基于集成学习思想的深度图像遮挡边界检测方法

摘要

针对现有深度图像遮挡检测方法不能有效地检测出深度信息变化不明显的遮挡边界点的状况，提出了8邻域总深度差特征和最大面积特征，并定义了计算方法。在此基础上，提出一种新的基于集成学习思想的深度图像遮挡边界检测方法，该方法结合所提新特征及现有遮挡相关特征训练基于决策树的AdaBOOst分类器，完成对深度图像中遮挡边界点及非遮挡边界点的分类，实现对深度图像中遮挡边界的检测。实验结果表明，同已有方法相比，所提方法具有较高的准确性和较好的普适性。%The existing OccIusiOn detectiOn methOd fOr depth image can nOt effectiveIy detect the OccIusiOn bOundary pOint with Iess ObviOus depth change,this status shOuId be changed. The eight neighbOrhOOd tOtaI depth difference feature and maximaI area feature are prOpOsed firstIy,and then the caIcuIatiOn methOds fOr these twO new features are defined. On this basis,a new OccIusiOn detectiOn apprOach based On ensembIe Iearning is prOpOsed,which cOmbines the prOpOsed features and existing OccIusiOn reIated features tO train the decisiOn tree-based AdaBOOst cIassifier tO cIassify the pixeI Of depth image intO OccIusiOn bOundary pOint Or nOn-OccIusiOn bOundary pOint. The experimentaI resuIts shOw that,cOmpared with the existing methOds,the prOpOsed apprOach has higher accuracy and better universaIity.