Analysis of Feature Detectors and Descriptors in a Feature-Based Stitching of UAV Images with Perspective Scene Distortion

Abstract

"Feature-based stitching algorithms are essential for creating seamless mosaics from overlapping images, particularly in UAV imagery. However, stitching UAV images presents unique challenges due to perspective distortions caused by varying altitudes, angles, and the dynamic nature of UAV flight paths. This study thoroughly examines various feature descriptors, focusing on their effectiveness in handling perspective scene distortion during UAV image stitching. The findings reveal that conventional feature detectors and descriptors, such as those used in standard stitching software, often fall short in managing the complexities of perspective distortions in pairwise image stitching. To address this, significant improvements can be achieved by refining features through enhanced sub-tasks integrated within the stitching pipeline, such as advanced matching techniques or adaptive warping strategies. Among the algorithms evaluated, SIFT emerges as a particularly strong candidate for feature detection due to its robustness against scale and rotation, while ORB and Harris also show solid performance. These results underscore the importance of selecting and refining feature descriptors to enhance the quality of stitched panoramas, particularly in challenging scenarios where perspective distortions are prevalent. The insights gained from this study are crucial for advancing UAV image stitching techniques, providing valuable guidance on overcoming the limitations of existing methods in complex environments."