基于自适应特征融合的无人机小目标检测算法
Uav Small Target Detection Algorithm Based on Adaptive Feature Fusion
赵滨淋 1陈功 1李胜2
作者信息
- 1. 成都信息工程大学通信工程学院,四川 成都 610200
- 2. 成都信息工程大学统计学院,四川 成都 610200
- 折叠
摘要
本文提出了一种针对无人机捕获图像的目标检测算法,旨在解决无人机视角下小目标检测存在漏检现象严重、检测精度低的问题.主要改进包括重新设计聚类算法生成更准确的先验框,引入自适应特征融合模块以使模型更灵活地学习上下文特征信息,更改检测头在较大的特征层上进行目标检测并解耦分类和回归任务.通过在VisDrone2019数据集上进行广泛实验,改进后的YOLOv5s模型相较于基准模型在mAP50上提高了5.8%,并且保持了较高的帧率(67 FPS).实验结果表明该改进方法能够显著提高模型的检测性能,使其适用于复杂的无人机捕获场景.
Abstract
This paper proposes a target detection algorithm for images captured by drones,aiming to solve the issues of serious missed detection and low detection accuracy for small objects from a drone's perspective.The main improvements include redesigning the clustering algorithm to generate more accurate prior boxes,introducing an adaptive feature fusion module to enable the model to more flexibly learn contextual feature information,and modifying the detection head to perform object detection on larger feature maps while decoupling classification and regression tasks.Through extensive experiments on the VisDrone2019 dataset,the improved YOLOv5s model showed a 5.8%increase in mAP50 over the baseline model and maintained a high frame rate(67 FPS).The experimental results demonstrate that the proposed improvements significantly enhance the model's detection performance,making it suitable for complex scenarios captured by drones.
关键词
小目标检测/检测头/特征融合/聚类算法Key words
Small target detection/detection head/feature fusion/clustering algorithm引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据