A Real-time Stitching method based on improved FLANN algorithm of Multi-UAVs for Remote Sensing Image

Abstract

The proliferation of low-altitude economic activities has precipitated a significant expansion in the application domains of Unmanned Aerial Vehicles (UAVs), transcending the constraints of singular platform operation. Multi-UAVs cooperative online remote sensing image stitching involves several challenges，include image blurriness and deformation caused by platform vibrations, as well as the management of vast data and transmission delays. This paper introduces an expedited alignment methodology that leverages the parameters garnered from UAV aerial photography to address the inefficiencies of the SIFT algorithm in image alignment. Data from the onboard GPS and IMU are used to estimate how much two aerial images overlap. Subsequently, the GPU within an edge computing platform is tasked with swiftly detecting SIFT features within the overlapping image sectors. Feature point matching is then performed using an improved matching algorithm, and finally a homography matrix is computed to facilitate fast stitching of aerial images. The results demonstrate that the RK3588-based edge computing platform exhibits rapid image decoding and stitching capabilities, with an average stitching latency of 36.43ms. Furthermore, this study developed a ground-based human-machine interaction system to present stitched video streams in real-time, thereby improving the operational efficiency and visual quality of Multi-UAVs aerial surveillance.