A Temporal and Self-Attention Based Method for Small Object Detection in UAV Imagery

Unmanned aerial vehicle (UAV) images are widely applied in real-world applications. However, traditional object detection methods have encountered a sharp performance drop on UAV images, largely due to the presence of numerous small objects with low resolution, sparse textures, and blurred features. In this paper, we propose a novel Adaptive Anchor Detection Network (AADN) that dynamically generate anchors through deep neural network. Subsequently the proposed Temporal And Self-Attention (TAA) module enhances the extracted image features by capturing temporal and internal correlation information. Furthermore, we introduce a data augmentation method based on category k-means clustering, designed to reduce irrelevant background regions and improve the network’s ability to distinguish objects with similar categories. Extensive experiments show the superior performance of AADN on both the VisDrone2019 and TinyPerson dataset, achieving an impressive 2% increase in mean average precision (mAP) compared to state-of-the-art methods.