Real-Time Image Matching and Tracking for Autonomous Quadrotor Helicopters

An algorithm for real-time matching and tracking over image sequences is presented. The goal is that a quadrotor helicopter autonomously approaches a selected target. Therefore, it is necessary to locate the target in live images. In this paper, the target is a point in a window which is selected by an operator in a reference image. This reference point is projected into the live image and tracked over consecutive images. In order to reliably project the reference point into the live image, a robust matching algorithm is proposed. Projective transformations are applied to the reference image to extract additional keypoints and to gain invariance to out-of-plane perspective transformations. A rating based on a color descriptor and on the cross correlation values of regions around inlier keypoints is used to evaluate the quality of the estimated homography matrix. Since the matching algorithm does not exploit the beneficial characteristics of image sequences and because its processing time is not short enough for high frame rates, a tracking algorithm is presented. In order to avoid divergence, tracking is re-initialized based on the rating of the matching result. High detection rates even in image sequences with large viewpoint changes are achieved. Moreover, through parallelization the algorithm offers very low processing times. It is shown that images can be processed at high frame rates in real-time.