基于端到端学习的道路配准与匹配算法及其在非GPS无人机定位中的应用

Registration and matching of road landmarks with aerial images are the critical techniques to achieve accurate UAV (Unmanned Aircraft Vehicle) localization in GPS-denied urban environments. Designing high-precision road registration and matching algorithms can help UAV more accurately estimate its PVA（Position, Velocity and Altitude）state, which are expected to promote UVA technology to more GPS-denied scenarios. Based on the fact that the traditional multi-stage road recognition algorithms using hand-crafted features suffer from low precision, low computing efficiency and low noise resistance, this research project proposes a novel end-to-end-learning framework for road registration and matching, aiming to build the mapping directly from the input space to the output space. From the aspects of theory, the cross-domain deep feature embedding between binary road images and gray aerial images, high-precision road registration based on saliency map and geometric consistency, as well as road matching based on spatial feature transformation are intended to systematically studied. Finally, a simulation platform will be built to test the precision and efficiency of the road landmark-based UAV localization algorithm. This research project is intended to provide a novel technical approach for GPS-denied UAV localization. In addition, we quantitatively establish the relationship between road matching accuracy and UAV flight height, which provide the theory basis and guidance for using the vector road map data more efficiency in future.

道路航标与航拍图像之间的配准与匹配是实现GPS不可用环境下无人机自主定位的关键技术，建立高精度道路航标配准与匹配算法有助于精确解算无人机位姿状态，从而有望将无人机技术推广到更多GPS不可用的应用场景。本课题针对传统基于手工设计特征的多级联道路识别算法存在鲁棒性差、精度低、时效性差等诸多不足，本课题提出基于端到端学习的道路航标配准与匹配新方案，力图采用深度学习技术直接建立从输入空间到决策空间的映射关系, 从理论方面对跨域深度特征表达、融合视觉显著性与几何一致性检测的道路航标配准、以及基于特征空间变换的道路航标匹配技术进行研究，最后通过构建实验平台验证基于道路航标匹配的无人机定位算法的精度及效率。本课题的研究将为无人机定位中的道路航标配准与匹配研究提供新思路、新方法。另外，本课题对道路航标匹配精度与飞行高度之间的相关性进行定量分析，为未来更高效地使用矢量道路地图数据辅助无人机定位提供理论依据。

本项目，以非GPS环境下无人飞行系统的自主定位为背景，采用深度学习技术，深入研究了基于端到端学习的道路航标配准与匹配理论技术，主要研究成果有：. （1）针对无人系统机载计算机计算能力及存储空间有限等问题，采用可分离卷积，构建了一种轻量级双通道卷积神经网络结构，实现了高效的跨域特征表达，在航标匹配任务上取得了优于现有算法的匹配精度。. （2）受传统基于特征点匹配的图像配准算法启发，通过构建匹配层衡量深度特征点间的相似度，并设计双分支注意力模块筛选有效匹配对，提出一种高效的道路航标配准算法，有效解决了大尺度旋转、平移下的道路航拍配准问题。. （3）针对航拍图像中通常存在大量运动目标，容易导致航标匹配及配准精度下降等问题，创新性的从图像修复角度出发，提出一种由粗到精的动静场景翻译模型，有效提升了复杂动态环境下无人系统的定位精度。. 本项目所研究的新颖的道路航标配准及匹配算法，可有效增强GPS不可用环境下无人系统的自主定位导航能力，为无人飞行系统的视觉导航提供了新的理论基础和技术参考。