摘要
针对无人机远距离跑道线检测时有效信息少且定位困难的问题,提出了一种基于并行反向注意网络的跑道线检测方法。并行反向注意网络采用Res2Net作为主干网络,首先采用并行融合编码器将低级特征与高级特征融合从而获取跑道线的初始轮廓图。在此基础上,融合通道特征金字塔和轴向反向注意力机制来增强图像中的全局和局部特征信息的表达能力。基于无人机着陆图像数据集的仿真试验结果表明所提出的算法有效地检测出跑道线,图像语义分割平均交并比达到86.3%,单帧处理时间25 ms,对于远距离小目标检测有明显的优势。
Aiming at the problem that there is little effective information and difficult positioning when the long-distance runway line of the UAV is detected,a parallel anti-attention network is proposed.The network uses Res2Net as the backbone network,and first of all,the low-level features are fused with the high-level features by using a parallel fusion encoder to obtain the initial contour diagram of the runway line.On this basis,the channel feature pyramid and the axial reverse attention mechanism are fused to detect global and local feature information in the image to enhance the expression ability of the feature.The simulation results based on UAV landing image dataset show that the proposed algorithm can effectively detect the runway line,the MIoU reaches 86.3%,and the processing time of single frame is 25 ms,which has obvious advantages for the detection of long-distance small targets.
作者
白俊卿
张文静
BAI Junqing;ZHANG Wenjing(Xi’an Shiyou University,Xi’an 710065,China)
出处
《中国惯性技术学报》
EI
CSCD
北大核心
2022年第5期609-614,共6页
Journal of Chinese Inertial Technology
基金
陕西省重点研发计划项目(2022GY-031)
陕西省教育厅科研计划项目(22JK0503)。
