融合SuperGlue方法的深度单目视觉惯性导航算法

A deep monocular visual-inertial navigation algorithm with SuperGlue

针对图像的深度学习方法是解决传统视觉定位算法在复杂环境下特征提取不稳定、跟踪丢失等难题的有效途径。本文在VINS-Mono的基础上引入基于深度学习的SuperPoint特征提取方法和SuperGlue特征匹配方法,建立了一种融合SuperGlue方法的单目视觉惯性导航算法,并采用开源数据集和实际试验数据进行了评估。结果表明,该方法有效提升了复杂环境下单目视觉惯性算法的稳定性和精度,精度提升幅度可达26%。

The deep learning method for images is an effective way to solve the problems of unstable feature extraction and tracking loss of traditional visual positioning algorithms in complex environments.In this paper,we propose a visual-inertial navigation algorithm based on VINS-Mono,which using SuperPoint to get feature points and track them by using SuperGlue.And evaluate it using Open-source dataset and real world experiments.Experimental results show that our algorithm has a significant improvement in positioning accuracy and stability compared with VINS-Mono,and the accuracy improvement can reach 26%.