Diseño de un estimador de movimiento en secuencias de imágenes

Abstract

In an application for mosaicing individual images captured by cuadracopters, it is necessary to implement a motion estimation stage to accelerate the estimation of the final image geometrical transformations used in the mosaic. This work proposes the computation of the optical flow for a sequence of captured images in order to do a preliminary tracking of specific image locations that will be used in the final computation of the similarity transformation matrices. The implementation of two classical optical flow algorithms, specifically ones proposed by Lucas-Kanade and Horn-Schunck, serves as reference to compared with a third multiscale method for local optical flow accelerated with integral images. A genetic algorithm is used to find optimal parameterizations of the optical flow algorithms with respect to a set of fitness metrics such as the computation time and the accuracy of the results. It is possible to reduce the time of selection of correspondences points by 95.5% with respect to an exhaustive search algorithm of correspondences on a number of 1048 interest points taken of images of 320x240 pixels.