Professional Documents
Culture Documents
Y
X
Y
X
Z
f
y
x
mean
where
mean
Z
denotes the average depth of the scene and f the focal length of the camera.
Then it is possible to deduced an affine 2d motion model, which is free from the depth parameter of the scene and
which is able to handle translation, rotation and scaling:
AMAA_200311 11 / 15
(5) ) ) (a , y) (x, m(
by x
x
i i
3 5 4
0 2 1
=
+ +
+ +
=
a a a
a y a a
y
x
&
&
3.2 Regression criterion
Camera motion estimation can be performed as any common regression problem [15], finding the parameters
i i
a ) (
of the model that best matches two images ) (t I and ) 1 ( + t I i.e. such that
(6) ) t X, I( ) 1 t , ) ) (a , X ( m X I( min arg ) (
I x
1..6 i i
=
+ + =
r
i i
a
(Least square formulation)
The main drawback of the least square formulation is that the less conform the data are, the more influent they are,
and they substantially biased the final solution. Robust estimation permits to overcome this drawback.
3.3 Robust formulation
The field of robust statistics enables them to cope with huge errors which do not conform to the model assumptions
[18,15,14]. The principle is to assign weights to data according to their adequation to the model and/or to their
homogeneity with the rest of the dataset. This is achieved using a redescending function ) , ( x for which
influence of outliers tends to zero. is a scale parameter controlling the shape of the function and so the outliers
selection process.
Then the regression criterion is reformulated as:
(7) ) , ) t X, I( ) 1 t , ) ) (a X, ( m X I( ( argmin ) (a
I x
1..6 i i i i
=
+ + =
r
This can be easily converted into an equivalent iteratively re-weighed least square (IRLS) problem ( see [14,18] for
details)
(8) ). w(r, argmin ) (a
I x
i i
r
=
with
) t X, I( ) 1 t , ) ) (a X, ( m X I(
1..6 i i
+ + =
=
r
r
the residuals and
r
r
r
r w
) , (
) , (
=
the weights at the current
estimation step.
The general idea in IRLS techniques is to alternatively and iteratively: estimate a solution, assign weights to the
residuals of this current solution and perform a new least square regression using these new weights.
3.4 Incremental estimation
This criterion is non linear as referring to the parameters to estimate and might be non convex. To handle this
problem, a common solution is to use an incremental minimisation scheme such as in [14,15]. The principle of those
approaches is to suppose that an intermediate solution
k
u is known, then the full displacement u between first and
second image can be decomposed in
k k
du u u + =
. Using this remark the criterion can be linearized at point
k
u X +
and the regression is done to find
k
du . In practice this is achieved by constructing an intermediate image
k
I
warping the image at t+1 with the current motion estimation
k
u and then re-estimating motion between
k
I and the
image at t.
AMAA_200311 12 / 15
3.5 Large motion
To cope with large motions a coarse to fine strategy is used: a pyramid of low-pass filtered and sub-sampled images
is constructed. At each level the image resolution is cut by half and low motion is first estimated at coarse level
using the iterative scheme describe above. The computed flow field is then projected to the next level of the pyramid
(rescaled as appropriate [15]) and the estimation begins again using this projection for initialisation at this level. The
process is repeated until the flow has been computed at the full resolution. As scale and estimation evolve between
each level and even at each step of the algorithm, the parameter of the function is reduced progressively during
the process.
3.6 Results
To illustrate the algorithm results, several experiments have been performed in various situations. Each time the
dominant motion between two successive frames is estimated, the first image is then warped and subtracted to the
second to obtain the image of residuals. This one is then binarised [14] to build the detection map representing areas
where non conform motions are detected. Intermediate results are detailed only for the first sequence, for the others,
only the original image and the detection map are shown.
Experiment on a scene close to the camera:
Fig. 15. First image t Fig. 16. Second image t+1
Fig. 17. Warped image, residuals and detection map
Experiment with far and small targets:
Fig. 18. First image Fig. 19. Detection map
Experiment with a low contrasted pedestrian (in the top left corner):
AMAA_200311 13 / 15
Fig. 20. First image Fig. 21. Detection map
Experiment on a scene with multiple moving objects with different speeds and depths:
Fig. 22. First image Fig. 23. Detection map
The main drawback of the algorithm is the planar hypothesis because large depth ranges in the scene will likely lead
to false alarms. In fact in most of urban situations, the planar assumption is violated. In order to reduce these false
alarms, some other approaches also use the outliers map [21] or the residuals [22]. Otherwise, it is also possible to
build a more complex motion model to improve the compensation [17,20].
4 Conclusion
This paper has put emphasis on the main features of CEA/LETI infrared microbolometer technology. One of the key
points has been to elect a sensitive material made from amorphous silicon that features absolute compatibility with
standard silicon processing. This basic option leads to high performance and low cost infrared imaging systems
particularly suited for large market distribution such as automotive applications. This technology is now
commercialised in mass production by the French company ULIS, while a brand new advanced technological
arrangement has been demonstrated at CEA/LETI. The advent of this second generation of the technology results in
a fivefold performance improvement compared to the current industrial process and NETD of 56 mK obtained from
35m pitch, 160x120 IRFPA has been demonstrated.
The results of detection by image processing techniques on infrared video sequences indicate that the approach
gives better results compared to those obtained with images coming from video cameras working at visible
wavelength. One reason seems to be that IR images present less detail but they are nevertheless textured enough for
motion estimation. From that point they represent a very good alternative to visible wavelength sensors.
References
[1] F. Bertrand, J.L. Tissot, G. Destefanis, "Second generation cooled infrared detectors state of the art and
prospects", 4
th
International workshop on advanced infrared technology and applications, Florence - Italie, 15 -
16 septembre 1997
[2] R.A. Wood, Uncooled thermal imaging with monolithic silicon focal planes, Proceedings of SPIE Infrared
Technology XIX, Vol. 2020, pp322-329, (1993)
AMAA_200311 14 / 15
[3] J.L. Tissot, F. Rothan, C. Vedel, M. Vilain, J.J. Yon, LETI/LIRs amorphous silicon uncooled microbolometer
development, Proceedings of SPIE Infrared Detectors and Focal Plane Arrays V, Vol. 3379, pp139-144,
(1998)
[4] J.P. Chatard, X. Zhang, E. Mottin, J.L. Tissot, Infrared Camera for cAR ICAR a EURIMUS Project for
Driver Vision Enhancement, AMAA yearbook 2002, Sven Krueger, Wolfgang Gessner Eds, pp141-145, ISBN
3-540-43232-9
[5] http://www.save-u.org
[6] C. Hanson, Uncooled thermal imaging at Texas Instruments, Proceedings of SPIE Infrared Technology XIX,
Vol. 2020, pp330-339 (1993)
[7] R. Watton, P.N.J. Dennis, J.P. Gilham, P.A. Manning, M.C.J. Perkins, M.A. Todd, IR bolometer arrays; the
route to uncooled, affordable thermal imaging, Proceedings of SPIE Infrared Technology XIX, Vol. 2020,
pp379-390 (1993)
[8] T. Kanno et al, Uncooled Infrared Focal Plane Array having 128x128 Thermopile Detector Elements,
Proceedings of SPIE Infrared Technology XX, Vol. 2269, pp450-459, (1994)
[9] Tomohiro Ishikawa et al, Low cost 320x240 uncooled IRFPA using conventional silicon IC process,
Proceedings of SPIE Infrared Technology and applications XXV, Vol. 3698, pp556-564, (1999)
[10] P.I. Olden, E.A. Wachter, P.G. Datskos, T. Thundat, R.J. Warmack, Optical and infrared detection using
microcantilevers, Proceedings of SPIE Infrared Technology and Applications XXII, Vol. 2744, pp345-354,
(1996)
[11] J.L. Tissot, J.P. Chatard, E.Mottin, Technical trends in amorphous silicon based uncooled IR focal plane
arrays, Proceedings of SPIE Infrared Technology and Applications XXVIII, Vol. 4820, pp220-226, (2002)
[12] D. Murphy, M. Ray, R. Wyles, J. Asbrock, N. Lum, J. Wyles, C. Hewitt, A. Kennedy, D. Van Lue, High
sensitivity 25m microbolometer FPAs, Proceedings of SPIE Infrared Technology and Applications XXVIII,
Vol. 4820, pp208-219, (2002)
[13] R. Gooch, T. Schimert, Low-cost Wafer-Level Vacuum Packaging for MEMS, MRS bulletin January 2003,
pp55-59
[14] M.J Black.and P. Anandan, The robust estimation of multiple motions: Parametric and piecewise-smooth flow
fields, Computer Vision and Image Understanding, CVIU, 63(1), pp. 75-104, Jan. 1996.
[15] J.M. Odobez and P. Bouthemy, Robust multi-resolution estimation of parametric motion models
Journal of Visual Communication and Image Representation, Vol. 6, N
o
4, pages 348-365, dcembre 1995.
[16] Subhashis Banerjee, Camera models and affine Multiple Views Geometry
http://www.cse.iitd.ernet.in/~suban/vision/affine.pdf
[17] M. Irani, B. Rousso, S. Peleg , Recovery of Ego-Motion Using Region Alignement
IEEE Trans. on Pattern Analysis and Machine Intelligence (PAMI), Vol. 19, No. 3, pp. 268 272, March 1997.
http://www.wisdom.weizmann.ac.il/%7Eirani/abstracts/pami97.html
[18] P.J Rousseuw, A.M Leroy, Robust regression and outlier detection Wiley, New York, 1987.
[19] H. S. Sawhney, S. Ayer and M. Gorkani, Model-based 2D&3D dominant motion estimation for mosaicing and
video representation In Proc. Intl. Conf on Computer Vision, pages 583-590, 1995
http://www.sarnoff.com/products_services/government_solutions/vision_technology/technical_papers/95sawh.
pdf
[20] J. Y. A. Wang and E. H. Adelson, Layered Representation for Motion Analysis, Proceedings of the IEEE
Computer Society Conference on Computer Vision and Pattern Recognition 1993, pp. 361-366, New York,
June 1993. http://persci.mit.edu/people/jyawang/wang.html#Publications
[21] G. Lefaix, E. Marchand, P. Bouthemy, Motion-based Obstacle Detection and Tracking for Car Driving
Assistance IAPR Int. Conf. on Pattern Recognition, ICPR2002,volume 4, pages 7477, Quebec, Canada,
August 2002 http://www.irisa.fr/vista/Papers/2002/2002_ICPR_lefaix.pdf
AMAA_200311 15 / 15
[22] M. Irani, B. Rousso, and S. Peleg, Computing Occluding and transparent motion
International Journal of Computer Vision (IJCV), Vol. 12, No. 1, pp. 5-16, February 1994.
http://www.wisdom.weizmann.ac.il/%7Eirani/abstracts/multipleMotions.html