부피 특징 벡터와 3차원 Haar-like 필터를 이용한 조명과 포즈에 강인한 손 검출

Abstract

This paper proposes an illumination and pose robust hand detection method using volumetric features and 3D Haar-like filters, which are obtained from AdaBoost learning. Existing hand detection methods using skin color or 2D appearance have the drawback that they are not robust to illumination changes and pose variations. We take volumetric features to getrobustness about illumination changes and propose 3D Haar-like filters toget robustness about pose variations. We consider two types of hand detectors that use different 3D Haar-like filters, where one is intended to detectas many hands as possible and the other is intended to verify the detectedhands and combine them by a logical AND operation to improve the handdetection performance. We performed several experiments using 2,500 handimages to validate the proposed hand detection method. The experimentalresults show that (1) the classification performance of volumetric featuresoutperforms those of other features like color and pattern by more than 20%,(2) the classification performance of the combined hand detector outperforms that of the single hand detector by more than 5%, and (3) the detection performance of the proposed hand detector with 3D Haar-like filters is about90%.