A Multiple-Scale Vision-Based Object Detection Algorithm for UAVs Based on the Local Feature Matching with Transformer Model

With the addition of UAVs, the perception and recognition of ground targets have also advanced significantly. UAV photography has gradually become an integral part of this field. Due to the flexibility and lightness of UAVs, they can freely adjust their height and angle. Moreover, multiple UAVs can perceive and recognize the same target from different directions, which has led to increased attention toward collaborative perception using multiple viewpoints of UAVs. Given the diverse image information captured by UAVs from various angles, heights, and environments, i.e., multi-scale conditions, this paper addresses the challenge of target detection by leveraging the Local Feature Matching with Transformer (LoFTR) approach. The paper also explains the process of coordinating transformations for different images captured by multiple UAVs to ensure data consistency and synergy. Finally, experimental validation using the VisDrone dataset demonstrates that the developed detection method can reliably and efficiently operate on targets at different angles and altitudes.