Hand Segmentation Using Randomized Decision Forest Based on Depth Images

The segmentation of human hand is still challenged by its complex geometry and similar optical characteristics with other parts of human body. This paper presents an improved strategy for hand segmentation using the randomized decision forest framework based on depth images. In the proposed method, a new depth feature derived from the central point of hand structure, is induced to strengthen the ability of generalization of depth feature as well as reduce the requirement for training dataset, while not sacrifice the accuracy of hand segmentation. The experiments conducted on test dataset and real data showed that the improved method is relatively more efficient to be implemented and holds a good robustness to predict unknown hand pose.