REGULAR SPATIAL B-SPLINE ACTIVE CONTOUR FOR FAST VIDEO SEGMENTATION F. Precioso and M. Barlaud laboratoire I3S - UPRES-A 6070 CNRS Universite de Nice - Sophia Antipolis 2000 route des Lucioles - F-06903 Sophia-Antipolis FRANCE
[email protected] -
[email protected]
ABSTRACT This paper deals with fast video segmentation using active contours. Region-based active contours is a powerful technique for video segmentation. However most of these methods are implemented using level-sets. Although level-set methods provide accurate segmentation, they suffer from large computational cost. The proposed method uses BSpline parametric method to highly improve the computation cost. Our method removes irregular sampling assumption. It combines multi-resolution regular sampling and length penalty.
In section 3 we present the new method and its implementation and finally we present some experiments. 2. B-SPLINE REGION BASED ACTIVE CONTOUR 2.1. Criterion and velocity Let first define a general segmentation criterion. We search and object regions or a a background region in each frame n of a video sequence minimizing contour a criterion including region functionals.
!
#
%
&
'
)
+
!
#
%
3
#
)
+
1. INTRODUCTION: A REGION-BASED ACTIVE CONTOUR
,
.
0
,
(1) .
+
is the descriptor of the background doIn this criterion, main , and is the descriptor of the object domain . Using the variational Eulerian method proposed by JehanBesson & al [12, 13] for region-based active contour segmentation, we can compute the derivative of the criterion. Thus the active contour evolves from an initial position towards the objects with a velocity where is the inward normal of the active contour: +
Video segmentation refers to the process of extracting arbitrarily shaped image regions corresponding to specific semantic objects in either a supervised or completely automatic manner. Image and Video segmentation can be cast in a framework of minimization of a criterion including region and boundary functionals. Boundary functionals were first proposed by [1] and geodesic active contours [2] for active contour segmentation. Region based active contour first introduced by [3, 4]. Then [5, 6, 7, 8, 9] introduce region based statistic descriptor for image or video segmentation. Finally [10] addresses the segmentation problem where features of the region are embedded in region functionals. All these contour or region based methods use level set approach which is accurate but time consuming . In this paper we are going to address this problem. We use a parametric active contour evolution based on a cubic regularity in each B-spline interpolation preserving a point of the contour [11]. We propose a multi-resolution cubic B-splines to interpolate the active contour and a FIR implementation for the computation of the polynomial coefficients. In section 2 we present a survey of region based criterion, the derivation of the velocity vector and B-spline implementation.
0-7803-7622-6/02/$17.00 ©2002 IEEE
7
>
9
;
#
C
D
(2)
@
;
and
J
are:
a
}
X
d
U
K
K
9
K
U
U
J
U
N
9
k
}
X
O
P
0
X
0
X
e
P
(6)
U
0
U
U
(11)
¡
C
§
3
d
C
d
§
3
The contour is a closed curve. Thus the first and the last point and are defined as follows: X
H
X
O
¨
¥
C
§
D
©
d
C
d
§
3 J
>
J
J
J
> J
J
J
D
e
¥
D
(12)
¦
¦
©
9
U
£¤
¦
P
X
O
P
X
X
H
(7)
II - 762
»
3
>
¨
O
§
§
¦
P
X
C
3
C
3
O
D
©
D
¦
e
¥
D
¦
4. EXPERIMENTS All the following experiments have been computed over 100 resolution seframes of the Akiyo sequence (a quence) with a PIII 850 MHz PC.
Figure (Fig.2) shows the CPU Time related to the regularization parameter . The slowest segmentation corresponds to the highest resolution of 128 points. Nevertheless the segmentation is achieved in real time (less than 40 ms per frame).
4.1. Accuracy versus resolution and spatial regularization
4.3. Accuracy versus time for 3 resolutions and different spatial regularizations
Fig. 1. Accuracy versus Regularization
Fig. 3. Accuracy versus Time
Figure (Fig.1) illustrates the accuracy of the segmentation depending on the regularization parameter values. The accuracy is evaluated with the COST 211 quality distance ([17]) for different resolutions of the contour. The optimal value of the parameter is around 0.8. Thus adding in the criterion (9) the regularization to the intrinsic continuity of the B-splines increases the accuracy of the segmentation.
Figure (Fig.3) shows an evaluation of the segmentation related to the computation time. The optimal value of the ragularization parameter is 0.8 related to the accuracy of the segmentation (Fig.1). The segmentation is twice faster with a resolution of 64 points than with 128 points (Fig.2). Hence, according to the loss of the quality between resolution of 64 points and 128 points, 64 points appear as the most efficient resolution.
p
p
®
®
4.2. CPU Time versus resolution and spatial regularization
4.4. Accuracy per frames for different resolutions
Fig. 2. CPU Time versus Regularization
Fig. 4. Accuracy per frame
II - 763
Figure (Fig.4) shows the segmentation accuracy per frame. We fix the criterion parameters (9) for all the sequence. Thus, the parameters are more or less adapted to the current frame. The accuracy variations depends on the resolution as well as the frame considered. 4.5. Comparison with level set approach Let compare the average CPU Time presented in figure (Fig.2) to the computation time of level set approaches. Indeed, the highest experimented resolution (128 points) for our method provides the slowest segmentation with less than 40 ms per frame average. The computation time of the fast level set method for propagating interfaces reported in the literature [18] is about few seconds. The level set method developped by [19] for fast segmentation is achieved in more than one second. Compared to these level set approches, our B-spline active contour method runs up to twenty-five to thirty orders of magnitude faster. 5. CONCLUSION Our method remove irregular sampling assumption. It combines multi-resolution regular sampling and length penalty. A regular sampling of the curve where the number of points is defined according to the resolution. A local regularization over the smoothness of the contour is added to the regularity of the B-splines. intrinsic Experimental results on real video sequences show the accuracy and the speed of the proposed method. k
6. REFERENCES [1] M. Kass, A. Witkin, and D. Terzopoulos, “Snakes: Active contour models,” IJCV, vol. 1, pp. 321–332, 1988. [2] V. Caselles, R. Kimmel, and G. Sapiro, “Geodesic active contours,” IJCV, vol. 22, no. 1, pp. 61–79, 1997. [3] R. Ronfard, “Region-based strategies for active contour models,” IJCV, vol. 13, no. 2, pp. 229–251, 1994. [4] L. Cohen, E. Bardinet, and N. Ayache, “Surface reconstruction using active contour models,” in SPIE Conf. on Geometric Methods in Computer Vision, San Diego, CA, 1993.
[9] E. Debreuve, M. Barlaud, G. Aubert, and J. Darcourt, “Space time segmentation using level set active contours applied to myocardial gated SPECT,” IEEE Transactions on Medical Imaging, vol. 20, no. 7, pp. 643–659, july 2001. [10] S. Jehan-Besson, M. Barlaud, and G. Aubert, “Video object segmentation using eulerian region-based active contours,” in ICCV, Vancouver, Canada, 2001. [11] J. Kybic, P. Thevenaz, and M. Unser, “Multiresolution spline warping for epi registration,” SPIE, vol. 3813, pp. 571–579, July 1999. [12] S. Jehan-Besson, M. Barlaud, and G. Aubert, “DREAM S: Deformable regions driven by an eulerian accurate minimization method for image and video segmentation,” Tech. Rep. RR 2001-14, Laboratoire I3S, 2001. ¯
[13] S. Jehan-Besson, M. Barlaud, and G. Aubert, “Dreams: Deformable regions driven by an eulerian accurate minimization method for image and video segmentation. application to face detection in color video sequences,” in ECCV, Copenhagen, Denmark, May 2002. [14] M. Wang, J. Evans, L. Hassebrook, and C. Knapp, “A multistage, optimal active contour model,” IEEE Transaction on Image Processing, vol. 5, no. 11, pp. 1586–1591, november 1996. [15] F. Precioso and M. Barlaud, “B-spline active contours for fast video segmentation,” in ICIP, Thessaloniki, Greece, 2001. [16] R.H Bartels, J.C Beatty, and B.A Barsky, An introduction to Splines for use in Computer Graphics and Geometric Modeling, Morgan-Kaufmann, Los Altos, Californie, 1987. [17] A. Alatan, L. Onural, M. Wollborn, R. Mech, E. Tuncel, and T. Sikora, “Image sequence analysis for emerging interactive multimedia services - the european COST 211 framework,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 8, no. 7, pp. 802–813, November 1998. [18] D. Adalsteinsson and J. A. Sethian, “A fast level set method for propagating interfaces,” Journal of Comp. Phys., vol. 118, no. 2, pp. 269–277, May 1995. [19] J. Weickert, “Fast segmentation methods based on partial differential equations and the watershed transformation,” in DAGM-Symposium. 1998, pp. 93–100, Springer.
[5] T. Chan and L. Vese, “An active contour model without edges,” in Scale-Space Theories in Computer Vision, Corfu Greece, 1999. [6] S. Zhu and A. Yuille, “Region competition: unifying snakes, region growing, and bayes/MDL for multiband image segmentation,” PAMI, vol. 18, pp. 884–900, september 1996. [7] O. Amadieu, E. Debreuve, M. Barlaud, and G. Aubert, “Inward and outward curve evolution using level set method,” in ICIP, Kobe, Japan, 1999. [8] N. Paragios and R. Deriche, “Geodesic active regions for motion estimation and tracking,” in ICCV, Corfu Greece, 1999.
II - 764
