基于轮廓基元的目标表示及检测方法

摘要

为了从复杂真实的场景中提取筛选出目标本质特征,训练目标模型,进行有效的目标检测,提出了一种基于轮廓基元的目标表示及检测方法.采用轮廓基元进行部件建模,提出了从复杂真实场景中进行特征筛选的基本准则,无需对训练图像进行分割,使用外观聚类、位置聚类.AdaBoost三层筛选框架,建立部件模型并学习获得集成分类器.实验结果表明,该方法对复杂背景、局部遮挡和姿态变化具有较强的鲁棒性,对尺度变化具有不变性.%To extract local features which are used to model and detect a category of objects from complex, real images, an object representation and detection method based on contour primitive is proposed which uses contour primitives to build the part based model,some principles in contour primitive selection from complex real images are introduced, and a feature selection framework consist of appearance cluster, location cluster and AdaBoost is used without the pre-segment of the training images. Experimental results are presented and the robustness to complex background, partial occlusion and posture change, and the invariance to scale changes are confirmed.