Maximizing feature detection in aerial unmanned aerial vehicle datasets

This paper compares several feature detectors applied to imagery from an unmanned aerial vehicle to find the best detection algorithm when applied to datasets that vary in translation and have little or no image overlap. Metrics of inliers and reconstruction accuracy of feature detectors are considered with respect to three-dimensional reconstruction results. The image matching results are tested experimentally, and an approach to detecting false matches is outlined. Results showed that although the detectors varied in the number of keypoints generated, a large number of inliers does not necessarily translate into more points in the final point cloud reconstruction and that the process of comparing a large quantity of redundant keypoints may outweigh the advantage of having the extra points. The results also showed that despite the development of keypoint detectors and descriptors, none of them consistently demonstrated a substantial improvement in the quality of structure from motion reconstruction when applied to a wide range of disparate urban and rural images.