Speeding up probabilistic inference of camera orientation by function approximation and grid masking

This article presents modifications to an existing technique for camera orientation estimation intending to make it faster for use in real time applications and also for analysis of large image sets. The technique is based on likelihood maximization of a probability function that has the image gradient as the observed data and the camera orientation as parameter values. The camera orientation is inferred from the vanishing points of the image, and the directions of the edges in the environment are assumed to be in three mutually orthogonal directions. The first proposed modification is to substitute the expression that is calculated at each pixel by a computationally lighter approximation. The second proposal is to take in consideration only a few of the pixel lines and columns of the image during the calculations, performing a grid windowing of the image. This article presents the derivation and reinterpretation of the likelihood function approximation and also a performance evaluation