SEGL-SLAM: A Visual SLAM With Segformer Segmentation and Line Features Enhancement in Dynamic Environments

摘要

Visual simultaneous localization and mapping (VSLAM) is a crucial technology for autonomous navigation in robotics and remains a prominent area of research. The presence of dynamic objects significantly impacts the localization accuracy of visual SLAM systems, particularly traditional systems that rely on the assumption of a static environment. These systems often struggle to estimate positions accurately when dynamic objects are present in the scene. To address this issue, we propose a dynamic RGB-D SLAM system called SEGL-SLAM, built upon ORB-SLAM3. Our approach integrates a semantic segmentation network to identify objects in dynamic environments, thus extracting semantic information for enhanced localization. We then apply the epipolar constraint method to accurately identify and remove dynamic object features, reducing their interference with both localization and mapping. After removing dynamic feature points, we extract line features from each frame using the LSD algorithm, compensates for the reduced number of features, preventing trajectory estimation errors or map interruptions that could result from the scarcity of feature points. Using both static features and semantic labels, SEGL-SLAM generates dense semantic maps. We evaluate the proposed algorithm on the TUM RGB-D dataset and in real-world dynamic environments. Experimental results demonstrate that our method effectively mitigates the impact of dynamic objects in complex scenarios, exhibiting superior robustness and improved localization accuracy.