基于改进YOLOv5的Video SAR动目标检测算法  

Video SAR Moving Target Detection Algorithm Based on Improved YOLOv5

作　　者：白浩琦 李和平 BAI Haoqi;LI Heping

机构地区：[1]中国科学院空天信息创新研究院,北京100094 [2]中国科学院大学电子电气与通信工程学院,北京100049

出　　处：《科技创新与应用》2024年第26期54-59,共6页Technology Innovation and Application

摘　　要：视频合成孔径雷达图像中动目标阴影可反映其真实位置,针对目前利用深度学习检测阴影的方法来检测动目标存在检测概率低和虚警等问题,提出一种改进的YOLOv5模型。该算法以YOLOv5框架为基础,一方面添加小目标检测层、引入兼顾通道和空间的三维注意力机制,并调整损失函数计算方法为最小点距交并比方法,来提升网络对动目标阴影的感知能力;另一方面增加道路提取作为判定条件,来降低道路外侧静止物体阴影对检测动目标阴影的干扰。经实验验证,该算法与YOLOv5相比在平均精度上提高7.12%,满足动目标检测需求。The shadows of moving targets in video synthetic aperture radar(Video SAR)images can reflect their real positions,and an improved YOLOv5 model is proposed to address the problems of low detection probability and false alarms in the current method of detecting shadows using deep learning to detect moving targets.The algorithm is based on the YOLOv5 framework.On the one hand,it adds a small target detection layer,introduces a three-dimensional attention mechanism that takes into account the channel and space,and adjusts the loss function calculation method to the intersection over union based on minimum point distance method to improve the network's ability to detect shadows of moving targets;on the other hand,it adds the road extraction as a decision condition to reduce the interference of the shadows of static objects outside the road in detecting the shadows of moving targets.After experimental verification,the algorithm improves the average accuracy by 7.12%compared with YOLOv5,which meets the requirements of moving target detection.

关 键 词：视频SAR 运动目标检测 深度学习 YOLOv5 阴影检测 

分 类 号：TN951[电子电信—信号与信息处理]