动态场景下移动侦测的改进方法  

Improvement Method of Motion Detection in Dynamic Scenes

作　　者：刘敏[1] 武明虎[1] 王娟[1] 

机构地区：[1]湖北工业大学太阳能高效利用湖北省协同创新中心,武汉430068

出　　处：《计算机工程》2017年第5期290-293,共4页Computer Engineering

基　　金：国家自然科学基金面上项目(61471162);湖北省科技支撑计划项目(2015BAA115)

摘　　要：针对动态场景移动侦测误报率较高的问题,提出一种改进的移动侦测方法。基于动态纹理(DT)模型,采用滑动窗口机制对视频序列进行微视频元(MVE)划分,并对每一个MVE进行DT建模。为减少奇异值分解的次数,使用Batch-PCA组合增量主成分分析的方式进行DT模型参数的学习与更新。引入控制理论中观测性概念,运用观测性测量方法实现显著性运动物体的侦测。实验结果表明,对于存在摇摆的树枝、飘雪或起伏的水面等背景元素变化的场景,改进方法检测准确率高,鲁棒性好,实时性强。Aiming at the problem of the high false alarm rate of moving object detection in the dynamic scenes, this paper proposes an improved moving objects detection method. Based on Dynamic Texture （ DT）, the video sequence is divided into many Micro Video Elements（MVE） by the sliding window technique. Each MVE is modeled by DT model. In order to reduce the frequencies of Singular Value Decomposition （SVD）, the Batch Principal Component Analysis （Batch-PCA） and the Increment Principal Component Analysis（IPCA） are combined to estimate the DT parameters in a MVE group. The observability measure approach of control theory is introduced to detect the salient moving objects. Experimental results show that the method has the better detection rate, robustness and real-time performance in the dynamic scenes with the swaying branches, falling snow or waving water.

关 键 词：动态纹理 批量主成分 增量主成分 奇异值分解 观测性 显著性运动 

分 类 号：TP391.72[自动化与计算机技术—计算机应用技术]