| 摘要: |
| 无人机航拍场景由于目标尺度差异大、分布密集且部分目标特征不明显，导致通用的目标检测算法难以有效捕获到目标区域，检测精度低。为解决该问题，基于YOLOv8s提出一种改进的无人机航拍目标检测方法。通过引入可变形注意力机制，实现感受野自适应调整，提升模型对目标几何形状的感知；使用尺度特征交互模块进行特征增强，动态聚合特征局部与全局信息，增强模型的特征表达能力；进一步采用EIoU损失函数，联合优化目标位置与形状，降低漏检和误检率。实验结果表明，在公开数据集VisDrone2019上，该算法相较于YOLOv8s，精确率提升2.5%，召回率提升1.7%，平均精度mAP提升2.3%；对比其他主流模型也拥有更好的检测性能。通过可视化效果验证，该算法在密集场景、多尺度目标和夜间场景下检测能力均有提升。 |
| 关键词: 无人机航拍场景目标检测 可变形注意力 尺度特征交互 |
| DOI：10.20079/j.issn.1001-893x.250326001 |
|
| 基金项目: |
|
| An Improved YOLOv8s-based Object Detection Method for Unmanned Aerial Vehicle Aerial Scenarios |
| ZHANG Min |
| (1.Southwest China Institute of Electronic Technology，Chengdu 610036，China;2.National Key Laboratory of Complex Aviation System Simulation，Chengdu 610036，China) |
| Abstract: |
| Object detection in unmanned aerial vehicle(UAV) aerial scenarios is challenging due to the wide variation in object scales，dense distributions，and some object features are not obvious.General object detection algorithms struggle to accurately capture the object regions in such scenes，leading to low detection accuracy.To address this problem，an improved object detection method based on YOLOv8s is proposed for UAV aerial scenarios.By introducing a deformable attention mechanism，the receptive field is adaptively adjusted to improve the model’s perception of object geometry.Feature enhancement is achieved through the attention-based intra-scale feature interaction module，which dynamically integrates local and global feature information，strengthening the model’s feature representation.Additionally，the EIoU loss is adopted to jointly optimize object position and shape，reducing false positives and missed detections.Experimental results on the VisDrone2019 dataset demonstrate significant improvements，with precision increased by 2.5%，recall improved by 1.7%，and mean average precision(mAP) enhanced by 2.3% compared with YOLOv8s.Additionally，the proposed method outperforms other state-of-the-art models.Through the visualization effect verification，the proposed method shows superior performance in dense scenes，multi-scale objects and night scenes. |
| Key words: UAV aerial scenarios object detection deformable attention mechanism attention-based intra-scale feature interaction |
《电讯技术》编辑部 