Robustness Analysis of Visual SLAM Algorithms with Neural Implicit Map Representation

The study analyzes the robustness of implicit maps to external and internal disturbances in the task of simultaneous localization and mapping (SLAM) based on images from a depth camera (RGB-D). Using Co-SLAM as a representative method, we examined the sensitivity to three realistic perturbations: local bright spots and glare, inaccuracies in the calibration of camera intrinsics, and additive noise in the depth channel. For quantitative robustness evaluation, the root mean square error of the camera position is used after data alignment using the Kabsch-Umeyama algorithm. The results of the simulations with disturbances showed high robustness to local light flares, moderate to small inaccuracies in camera calibration, and low to noise in the depth channel. This study also proposes a new method for determining the initial position and orientation of the camera based on the linear and angular velocities from visual/visual-inertial odometry algorithm, which provides increased accuracy of the localization in the situation with noisy depth channel without a significant increase in computational complexity.