视频中稳定的跨场景前景分割  

Stable Cross-scene Foreground Segmentation in Video

作　　者：魏宗琪 梁栋[1] WEI Zong-qi;LIANG Dong(School of Computer Science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 211100,China)

机构地区：[1]南京航空航天大学计算机学院,江苏南京211100

出　　处：《计算机技术与发展》2022年第12期37-42,共6页Computer Technology and Development

基　　金：国家自然科学基金资助项目(61772268)。

摘　　要：通过训练单个模型进行跨场景前景分割是一项具有挑战性的任务,特别是对于大规模视频监控,因为现有的模型通常严重依赖特定场景的信息。光流是描述前景目标的运动信息,但现有的光流机制方法只能表示瞬时特征,对开放的环境变化不具有鲁棒性。为了通过细粒度的运动特征表示并适应场景实现前景分割,一种间隔光流的新模块被设计出来,并使用注意力模块将运动特征融合到模型中。基于这种互补机制,可以构建实现运动和外观特征交互的跨模态动态特征滤波器。与现有方法相比,提出的模块倾向于在前景和背景区域的运动模式之间学习更多的语义信息,从而获得更好的跨场景适应性和鲁棒性。此外,由于数据集偏差问题,在跨场景前景分割任务中小目标的分割结果不佳,因此进一步设计了一个类内尺度的焦点损失函数来平衡前景目标的大小多样性。提出的模块可以即插即用到任意视频监控识别框架中,提高了跨场景前景分割结果的质量。Cross-scene foreground segmentation by training a single model is a challenging task,especially for large-scale video surveillance,because off-the-shelf models usually rely heavily on specific scene information.Optical flow is the motion information describing the foreground target.However,the existing optical flow mechanism methods can only represent the instantaneous motion and are not robust to open set.In order to achieve scene adaptation for foreground segmentation through fine-grained motion feature representation and interaction,interval optical flow was designed to combine fine-grained motion features with the attention module.Based on this,a cross-modal dynamic feature filter that realizes the interaction of motion and appearance features can be constructed.Compared with existing methods,the proposed module tends to learn more semantic information between the motion patterns of the foreground and background,so as to obtain better cross-scene adaptability and robustness.In addition,because the data set deviation usually misses small objects in cross-scene foreground segmentation tasks,a focus loss function of classification scale is further designed to balance the size diversity of foreground instances.The proposed module can be plug-and-played into any video surveillance recognition framework to improve the quality of the cross-scene foreground segmentation mask.

关 键 词：前景分割 双重模态 注意力 跨场景 自适应 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]