Vision Transformer-based High Precision Semantic Segmentation for Radio SLAM

Abstract

In this paper, we propose a vision transformer-based semantic segmentation for radio simultaneous localization and mapping (SLAM). To take advantage of a high-level understanding of surroundings, radio SLAM can provide additional landmark features such as materials besides geometric information. It has been extensively studied to extract landmark attributes using radio cross section (RCS) of materials. However, it is challenging to construct a semantic map for more diverse environments. To address this issue, this paper proposes a vision transformer-based semantic segmentation using signal power maps to effectively classify object materials in more general situations. The proposed method achieves high precision of segmentation performance in dynamic scenarios with different placements, rotations, and signal power. By numerical evaluation, the proposed method is verified to attain more than 95% of intersection over union (IoU) in 3 materials and the background.