基于光流优化的堆叠Unet背景建模神经网络

Optical flow based stacked Unet background modeling network

针对现有背景建模算法难以处理复杂前景及间歇性运动前景的问题,提出了一种基于非监督学习的背景建模算法(改进的BM-Unet算法)。该算法结合光流法和Pearson相关系数在视频帧上提取背景关注区域,以此优化网络训练集和损失函数,从而有效提高了该算法在复杂前景情况和前景停留情况下的适应性;在此基础上,为进一步提高背景生成的精确度,又提出了一种堆叠Unet网络架构BM-SUnet(background modelling stacked Unet)。在SBMnet数据集上与现有算法在可视化效果和评估参数两方面的比较结果表明,所提算法在复杂前景和间歇运动前景情况下建模准确性好且鲁棒性高的结论。

To tackle the cluttered foreground and the intermittent foreground motion challenges of background modeling algorithms,this paper proposed an unsupervised background modelling method( improved BM-Unet). The proposed method combined optical flow and Pearson correlation coefficient to extract the attention region in frames. In addition,the attention region contributed to modify the training set generation and loss function to improve the ability to adapt to the circumstance including the cluttered and the intermittent motion of foreground. In a further step,this paper proposed a stacked Unet architecture( BMSUnet) to enhance the accuracy of background generation. Experiments on dataset SBMnet show that the algorithm is not only robust to the cluttered foreground and the intermittent motion challenge,but also outperforms the latest methods in terms of both qualitative and quantitative evaluation.