自然场景下运动目标检测与阴影剔除方法  

A Method for Detecting Moving Objects and Eliminating Shadow in Natural Scene

作　　者：屠礼芬[1] 仲思东[1] 彭祺[2] 

机构地区：[1]武汉大学电子信息学院,武汉430072 [2]湖北工程学院物理与电子信息工程学院,湖北孝感432000

出　　处：《西安交通大学学报》2013年第12期26-31,83,共7页Journal of Xi'an Jiaotong University

基　　金：国家"973计划"资助项目(2012CB725301);中央高校基本科研业务费专项资金资助项目(T201221207)

摘　　要：针对自然场景下运动目标检测中投射阴影使前景掩模检测精度低的问题,提出一种基于双阈值多分辨率的运动目标检测和多属性阴影剔除方法(MRPS)。首先根据不同分辨率下点的颜色属性,用双阈值混合高斯模型来检测运动掩模,用计算颜色模型来克服光照变化和轻度阴影的影响;然后根据阴影的光学属性,确定运动掩模中的潜在阴影区域,根据该区域外轮廓的位置属性对其分类;最后通过分析边缘属性恢复与阴影相连的目标背光区域,保留运动掩模中真实的运动目标区域。用公共测试图像的shadow序列对该方法进行了验证,并与狄利克雷过程(DPGMM)方法进行了比较,结果表明,该方法获取的前景掩模能更加有效地克服投射阴影的影响,前景检测精度提高了7.08%。A method for detecting moving objects based on dual-threshold and multi-resolution Gaussian mixture models and for eliminating shadow based on multi-attribute （named MRPS） is proposed to address the problem of low foreground mask detection accuracy caused by cast shadows under natural scene in detecting moving objects. First, the MRPS detects moving motion mask using the dual-threshold and multi-resolution Gaussian mixture models and overcomes the influence of illumination change and light shadows using the computational color model according to the point＇s color properties under different resolutions. Then, the potential shadow areas are determined in the motion mask from the optical properties, and the properties of the potential shadow areas are classified through analyzing the location properties for their outer contours. Finally, the baeklighting object areas connecting with the shadows is recovered and the real object areas in the moving mask are retained by analyzing edge properties. The proposed algorithm is verified using the shadow sequences in the public test image sequences and compared with Dirichlet processes Gaussian mixture models （DPGMM） method. Experimental results show that the foreground obtained by the MPRS overcomes the influence of cast shadows more effectively, and the foreground detection precision increases by 7.08 %.

关 键 词：运动目标检测 阴影剔除 双阈值 多分辨率 多属性 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]