| 摘要: |
| 针对复杂电磁作战环境下无人机自主着陆应用场景，提出了一种基于图像语义分割的机场跑道检测算法，并设计构建了轻量高效的端到端跑道检测神经网络RunwayNet。在特征提取部分，使用空洞卷积对ShuffleNet V2进行改造，得到了输出特征图分辨率可调的主干网络，并利用自注意力机制设计了自注意力网络模块，使网络具备全局跑道特征提取能力。设计了解码器模块将网络浅层丰富的细节、空间位置信息与顶层粗略、抽象的语义分割信息相融合，从而获得精细的跑道检测输出结果。实验结果表明，RunwayNet网络在无人机着陆全过程都可以对跑道区域进行精准的分割识别，并且在嵌入式计算平台上能达到接近实时的处理速度，具有很强的实用价值。 |
| 关键词: 语义分割 机场跑道检测 自注意力模块 主干网络 |
| DOI： |
|
| 基金项目:国家自然科学基金重大科研仪器研制项目（41527803） |
|
| An Airport Runway Detection Algorithm Based on Semantic Segmentation |
| WANG Liu-jun,JIANG Hai-tao,LIU Chong-liang,PEI Xin-kai,QIU Hong-bo |
| (Beijing Institute of Automatic Control Equipment, Beijing 100074, China;The Third Representative Office of the Naval Equipment Department in Beijing Area, Beijing 100074, China) |
| Abstract: |
| For the application scenarios of autonomous landing of UAVs in complex electromag-netic combat environments, an airport runway detection algorithm based on image semantic segmentation is proposed, and a lightweight and efficient end-to-end runway detection neural network called RunwayNet is designed and constructed. In the feature extraction part, ShuffleNet V2 is modified by using atrous convolution to obtain a backbone network with adjustable output feature map resolution. Self-attention module is designed using the self-attention mechanism to make the network capable of global runway feature extraction. And the decoder module is designed to fuse the rich details, the spatial location information of the low-level layers, and the rough, abstract semantic segmentation information of the high-level layers to obtain a fine runway detection output. The experimental results show that RunwayNet can accurately segment the runway area during the entire landing of the UAVs, and can achieve near real-time processing speed on the embedded computing platform, which has strong practical value. |
| Key words: Semantic segmentation Airport runway detection Self-attention module Backbone network |
