摘要
针对图像序列运动遮挡检测的准确性与鲁棒性问题,提出一种基于光流与Delaunay三角网格的图像序列运动遮挡检测方法.首先构造基于非局部约束的TV-L1光流估计模型;然后根据图像Delaunay三角网格划分与光流估计结果对图像序列帧间对应像素点和局部三角形进行运动遮挡判断并检测遮挡区域;最后采用MPI Sintel和Middlebury数据库提供的测试图像集对本文方法与SMOD、GOSF等代表性方法进行对比测试.实验结果表明,本文方法相对于SMOD和GOSF方法在十组测试图像集的平均漏检率和误检率分别降低15.21%与30.57%,说明本文方法针对非刚性运动、复杂场景、弱纹理、光照阴影以及大位移等类型图像序列均具有较高的检测精度和较好的鲁棒性.
For the accuracy and robustness of the motion occlusion detecting from image sequence, this paper proposes a novel occlusion detection method based on the optical flow and Delaunay triangulation. Firstly, a TV-L1 optical flow model based on the non-local constraint is presented. Secondly, according to the results of the Delaunay triangulation and op- tical flow result of image sequence, the occlusion of the corresponding pixels and local triangles between the continuous frames is located and the motion occlusion regions could be detected. Finally, the evaluation sequences of the MPI Sintel and Middlebury databases are employed to test the performance of the motion occlusion detecting between the proposed method, the SMOD and GOSF methods. The experimental results show that the average omission rate and average false rate of the proposed method on the ten test image sequences are reducing 15.21% and 30.57% compared to the SMOD and GOSF methods, which indicates the proposed method has the higher accuracy and better robustness of the motion occlusion detecting under the non-rigid motion, complex scene, weak texture, brightness shadow and large displacement.
出处
《电子学报》
EI
CAS
CSCD
北大核心
2018年第2期479-485,共7页
Acta Electronica Sinica
基金
国家自然科学基金(No.61401190
No.61462062)
航空科学基金(No.2015ZC56009)
江西省重点研发计划项目(No.20161BBE50080)
江西省图像处理与模式识别重点实验室开放基金(No.TX201604001)
无损检测技术教育部重点实验室开放基金(No.ZD201529001)
