Image Registration for Virtual Drone Flight Simulation

In the past decade, drones have expanded their applications, such as surveying, search and rescue, and last-mile delivery. Drone autonomy is a growing focus in these areas, requiring test environments for both drone operators and autonomous algorithms. This work addresses the creation of a realistic test environment for accurate position perception, motion planning, and drone control. Our approach uses RGB-D sensors and image registration techniques to construct a 3D map of the real environment, enabling drone navigation simulation. We adapt the Iterative Closest Point (ICP) algorithm, incorporating feature point filtering and local optical flow to expedite 3D point processing, particularly in static scenes. A challenge in the standard ICP algorithm is selecting suitable initial points for effective matching between consecutive point clouds. Our modified approach tackles this issue, optimizing the algorithm for 3D map construction. Since our application involves RGB-D image sequences with minimal variation between successive key frames, we use a scale-space pyramid and FAST techniques to select a limited number of key points, reducing the computational load. Moreover, by employing a local optical flow-based point-matching approach, we eliminate resource-intensive operations in each ICP optimization iteration. We demonstrate the practical utility of our method through a virtual 3D drone flight simulator.