Tracking of Facial Regions Using Active Shape Models and Adaptive Skin Color Modeling

Abstract

It is widely accepted that skin-color is an effective and robust cue for face detection, localization and visual tracking. Well-known methods of color modeling, such as histograms and Gaussian mixture models enable creation of appropriately exact and fast detectors of skin. In particular, skin color-based methods are robust to changes in scale, resolution and partial occlusion. In real scenarios an object undergoing tracking may be shadowed by other objects or even by the object itself. However, many color-based tracking approaches assume controlled lighting. These methods construct or learn models in advance and then use them in tracking, without adaptation to suit new conditions. Consequently, these techniques usually fail or have significant drifts after some period of time, mainly due to variation of lighting in the surrounding. Thus, such techniques are not as good as can be for use in real environments because skin-color perceived by a camera usually changes when the lighting conditions vary. Therefore, for reliable detection of skin pixels a dynamic color model that can cope with nonstationary skin-color distribution over time should be applied in vision systems. Two types of information are typically used to perform segmentation during face tracking. The first is color information (Bradski, 1998; Comaniciu et al., 2000; Fieguth & Terzopoulos, 1997; Perez et al., 2002; Sobottka & Pitas, 1996). The second is the geometric configuration of the face shape (Chen et al., 2002). It is often not easy to separate skin colored objects from non-skin objects like wood, which can appear to be skin colored. Therefore, both skin-color modeling and contours are used to separate the facial region undergoing tracking (Birchfield, 1998). The oval shape of the head is often approximated by an ellipse (Birchfield, 1998; Srisuk et al., 2001). To cope with varying illumination conditions the color model is accommodated over time using the past color distribution and newly extracted distribution from the ellipse's interior. However, such tracker pays little attention to what lies inside the ellipse and what is utilized to accommodate the color model. The kernel density-based tracking has recently emerged as robust and accurate method due to its robustness to appearance variations and its low computational complexity (Bradski, 1998; Comaniciu et al., 2000; Perez et al., 2002). Due to the use of a simple pixel-based representation as well as reduced adaptation capabilities of Mean-Shift methods the algorithm performs poorly under large illumination change. Updating the color model is one of the crucial issues in color-based tracking. A technique for color model adaptation was addressed in (Raja et al. 1998). A Gaussian mixture model was