中文摘要：

针对现有研究对目标空间信息的普遍忽视及其对高密度群体精确计数的困难,提出对输入图像估计一个密度函数.通过该函数在任意图像区域上的积分得出该区域中的目标个数.经过数学推导,得到密度函数的参数化模型,分析特征向量需要满足的条件以及加入图像分割对结果的影响.由正则化风险最小化原理求取密度函数模型的参数,将密度函数的经验风险最小化问题简化为一个线性规划问题.实验表明,该方法只需少量图像进行训练,就可以准确地估计测试图像的目标数目.对于高密度群体,该方法能够给出目标计数,而不仅是密度等级估计.

英文摘要：

Current studies of object counting commonly ignore the spatial information of objects and encounter difficulties when dealing with high density object groups.An object counting approach was presented which estimated a density function for every input image,whose integral over any image region gives the count within that region.A parametric model of density function was built by mathematical derivation.The conditions that feature vectors should satisfy and the effects of image segmentation were analyzed.The parameters in the model of density function were estimated by the principle of regularized risk minimization,and the density function empirical risk minimization can be boiled down to a linear program.Experimental results show that the method can accurately estimate the object counts for testing images with only a few training images.For high density object groups,the approach also gives counts,not only density levels.