Development of scale and illumination invariant feature detector with application to UAV attitude estimation

Abstract

Feature detection has great importance in many applications such as vision navigation. Examining the developed detectors, it is found in many recent studies that most of the scale-invariant detectors are sensitive to illumination. In this work, we propose a novel detector that has good robustness to both scale and illumination. Motivated by the good robustness of Log-Gabor kernels toward light changes, we employ these kernels as a basis to construct the scale space. To detect potential features, we develop an effective interest points measure which is motivated by the concept of the autocorrelation and Hessian matrices. To confirm the good performance of our detector, we hold experiments on many datasets and with comparisons to common state-of-the-art methods. Furthermore, we evaluate the saliency of the detected features on a UAV attitude estimation task.