摘要
针对非刚性大位移运动场景的光流计算准确性与鲁棒性问题,提出一种基于深度匹配的由稀疏到稠密大位移运动光流估计方法.首先利用深度匹配模型计算图像序列相邻帧的初始稀疏运动场;其次采用网格化邻域支持优化模型筛选具有较高置信度的图像网格和匹配像素点,获得鲁棒的稀疏运动场;然后对稀疏运动场进行边缘保护稠密插值,并设计全局能量泛函优化求解稠密光流场.最后分别利用MPI-Sintel和KITTI数据库提供的测试图像集对本文方法和Classic+NL,DeepFlow,EpicFlow以及FlowNetS等变分模型、匹配策略和深度学习光流计算方法进行综合对比与分析,实验结果表明本文方法相对于其他方法具有更高的光流计算精度,尤其在非刚性大位移和运动遮挡区域具有更好的鲁棒性与可靠性.
In order to improve the accuracy and robustness of optical flow estimation under the non-rigid large displacement motion,we propose a sparse-to-dense large displacement motion optical flow estimation method based on deep matching.First,we utilize the deep matching model to compute an initial sparse motion field from the consecutive two frames of the image sequence.Second,we adopt the gridded neighborhood support optimization scheme to extract the image grids and matching pixels which have the high confidence,and acquire the robust sparse motion field.Third,we interpolate the sparse motion field to obtain the dense motion field and plan a global energy function to estimate the optimized dense optical flow.Finally,we respectively employ the MPI-Sintel and KITTI datasets to compare the performance of the proposed method with several variational,region matching and deep-learning based optical flow approaches including Classic+NL,DeepFlow,EpicFlow and FlowNetS models.The experimental results indicate that the proposed method has the higher computational accuracy compared with those of the other state-of-the-art approaches,especially owns the better robustness and reliability in the areas of non-rigid large displacements and motion occlusions.
作者
陈震
张道文
张聪炫
汪洋
CHEN Zhen;ZHANG Dao-Wen;ZHANG Cong-Xuan;WANG Yang(Key Laboratory of Nondestructive Testing,Ministry of Education,Nanchang Hangkong University,Nanchang 330063)
出处
《自动化学报》
EI
CAS
CSCD
北大核心
2022年第9期2316-2326,共11页
Acta Automatica Sinica
基金
国家自然科学基金(61866026,61772255)
江西省优势科技创新团队(20165BCB19007)
江西省杰出青年人才计划(20192BCB23011)
航空科学基金(2018ZC56008)资助。
关键词
稠密光流
深度匹配
邻域支持
图像网格
全局优化
Dense optical flow
deep matching
neighborhood support
image grid
global optimization
