月面着陆多尺度边缘光流特征提取与跟踪方法  

Multi-Scale Edge Flow Based Feature Extraction and Tracking Method for Lunar Landing

作　　者：李云天 穆荣军[1] 单永志 崔乃刚[1] LI Yun-tian;MU Rong-jun;SHAN Yong-zhi;CUI Nai-gang(Department of Aerospace Engineering,Harbin Institute of Technology,Harbin 150001,China;Aviation Ammunition Institute,NORINCO Group,Harbin 150001,China)

机构地区：[1]哈尔滨工业大学航天工程系,哈尔滨150001 [2]中国兵器工业集团航空弹药研究院,哈尔滨150001

出　　处：《宇航学报》2020年第10期1305-1313,共9页Journal of Astronautics

基　　金：载人航天领域第四批预先研究项目(060201)。

摘　　要：针对月面着陆器动力下降过程中,基于描述子的传统视觉特征提取和跟踪方法耗时长、误匹配率高等问题,提出一种采用边缘光流的多尺度特征提取与跟踪方法。首先通过构建序列图像金字塔和应用多级掩膜提取方法,改善了图像平面上特征点分布的均匀性;在此基础上,利用边缘直方图灰度差平方和(SSGD)滑窗搜索算法,将光流计算由二维迭代简化为一维匹配,大幅缩短了算法耗时;进一步利用多尺度边缘直方图迭代搜索算法在改善大尺度运动下特征跟踪鲁棒性的同时,将光流计算精度拓展至亚像素级。仿真结果表明,本方法计算耗时不超过描述子方法的50%,非大尺度运动下具有更高的稳定跟踪特征点数目,大尺度运动下相比描述子方法下降不超过15%,在跟踪效率和稳定性方面取得了较好的平衡。A Multi-scale feature extraction and tracking method based on edge-flow for visual navigation during powered descent of lunar landing is proposed in this paper to improve the computational efficiency and tracking accuracy.The distribution of feature points is optimized by utilizing the image pyramid constructed from the sequential images combined with multi-level mask.Meanwhile,traditional 2D iterative optical flow calculation is reduced to 1D matching by performing the sum of square grayscale difference(SSGD)based sliding window search algorithm on the edge histogram,which eases the computational cost significantly.In order to improve the robustness during aggressive movement and extend the optical flow to subpixel,a multi-scale iterative search method is adopted.Results from numerical simulation indicate that more stable tracked feature points are maintained with only half computational cost in unaggressive movement and the lost feature points are less than 15%in aggressive movement while comparing with traditional descriptor based methods,which means an effective balance between tracking robustness and efficiency is kept.

关 键 词：边缘直方图 光流 多尺度 特征跟踪 月面着陆 

分 类 号：V488.2[航空宇航科学技术]