Monocular Vision Based Relative Depth Estimation for Hand Gesture Recognition

In this paper, we focus on real-time relative hand depth estimation with cluttered backgrounds and variable illumination for a target application to hand pushing and pulling detection. The task is characterized by a lack of consistent internal contrast in the hand combined with the complex background. We adopt a detection-and-registration strategy to predict frame-toframe scaling factor and make decision by the cumulated relative scale changes. The registration accuracy is enhanced by an adaptive skin color segmentation step. Our system works effectively in real time and gives a low false alarm rate.