In this paper, we describe an approach to estimate optic flow from an image sequence based on Support Vector Regression (SVR) machines with an adaptive ɛɛ-margin. This approach uses affine and constant models for velocity vectors. Synthetic and real image sequences are used in order to compare results of the SVR approach against other well-known optic flow estimation methods. Experimental results on real traffic sequences show that SVR approach is an appropriate solution for object tracking.
In the real world scene, we have the ability to detect and identify objects with different motions instantaneously. This is due to our perception system which has several levels of image processing. Thus, thanks to this faculty perception, we drive our car with much ease.
In automatic drive assistance domain, an incar camera is used to observe the environment scene and to estimate the car movements. Optic flow estimation method is usually used in order to compute the apparent motion in an image sequence. Indeed, optic flow method is sensitive to noise, large displacements, intensity discontinuities, etc. (Barnum et al., 2003). Moreover, it is considered as an ill-posed problem (aperture problem). Therefore, it is important to develop robust estimation algorithms to overcome these problems.
In computer vision literature, there are three families of optic flow estimation algorithms: frequency–domain motion, block-based motion and gradient-based motion (Barron et al., 1994). Since 1981, many robust estimators have been implemented to compute optic flow like: least squares (LS) (Lucas and Kanade, 1981), least trimmed squares (LTS) (Ye and Haralick, 2000), least median squares (LMS or LMedS) (Bab-Hadiashar and Suter, 1998), M-estimator (Black and Anandan, 1993), quick maximum density power estimator with a variable bandwidth (vbQMDPE) (Wang and Suter, 2003), etc.
Support Vector Machines (SVM) have been widely used in classification domain. In this paper, we propose to use its extension Support Vector Regression (SVR) machines to estimate optic flow in dynamic vision. Compared with classical local estimators, SVR possesses two advantages: a unique optimization solution and a large robustness against outliers (Vapnik, 1995 and Vapnik, 1998).
The following section briefly describes optic flow principle. Section 3 presents the SVR approach for optic flow estimation purpose. In Section 4, we are going to evaluate the performance and show some results of the proposed approach using synthetic and real image sequences.
We have presented SVR machine as an interesting robust approach to estimate optic flow. Thanks to its adaptive ɛɛ-margin, it can automatically detect and eliminate outliers to refine its estimate. The proposed SVR approach has been compared to classical methods using the well-known synthetic Yosemite sequence. Preliminary experimental results on real image sequences representing traffic road scenes show that SVR approach, although it requires huge computing time, is promising for object tracking.
