LIG-SLAM: A Lightweight Visual RGB-D SLAM for Indoor Dynamic Environments Leveraging Instance Segmentation and Geometric Information

Traditional visual Simultaneous Localization and Mapping (SLAM) systems achieve high accuracy in static environments. However, in indoor dynamic scenes with frequent object motions, the presence of moving objects severely violates the scene rigidity assumption, often leading to significant performance degradation and tracking instability. To explicitly address this challenge, this paper introduces LIG-SLAM, a resource-efficient visual SLAM solution that extends the ORB-SLAM3 architecture. By incorporating dynamic object perception and geometric constraints, the system achieves robust localization in dynamic indoor environments, while its inference efficiency is significantly enhanced through targeted optimization. Specifically, a YOLOv5-based instance segmentation network is employed to obtain pixel-level segmentation of dynamic regions. To mitigate the erroneous rejection of static feature points, epipolar geometric constraints are incorporated to improve the accuracy of dynamic feature selection. Furthermore, a RANSAC-based depth consistency check is adopted to further enhance accuracy and alleviate the effects of epipolar degeneracy. Unlike conventional semantic SLAM frameworks, the proposed system incorporates ONNX-based optimization, thereby accelerating inference and improving real-time performance. Empirical evaluations conducted on TUM dynamic datasets indicate that the developed approach surpasses ORB-SLAM3 by a substantial margin, achieving a reduction of over 90% in terms of the Absolute Trajectory Error (ATE). Compared with existing semantic SLAM approaches, it achieves improvements in both accuracy and real-time performance, particularly in challenging indoor dynamic scenarios.