 | 肌色情報に基づいた顔検出と手サイン認識に関する研究
|
| | Jean-Christophe TERRILLON1), 渡辺 博己1) |
| | | Abstract Human skin color is a powerful fundamental cue that can be used in particular, at an early stage, for the important applications of face and hand detection in color images or video sequences, and ultimately, for meaningful human-machine interactions [1]. One important first issue that we addressed during both Phase I and Phase II of the HOIP project is the selection of an efficient chrominance (color) space, because the performance of face and hand detection depends critically on the performance of the initial steps of skin pixel detection and of image segmentation, which in turn ultimately depends on the chrominance space that is used [2], [3]. Secondly, a suitable image segmentation model is applied depending on the selection of the color space [2]. The extraction of binary global facial feature information is then performed by use of fully translation-, scale- and in-plane rotation-invariant Fourier-Mellin moments, in order to ensure invariant face detection [4]. For subsequent binary face/non-face classification in complex scenes, we applied two different statistical learning techniques: a multiplayer perceptron Neural Network (NN) [4], or alternately, Support Vector Machines (SVM), that may [5], [6] or may not [7] rely on color as a first cue. Finally, hand posture recognition of the Japanese Sign Language (JSL) is performed by use of the Phase-Only correlation Filter (POF) [1], that yields an efficient discrimination both between different hand postures and between hand postures and scene background areas that have incorrectly been detected as skin. This report presents, for Phase I of the HOIP project, a detailed overview and experimental results of the skin color analysis and of the face detection and hand posture recognition system based on skin color (for convenience, later results that were obtained during Phase II are included in this section). The part of the report devoted to Phase II involves the integration of the system based on skin color into a more global, real-time system that is capable to simultaneously detect or track multiple faces as well as recognize hand postures of the JSL in color video sequences. |
| | | |
