CAJ | 학술논문

视线跟踪作为一种重要的人机接口模式,能够提供丰富的人机交互信息.提出了基于单目视觉的视线跟踪方法( Monocular Vision Approach,MVA).从眼部图像提取的表观特征,再经过支持向量回归( Support Vector Regression,SVR)计算实现可头部动作的注视方向估计.本方法仅用一个摄像机采集一副人脸图像作为输入数据,输出的计算结果是人的头部姿态和注视方向,以摄像机坐标系为参照系.采用的表观特征是基于方向二值模式( Directional Binary Pattern,DBP)算法,解析瞳孔在眼窝中运动引起的图像纹理变化.视线跟踪方法首先将双眼分割出来,并编码成高维的方向二值模式特征,最终通过支持向量回归作为匹配函数计算注视视角.共有11个人共23 676回归样本,按照姿态分成5个聚类集合.实验结果显示,基于本方法进行注视方向估计可以获得3°的测试误差.
시선근종작위일충중요적인궤접구모식,능구제공봉부적인궤교호신식.제출료기우단목시각적시선근종방법( Monocular Vision Approach,MVA).종안부도상제취적표관특정,재경과지지향량회귀( Support Vector Regression,SVR)계산실현가두부동작적주시방향고계.본방법부용일개섭상궤채집일부인검도상작위수입수거,수출적계산결과시인적두부자태화주시방향,이섭상궤좌표계위삼조계.채용적표관특정시기우방향이치모식( Directional Binary Pattern,DBP)산법,해석동공재안와중운동인기적도상문리변화.시선근종방법수선장쌍안분할출래,병편마성고유적방향이치모식특정,최종통과지지향량회귀작위필배함수계산주시시각.공유11개인공23 676회귀양본,안조자태분성5개취류집합.실험결과현시,기우본방법진행주시방향고계가이획득3°적측시오차.
As an important modality in Human - computer Interaction ( HCI),eye gaze provides rich information in communications.A Monocular Vision Approach (MVA) was proposed for gaze tracking under allowable head movement based on an appearance -based feature and Support Vector Regression (SVR).In MVA,only one commercial camera is used to capture a monocular face image as input,and the outputs are the head pose and gaze direction in sequence with respect to the camera coordinate system.This appearance -based feature employs a novel Directional Binary Pattern (DBP) to calculate the texture change relative to the pupil movement within the eye socket.In this method,the cropped two eye images are encoded into the high -dimensional DBP feature,which is fed into Support Vector Regression (SVR) to approximate the gaze mapping function.The 23 676 regression samples of 11 persons are clustered related to five head poses.Experimenta1 results show that this method can achieve the accuracy less than.