| 基于机器视觉的无人机避障系统研究 |
| |
| 引用本文： | 杨娟娟,高晓阳,李红岭,贾尚云.基于机器视觉的无人机避障系统研究[J].中国农机化学报,2020(2):155-160. |
| |
| 作者姓名： | 杨娟娟 高晓阳 李红岭 贾尚云 |
| |
| 作者单位： | 甘肃农业大学机电工程学院 |
| |
| 基金项目： | 国家自然科学基金项目(61164003)。 |
| |
| 摘 要： | 无人机避障不及时造成的人员伤亡及财产损失是阻碍无人机发展应用的重要原因之一,实时性好、准确率高的避障系统可降低无人机的运行风险。提出基于目标检测的智能避障系统,以one stage与two stage目标检测方法相结合的方式改进目标检测模型YOLOv3。其中,障碍物检测分三部分完成:基于darknet-53进行三个不同尺度的特征提取、RPN根据ground truth筛选感兴趣区域和yolo层多尺度特征融合预测障碍物的位置和分类。然后,在该文数据集的基础上将训练好的障碍物检测模型进行测试,测试结果表明:改进模型的障碍物检测速率为25帧/s,mAP为95.52%,与现有的目标检测模对比结果表明:本研究改进的目标检测智能避障算法,比Faster R-CNN的mAP提高17.2%,检测速率加快14个FPS;并在保证实时性的同时,mAP比YOLO2提高23.3%,比YOLOv3提高6.25%。最后,将目标检测模型应用于无人机避障系统中提出实现方案,进一步为无人机安全运行提供新的方法。
|
| 关 键 词： | 目标检测 YOLO 无人机 RPN 卷积神经网络 |
Research on UAV obstacle avoidance system based on machine vision |
| |
| Yang Juanjuan,Gao Xiaoyang,Li Hongling,Jia Shangyun.Research on UAV obstacle avoidance system based on machine vision[J].Chinese Agricultural Mechanization,2020(2):155-160. |
| |
| Authors: | Yang Juanjuan Gao Xiaoyang Li Hongling Jia Shangyun |
| |
| Institution: | (College of Mechanical and Electrical Engineering,Gansu Agricultural University,Lanzhou,730070,China) |
| |
| Abstract: | The casualties and property losses caused by the delay of UAV obstacle avoidance are one of the important reasons hindering the development and application of UAV. The real-time and high accuracy obstacle avoidance system can reduce the operational risk of UAV. This paper proposes an intelligent obstacle avoidance system based on object detection, which improves the object detection model YOLOv3 by combining one stage with two stages. Among them, obstacle detection is completed in three parts: feature extraction based on darknet-53 at three different scales, RPN screening regions of interest and Yolo layer multi-scale feature fusion based on ground truth to predict the location and classification of obstacles. Then, on the basis of this data set, the trained obstacle detection model is tested. The test results show that the obstacle detection rate of the improved model is 25 frames/s and the mAP is 95.52%. Compared with the existing target detection model, the improved intelligent obstacle avoidance algorithm of target detection in this study is 17.2% higher than the mAP of Faster R-CNN, and the detection rate is 14 FPS faster and guaranteed. At the same time, the real-time performance of mAP is 23.3% higher than YOLO2 and 6.25% higher than YOLOv3. Finally, proposed the realization scheme that the object detection model is applied to the obstacle avoidance system of UAV, which provides a new method for the safe operation of UAV. |
| |
| Keywords: | object detection YOLO UAV RPN Convolution Neural Network |
| 本文献已被 维普 等数据库收录！ |
|
