High accuracy stereo Structure from Motion through symmetry cost functions : Using cost functions with symmetry terms for high accuracy relative motion estimation

Abstract: A large effort, in the field of Computer Vision in general and Structurefrom Motion (SfM) in particular, has previously b een put into developing algorithms for estimating camera p ose and triangulate 3D features. Many of thesealgorithms however are based up on maximum p osterior likeliho o d cost functions which are efficient in terms of time consumption rather than in accuracy.The implementations are thus based on a simplified mo del in order to makethe algorithm more efficient. This pro ject aimed to lo ok at a particular problem, stereo cameras with approximately known relative motion, and the goalwas to develop a more accurate cost function. The algorithms presented usesthe symmetry of the measured noise in order to pro duce an estimated relativemotion which outp erforms the least square(L2) method with outlier rejection.The project did include both synthetic and real test data sets. The project wasdone at the request of a company and the testing sets were based on typicalproperties of a projectile tracking system used by the company.