P‐4.9: A Study on Performance Improvement of Rover‐SLAM in Dynamic Environments Based on YOLOv8 Instance Segmentation
TL;DR
An improved VSLAM method with the fusion of deep learning and geometric principles is proposed: it integrates SuperPoint with LightGlue to enhance the stability of feature point extraction and matching, combines the YOLOv8 and Farneback algorithms to eliminate dynamic feature points, and optimizes both the velocity model and pose estimation strategy.
Abstract
Visual Simultaneous Localization and Mapping (VSLAM) is widely applied in fields such as robot navigation due to its low cost and intuitive detection performance. Nevertheless, classic VSLAM algorithms are built on the static world assumption, making them vulnerable to accuracy impairment induced by factors including dynamic objects. Based on the Rover‐SLAM framework, this paper proposes an improved VSLAM method with the fusion of deep learning and geometric principles: it integrates SuperPoint with LightGlue to enhance the stability of feature point extraction and matching, combines the YOLOv8 and Farneback algorithms to eliminate dynamic feature points, and optimizes both the velocity model and pose estimation strategy. Verified on the dynamic sequences of the TUM dataset, the proposed method effectively improves the localization accuracy in dynamic environments while preserving the realtime performance of the algorithm. This study has a limitation in that the real‐time performance is still affected to a certain extent; in future research, we will further optimize the universality and real‐time performance of the method.