光学 精密工程, 2017, 25 (2): 529, 网络出版: 2017-03-29
张量低秩表示和时空稀疏分解的视频前景检测
Video foreground detection of tensor low-rank representation and spatial-temporal sparsity decomposition
视频 前景检测 低秩 时空稀疏分解 张量低秩表示 非精确增广拉格朗日乘子 video foreground detection low rank spatial-temporal sparsity decomposition tensor low-rank representation inexact augmented Lagrange multiplier
摘要
针对视频中前景检测的问题, 提出了一种基于张量低秩表示(Tensor Low-Rank Representation ,TLRR)和时空稀疏分解的检测方法。由于视频序列中的前景除具有稀疏性外, 本身还具有空间上的连续性以及时间上的持续性, 本文提出采用时空稀疏范数对前景特性进行深入发掘。利用张量低秩表示方法将原始视频用张量形式进行分解, 充分利用了原始数据的行信息和列信息, 且将原始的背景、前景二分解泛化为背景、前景和噪声的三分解, 使用非精确增广拉格朗日乘子(Inexact Augmented Lagrange Multiplier ,IALM)方法进行最优化求解, 并对算法进行了分析。设计实验对本文新方法的有效性进行了验证和比较, 并对影响算法效果的重要参数ρ进行了进一步研究实验。实验结果表明: 该方法能够有效检测出视频中的运动前景, 其准确性相对已有方法有一定提高。
Abstract
A detection method based on Tensor Low-Rank Representation (TLRR) and spatial-temporal sparsity decomposition was proposed to detect foreground targets in video sequences. Since foreground in video sequence has sparsity inherently besides spatially continuous and temporally continuous, this paper put forward spatial-temporal sparsity-inducing norm to perform deep research on property of foreground. Original video was decomposed in tensor representation formed by tensor low-rank representation method, line information and column information of original data were fully used, and two-stage decomposition of original background and foreground was generalized to three-stage decomposition of background, foreground and noises. Optimization solution was performed with Inexact Augmented Lagrange Multiplier (IALM) method.Verification and comparison experiment was established, and further research experiment was performed to research how ρ affecting performance of algorithm. Experimental results show that the method can detect moving foreground in video effectively and improve accuracy when compared with existing methods.
隋中山, 李俊山, 张姣, 樊少云, 孙胜永. 张量低秩表示和时空稀疏分解的视频前景检测[J]. 光学 精密工程, 2017, 25(2): 529. SUI Zhong-shan, LI Jun-shan, ZHANG Jiao, FAN Shao-yun, SUN Sheng-yong. Video foreground detection of tensor low-rank representation and spatial-temporal sparsity decomposition[J]. Optics and Precision Engineering, 2017, 25(2): 529.