基于改进YOLOv5s的四旋翼自主降落标识检测算法

无人机自主降落是无人机领域研究的热点之一，导航信息在自主降落过程中又起到至关重要的作用，而视觉导航相较于传统导航方式可以提供更多环境信息，有利于提高无人机着陆安全性。当无人机飞行高度越高，机载相机捕获到的降落标识物就越小，为了提升无人机识别标识物的能力，基于YOLOv5s算法提出了一种改进的无人机实时小目标检测算法。首先，为了检测到更小尺度的目标在原算法基础上新增一个检测头；然后采用BiFPN代替原先PANet结构，提升不同尺度的检测效果；最后将EIoU Loss替换CIoU Loss作为算法的损失函数，在提高边界框回归速率的同时提高模型整体性能。将改进算法应用于无人机自主降落场景下的二维码降落标识检测，实验结果表明改进后的算法在小目标检测中相比于原始YOLOv5s算法的特征提取能力更强、检测精度更高，证明了改进算法的优越性。

Object Detection Based on Improved YOLOv5s for Quadrotor UAV Auto-Landing

Nowadays, the research of UAV auto-landing is quite hot. Navigation information plays an important role in the process of autonomous landing. And compared with traditional navigation methods, visual navigation can provide more environmental information, which is conducive to improving the landing safety of UAV. The higher the flying height of UAV, the smaller the landing marker captured by airborne camera. In order to improve the ability of UAV to recognize small landing marker, an algorithm based on yolov5s was proposed. Firstly, a prediction head was added to detect smaller targets. Secondly, the Bi-FPN was used to replace the PANet. Finally, EIoU Loss was used to replace CIoU Loss. The improved algorithm was applied to the landing marker detection. The results show that our algorithm has stronger feature extraction ability and higher detection accuracy in small object detection, which proves the superiority of the improved algorithm.