Research on Low-Light Environment SLAM System via Retinex Enhancement and Deep Feature Optimization

With the rapid advancements in autonomous driving, robot navigation, and related fields, Visual simultaneous localization and mapping (Visual SLAM) technology plays a vital role in real-time positioning and map creation. However, in low-light environments such as coal mines, tunnels, and underground parking lots, challenges like poor image quality, sparse feature points, and color distortion significantly hinder extracting and matching these feature points in Visual SLAM. As a result, the positioning performance in such environments declines sharply. To this end, this article proposes an improved SLAM method that integrates deep learning techniques, specifically targeting the issues of feature point extraction and matching in low-light environments. The core idea of this method is to introduce Retinexformer, which is based on the Retinex principle, at the front end to enhance low-light images based on the ORB-SLAM3 algorithm framework. The proposed approach improves image clarity and contrast by preprocessing the input images, enhancing the SLAM system’s perception in low-light conditions. Furthermore, to address the issue of sparse feature points in low-light environments, this article proposes an efficient feature extraction and matching module, which further improves the map construction and positioning accuracy of the SLAM system while improving computational efficiency. We conducted extensive experiments on public datasets and real low-light scenarios. The results demonstrate that the proposed algorithm exhibits greater robustness and higher accuracy in low-light environments than traditional SLAM algorithms. The proposed algorithm effectively improves SLAM systems’ perception and positioning performance in low-light conditions by enhancing image quality and strengthening feature point processing capabilities.