聚类差分图像核密度估计前景目标检测

Clustering Difference Image Kernel Density Estimation for Foreground Object Detection

针对非参数核密度估计学习阶段信息冗余与重复计算,估计阶段的估计错误噪声和计算量大的问题,提出了一种基于聚类分析的差分图像核密度估计前景目标检测算法。该方法在非参数核密度估计的学习阶段基于最大最小聚类原理从原采样全样本中提取那些具有较高频度和多样性的小样本来包含尽可能多的关键样本信息,在估计阶段采用基于自适应阈值的图像差分滤去非典型的运动像素,再利用高斯核密度估计进行运动像素分类。实验结果表明该方法限制了非典型运动像素估计错误产生的噪声,并减少了核密度估计计算量,提高了算法的实时性。

For non-parametric kernel density estimation information redundancy and repetition computation in the training stage estimate error and large amount of calculation in the estimated phase, this paper proposed a method of clustering difference image kernel density estimation for foreground object detection. We first choose those samples that have higher frequency and diversity to contain important information based on max-min distance clustering in training sequence. A Gausisian KDE is built to estimatea motion object after adaptive threshold image difference calculation. Experimental results were given to demonstrate that the proposed algorithms are elimination of the typical non-movement noise point for estimated error and improving real-time capability.