摘要
针对无人机全自主飞行对目标检测的实时性与准确性需求不断提升的现状,对现有YOLOv4网络进行优化,提出采用轻量型MobilenetV3网络取代原始模型中的主干特征提取网络,并在特征金字塔结构中利用深度可分离卷积模块取代传统卷积,实现了保证模型检测精度的同时减少模型参数的目的。通过采用CIOU位置回归损失函数,促使目标框回归变得更加稳定,采用的数据增强方法进一步提高了目标检测算法的鲁棒性。在相同配置条件下的对比实验结果表明,改进YOLOv4模型损失小幅精度却实现检测速度的大幅提升,其中参数容量减少82%,仅44.74 M,FPS提升69%并达到22帧/s,验证了所提算法的有效性。
In view of the increasing demand for the real-time property and accuracy of target detection in the full autonomous flight of UAV,the existing YOLOv4 network is optimized,the lightweight MobilenetV3 network is proposed to be used to replace the backbone feature extraction network in the original model,and the deep separable convolution module is used to replace the traditional convolution in the feature pyramid structure,so as to ensure the accuracy of model detection and reduce the model parameters at the same time.By using the CIOU position regression loss function,the target frame regression becomes more stable,and the data enhancement method further improves the robustness of the target detection algorithm.The comparative experimental results under the same configuration conditions show that the improved YOLOv4 model has a small loss of accuracy,but the detection speed is greatly improved,in which the parameter capacity is reduced by 82%,to only 44.74 M,and the FPS is increased by 69%to 22 frames/s,which verifies the effectiveness of the proposed algorithm.
作者
孙伟
潘森
黄恒
SUN Wei;PAN Sen;HUANG Heng(School of Geomatis and Geoscience,Liaoning Technical University,Fuxin Liaoning 123000,China)
出处
《传感技术学报》
CAS
CSCD
北大核心
2023年第3期456-461,共6页
Chinese Journal of Sensors and Actuators
基金
2019辽宁省“兴辽英才计划”青年拔尖人才(XLYC1907064)
辽宁工程技术大学学科创新团队资助项目(LNTU20TD-06)
2018年度辽宁省“百千万人才工程”人选科技活动资助项目(辽百千万立项[2019]45号)
2021辽宁省普通高等学校本科教学改革研究项目(辽教办[2021]254号)。
