On the performance of several stereo calibration methods and models

Best Conference Paper (Poster Session) Stereo vision is a well-known technique which relies on a pair of cameras in order to reconstruct the shape and position of a generic object, without any additional geometrical and/or parametric information. The central issue in the set-up of a stereo vision system is two-fold: as first, removing the geometrical distortion caused by camera lens from images, then, make cameras aware of their own relative position in space. This paper is aimed to test the influence of different components of a calibration routine. In particular the goal is to compare the performance of several optimization algorithms and a number of alternative implementations of the pin-hole model. A main difference between this work and other tests present in literature is that the calibration performance is evaluated with respect to the measurement accuracy of the system, rather than by means of estimated reconstruction errors. In such a way, we get rid of theoretical errors, which do not represent any real application case, and we evaluate the accuracy on-the-field, facing with real-world issues. The obtained results show that, even though very complex equations can be used in order to represent cameras, usually, simpler pin-hole models remain competitive and robust, while refinements can be attained by using more powerful operational research algorithms.