准临界状态下的场景三维重建研究

Critical configuration of 3D reconstruction is meant those conditions under which 3D scene reconstruction from 2D images is theoretically of impossibility due to the speciality of viewpoint locations or/and scene shape,. Quasi-critical configuration is meant those configurations which are close to, but not strictly of, critical one. The 3D scene reconstruction under quasi-critical configuration usually exhibits wild unstablity or extreme sentitivity to noise. In practice, various quasi-critical cases occur frequently, for example, when the camera undergoes a quasi-cicular motion or radial motion during image acqusition, or the scene is composed of a dominant planar one but a general scene shape assumption is presumed in reconstruction. Special care should be paid to such quasi-critical configurations, otherwise, gross reconstruction errors will occur at best, or the system could fail completely. This proposal is intended to explore and investigate theories and algortihms for quasi-critical configuration handling, in particular, how to use possible scene or motion qualitative a priori knowledge to improve the reconstruction robustness. The outcome could enrich the 3D reconstruction theory and methodolody, and could provide some theoretical quidance for complex scene reconstrction under constrained acqusition conditions.

所谓三维重建的临界状态，是指由于待重建场景的特殊性或视点轨迹的特殊性使得从这些视点获取的二维图像从理论上来说无法重建三维场景的情况。准临界状态是指当场景形状或相机轨迹很靠近临界状态但不严格位于临界状态时的状况，此时重建过程呈现极端的不鲁棒性。从图像重建三维场景的过程中经常会遇到准临界状态下的重建问题，例如当拍摄条件受到限制，拍摄图像时相机轨迹主要为旋转运动、径向运动，以及当场景主体为平面但仍基于一般场景框架下的重建等。准临界状态下的三维重建，如不采取特殊处理，会导致重建算法产生巨大的误差，甚至彻底崩溃。本项目旨在探索如何提高准临界状态下三维重建鲁棒性的理论框架和计算方法，特别是基于定性先验知识的重建方法。本项目的研究结果可望从一定程度上丰富三维重建的理论研究，同时为复杂、受限场景的三维重建提供一些理论指导和处理方法。

三维重建的精度和鲁棒性受相机之间的姿态影响。在一些特殊姿态下，理论上无法进行三维重建，这类姿态称之为“临界状态”。在实际应用中，当相机之间的姿态接近这种“临界状态”时，本项目称之为“准临界状态”，三维重建的精度和鲁棒性会剧烈下降。本项目针对相机径向运动，近似纯旋转运动，或场景仅有若干个主平面这些“准临界状态”下的三维重建进行了研究。采取的主要途径是：（1）：利用多幅图像之间的信息互补性；（2）：利用场景或相机运动的一些先验知识；（3）：部分信息的人机交互。项目执行期间，共发表论文11篇，培养博士：4名，硕士：3名。参加国际会议2次，组织国内学术研讨会 4次。 总体上说，按计划取得了一些相关进展，但进展不是特别显著。