Robust hand tracking with posture recognition via online learning

Robust hand tracking is a great challenge due to human hand´s small size and drastic appearance changes. Recently, machine learning especially online learning methods have shown their promising ability in object tracking. This paper successfully achieved hand tracking under a lately popular online learning framework named Tracking-Learning-Detection (TLD) by a win-win thought that hand tracking and posture recognition can benefit from each other. Firstly, the object model is extended in order to import posture recognition which is done without extra recognition algorithms. In turn, the introducing of hand postures enhance hand tracking since the tracker is adaptive to different hand shapes. At last, skin color is sufficiently applied in every module (tracking, learning and detection) of TLD further improving the speed and accuracy of tracking. Experiments show that the proposed method works well on hand tracking with the additional ability to recognize some given hand postures under various difficulties.