复杂场景下基于动态纹理的运动分割和背景估计方法  

Motion segmentation and background estimation method based on dynamic texture in complex scenes

作　　者：夏蕾[1] 罗佳[2] 

机构地区：[1]武汉商学院信息工程系,湖北武汉430056 [2]清华大学深圳研究生院,广东深圳518055

出　　处：《现代电子技术》2016年第11期63-69,共7页Modern Electronics Technique

基　　金：国家自然科学基金(61273072)

摘　　要：针对现有的运动分割和背景估计方法无法分割停止运动的对象、不适用于复杂动态场景等不足,首先提出一种基于动态纹理(DT)的背景-前景混合模型(FBM),实现动态场景下前景和背景的联合表示。FBM包括一组关于位置的DT成份和一组全局DT成份,前者用于模拟本地背景运动,后者用于模拟持续性的前景运动。其次,提出一种可学习FBM参数的EM算法及变分近似策略,使得FBM在不需要人工选择阈值和不需要单独训练视频的前提下,实现多种运动复杂场景下的前景运动分割,并检测出停止运动的对象。仿真实验结果表明,与当前最新的运动分割和背景估计方法相比,该方法可显著提升背景估计和运动分割的精度。Since the available motion segmentation and background estimation method can′t segment the static object,and is unsuitable for complex dynamic scenes,a foreground-background hybrid model（FBM）based on dynamic texture（DT）is proposed to realize the union expression of background and foreground in dynamic scenes. The FBM is composed of a set DT components about location and a set global DT components. The former is used to simulate the local background motion,and the latter is used to simulate the persistent foreground motion. The EM algorithm for learning FBM parameters and variation approximation strategy are proposed,which can realize the foreground motion segmentation in various moving complex scenes and detect the static object while FBM needn′t threshold artificial selection and individually training video. The simulation experiment results show that,in comparison with the latest motion segmentation and background estimation method,the proposed method can significantly improve the accuracy of background estimation and motion segmentation.

关 键 词：运动分割 背景估计 动态纹理 EM算法 阈值 分割精度 

分 类 号：TN911.7[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]