改进反向投影的Camshift人脸跟踪算法

对人脸进行检测与跟踪是诸如人机交互、视频监控等众多应用的基础.在众多方法当中,连续自适应均值偏移(Con-finuously Adaptive Mean Shift,简称Camshift)算法在兼具良好跟踪性能的同时做到了较低的计算成本.然而在经典Camshift算法中,反映像素类肤色概率的"反向投影图"会受到初始搜索框内背景像素的影响,是几乎所有基于经典Camshift的算法中普遍存在的一个问题.针对反向投影图的原理进行分析,并采用人脸检测结果作为替代方案,从而对传统Camshift算法进行改进.同时,对YCrCb色彩空间中的人脸检测进行多时段分析,并借此自动确定初始跟踪区域,较传统Camshift算法具有更好的效果.

An Improved Back-projection Based Camshift Algorithm for Face Tracking

The detection and tracking of face play a basic role in many applications such as interaction between computer and human,video surveillance and so on.Despite so many methods of tracking,Camshift meets not only the requirement of high quality but also the need of low complexity at the same time.However,the back-projection which shows the skin-like probability would be affected by pixels belonging to the background in initial searching window.And this is an universal problem existing in most of Camshift-based...