含有平面结构场景的捆绑调整

摘要

Bundle adjustment has been considered as one of the most important components in computer vision systems where three-dimensional structures are needed. A general bundle adjustment can optimize the coordinates of space points independently. But for a scene composed of both natural and structured objects, this often leads to an over parametrization the result and which deviates from truth. In this paper, a bundle adjustment with planar constraints and angle constraints is proposed for recovering the structures of environments with planes. By the aid of a new parametrization, the optimization remains an unconstrained non-linear least squares problem even if these two kinds of constraints are added. Experiments show that this new bundle adjustment method with prior knowledge provides an accurate estimation of the structures. Since prior information is added, the dimensionality of the augmented normal equation increases. A sparse solver is used after preconditioning in order to alleviate this problem. Moreover, a graph based angle constraint inference is devised for automatically finding constraints in a greedy manner once all planes are identified. This greedy method can preserve as many angle constraints as possible.%捆绑调整是计算机视觉中三维结构恢复过程的重要步骤。捆绑调整通常将空间中点与点坐标的调整视为相互独立的过程，但是在包含有自然物和人工物的场景中，由于存在多余的自由度，这种调整方法会导致优化结果偏离真值。提出了一种带有共面约束和平面夹角约束的捆绑调整，用于优化带有平面的场景。借助新的参数化方法，共面约束和夹角约束可以方便地进行表示，并且带有这两类约束的捆绑调整求解过程，仍然是一个无约束的非线性最小二乘问题。实验结果表明，这种带有先验信息的捆绑调整提供了对结构的更准确估计。由于先验信息的加入，增强型法方程的维度变高，借助了稀疏的求解技术和预条件子方法，大大降低了求解时间。最后，为了在实际应用中能够自动寻找出夹角约束，提出了一种基于最大完全图的贪心方法，该方法尽可能多地保留了夹角约束。