抗遮挡和模糊的平面目标鲁棒追踪算法研究
Research on Robust Tracking Algorithm for Planar Object Against Occlusion and Blur
方晗 1姚剑2
作者信息
- 1. 武汉大学遥感信息工程学院,湖北武汉,430079
- 2. 武汉大学遥感信息工程学院,湖北武汉,430079;广东开放大学人工智能应用创新研究中心,广东广州,510091
- 折叠
摘要
本文提出一种基于渲染的数据生成方式,可模拟相机运动导致的运动模糊、散焦模糊等效果,并由此构建了一个新的数据集,以支持平面目标追踪算法的研究.此外,基于图像光流和平面掩码多任务学习,提出一种新的平面目标追踪模型.具体来说,该方法通过估计的平面掩码剔除背景和遮挡干扰,筛选出可靠的稠密光流估计,以提高追踪算法的鲁棒性.通过在公共数据集和补充数据上进行验证,该算法精度和鲁棒性都有所提升,同时在模糊、遮挡以及纹理缺乏等情况下都可进行有效追踪.
Abstract
This paper proposes a rendering-based data genera-tion method that can simulate motion blur,defocus blur and other effects caused by camera movement,and builds a new data set based on this to support the research of planar target tracking algorithms. In addition,this new planar object track-ing model based on image optical flow and planar mask multi-task learning. Specifically,the method eliminates background and occlusion interference through the estimated plane mask,and screens out reliable dense optical flow estimates to im-prove the robustness of the tracking algorithm. The applica-tion of the planar target tracking algorithm on public datasets and supplementary data reveals that its accuracy and robust-ness improve,and it can be effectively tracked under condi-tions such as blur,occlusion,and lack of texture.
关键词
平面目标追踪/光流估计/掩码估计/数据集构建Key words
planar object tracking/optical flow estimation/mask estimation/dataset construction引用本文复制引用
基金项目
国家自然科学基金(42271445)
深圳市科技计划资助(JCYJ20220530140618040)
深圳市中央引导地方科技展专项资金资助(2021Szvup100)
出版年
2024
NETL
NSTL
万方数据