Embedded mosaic generation using aerial images

Real-time processing of sensor data collected from moving vehicles is crucial to the development of active navigational systems. In this paper we develop a fast solution to the task of mosaic construction from aerial images, that allows the embedded generation of a faithful representation of the surveyed area. Most current solutions are off-line (i.e. all data is first collected and then processed in a single batch), and therefore do not allow decision making during data collection. We show that, by applying a sparse variation of the Bundle Adjustment algorithm to minimize reprojection error and using georeferenced images to determine overlapping areas, thus decreasing the computational time required for feature extraction and matching, we can achieve near real-time results on embedded hardware. The proposed methodology was tested using a DJI Phantom 3 drone with a Qualcomm SnapDragon 805 processor, and showed a high accuracy in comparison with standard off-line solutions, at a fraction of the computational time.