基于迭代张量高阶奇异值分解的运动目标提取

Moving object extraction based on iterative tensor high-order singular value decomposition

将基于凸优化的低秩矩阵恢复(LRMR)理论用于背景建模,当背景不稳定时,这种方法提取运动目标的效果不佳。由于矩阵的数据表示形式破坏了视频在时间和空间上的原始结构,采用张量表征视频的高维结构特性,提出了一种基于迭代张量高阶奇异值分解(HOSVD)的运动目标提取方法。用高阶奇异值分解代替LRMR中的矩阵奇异值分解(SVD),利用增广拉格朗日乘子法重建出三维视频张量的背景部分和运动目标部分,并进一步对运动目标部分进行形态学开闭运算。实验结果证明,相比常用方法,该方法错分率更低,能更准确完整地提取运动目标。

The method using the theory of low-rank matrix recovery( LRMR) based on convex optimization to model the background can't extract the moving object effectively when the background is unstable. Since matrix representation of video damaged the original temporal and spatial structure of data,this paper proposed an approach of moving object extraction based on iterative tensor high-order singular value decomposition( HOSVD),which utilized the tensor to characterize the high dimensional spatiotemporal structure of video. The proposed approach used HOSVD to replace the singular value decomposition( SVD) in LRMR,and utilized the augmented Lagrange multipliers to reconstruct the background portion and moving object portion of three-dimensional video tensor. Additionally,the proposed approach applied morphological opening and closing operations to the moving objects. Comparing with traditional methods,the experimental results show the proposed method can extract the moving objects more accurately and completely with lower misclassification rate.