð24Þ
Given that the depth ZðiÞ is unknown, we can only solve for
either g ¼ Ty=Tx or g ¼ Tx=Ty: This parameter is related to
the direction of the translation in the image plane, whose
angle to the horizontal axis is given by arctanðTy=TxÞ: We
achieve the estimation of this parameter by projecting the
observed process on a deterministic process eðiÞ that is to be
specified later. This projection will yield:
u1 ¼X
i
uðiÞeðiÞ ¼X
i
mðiÞeðiÞ þX
i
N1ðiÞeðiÞ ð25Þ
v1 ¼X
i
vðiÞeðiÞ ¼X
i
nðiÞeðiÞ þX
i
N2ðiÞeðiÞ ð26Þ
As a consequence of the above assumptions, the mean
values of variables u1 and v1 are:
E{u1} ¼ 2TxfX
i
eðiÞ
ZðiÞand E{v1} ¼ 2Tyf
Xi
eðiÞ
ZðiÞð27Þ
Their variances are given by:
var{u1} ¼ s2X
i
m2ðiÞe2ðiÞ and var{v1} ¼ s2X
i
n2ðiÞe2ðiÞ
ð28Þ
We propose to estimate g ¼ Ty=Tx if u1 . v1; or g ¼ Tx=Ty
otherwise. Without loss of generality, we consider the first
case, and the estimate will be g ¼ v1=u1:We will now consider the choice of the axis of projection
{eðiÞ}: A possible criterion is the maximization of the signal
to noise ratio of the denominator variable, i.e. u1 :
r ¼
Pi
eðiÞ
ZðiÞ
� �2
Pi
e2ðiÞ
Z2ðiÞ
Note that the criterion may be computed for the numerator
as well. This ratio is maximized if eðiÞ ¼ lZðiÞ: As ZðiÞ is
unknown but always positive, we propose to choose the
simplest one, that is, eðiÞ ¼ 1=K; where K is the number of
IMAVIS 1900—22/7/2002—17:43—DMESSENGER—51394— MODEL 5
C. Garcia, G. Tziritas / Image and Vision Computing xx (0000) xxx–xxx 5
ARTICLE IN PRESS
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
UNCORRECTED PROOF
points. The estimate is then given by:
g ¼
PivðiÞPiuðiÞ
ð29Þ
Let us now consider the ideal choice eðiÞ ¼ ZðiÞ=K: We
obtain:
E{u1} ¼ 2Txf ; E{v1} ¼ 2Tyf ð30Þ
var{u1} ¼s2
KT2
x f 2; var{v1} ¼s2
KT2
y f 2 ð31Þ
The last equations show the very important reduction of the
noise disturbance in estimating g; in this ideal case. Indeed,
it is known that under the above conditions the estimator is
unbiased and efficient, with a variance equal to s2=K: In our
case, by selecting eðiÞ ¼ 1=K; the estimator is still unbiased
but with a variance proportional to s2=K with a factor of
1 þ
var1
Z
� �2
1
Z0
� �2
0BBB@
1CCCA;
where Z0 is the mean depth of the scene. Thus, the efficiency
of the estimator depends on the variation of the depth of the
scene with respect to its mean value. If the noise is spatially
correlated, another factor increases the estimate variance.
The stronger the correlation coefficient is, the greater the
value of this factor will be. A very important property is that
our estimator is unbiased and the associated error is
proportional to s2=K:For a constant noise model, in accordance with the
previous approach, it will be necessary to weight the
measurements in order to obtain at each point approxi-
mately the same signal-to-noise ratio. As the signal-to-noise
ratio is now proportional to the real motion vector
magnitude, we suggest that the weight of the measurements
should be the measured motion vector magnitude itself.
Therefore we propose the following estimate:
g ¼
PivðiÞ
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiu2ðiÞ þ v2ðiÞp
PiuðiÞ
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiu2ðiÞ þ v2ðiÞp ð32Þ
4.3. Translation non parallel to the image plane
We consider the general case where the 3-D translation is
not parallel to the image plane ðTz – 0Þ: We aim at
estimating the FOE which is the point ða;bÞ ¼ðTxf=Tz; Tyf=TzÞ in the image plane. According to Eq.
(10), we can write:
nðiÞa2 mðiÞb ¼ nðiÞxðiÞ2 mðiÞyðiÞ
From the overdetermined set of equations with noisy
coefficients computed from the motion field, we propose
to obtain two equations by projecting into two deterministic
fields e1ðiÞ and e2ðiÞ: These two equations are:
v1a2 u1b ¼ w1
v2a2 u2b ¼ w2
where, for k ¼ 1; 2 :
uk ¼X
i
uðiÞekðiÞ; vk ¼X
i
vðiÞekðiÞ; wk
¼X
i
ðvðiÞxðiÞ2 uðiÞyðiÞÞekðiÞ:
We therefore obtain the estimate of the position of the FOE:
ða; bÞ ¼u1w2 2 u2w1
u1v2 2 u2v1
;v2w1 2 v1w2
u1v2 2 u2v1
� �ð33Þ
According to the assumptions on the noise model, we have:
E{u1v2 2 u2v1} ¼
�Xi
mðiÞe1ðiÞ
��Xi
nðiÞe2ðiÞ
�
2
�Xi
mðiÞe2ðiÞ
��Xi
nðiÞe1ðiÞ
�
Let us consider the mean values of the two numerators of
Eq. (33):
E{u1w2 2 u2w1} ¼
�Xi
mðiÞe1ðiÞ
�
£
�Xi
ðnðiÞxðiÞ2 mðiÞyðiÞÞe2ðiÞ
�
2
�Xi
mðiÞe2ðiÞ
�
£
�Xi
ðnðiÞxðiÞ2 mðiÞyðiÞÞe1ðiÞ
�
and
E{v2w1 2 v1w2} ¼
�Xi
nðiÞe2ðiÞ
�
£
�Xi
ðnðiÞxðiÞ2 mðiÞyðiÞÞe1ðiÞ
�
2
�Xi
nðiÞe1ðiÞ
�
£
�Xi
ðnðiÞxðiÞ2 mðiÞyðiÞÞe2ðiÞ
�
We supposeP
xðiÞ ¼P
yðiÞ ¼P
xðiÞyðiÞ ¼ 0: If this is not
the case, xðiÞ and yðiÞ are expressed in a new coordinate
system centered at their centroid and whose orthonormal
axes are the first and second principal axes of the
distribution of the points. Principal component analysis is
used for achieving these estimations. Therefore, if we set
IMAVIS 1900—22/7/2002—17:43—DMESSENGER—51394— MODEL 5
C. Garcia, G. Tziritas / Image and Vision Computing xx (0000) xxx–xxx6
ARTICLE IN PRESS
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
UNCORRECTED PROOF
e1ðiÞ ¼ lxðiÞZðiÞ and e2ðiÞ ¼ lyðiÞZðiÞ; we have:
E{u1v2 2 u2v1} ¼ ðlTzf Þ2X
i
x2ðiÞX
i
y2ðiÞ ð34Þ
E{u1w2 2 u2w1} ¼ ðlTzf Þ2aX
i
x2ðiÞX
i
y2ðiÞ ð35Þ
E{v2w1 2 v1w2} ¼ ðlTzf Þ2bX
i
x2ðiÞX
i
y2ðiÞ ð36Þ
These relations prove that the proposed estimators are
unbiased. Indeed, the quotient (35)/(36) is a and the
quantity (36)/(34) is b: We may also prove that Eqs. (34)
and (35) (idem for Eqs. (36) and (34)) are decorrelated and
that the signal-to-noise ratio for both numerator and
denominator is approximately K, the number of points. As
the depth ZðiÞ is unknown, we propose to choose as basis
e1ðiÞ ¼ xðiÞ and e2ðiÞ ¼ yðiÞ: As in the fronto-parallel
translation case, the effectiveness of this choice depends
on the variation of respect to its mean value, and also on the
spatial noise correlation.
4.4. Robustness against mean deviations
The estimated 2-D motion or optical flow field could be
affected by some systematic errors, that is, errors on the
mean value of the field. In other words, the disturbing noise
may not be zero-mean. A similar effect occurs if the two
noise components are correlated, in which case the
estimates might be biased. In the following we propose a
technique for limiting the effects of mean deviation, or
‘correcting’ the bias of the estimator.
As shown in Sections 4.1–4.3, under some assumptions,
the proposed estimators are unbiased. Therefore if the
obtained estimation does not validate this property, we can
conclude that the noise model was not suitable. The test
should be established on measurements which would not
necessitate either knowledge of the real parameter values or
knowledge of the depth. We propose to measure and test the
angle between the direction of the motion field and the
direction suggested by the estimated FOE. If the noise
model was as assumed, the parameter estimators should be
unbiased, and the average angle should be near zero. If a
significant difference is observed, we can conclude that the
noise might not be zero-mean or inter-component
correlated.
Next we illustrate a way of limiting this kind of noise
effect. We propose to ‘correct’ the motion field by rotating it
according to the observed average deviation. In addition, as
a percentage of outliers might also exist, we propose an
iterative algorithm for correcting the motion field and
rejecting the outliers. We iteratively first estimate the
amount of correction and the scale of acceptable inliers, and
then solve, for the points selected and the corrected motion
vectors, the equations presented in Sections 4.1–4.3. We
give a detailed description of the proposed procedure.
Let us denote by m the iteration index. The average angle
deviation for the full translation case is given by
um ¼1
n
Xni¼1
arctan
£ðxðiÞ2 am21Þum21ðiÞ2 ðyðiÞ2 bm21Þvm21ðiÞffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ððxðiÞ2 am21Þ2 þ ðyðiÞ2 bm21Þ
2Þðu2m21ðiÞ þ v2
m21ðiÞÞq
The correcting angle is a filtered version of the estimated
deviation
fm ¼ ltmum þ ð1 2 ltmÞfm21; f0 ¼ 0; 0 , t , 1:
The motion or optical flow vectors are rotated by fm;
umðiÞ
vmðiÞ
" #¼
cos fm sin fm
2sin fm cos fm
" #uðiÞ
vðiÞ
" #:
The rejection of outliers is based on the residuals of the
depth deviation
rmðiÞ ¼xðiÞ2 am
umðiÞ2
yðiÞ2 bm
vðiÞ:
The scale of the residuals is estimated by
cm ¼ 1:5 mediani{lrmðiÞl}:
Therefore points for which lrmðiÞl . cm are rejected as
outliers.
The resulting iterative algorithm is obvious. At each step,
the average angle deviation and the scale of the residuals are
calculated. Then the outliers are rejected and the motion
field is rotated. Finally the 3-D motion parameters are
estimated according to the equations of Section 4.3. The
procedure stops when the actual average angle deviation is
less than the candidate angle correction. Indeed, the angle
correction is designed to give an increasing sequence of
values which are smaller in absolute value than the
calculated average angle deviation.
5. Experimental results
5.1. Results from simulated realistic data
In order to compare the different methods and to study
the effect of noise on their accuracy, we use synthetic optical
flow fields which are contaminated by different amounts of
noise. The simulated optical flow fields are generated using
range images from the MSU/WSU Range Image Database,
available online at http://www.eexs.wsu.edu/IRL/RID/.
Given appropriate values for the intrinsic parameters of
the simulated camera (focal length and principal point), the
dimensions of the retina and the true parameters for
IMAVIS 1900—22/7/2002—17:43—DMESSENGER—51394— MODEL 5
C. Garcia, G. Tziritas / Image and Vision Computing xx (0000) xxx–xxx 7
ARTICLE IN PRESS
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
UNCORRECTED PROOF
translation, the optical flow is generated by using Eqs. (7)
and (8), and the depth value Z of each point of the range
image.
To simulate a realistic flow field, noise is introduced into
the synthetic optical flow vectors. In Ref. [8], various noise
models and levels have been tried, such as uniform noise in
a range, Gaussian noise with zero mean and specified
variance and Gaussian noise with zero mean and specified
fractional variance. Being similar to the one used in Ref.
[12], we chose the following model, which is compatible
with the basic assumptions described in Section 4:
u ¼ mþ Nð0; bÞ p 0:01 p m; v ¼ nþ Nð0; bÞ p 0:01 p n
ð37Þ
where Nð0; bÞ is a Gaussian random variable with mean 0
and standard deviation b. This noise will be referred as
‘Gaussian noise with mean 0 and standard deviation b%‘.
As noted in Ref. [12], this error model provides the ability to
synthesize errors whose range is similar to that produced by
optical flow estimation techniques. A parameter is also used
to control the percentage of optical flow points affected by
the selected noise.
We performed different experiments using the range
image ‘Blocks’ extracted from MSU/WSU range image
database, which is shown in Fig. 2. The image size is 128 £
128 pixels, the principal point being assumed to be in the
center of the image. We chose a value of 96 for the focal
length of the virtual camera, giving a field of view (FOV) of
67.48. All depths in the range image have been uniformly
scaled to the range [2500, 7000] in order to be viewed by the
virtual camera and produce meaningful optical flow.
The chosen motion parameters are ðTx; Ty; TzÞ ¼
ð45;235; 100Þ in pixels/frame for the full translation case
and ðTx; Ty; TzÞ ¼ ð45;235; 0Þ for the fronto-parallel trans-
lation case. In the case of full translation, the maximum
absolute horizontal velocity field is 5.34 with an average of
1.40. Similar values of 4.10 and 1.35 are found for the
absolute vertical velocity field. The true FOE is (43.2, 2
33.6). In the case of fronto-parallel translation, the
maximum absolute horizontal velocity field is 2.16 with
an average of 1.31. Similar values of 1.68 and 1.03 are
found for the absolute vertical velocity field. The true
fronto-parallel translation direction angle is g ¼ 219:38;i.e. 219:3 þ 180 ¼ 160:78 in the image coordinate system.
Fig. 3 shows the scaled and subsampled ideal synthesized
optical flow and some noisy versions of it. The first row
corresponds to the case of full translation whereas the
second row corresponds to the case of fronto-parallel
translation. The ideal optical flow is displayed in the first
column. The second and third columns show perturbed
optical flows with noise Nð0; 50Þ and Nð0; 100Þ with a
percentage p ¼ 100% of affected pixels.
In the case of full translation, the criterion that we use for
assessing noise tolerance is the error in the FOE estimation.
This error is defined as the angle in degrees between the
vectors ða;b; f Þ and ða; b; f Þ; given by:
ErrFOE ¼ arccosaaþ bbþ f 2ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ða2 þ b2 þ f 2Þða2 þ b2 þ f 2Þ
q0B@
1CA ð38Þ
where ða;bÞ is the true FOE and ða; bÞ is the estimated one.
In the case of fronto-parallel translation, the criterion is
the error in degrees in the direction of the translation in the
image plane. This error is defined as:
Errg ¼ larctanðgÞ2 arctanðgÞl ð39Þ
The methods which have been implemented for comparison
are Least Squares (LS), Total Least Squares (TLS), M-
estimators reweighted least squares with c0 ¼ 4:5 (RLS-4.5)
and the proposed method (PROJ).
These different methods are compared for noise tolerance
by computing ErrFOE or Errg; depending on the type of
translation, versus the noise standard deviation b, varying
from 0 to 100% for p ¼ 100% of affected pixels. For each
noise level, the computed values are average values over 50
runs. Fig. 4(a) plots the values of ErrFOE in the full
translation case and Fig. 4(b) plots the values of Errg in the
fronto-parallel translation case. It can be seen from these
two graphs that our method is far more efficient than the
other three, giving maximum errors of 0.67 in the full
translation case and 0.248 in the fronto-parallel translation
case. As expected, LS is the least tolerant to noise with
respective maximum errors of 24.61 and 16.57. RLS-4.5 is
much more efficient than TLS in the case of full translation.
It is slightly less efficient than TLS in the fronto-parallel
translation case. RLS-4.5 achieves good performance by
rejecting outliers, up to 50% of the data with c0 ¼ 1: In theFig. 2. The range image Blocks used for synthesizing flow field (intensity is
proportional to depth, the closest points being brighter).
IMAVIS 1900—22/7/2002—17:43—DMESSENGER—51394— MODEL 5
C. Garcia, G. Tziritas / Image and Vision Computing xx (0000) xxx–xxx8
ARTICLE IN PRESS
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
UNCORRECTED PROOF
present case, with c0 ¼ 4:5; approximately 20% of outliers
were rejected. On the other hand, being iterative, this
method is more time-consuming.
The proposed method proves to be very tolerant to the
noise model we applied which has been found to be close to
the one affecting real optical flows. Moreover, it offers the
advantage of being very fast and easily implemented, since
it consists primarily of projection and summation. Con-
sidering this particular aspect, TLS is more computationally
expensive, performing singular value decomposition.
Eigenanalysis is also proved to be unstable when the matrix
is ill-conditioned which may happen if the amount of noise
is large.
5.2. Results from real data
The algorithms were applied to the well-known ‘marbled
block’ and ‘flower garden’ sequences, with known ground
truth values. The marbled block sequence was captured by a
robot arm moving in full translation over a textured floor
[14]. Four dark blocks lying on the floor are stationary while
a white block in the middle of the scene is moving
independently. The standard sequence flower garden was
shot from a camera placed on a driving car and corresponds
to a fronto-parallel translation along the horizontal axis Tx
of the camera. The scene contains a tree in the foreground, a
textured garden, and a house in the background. The
marbled block sequence contains many sharp discontinu-
ities in depth and the flower garden sequence presents some
non-textured areas that cause problem for the optical flow
computation, giving rise to an important number of outliers.
Optical flows are computed using the algorithm of Anandan
[2]. They are displayed in Fig. 5.
It may be noticed that these real cases differ from the
synthetic cases because of the presence of strong outliers.
Our method, in the original form (PROJ), has not been
designed explicitly to be optimal in that case. To better
handle the outliers and the failures of the model, the
technique presented in 4.4 is applied. The robust version of
our method is referred as ROB-PROJ.
Table 1 gives the results of the different algorithms on
these two sequences. The proposed method is the most
efficient of the set on these real examples as well, especially
in the fronto-parallel translation case. These results tend to
Fig. 3. Ideal and noisy synthesized optical flows in the two cases of translation.
IMAVIS 1900—22/7/2002—17:43—DMESSENGER—51394— MODEL 5
C. Garcia, G. Tziritas / Image and Vision Computing xx (0000) xxx–xxx 9
ARTICLE IN PRESS
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
UNCORRECTED PROOF
Fig. 4. Comparison of the four methods for noise tolerance in the cases of (a) full translation and (b) fronto-parallel translation.
IMAVIS 1900—22/7/2002—17:43—DMESSENGER—51394— MODEL 5
C. Garcia, G. Tziritas / Image and Vision Computing xx (0000) xxx–xxx10
ARTICLE IN PRESS
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
UNCORRECTED PROOF
show that the assumptions made on the noise model and on
the criteria of selection of the projection bases were
generally valid. As another source for comparison, Danii-
lidis reported a result with ErrFOE ¼ 7:248 for the marbled
block sequence [6].
6. Conclusion
In this paper, we have presented a novel method for
estimating the parameters of translational motion from
optical flow. Our results on synthetic and real optical flows
are more accurate than the other tested methods. This is due
to the fact that our scheme, unlike the other methods, is
based on an unbiased estimator that strongly reduces the
influence of noise contamination in the data. Moreover,
computational requirements are low, making this method
very attractive for fast 3-D translational motion parameter
estimation. In order to handle outlier suppression, an
iterative scheme gives even better results within a few
Fig. 5. Original frames and optical flows for (a) marbled block and (b) flower garden.
IMAVIS 1900—22/7/2002—17:43—DMESSENGER—51394— MODEL 5
C. Garcia, G. Tziritas / Image and Vision Computing xx (0000) xxx–xxx 11
ARTICLE IN PRESS
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
UNCORRECTED PROOF
iterations. We are currently working on the extension of this
method to the general case of 3-D motion.
7. Uncited reference
[19].
References
[1] G. Adiv, Determine three-dimensional motion and structure from
optical flow generated by several moving objects, IEEE Transactions
on pattern Analysis Machine Intelligence 7 (1985) 384–401.
[2] P. Anandan, A computational framework and an algorithm for the
measurement of visual motion generated by several moving objects,
International Journal of Computer Vision 2 (1989) 283–310.
[3] J.L. Barron, D.J. Fleet, S.S. Beauchemin, Performance of optical flow
techniques, International Journal of Computer Vision 12 (1994)
43–77.
[4] P. Bouthemy, E. Francois, Motion segmentation and qualitative
dynamics scene analysis from an image sequence, International
Journal of Computer Vision 10 (1993) 157–182.
[5] A.R. Bruss, B.K.P. Horn, Passive navigation, Computer Vision,
Graphics and Image Processing 21 (1983) 3–20.
[6] K. Daniilidis, Fixation simplifies 3-D motion estimation, Computer
Vision and Image Understanding 68 (1997) 158–169.
[7] A.M. Earnshaw, S. Blostein, The performance of camera translation
direction estimators from optical flow, comparison, and theoretical
limits, IEEE Transactions on Pattern Analysis and Machine
Intelligence 18 (1996) 927–932.
[8] N. Gupta, L. Kanal, 3-D motion estimation from motion field,
Artificial Intelligence 78 (1995) 45–86.
[9] D.J. Heeger, A.D. Jepson, Subspace methods for recovering rigid
motion i: algorithm and implementation, International journal of
Computer Vision 7 (1992) 95–117.
[10] P. Huber, Robust Statistics, Wiley, New York, 1981.
[11] N. Komodakis, G. Tziritas, Robust 3-D motion estimation and depth
layering, Proceedings of International Conference on Digital Signal
Processing, Santorini, Greece (1997) 425–428.
[12] N. Da Voctoria Lobo, J.K. Tsotsos, Computing egomotion and
detecting independent motion from image motion using collinear
points, Computer Vision and Image Processing 64 (1996) 21–52.
[13] M. Muhlich, R. Mester, The role of total least squares in motion
analysis, Proceedings of European Conference on Computer Vision,
Freiburg, Germany (1998) 305–321.
[14] M. Otte, H. Nagel, Estimation of optical flow based on higher-order
spatiotemporal derivatives in interlaced and non-interlaced image
sequences, Artificial Intelligence 78 (1) (1995) 5–43.
[15] K. Prazdny, Estimation and relative depth from optical flow,
Biological Cybernetics 36 (1980) 87–102.
[16] P.J. Rousseeuw, Least median of squares regression, Journal of
American Statistical Assocation 84 (1984) 871–880.
[17] P.J. Rousseeuw, A.M. Leroy, Robust Regression and Outlier
Detection, Wiley-Interscience, New York, 1987.
[18] E. Simoncelli, E. Adelson, D. Heeger, Probability distributions of
optical flow, Proceedings of Computer Vision and Pattern Recog-
nition (1991) 310–315.
[19] G. Tziritas, C. Labit, Motion analysis for image sequence coding,
Elsevier, Amsterdam, 1994.
[20] Z. Zhang, Prmeter estimation techniques: a tutorial with application to
conic fitting, Image and Vision Computing 15 (1997) 59–76.
Table 1
Comparative results on real optical flows
Sequence Marbled block Flower garden
Type Full translation Fronto-parallel Translation
Truth ða;bÞ ¼ ð2777:0; 95:6Þ g ¼ 08
LS ErrFOE ¼ 7:588 Errg ¼ 2:738
TLS ErrFOE ¼ 5:258 Errg ¼ 2:678
RLS-4.5 ErrFOE ¼ 5:428 Errg ¼ 2:638
PROJ ErrFOE ¼ 4:948 Errg ¼ 1:428
ROB-PROJ ErrFOE ¼ 3:658 Errg ¼ 0:928
IMAVIS 1900—22/7/2002—17:43—DMESSENGER—51394— MODEL 5
C. Garcia, G. Tziritas / Image and Vision Computing xx (0000) xxx–xxx12
ARTICLE IN PRESS
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344