DR-TC-SLAM: a dynamic-interference-aware tightly coupled visual-LiDAR-inertial SLAM framework

Abstract

In dynamic environments, the localization accuracy and mapping consistency of robotic systems may degrade when visual and LiDAR measurements are contaminated by moving objects. To improve the reliability of multi-modal state estimation under such interference, this paper presents DR-TC-simultaneous localization and mapping (SLAM), a dynamic-interference-aware tightly coupled visual-LiDAR-inertial SLAM framework. The system contains a dynamic-aware visual frontend that combines a multi-scale SuperPoint feature extractor with a YOLOv11-assisted extended Kalman filter tracker to reduce the influence of features located on moving objects. To enhance geometric reliability, the RG-GLO module constructs LiDAR odometry and loop-closure constraints through feature-based registration and a robust loop registration pipeline consisting of maximum clique inlier selection, graduated non-convexity rotation estimation, and covariance-weighted translation estimation. These visual, LiDAR, inertial, and loop constraints are incorporated into a unified factor-graph optimizer. Experiments on KITTI, M2DGR, and custom UAV datasets show that the proposed framework improves global trajectory accuracy and mapping consistency in the tested dynamic-interference scenarios. Meanwhile, the results also indicate that aggressive dynamic feature removal may weaken short-term constraints in loop-free sequences, which motivates the additional analysis provided in this paper.