Hands Tracking from Frontal View for Vision-Based Gesture Recognition

We present a system for tracking the hands of a user in a frontal camera view for gesture recognition purposes. The system uses multiple cues, incorporates tracing and prediction algorithms, and applies probabilistic inference to determine the trajectories of the hands reliably even in case of hand-face overlap. A method for assessing tracking quality is also introduced. Tests were performed with image sequences of 152 signs from German Sign Language, which have been segmented manually beforehand to o er a basis for quantitative evaluation. A hit rate of 81.1% was achieved on this material.