To address the issue of declining detection accuracy caused by arbitrary object rotation, complex backgrounds, and small target sizes in UAV aerial images, this paper proposes YOLO-MAT, a lightweight object detection network based on the YOLOv12 architecture that integrates rotation-aware attention and adaptive feature filtering. First, a multi-path rotation-aware attention C2f module (MRAC2f) is proposed, which enhances the model's robustness in representing rotated targets through a rotation-invariant attention mechanism (RAM). Second, an adaptive multi-scale-feature filtering and fusion module (AMFF) is designed, which integrates a dual domain collaborative attention (DDCA) mechanism and a Laplace edge enhancer (LEE) to suppress background noise in shallow features while enhancing high-frequency details, and employs learnable weights to achieve adaptive fusion of multi-scale features. Third, a high-resolution tiny head is constructed to further improve the detection performance for small targets. Experimental results on the VisDrone2019 and NWPU VHR-10 datasets demonstrate that compared to the baseline model YOLOv12, the YOLO-MAT reduces the number of parameters by 3.2% while increasing the mean average precision (mAP) by 6.7% and 9.3%, respectively, achieving an effective balance between lightweight design and detection accuracy. Compared with other mainstream detection algorithms, the YOLO-MAT exhibits clear advantages in detection accuracy, offering an efficient solution for real-time UAV object detection tasks.