摘要
针对广泛存在的建筑物场景,提出了一种基于视频的平面检测与跟踪算法.算法分为3步,给定初始帧图像中的平面边界,第1步通过相邻2帧图像间的单应性矩阵跟踪平面边界,由于单应性矩阵跟踪会导致累积误差,首先采用RANSAC算法拟合边界线段,再使用绝对二次曲线约束和LM算法优化平面边界.第2步检测新出现的平面,根据原平面边界是由两组不同的平行线段组成提取该平面外围的第三类平行线段,由此判断是否有新的平面出现并检测新平面边界.第3步在检测到新平面边界后跟踪当前帧的所有平面到下一帧,当新出现的平面区域有增加或减少时,根据同一平面材质相似这一信息使用漫水填充和图像分割更新平面边界.最后,采用5个真实场景进行了实验,结果表明,文中算法能准确、稳定地在线检测与跟踪视频中的平面边界.
In this paper,an algorithm of tracking and detecting planar surface boundaries from videos is proposed for architecture scenes.The algorithm was divided into three steps.The planar surface boundaries were given for the first frame.In the first step,the homography matrix between the current frame and its previous frame was used to track planar boundaries.Considering that the tracking process inevitably generates cumulative errors,the RANSAC algorithm was used firstly to fit the lines on planar boundaries,and then the absolute quadratic curve was incorporated as a constraint and finally the LM algorithm was applied to optimize line segments.The second step was to detect the newly appearing planar surfaces.Assuming that the current planar boundaries are composed of two different groups of parallel line segments,the third group of parallel line segments outside the current planar boundaries was extracted,and then the newly appearing planar boundaries were detected.Finally,after detecting the newly appearing planar boundaries,all the planar boundaries of the current frame were tracked to the next frame.When the newly appearing planar areas increased or decreased,the line segments on the planar boundaries were updated by combining the Flood-Filling algorithm and image segmentation based on the assumption that reflectance are similar on the same plane.Experimental results of five realistic scenes show that the proposed algorithm can detect and track the planar surfaces accurately and robustly.
作者
刘艳丽
李志明
赵晓莹
邢冠宇
Liu Yanli;Li Zhiming;Zhao Xiaoying;Xing Guanyu(College of Computer Science,Sichuan University,Chengdu 610000;National Key Laboratory of Fundamental Science on Synthetic Vision,Sichuan University,Chengdu 610000)
出处
《计算机辅助设计与图形学学报》
EI
CSCD
北大核心
2019年第12期2074-2081,共8页
Journal of Computer-Aided Design & Computer Graphics
基金
国家自然科学基金(61572333,61972271)
光学辐射重点实验室开放基金(61424080109)
关键词
移动视点
平面检测
平面跟踪
moving viewpoint
plane detection
plane tracking
