Real Time Appearance Based Hand Tracking实时基于外观的手跟踪.pptVIP

Real Time Appearance Based Hand Tracking The 19th International Conference on Pattern Recognition (ICPR) December 7-11, 2008, Tampa Convention Center, Tampa, FL, USA 報告者：彭成瑋 日期：2009/12/29 指導教授：陳立祥 教授 實驗室：網際網路多媒體應用實驗室 Outline Introduction Tracking method Experiments Conclusion QA Introduction Hand tracking is an important problem in the field of human-computer interaction. Application：sign language recognition or controlling computer games. Model-based(3D model) and Appearance-based (Image features) Introduction（ Cont. ） the hand presents a motion of 27 degrees of freedom (DOF), 21 for the joint angles and 6 for orientation and location[11, 10]. Substantial problems：out-of-plane rotations scale changes, self-occlusions or segmentation accuracy. Real-time tracking performance Maximally Stable Extremal Region (MSER) tracking algorithm. Tracking method Novel tracking method Multivariate Gaussians with the Kullback-Leibler distance Color likelihood calculate a probability value p(O|xi) for every pixel in the current frame object-to-be-tracked (hand) O Kullback-Leibler distance instead of the Bhattacharyya distance The integral image for Bhattacharyya distance calculation Color likelihood （ Cont. ） Mahalanobis Distance Bhattacharyya Distance Color likelihood （ Cont. ） color likelihood value -- p(O|xi) every pixel – xi r × c window color distribution of the hand O in the frame t?1 -- Gaussian 3×1 mean vector – μO 3×3 covariance matrix -- Gaussian multivariate Gaussian -- Maximally Stable Extremal Region (MSER) tracking (a) Input Image (b) Image histogram (c) MSER result Modified MSER tracking (a) Color likelihood (b) MSER detection result Experiments 25 frames per second on a 320 × 240 video sequences Experiments（Cont.） A simple gesture recognition allows to use the tracker for controlling the mouse pointer and activating mouse-clicks. Conclusion Novel real time method for tracking hands through image sequences Efficiently calculated