An end-to-end SLAM with Gaussian Splatting in low light environments

Purpose In recent years, three-dimensional (3D) Gauss-based visual simultaneous localization and mapping (SLAM) technology has been developed rapidly. However, achieving realistic scene reconstruction in dimly lit environments remains a major challenge. To solve this problem, the purpose of this paper is to enhance the perceptibility of the environment by re-illuminating the scene, so as to build a 3D map that reflects the detailed density information of the scene. Design/methodology/approach This paper proposes an end-to-end tracker-based SLAM system for dimly lit environments, using 3D Gaussian Splatting as the core representation for effective scene modeling. This system enhances the perceptibility of the environment by re-illuminating the scene, thereby constructing a 3D map that reflects the detailed density information of the scene. Additionally, to improve the map representation, the system implements appropriate point densification and opacity resetting operations in error-prone 3D regions of the Gaussian map, further optimizing its accuracy and detail. Findings The algorithm in this paper is compared with other most advanced Gaussian SLAM algorithms on public data set ETH3D and real data set, and it gets better rendering effect and, finally, realizes the scene reconstruction after re-lighting in dark environment. Originality/value Compared to prior studies, this algorithm achieves superior performance in dimly lit environments through scene re-illuminating. The authors propose a method that identifies erroneous 3D regions and performs adaptive point densification and opacity resetting operations, jointly optimizing system errors and texture loss induced by relighting.