AE-DETR: Altitude-Environment Aware Vehicle Detection for UAVs

To address the challenges of UAV object detection, including feature degradation in low-light or hazy conditions, cross-scale conflicts from dynamic viewpoints, and insufficient feature adaptation to altitude variations, this paper proposes the AE-DETR framework with three core innovations: 1) DATIFI Module: Replaces conventional attention mechanisms with deformable sparse attention to enhance blurred contour capture in foggy or nighttime scenarios. 2) LHSFPN Module: Optimizes cross-scale feature fusion through dynamic weighting strategies, effectively suppressing background interference caused by abrupt viewpoint changes. 3) DRepC3 Module: Adaptively adjusts receptive fields via multi-branch dilated convolutions to accommodate flight altitude variations. On the refined UAVDT dataset, the proposed model achieves state-of-the-art detection accuracy with $37.3 \%$ mAP50-95, marking a $13.0 \%$ improvement over the baseline RT-DETR. Notably, it demonstrates superior performance in high-altitude operations and complex illumination environments, validating its robustness and adaptability in UAV-based perception tasks.