基于图半监督学习的移动相机背景减除

Mobile Camera Background Subtraction Based on Graph-Based Semi-Supervised Learning

在对移动相机拍摄的视频进行背景减除时,已有的无监督和监督学习模型的泛化能力都比较差。为此提出一种基于图表示和半监督学习的移动相机背景减除模型。首先提出了一种基于凸非凸图全变差正则的半监督学习模型。模型利用L1范数与其广义Moreau包络的差来构造非凸图全变差正则,可避免图全变差中L1正则项带来的有偏估计,并且在理论上可以保证模型中目标函数的整体凸性,进而可以利用交替方向乘子法对模型进行求解。数值实验中,将新模型应用到背景减除中,并在CDnet2014数据集的PTZ挑战上进行了比较实验。实验结果表明,对移动相机视频序列进行背景减除时,新模型在视觉效果和数值指标上都要优于已有的无监督和监督学习模型。

When performing background subtraction on videos captured by mobile cameras,existing unsupervised and supervised learning models have poor generalization ability.To solve this problem,a background subtraction model for a moving camera is proposed based on graph representation and semi-supervised learning.Firstly,a semisupervised learning model based on the convex non-convex graph total variance regularization is proposed.This model constructed a non-convex graph total variance regularization by using the difference between the L norm and its generalized Moreau envelope,which can avoid the biased estimation caused by the L,regularization.The global convexity of the objective function under certain condition was also theoretically proved.Furthermore,the alternating directional multiplier method algorithm was used to solve the proposed model.In numerical experiments,the proposed model was applied to background subtraction,and the comparison experiments were conducted on the PTZ challenge of CDnet2014 dataset.The experimental results show that the proposed model outperforms the existing unsupervised and supervised learning models in both visual effects and numerical criteria for background subtraction for a moving camera.