基于高阶阈值算子与张量全变分的前景检测

Foreground detection based on tensor total variation and threshold operator

基于鲁棒主成分分析(RPCA)的方法有一个潜在的假设,即场景中移动物体的像素是稀疏离群值,其往往忽略了物体的时间和空间结构,导致这些方法在动态背景、遮挡、光照变化等场景下检测效果降低。针对这一问题,提出了一种基于张量鲁棒主成分分析(TRPCA)的张量非凸稀疏模型。首先,利用三种常见收缩算子的优点,引入了二阶广义收缩阈值算子(GSTO),探索适用于高阶张量数据的高阶广义阈值收缩算子(HoGSTO),进而提升背景建模的鲁棒性;然后,为了表征视频前景中移动目标像素之间的相关性,在视频前景建模的过程中,利用张量全变分正则化(TTV)增强前景的时空连续性;接着,通过自适应l1范数对视频中的动态成分建模,避免了对前景建模产生干扰。多个视频帧的实验结果表明,该方法在移动目标检测任务中优于现有的方法,能够更好地分离前景和背景。

Robust Principal Component Analysis(RPCA)methods have an underlying assumption that the pixels of moving objects in the scene are sparse outliers,which often ignore the temporal and spatial structure of objects,leading to the reduced detection effect of these methods in dynamic background,occlusion,lighting changes and other scenses.To solve this problem,a tensor nonconvex sparse model based on tensor robust principal component analysis(TRPCA)is proposed.Firstly,take advantage of the advantages of three common contraction operators,a second-order generalized shrinkage is introduced.Generalized Shrinkage Threshold Operator(GSTO)explore the high-order generalized threshold shrinkage operator(HoGSTO)suitable for high-order tensor data,and then improve the robustness of background model⁃ing;then,in order to characterize the correlation between moving target pixels in the video foreground in the process of video foreground modeling,tensor total variational regularization(TTV)is used to enhance the temporal and spatial conti⁃nuity of the foreground;then,the dynamic components in the video are modeled through the adaptive l1 norm,which avoids modeling the foreground Produce interference.Experimental results in multiple video frames show that the pro⁃posed method is superior to the existing methods in moving target detection tasks,and can better separate the foreground and background.