MAD-YOLO: multi-dimensional attention and dynamic feature aggregation for UAV small object detection

Abstract

Object detection in UAV aerial imagery presents significant challenges due to the small size of targets, dense spatial distributions, and complex background interference, all of which often result in missed detections and unstable feature representations. To address these challenges, this paper proposes MAD-YOLO, an efficient object detection framework aimed at enhancing small-object perception and multi-scale feature interaction in complex aerial scenes. The proposed framework adopts the existing mixed local channel attention mechanism and embeds it into the C2PSA architecture to enhance local-global feature representation. Additionally, an information-aware ADown-style module is incorporated to alleviate downsampling-induced feature degradation, and a high-resolution P2 detection layer is introduced to optimize the utilization of shallow spatial features for small-object detection. Furthermore, a lightweight dynamic feature aggregation head is developed to adaptively model scale-aware, spatial-aware, and task-aware feature dependencies, thereby improving robustness across varying object distributions. Experimental results on the VisDrone2019 dataset demonstrate that MAD-YOLO(n) achieves significant performance enhancements over YOLOv11n. Specifically, mean Average Precision (mAP) at 0.5 and mAP at 0.5:0.95 increase by 9.4 and 6.0 percentage points, respectively, with an increase of 0.21 M parameters compared with YOLOv11n. Although the proposed model introduces additional computational overhead, it achieves 37.0 FPS with an average latency of 27.0 ms on the tested hardware platform, demonstrating a favorable accuracy–efficiency trade-off. Precision and Recall also improve by 11.5 and 6.3 percentage points, respectively, indicating enhanced detection capabilities for small and densely distributed targets. Moreover, cross-dataset evaluations on the RS-STOD and TinyPerson benchmarks further validate the robustness and generalization capabilities of the proposed framework across diverse UAV imaging conditions.