UNF-SLAM: Unsupervised Feature Extraction Network for Visual-Laser Fusion SLAM

In this paper, we propose a new visual-laser fusion SLAM. Firstly, the dense depth map is obtained from the sparse laser information by depth completion. Then, a network that can extract fused features based on visual and laser information is proposed. After the depth completion which can complete depth and generate pixel-level confidence maps from lidar, RGB frames and completion confidence maps are fed into the network to create feature points that utilize more reliable information from lidar and are better for tracking. The Unsupervised Network for Fusion SLAM (UNF-SLAM) provides a new method of using a learning network for tight coupling of multi-source information. Experiments are conducted on the KITTI dataset and real scenarios. The results show that the proposed SLAM can achieve high accuracy and great robustness. Source code is available at https://github.com/salt0107fish/UNF_SLAM-master.