Optimization of SL-SLAM Algorithm Based on Nanodet + Epipolar Geometry Dual Verification

Abstract

SL-SLAM is a deep learning-enhanced visualinertial SLAM system based on ORB-SLAM3, which replaces traditional ORB features with the deep learning feature extractor SuperPoint and the matcher LightGlue, it performs well in lowlight, low-texture scenes, but its robustness is not very high for dynamic scenes. NanoDet is an ultra-lightweight target detection model that runs extremely fast. This article proposes to integrate NanoDet into the SL-SLAM system to detect dynamic features and perform weight reduction processing, and then use epipolar geometry to re-verify the weighted feature points (whether to remove or retain these feature points). The improved system was evaluated on four sequences in the TUM dataset. The root mean square error (RMSE) of absolute trajectory error (ATE) was reduced by $\mathbf{3 0 \%}$, 18% and 31% respectively compared to SL-SLAM in the walking_halfsphere, walking_xyz and walking_static sequences, while the RMSE was basically the same as that of SL-SLAM in sitting_rpy, which has almost no dynamic, these results are presented using a visual comparison of 2D and 6D trajectories.