LUD-YOLO: An Efficient and Balanced Object Detection Model for Unmanned Aerial Vehicles

Abstract

Unmanned aerial vehicles (UAVs) have been widely used in defense, precision agriculture, ecological monitoring, and intelligent transportation because of their compact size, high mobility, and flexible deployment. Object detection based on UAV imagery is a key technique for autonomous perception and mission execution. However, UAV images usually contain complex backgrounds, small objects, dense target distributions, occlusions, and large-scale variations, which make it difficult for existing detectors to achieve a good balance between detection accuracy and computational efficiency. To address these challenges, this paper proposes LUD-YOLO (Lightweight UAV Detection–YOLO), an efficient UAV object detection model built on YOLOv7. First, an InceptionNeXt module is introduced into the backbone network to enhance multi-scale feature extraction. Second, an Efficient Multi-Scale Attention (EMA) module is embedded in the feature fusion network to suppress background interference and strengthen discriminative target regions. Third, a P2 detection head is added to improve the sensitivity and localization accuracy of small objects. Experimental results on the VisDrone2019 and UAVDT benchmarks demonstrate the effectiveness of the proposed method. On VisDrone2019, LUD-YOLO-n improves mAP@0.5 from 31.2% to 38.2% and mAP@0.5:0.95 from 18.5% to 22.4% compared with YOLOv7-n. LUD-YOLO-s achieves an mAP@0.5 of 41.7%, outperforming YOLOv7-s by 8.9 percentage points. On UAVDT, LUD-YOLO achieves an mAP@0.5 of 68.2%, providing supplementary evidence of cross-dataset applicability on another UAV detection benchmark. The parameter count of LUD-YOLO-n increases by only 7.7%, from 6.89M to 7.42M, while its mAP@0.5 improves by 22.4% relative to the baseline. These results indicate that LUD-YOLO achieves a favorable balance between accuracy and computational cost, making it suitable for UAV platforms with limited computing resources.