提出一种针对实时视频流的快速且鲁棒的人眼定位方法,该方法直接应用于基于用户双眼精确定位的无辅助立体显示系统.在检测系统前端,结合稳定的肤色团块特征和积分图获取人脸检测候选区域；利用人脸-人眼两层分类器精确定位人眼,通过改进连续型AdaBoost算法中平滑因子ε的选取方法,在加快收敛速度的同时避免过学习,同时合理分布的样本保证了低误检率和鲁棒性；在后端,用卡尔曼滤波器预测人眼位置和克服跳变.实验证明,该方法可以快速精确定位人眼,满足无辅助立体显示的要求.%A fast and robust eye location method for real-time video stream is proposed, and it is directly applied in auto-stereoscopic display system which is based on accurate location of both eyes. In front stage of the detection system, stable mass features of the complexion and integral image are combined to obtain candidate areas for face detection; Two-layer face-eye classifier is used to locate eyes accurately, and by improving the selection way of the smoothing factor e in real AdaBoost algorithm, the overfitting is avoided while accelerating the convergence speed, meanwhile the reasonable distribution of samples insures low false alarm rate and robustness; In rear stage, Kalman filter is applied to predict eyes position and overcome sharp change of the detected positions. Experimental results show that this method can locate eyes fast and accurately and can meet the needs of auto-stereoscopic display.
展开▼
