Localization of Ground Targets by Unmanned Aerial Vehicles Based on BEBLID and Planar Perspective Transformation

The perspective of unmanned aerial vehicles (UAVs) often undergoes drastic changes during mission execution. However, the BRIEF descriptor used in the ORB algorithm is based on binary encoding of local image regions, which lacks robustness to viewpoint changes, affecting matching accuracy and leading to significant positioning errors. To address this issue, this study first introduces the Boosted Efficient Binary Local Image Descriptor (BEBLID), which exhibits good robustness to image rotation. Next, a template-based misalignment point rejection method is designed to effectively eliminate a large number of erroneous matching points generated by brute-force matching. The homography matrix H1 is obtained through brute-force matching, and then refined by incorporating the constraints of an onboard inertial measurement unit (IMU) using the Levenberg-Marquardt algorithm. This refinement allows for handling extreme viewpoint changes and obtaining a more accurate homography matrix. Finally, to mitigate the influence of low-precision onboard sensors, a planar perspective transformation is employed for real-time localization of ground targets.