Attention from Above: A Multimodal Model for Drone-Based Object Localization

As drone technology becomes integral to industrial and surveillance applications, the need for accurate, real-time object detection in complex aerial environments has grown. This paper addresses the challenge of detecting small, user-specified target objects in drone imagery by enhancing the existing YOLO-World multimodal framework, which combines computer vision with natural language processing.

The author proposes a modified backbone for the YOLO-World model. The core innovation is the replacement of standard C2f layers—the building blocks of the YOLOv8 architecture—with new 'A2C2f' layers. These layers integrate an Area Attention (A2) module, which allows the model to recalibrate feature responses by focusing on informative regions while suppressing background noise. Additionally, the model incorporates a C2PSA (Cross Stage Partial with Spatial Attention) module to explicitly learn the importance of spatial locations, helping the model better identify objects of varying scales and positions.

Comparative experiments conducted on the VisDrone dataset demonstrate that the proposed architecture consistently outperforms the original YOLO-World model. The modifications led to a precision increase from 43.0% to 45.1% and a recall increase from 32.8% to 35.0%. The F1 score improved from 37.2% to 39.4%, and the mean average precision (mAP@0.5) rose from 32.5% to 35.2%. These results indicate that the attention-based backbone is more effective at capturing local features, particularly for small objects, compared to the baseline model.

This research provides a practical, lightweight enhancement for drone-based object localization. By integrating attention mechanisms into a text-guided detection framework, the model enables more flexible and accurate identification of specific targets in aerial footage. This is particularly valuable for industries like logistics, inspection, and agriculture, where drones must reliably identify specific objects in diverse, cluttered, or large-scale environments.