Enhancing Concealed Drone Detection with Attention Mechanisms in RT-DETR

The detection of concealed drones in complex environments, such as urban areas or forests, remains a significant challenge for computer vision systems. While the Real-Time Detection Transformer (RT-DETR) achieves state-of-the-art performance in general object detection, its reliance on global self-attention may limit its effectiveness for targets that blend into the background. This work investigates the impact of integrating alternative attention mechanisms into the RT-DETR architecture to enhance the detection of concealed drones. We evaluate four distinct attention modules - MultiHead Self-Attention (MHSA), Convolutional Block Attention Module (CBAM), Window Attention (WA), and Local-Global Attention (LGA) - prior to the encoder, as well as Efficient Channel Attention (ECA) prior to the decoder. Using a test set biased towards camouflaged drones, we demonstrate that the placement and type of attention are critical to performance. Specifically, applying window attention to lower-level feature maps improves Average Precision (AP) by 1.5%, while the implementation of ECA at the encoder output yields a gain of 2.0% in AP. These findings suggest that for concealed drone detection, local and channel-selective attention mechanisms are superior to global self-attention, provided they are applied at semantically rich feature levels.