基于关键帧的核密度估计背景建模方法

Kernel density estimation background model method based on key frame

在充分研究现有运动目标检测算法的基础上,提出了一种新的非参数核密度估计背景模型。应用高斯核密度估计进行背景建模,不需要事先假定背景特征的密度分布,根据视频序列像素灰度的相似性原理从训练样本中提取关键帧,减小了密度估计的样本量。将剩余的灰度值按距离最近原则归并到关键帧中去,降低目标检测的虚警率和误检率。实验结果表明:该算法在检测精度影响极小的情况下,大大提高了原算法的速度,可用于室外的实时视频监控系统。

A new background model of non-parameter kernel density estimate is presented on the basis of abundant study on algorithms of moving object detection. Gauss kemd density estimate is used to build background model without assuming prior density distribution. According to pixels＇ similarity of video sequence, key frame are chosen from training samples, and the remaining gray value are merged to the key frames according to nearest interval principle. Samples of density estimate and false alarm rate as well as detection error rate are reduced. Experimental results show that the algorithm greatly enhances the speed of original algorithm in the case of minimal impact on measurement accuracy, which enables the algorithm to be used in outdoor environment surveillance systems.