基于多种类视觉特征的混合高斯背景模型

Mixture of Gaussian background modeling method based on multi-category visual features

Stauffer等人提出的混合高斯背景减除建模技术及其改进算法在真实场景的运动目标检测系统中取得了较好的检测效果且被人们广泛应用。然而,此类方法通常采用单一的颜色视觉特征进行建模。当运动目标的表观颜色和背景场景的表观颜色相近时,检测准确度会大大降低。对于场景亮度条件的突变而引起的前景噪声,即使采用模型更新机制,也不能有效及时的去除。针对这些不足,提出一种基于颜色、边缘和纹理视觉特征的混合高斯建模技术。新的建模特征能够很好的描述背景区域的本质,对前景目标有着非常好的区分力,并且采用准确率和召回率对实验结果进行定量分析。实验分析表明,新算法有效地解决了传统算法存在的问题。同时也为后继的高层视觉分析任务打下了良好的基础。

The mixture of Gaussian method （GMM） proposed by Stauffer et al and its improved versions achieve better performance and have been widely used in smart vision systems. Traditionally, GMM only uses a single color visual feature to model scenes, but it will fail when moving objects have the same color appearance with scenes＇ color appearance. In addition, traditional GMM can＇t effectively solve problems caused by lighting changes and especially lighting sudden changes. To overcome these weaknesses, in this paper, a new modeling method based on multi-category visual features （i. e. color, texture and edge） is proposed. Experiment results show that our new method＂ solves the above problems effectively,which will be a baseline for high-level vision analysis.