Real-Time Incremental and Geo-Referenced Mosaicking by Small-Scale UAVs

In the last decade, the use of small-scale Unmanned Aerial Vehicles (UAVs) is increased considerably to support a wide range of tasks, such as vehicle tracking, object recognition, and land monitoring. A prerequisite of many of these systems is the construction of a comprehensive view of an area of interest. This paper proposes a small-scale UAV based system for real-time creation of incremental and geo-referenced mosaics of video streams acquired at low-altitude. The system presents several innovative contributions, including the use of A-KAZE feature extractor in aerial images, a Region Of Interest (ROI) to speedup the stitching stage, as well as the use of the rigid transformation to build a mosaic at low-altitude mitigating in part the artifacts due to the parallax error. To prove the correctness of the proposed system at low-altitude, the public UMCD dataset and a simple metric based on the difference between image regions are presented. Instead, to show the overall effectiveness of the system, the public NPU Drone-Map dataset and a correlation measure are used. The latter metric evaluates the similarity between mosaics generated by the proposed method and those provided by a reference work of the current literature. Finally, the performance of the system compared with that of different modern solutions is also discussed.