Visual simultaneous localization and mapping (vSLAM) algorithm based on improved Vision Transformer semantic segmentation in dynamic scenes

Identifying dynamic objects in dynamic scenes remains a challenge for traditional simultaneous localization and mapping (SLAM) algorithms. Additionally, these algorithms are not able to adequately inpaint the culling regions that result from excluding dynamic objects. In light of these challenges, this study proposes a novel visual SLAM (vSLAM) algorithm based on improved Vision Transformer semantic segmentation in dynamic scenes (VTD-SLAM), which leverages an improved Vision Transformer semantic segmentation technique to address these limitations. Specifically, VTD-SLAM utilizes a residual dual-pyramid backbone network to extract dynamic object region features and a multiclass feature transformer segmentation module to enhance the pixel weight of potential dynamic objects and to improve global semantic information for precise identification of potential dynamic objects. The method of multi-view geometry is applied to judge and remove the dynamic objects. Meanwhile, according to static information in the adjacent frames, the optimal nearest-neighbor pixel-matching method is applied to restore the static background, where the feature points are extracted for pose estimation. With validation in the public dataset TUM (The Entrepreneurial University Dataset) and real scenarios, the experimental results show that the root-mean-square error of the algorithm is reduced by 17.1 % compared with dynamic SLAM (DynaSLAM), which shows better map composition capability.