A Robust Hand Detection Method Based on Skin Probability Map

Existing hand detection methods are usually implemented with Haar-like and HOG descriptors, which lack of robustness to complicate background and the change of the lighting conditions in real applications. In this paper, a robust hand detection method is presented by exploring various feature representations based on skin probability map (SPM) instead of conventional grayscale image space. We show that, in the generated SPM image, hand region can be greatly highlighted and the background regions can be suppressed. And then, the Haar-like and HOG features are calculated from the skin mask and probability maps respectively, and named as skin-Enhanced features. In the sliding window searching stage, an efficient 3-level detection framework is introduced. In the experiment, an extensive dataset is established for the training of classifiers. And the hand detection results show that, the proposed method is much more robust than conventional hand feature descriptors.