Oriented object detection in aerial and remote sensing images poses significant challenges due to arbitrary rotations, scale variations, and complex backgrounds. This paper presents a novel detection framework DAR-Det, with four key technical innovations to address these challenges. First, the paper proposes a Swin Transformer backbone architecture enhanced with position-sensitive attention mechanisms for capturing long-range dependencies. Second, the paper designs a Dynamic Attention Module (DAM), which adaptively recalibrates feature responses based on oriented object characteristics. Third, the paper presents a BiFPN architecture with learnable weighted connections, enabling robust multi-scale feature fusion. Fourth, a Deformable Convolution enhanced Rotated Region Proposal Network (DC-RRPN) is developed, leveraging deformable convolutions for precise angle prediction. Through extensive ablation studies and comprehensive experiments on the DOTA-v1.5, the experimental results demonstrate that the approach achieves state-of-the-art performance across various scales and angles. Notably, DAR-Det achieves an impressive 80.45% mAP on DOTA-v1.5, surpassing many of the current methods by a significant margin.