Paper ID: 2407.21348

SuperVINS: A visual-inertial SLAM framework integrated deep learning features

Hongkun Luo, Chi Guo, Yang Liu, Zengke Li

In this article, we propose enhancements to VINS-Fusion by incorporating deep learning features and deep learning matching methods. We implemented the training of deep learning feature bag of words and utilized these features for loop closure detection. Additionally, we introduce the RANSAC algorithm in the deep learning feature matching module to optimize matching. SuperVINS, an improved version of VINS-Fusion, outperforms it in terms of positioning accuracy, robustness, and more. Particularly in challenging scenarios like low illumination and rapid jitter, traditional geometric features fail to fully exploit image information, whereas deep learning features excel at capturing image features.To validate our proposed improvement scheme, we conducted experiments using open source datasets. We performed a comprehensive analysis of the experimental results from both qualitative and quantitative perspectives. The results demonstrate the feasibility and effectiveness of this deep learning-based approach for SLAM systems.To foster knowledge exchange in this field, we have made the code for this article publicly available. You can find the code at this link: https://github.com/luohongk/SuperVINS.

Submitted: Jul 31, 2024