基于混合高斯模型与核密度估计的目标检测

摘要

Background modeling and motion detection is an important step in video tracking. Non-parametric kernel density estimation (KDE) and Gaussian mixture model ( GMM) are two classic methods in this field. This paper first introduces the basic principles of KDE and GMM, and analyses the advantages and disadvantages of both methods. Then a 2-step KDE-GMM cascade algorithm is proposed. It first does a fast segmentation of foreground and background region with KDE. For the region which cannot be modeled accurately, it uses GMM for a second judgment. The proposal method effectively combines their respective advantages. The test results showed that the improved algorithm is better than traditional ones and has good real-time and robustness.%背景建模与目标检测是视频跟踪的重要步骤和基础,非参数核密度估计与混合高斯模型是背景建模与目标检测的经典方法.文中首先介绍了高斯模型与核密度估计的基本原理及各自的优缺点,然后提出了一种核密度-混合高斯模型级联算法,利用核密度估计快速分割前景与背景区域,再由混合高斯模型对于无法精确建模的区域进行二次判定,有效综合了二者各自的优点.仿真结果表明,该算法具有良好的实时性和鲁棒性.