Estudo de métodos de extração de características aplicados ao problema da estimação da posição de um VANT em navegação autônoma com visão computacional

Abstract

This work presents a comparative study of feature extraction methods applied to the problem of position estimation for autonomous navigation of UAVs. Based on an initial study of the existing methods, this work also proposes two novel feature extraction algorithms invariant to the affine transformation and with a low computational cost. The first proposal is a modification of ASIFT algorithm. The method utilizes the data provided by the INS to reduce its computational cost. The second proposal is alsobased on the affine invariance property of the ASIFT algorithm, but uses the SURF method for extracting candidate feature points. The proposed methods are compared with SIFT, SURF and ASIFT to aerial images registration taken from UAVs. The evaluation metrics used in the comparative study to measure the performance of feature extractors are the amount of extracted points in each image, the amount of matched points, the distance-error of matched points, and the computational costmeasured by the time of execution. The numerical experiments showed that the proposals are able to improve the accuracy and reduce the complexity of ASIFT algorithm, which means that can be used for autonomous navigation of UAVs.