Restricciones algebraicas epipolares para estimación visual eficiente de movimiento sin la estructura 3D = Algebraic epipolar constraints for efficient structureless multiview motion estimation.

Resumen

Visual reconstruction methods such as Structure from Motion (SfM) or visual SLAM can be successfully used nowadays in tasks such as autonomous robotic navigation, augmented reality, or 3D scene reconstruction. Increasing the computational efficiency of these methods has been a persistent interest in the research community. This led to important reductions in time and energy consumption, and increased the chances of their integration in smaller or cheaper hardware, such as lightweight robotic platforms, smartphones or low-end commodity hardware. An important time-consuming operation in incremental SfM is the bundle adjustment (BA) refinement. A large number of improvements have been proposed in the literature to speed up this operation, including structureless BA, where the cost optimized is not based on the re-projection error, but on multiple view relations such as the epipolar or trifocal constraints. This way the cost does not involve the structure parameters, thus improving the computational efficiency of its optimization. In this work we propose GEA (Global Epipolar Adjustment), a high-performance structureless BA correction method based on algebraic epipolar constraints. Due to the algebraic nature of the GEA cost, it can be optimized very efficiently, in most cases using a small fraction of the time required by BA to obtain the optimal configuration. Moreover, despite of this algebraic nature, under general circumstances the accuracy of the obtained camera poses is very close to that obtained with classical BA methods. We also propose a structureless incremental motion estimation procedure which uses GEA to obtain accurate initializations for the camera poses. This procedure does not require composing feature trackings or the triangulation of scene landmarks. Instead, it just requires pairwise feature correspondences detected between the input images with standard image matching methods. Both the incremental motion estimation method and GEA are designed to be robust against the unavoidable outliers found by these matching techniques. The resulting camera poses can be used afterwards to obtain highly accurate sparse or dense estimations of the scene structure. We demonstrate the advantages, computational efficiency and practical applications of the proposed technique on a large number of real reconstruction problems, with arbitrarily large sizes and near critical configurations, and discuss possible future research lines.