Abstract: The traditional background modeling algorithms based on Gaussian Mixture Models（GMM） or Codebook（CB） and improved G-KDE are used widely in moving objects detection, but they can not accurately detect moving objects in the scenes such as sudden illumination change, non-stationary background and objects move again after a short delay. In allusion to the problems mentioned above, a background subtraction algorithm for detecting moving objects in video sequences from a stationary camera is proposed. The method analyses statistically the pixel history as time series, kernel density estimation is used to decide whether the background pixels are subjected to moving objects or not, two consecutive stages of the K-means clustering algorithm are utilized to identify the most reliable background region, pixel updating is able to adapt to gradual illumination change, object based background updating mechanism is presented to cope with interference like sudden illumination change, non-stationary background and ghost appearance. The video taken actually is simulated, the results show that the robust and accuracy of proposed algorithm is better than other three algorithms for detecting moving objects.

Key words: background subtraction, motion detection, kernel density estimation, statistical modeling, K-means clustering, background updating

摘要： 基于混合高斯模型（Gaussian Mixture Models，GMM）或码书模型（Codebook，CB）的传统背景建模算法和改进后的G-KDE算法被广泛地运用于运动目标检测中，但是在光照突变、非静止背景和运动目标短暂停留再运动的场景中不能正确地检测出运动目标。针对以上问题，提出了一种从静止摄像机的视频序列中检测运动目标的背景减算法。通过统计像素的经历作为时间序列，利用核密度估计判断背景像素是否受到运动目标干扰，使用K-均值聚类算法的两个连续阶段来确定可靠的背景区域，通过像素更新适应渐进的光照变化，提出一种基于对象的背景更新机制适应突然的光照变化以及非静止背景、鬼影等干扰。对实际摄取的视频进行了仿真实验，结果表明该算法比其他三种方法检测运动目标鲁棒性更好，准确性更高。

关键词: 背景减法, 运动检测, 核密度估计, 统计建模, K-均值聚类, 背景更新