Light-SLAM: A Robust Deep-Learning Visual SLAM System Based on LightGlue under Challenging Lighting Conditions

Zhiqi Zhao,Chang Wu,Xiaotong Kong,Zejie Lv,Xiaoqi Du,Qiyan Li

2024-05-10

Abstract:Simultaneous Localization and Mapping (SLAM) has become a critical technology for intelligent transportation systems and autonomous robots and is widely used in autonomous driving. However, traditional manual feature-based methods in challenging lighting environments make it difficult to ensure robustness and accuracy. Some deep learning-based methods show potential but still have significant drawbacks. To address this problem, we propose a novel hybrid system for visual SLAM based on the LightGlue deep learning network. It uses deep local feature descriptors to replace traditional hand-crafted features and a more efficient and accurate deep network to achieve fast and precise feature matching. Thus, we use the robustness of deep learning to improve the whole system. We have combined traditional geometry-based approaches to introduce a complete visual SLAM system for monocular, binocular, and RGB-D sensors. We thoroughly tested the proposed system on four public datasets: KITTI, EuRoC, TUM, and 4Season, as well as on actual campus scenes. The experimental results show that the proposed method exhibits better accuracy and robustness in adapting to low-light and strongly light-varying environments than traditional manual features and deep learning-based methods. It can also run on GPU in real time.

Computer Vision and Pattern Recognition,Machine Learning,Robotics

What problem does this paper attempt to address?

### Problems the Paper Aims to Solve

The paper aims to address the challenges of ensuring robustness and accuracy in visual SLAM (Simultaneous Localization and Mapping) under challenging lighting conditions, such as low-light environments and drastic changes in lighting. Specifically:

1. **Limitations of Traditional Feature Point Methods**: In low-light environments or under drastic lighting changes, traditional handcrafted feature point methods (such as ORB, SIFT, etc.) struggle to recognize and associate feature information.
2. **Shortcomings of Deep Learning Methods**: Although some deep learning-based methods have shown potential, they still have significant drawbacks, particularly performing poorly in unknown environments, and most methods have not been comprehensively evaluated for their robustness in different scenarios.

To address these issues, the authors propose a new hybrid system—Light-SLAM, which is based on the LightGlue deep learning network. This system utilizes deep local feature descriptors instead of traditional handcrafted feature points and achieves fast and accurate feature matching through a more efficient and precise deep network. Experimental results show that Light-SLAM demonstrates better accuracy and robustness in low-light and drastically changing lighting environments.