多线索非参数化融合的单目视频深度估计  

Depth estimation of monocular video using non-parametric fusion of multiple cues

作　　者：刘天亮[1] 莫一鸣[1] 徐高帮[1] 戴修斌[1] 朱秀昌[1] 罗杰波 

机构地区：[1]南京邮电大学江苏省图像处理与图像通信重点实验室,南京210003 [2]罗彻斯特大学计算机科学系,美国罗彻斯特14627

出　　处：《东南大学学报（自然科学版）》2015年第5期834-839,共6页Journal of Southeast University：Natural Science Edition

基　　金：国家自然科学基金青年科学基金资助项目(61001152;31200747);国家自然科学基金资助项目(61071091;61071166;61172118);江苏省自然科学基金资助项目(BK2010523;BK2012437);南京邮电大学校级科研基金资助项目(NY210069;NY214037);国家留学基金委资助项目

摘　　要：为解决二维视频的三维转化问题,提出了一种基于非参数化学习和多线索融合的单目视频深度图提取方法.首先,利用单目图像的区域边界轮廓和几何透视结构线索,基于前景背景融合来估计单目视频中各帧的深度图像;然后,利用视频帧间空时相关性,借助非参数学习实现单目视频深度估计;最后,利用全局背景深度分段约束和去抖动来增强深度视频序列.实验结果表明,与其他现有方法相比,该方法能得到更为准确的单目视频深度图序列,无论在主观质量还是均方根误差(RMS)和结构相似性度量(SSIM)上,均能取得较好的效果.A depth estimation technique for monocular video based on non-parametric learning and fusion of multiple cues is proposed to solve the conversion from two-dimensional （2D） video to three-dimensional （3D）. First, according to the regional boundary contours and geometric perspective structures cues of the monocular image, the depth map of each frame of the monocular video is estimated by fusing the related foreground map and the background map. Then, the depth map sequence of the monocular video is estimated in a non-parametric learning framework with temporal and spatial relationship between inter-frames. Finally, the depth sequence is enhanced by exploiting global background depth frame constraint and deblurring. The experimental results show that, compared with the existing methods, the proposed technique can obtain the depth map sequence of monocular video with higher accuracy and better performance not only on subjective quality but also on the root mean square （RMS） and the structural similarity measure （SSIM）.

关 键 词：深度图 非参数化融合 多线索 线性透视 空时相关 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]