New Probability Models for Face Detection and Tracking in Color Images Marco J. Flores, José M. Armingol and Arturo de la Escalera Intelligent Systems Lab Department of Systems Engineering and Automation Universidad Carlos III de Madrid, 28911, Leganés, Madrid, SPAIN E-mail: {mjflores, armingol, escalera}@ing.uc3m.es Abstract - This paper presents a skin-color model and an automatic face detection system on color images. Three probability distribution functions are proposed to model the skin color: flexible generalized skew-normal distribution, skew generalized normal distribution and Gaussian mixture model, over three color spaces: CbCr, HS and H. The best model is chosen to build a system for detection and tracking face, using color information. The algorithm has been tested on several sequences of color images. Keywords - skin color, face detection, skew-normal, flexible generalized, Kalman filter. I. INTRODUCTION A. Motivation Face detection is a challenging and difficult task; however, it is important for built man-machine systems, for example, surveillance and security systems, tele-conference systems, virtual reality, monitoring systems etc. [2]. Among monitoring systems, we have driver’s monitoring systems for monitoring driver attention and state [7], [9]. For these reasons, many methods have been proposed among them, neural networks, support vector machines, color information, etc. Skin color offers useful and invaluable information for this task. Also, color allows fast processing and is highly robust to geometric variations of the face pattern due to different positions. To model skin color, there exist parametric and nonparametric models, in this paper; a parametric skin color models are studied. Some researchers [10], [15] and [16] model skin color through a multidimensional probability functions. Specifically, they use normal distribution functions. However, in this research the normality distribution of skin color is discarded, since statistical tests, on a random sample of skin color, did not show normal behaviour. For this we tried alternative distributions and they were compared among them on several color spaces. Our goal is to prove modern probability distribution functions for modelling skin color. B. Previous Work Many studies have been developed to analyze the skin color on several color spaces. Zarit et al in [19] have carried out comparative studies among several color spaces. The main conclusion is that the YCbCr color space shows worse results compared to the other four spaces: CIELAB, Fleck HS, HSV and normalized RGB. García et al in [4] conclude that the distribution of skin color presents higher consistency in the YCbCr space compared to HSV. Bradski in [2] choose the HSV space. He used H space instead of the normalized RGB because the pattern of skin color in the normalized space is more sensitive to changes of illumination and maintains saturation. On the other hand, Störring in [18] develops his work on the normalized RGB space. Plung et al in [16] show an extensive work where CbCr, HS, ab, YCbCr, CIE-Lab and HSV color spaces are compared. They have been studied three important issues: color representation, color quantization and classification algorithm. However, in [15] the most part of their work is development in YCbCr space, because this space is the most stable at illumination changing. The rest of this article is organized as follows. In section two the greater part of the work is presented, along with statistical reasons for choosing the distribution functions and the parameters of the proposed models are computed. Next in part three, the face detection and tracking system is described. Finally, in section four, the conclusions and future research are presented.