A robust approach for long-term hand tracking

Abstract

Considerable challenges arise in hand tracking due to background clutter, inconsistent lighting, scale changes, occlusions and total object disappearance. In order to cope with these difficulties, we present a robust hand tracker based on enhanced particle filters. To overcome the inherent shortcomings of the particle filter approach, we infuse it with the mean shift (CAMSHIFT) method, thereby reducing the number of the particles and improving their quality. Furthermore, by using detection based on the αHOG (Alpha Histograms of Oriented Gradients) feature proposed in our earlier work [1], our tracker ensures that the hand is re-tracked if it disappears from the image frame and re-enters it. We base our appearance model on RGB color histograms and our newly proposed RLBP (Rotation Invariant Noise Compensated Local Binary Patterns) descriptor, which is an improved version of the LBP (Local Binary Patterns) descriptor. By virtue of our experiments, we show that our proposed tracker shows excellent results in comparison to state-of-the-art trackers on various challenging hand sequences collected by us as well as benchmark object sequences generally used in object tracking. Moreover, our tracker shows strong reacquisition ability, something that the existing hand trackers are not equipped with.