YOLOv11-Hunter: Detection of Threats by Unmanned Aerial Vehicles in Complex Environments

The detection of small unmanned aerial vehicles (UAVs) within complex environments poses considerable challenges, primarily due to their diminutive size, intricate structural features, and visual similarity to background elements. Conventional detection techniques often exhibit suboptimal performance under adverse conditions, while loss functions based on Intersection over Union (IoU) demonstrate sensitivity to variations in scale and position. To mitigate these limitations, this study introduces three principal contributions: (1) EMO, a framework that combines residual architectures with self-attention mechanisms to enable effective multi-scale feature extraction; (2) SimAM, a neuron-aware mechanism aimed at enhancing the refinement of spatial and channel-wise features; and (3) a novel loss function that integrates Normalized Wasserstein Distance (NWD) with IoU to improve robustness. Empirical evaluations reveal that the proposed approach substantially enhances detection performance, yielding a 3.7\((%)\) improvement in mean Average Precision at 50\((%)\) IoU (mAP@50) on the YOLOv11 model, achieving 87.8\((%)\), alongside increased recall rates on the DUT Anti-UAV dataset. Moreover, the method demonstrates consistent detection accuracy under diverse lighting conditions, highlighting its potential applicability in practical security-related scenarios.