An Improved YOLOv8 UAV Target Detection Algorithm Based on Enhancing Feature Fusion and Extraction

To address the challenges in UAV-view target detection tasks-such as the insufficient feature extraction capability of existing models and their difficulty in adapting to complex scenarios like rapid changes in target size and position, as well as blurring and occlusion. This paper proposes a UAV target detection method based on an improved YOLOv8 model. This method integrates the Convolution and Attention Fusion Module (CAFM) with the C2f module of the YOLOv8 model to develop the C2f_AT module, which enhances the capture of both local and global features. Furthermore, this paper employs the Detect_ASFF detection head incorporated with the Adaptive Spatial Feature Fusion (ASFF) to further strengthen feature selection and boost model performance. Experimental results show that the improved model in this paper achieves better performance than the baseline on the MOT-FLY dataset. Specifically, the precision $(P)$ is increased by $2.7 \%$, the recall $(R)$ by $3.6 \%, \mathrm{mAP} \text{@} 0.5$ by $6 \%$, and $\mathrm{mAP} \text{@} 0.5: 0.95$ by $9.6 \%$, which effectively enhances the performance of UAV-view target detection in complex scenarios.