MSADM-DETR: Enhancing DETR with Multi-scale Semantic-Detail Alignment and Density-Aware Mechanisms for UAV Object Detection

Unmanned Aerial Vehicle Object Detection (UAV-OD) is widely used in areas such as public security, traffic management, and environmental monitoring. The DETR model removes traditional components like anchor boxes and Non-Maximum Suppression (NMS), simplifying the detection process and reducing manual design effort while enabling end-to-end object detection. However, such end-to-end models often show limited performance in complex UAV scenes with objects of different scales. To overcome these issues, this study proposes a new framework called MSADM-DETR (Multi-scale Semantic-Detail Alignment and Density-Aware Mechanisms) for UAV object detection. First, the Multi-scale Semantic-Detail Alignment (MSDA) module is designed to reduce object misalignment and detail loss during multi-scale feature fusion. Second, the Image-aware Object Density Estimation (IODE) module predicts the object density level in each image and adaptively adjusts the attention scale for different density conditions. Third, the Density-Aware Loss Weighting Mechanism (DA-Loss) dynamically adjusts the weights of classification and localization losses based on object density, improving the balance between tasks. Extensive experiments on the VisDrone2019 validation and test datasets show that MSADM-DETR achieves excellent mAP \(_{50}\) scores of 51.6% and 42.2%, respectively, outperforming existing state-of-the-art models.