Fingertip positioning and tracking by fusing multiple cues using particle filtering

Abstract

We present a vision-based approach for positioning and tracking fingertip in a video. Multiple cues are fused through defining the likelihood probability terms in particle filtering framework. Skin color has been proven its robustness toward hand region localization in complex background. Thus, we describe it by a Gaussian distribution and further define a skin-color likelihood term. For lighting invariance, we also incorporate the contour information to define two contour likelihood terms. They respectively model the fingertip contour and two-side boundaries of finger. However, the particle filtering generally has degradation problem. To overcome this, we embed the mean shift to the particle filtering framework for convergence consideration. Finally, we validate the proposed approach by providing some experimental results.