金字塔光流三维运动估计与深度重建直接方法  

Direct method for 3D motion estimation and depth reconstruction based on pyramid optical flow

作　　者：张聪炫[1,2] 陈震[2] 黎明[1,2] 

机构地区：[1]南昌航空大学测试与光电工程学院,南昌330063 [2]无损检测技术教育部重点实验室,南昌330063

出　　处：《仪器仪表学报》2015年第5期1093-1105,共13页Chinese Journal of Scientific Instrument

基　　金：国家自然科学基金(61401190;U1233125;61462062);江西省主要学科学术带头人培养计划(201208421);江西省自然科学基金重点(20133ACB20004);江西省优势学科团队建设计划项目(2015BCB24004);江西省科技落地计划(201408083);航空科学基金(2013ZC56005);南昌航空大学博士科研启动金(EA201508003)项目资助

摘　　要：针对基于图像序列光流的三维运动估计与深度重建问题,提出一种基于图像金字塔光流的三维运动估计与深度重建直接方法。首先根据光流计算亮度守恒假设和像素点光流与三维空间点运动的对应关系推导出基于图像亮度的三维运动守恒假设;然后借鉴变分原理,通过设计基于L1模型的鲁棒数据项以及图像与运动联合控制的平滑项构造基于变分光流的三维运动估计能量函数;为了应对图像序列中包含的大位移运动及运动遮挡问题,采用图像金字塔分层策略设计三维运动估计模型;最后根据图像三维运动估计结果重建图像中运动物体或场景的深度信息。实验表明,该方法能够较好地应对图像序列中光照变化、多目标大位移运动以及运动遮挡等情况,具有较高的三维运动估计与深度重建精度和较好的鲁棒性。Aiming at the problem of 3D motion estimation and depth reconstruction based on image sequence optical flow,this paper proposes a direct method for the 3D motion estimation and depth reconstruction based on image pyramid optical flow. Firstly,the 3D motion constant assumption is derived according to the optical flow calculation luminance constant assumption and the corresponding relationship between the 3D spatial point motion and the image point optical flow. Then,the 3D motion estimation energy function is constructed through designing the L1 model based robust data term and the image and motion combining controlled smooth term with the variational principle. To deal with the problems of large displacement and motion occlusion contained in the image sequence,the 3D motion estimation model is designed adopting the image pyramid hierarchical strategy. Finally,the depth information of the moving object or scene in the image could be reconstructed according to the image 3D motion estimation result.Experiment results show that the proposed method could better deal with the terrible influence caused by illumination variation,multi-object large displacement motion and motion occlusion in the image sequence; and has high accuracy and good robustness for the 3D motion estimation and depth reconstruction.

关 键 词：光流 三维运动估计 深度重建 金字塔分层策略 直接方法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TH701[自动化与计算机技术—计算机科学与技术]