Object Detection of Visdrone based on Attention Mechanism and FasterNet

Yun Bai,Bo Tao,Lei Kong,Yinchuan Wang,Linna Yang
DOI: https://doi.org/10.1109/CVIDL62147.2024.10603457
2024-04-19
Abstract:Object detection in Unmanned aerial vehicle (UAV) is an important foundation in various research fields. However, due to issues such as slow detection speed, more significant proportion of small targets, dense distribution, and instance overlap, drone target detection pose challenges. In this paper, an improved YOLOv8 architecture incorporating attention mechanism and FasterNet is proposed to enhance the detection performance of UVA images. First, FasterNet is applied in the backbone network of YOLOv8, which uses partial convolution (PConv) to better extract spatial features. Secondly, an attention mechanism called RepConv ShuffleNet (RCS)-based One-Shot Aggregation (RCS-OSA) module is employed to original neck which allows semantic information extraction. Experimental results on VisDrone dataset, indicate that the proposed method can effectively enhance the detection capability for drone targets, and upgrade the mean average precision by about $6 \%$.
Engineering,Computer Science
What problem does this paper attempt to address?