基于YOLOv5的单目视觉无人机检测与定位方法

Monocular vision detection and localization method of UAV based on YOLOv5

针对小型无人机的目标检测和定位问题,提出一种单目视觉检测与定位方法。首先,在YOLOv5目标检测算法中嵌入双注意力机制模块,通过对特征图在通道和空间层面进行重新加权,增强目标区域中小尺度无人机的细节特征权重。然后,引入考虑干扰的孪生区域提议网络(DaSiamRPN)作为动态目标跟踪方法,解决检测目标的运动模糊和遮挡问题。最后,在相机运动已知的条件下,利用运动前后两幅图像中匹配的目标检测框对,进行在线的无人机目标尺寸测量,进而实现目标定位。仿真结果表明,所提目标检测算法准确率达到98.3%,召回率为97.2%,优于原有YOLOv5算法,验证了所提算法的可行性。

A monocular vision-based detection and localization method is proposed to solve problems of target detection and localization for small UAVs.Firstly,a dual attention mechanism module is embedded in the YOLOv5 object detection algorithm.The weight of the detail features of small-scale UAVs in the region of interest is enhanced by re-weighting the feature map at the channel and spatial levels.Secondly,the DaSiamRPN is introduced as a dynamic target tracking method to solve the motion blur and occlusion of detected targets.Finally,if the camera's ego-motion has been known,the scale of the UAV target is measured online by the matched detection bounding box pair in images before and after the motion.The target localization is achieved by the measured target scale.The simulation results show that the proposed target detection algorithm has an accuracy rate of 98.3%and a recall rate of 97.2%,which is better than the original YOLOv5 algorithm.The simulation test results verify the feasibility of the proposed positioning algorithm.