融合全局灰度模板的改进CT算法

针对压缩跟踪（CT）算法在构建判别表观模型过程中提取背景像素稀疏Haar-like特征导致目标跟踪漂移加重的问题，提出一种融合归一化灰度直方图全局特征模板的改进算法。与局部特征模板相比，全局特征模板更适于对目标和背景进行判别。改进算法基于压缩感知理论提取局部稀疏Haar-like特征构建表观模型M1得到跟踪目标的第一个估计参数H（v），提取归一化全局灰度直方图特征构建表观模型M2得到跟踪目标的第二个估计参数HD，使用H（v）和HD的线性组合作为表观模型利用贝叶斯分类器进行目标跟踪。实验结果表明，改进的算法提升了算法的鲁棒性，减轻了漂移问题。

Advanced CT algorithm via global intensity template

An advanced CT algorithm based on global feature template of normalized intensity histogram is proposed in order to solve the worse drift problem of object trakcing caused by extracting sparse Haar-like features from pixels in background area during constructing discriminative appearance model in Compressive Tracking（CT）algorithm. Com-pared to local feature template, global feature template is sensitive to target and background. An appearance model named M1 based on compressive sensing is constructed using local sparse Haar-like features and thus one estimation value called H（v）is obtained. The second appearance model called M2 is constructed with normalized global intensity histogram and another estimation value named HD is obtained. Both values are used to track objects via naive bayes classifier. The result of experiments indicates that advanced algorithm improves robustness and alleviates drift problem.