A Semantic-Assisted LiDAR SLAM Based on Attention Feature Points

With the rapid development of deep learning, more researchers focus on leveraging richer semantic information to assist SLAM systems. However, due to the lack of color information and the sparsity of LiDAR, the semantic information obtained from it is limited. This paper proposes a semantic-assisted LiDAR SLAM framework based on features. We utilize semantic segmentation results to construct an adaptive weight for enabling the LiDAR odometry to focus on high-quality feature points, and the ground constraint is constructed by the ground normal fitted with ground points. In the mapping process, we use the high-quality feature points filtered by the adaptive weight to construct lightweight map. Additionally, due to the removal of ground points, the generated maps are navigation-friendly. We conducted comprehensive experiments on the KITTI public dataset that compared with other state-of-the-art methods, our method achieves better accuracy.