Building and Evaluation of a Mosaic of Images Using Aerial Photographs

To solve image mosaicing problem, this work first use the Harris-Laplace method to first find scaleinvariant keypoints. Through Gaussian-weighting the images intensities with a square window around each keypoint, it is possible to compute a finger print (descriptor) for each keypoint using the same method as the used in SIFT. Ransac, together with DLT algorithm, is used to robust estimate the projective transformations between two images from a set of putative correspondences. To find the putative correspondences between keypoints is used the ratio of the Euclidean distances of the first over the second nearest neighbor of each keypoint. Computed homographies are then used to build the mosaic of images. To study the robustness of the method a simulator was developed to take aerial photographies to an input image representing the Earth surface. Small shifting the parameters involved (e.g. quadrotor’s pitch) to obtain each photo, error measures, namely based on homography decomposition methods, were used to compare the results between the estimated and the exact mosaics. Results show the importance of the minimization of the tilt angles of the quadrotor, as well as the necessity for high overlapping percentages between images. Results also demonstrate that the images should be taken at high flight altitudes to avoid, on maximum, the parallax. However, the researched method has shown to be invariant to translations, rotations, scale variations, perspective transformations, contrast and brightness changes and even to the presence of some noise.