Hand tracking

Abstract

This thesis presents a hand detection and tracking system where hand is ini- tialized using the color clue and tracking is achieved with the integration of color and texture information. Three nonparametric skin classi cation schemes; histograms, kernel densities, voronoi tessellations are analyzed on six di erent colorspaces. The optimal fusion of color features is also investigated for illumination free skin classi cation. The texture and color cues are combined to track the hand through the course of action. Texture is de ned by Local Binary Patterns (LBP), which is a coarse estimation of joint probability of neighboring pixel values. By combining the color with texture more robust representation of hand is attained and meanshift algorithm is used to locate the hand in this representation space. The results show that texture-color combination can deal with face-hand overlaps and confusions of hand with other skin colored regions..