基于图像语义分割的无人机自主着陆导航方法

Image semantic segmentation-based navigation method for UAV auto-landing

针对复杂电磁作战环境下无人机自主着陆应用场景,提出了一种基于深度卷积神经网络图像语义分割的无人机自主着陆导航方法。首先设计了轻量高效的端到端跑道分割神经网络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;然后设计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的解码器模块,获取精细的跑道分割输出结果;最后设计了基于跑道分割区域的边线提取和位姿解算算法,实现相对位姿信息的解算。仿真和机载飞行实验结果表明,基于嵌入式实时计算平台可实现无人机着陆全过程跑道区域的精准分割识别,作用距离达到3 km、成功率接近90%,解决了着陆过程中跑道识别盲区和实时性等问题,显著提高了复杂环境下无人机着陆的鲁棒性。

A UAV auto-landing navigation method based on deep convolutional neural network image semantic segmentation is proposed for the application scenarios of UAVs auto-landing in complex electromagnetic combat environments. Firstly, a lightweight and efficient end-to-end runway detection neural network named RunwayNet is designed. In the feature extraction part ShuffleNet V2 is reformed by using void convolution to get a trunk network with adjustable output feature graph resolution. A self-attention module based on the self-attention mechanism is designed so that the network has global runway feature extraction capabilities. Secondly, a decoder module is designed by fusing the rich details, the spatial location information of the low-level layers with the rough, abstract semantic segmentation information of the high-level layers to obtain a fine runway detection output. Finally, an algorithm of edge line extraction and pose estimation based on the segmented area of runway is proposed to realize relative pose calculation. The results of simulations and airborn experiments show that the precise segmentation and recognition of the runway area during the landing of the drone can be realized by the embedded real-time computing platform. The operating distance can reach 3 km and the success rate is close to 90%. The problems of runway identification blind area and real time in the landing process is solved, and the robustness of UAV landing in complex environment is significantly improved.