Semantic SLAM for Mobile Robot with Human-in-the-Loop

Mobile robots are an important participant in today’s modern life, and have huge commercial application prospects in the fields of unmanned security inspection, logistics, express delivery, cleaning and medical disinfection. Since LiDAR is not affected by ambient light and can operate in a dark environment, localization and navigation based on LiDAR point clouds have become one of the basic modules of mobile robots. However, compared with traditional binocular vision images, the sparse, disordered and noisy point cloud poses a challenge to efficient and stable feature extraction. This makes the LiDAR-based SLAM have more significant cumulative errors, and poor consistency of the final map, which affects tasks such as positioning based on the prior point cloud map. In order to alleviate the above problems and improve the positioning accuracy, a semantic SLAM with human-in-the-loop is proposed. First, the interactive SLAM is introduced to optimize the point cloud pose to obtain a highly consistent point cloud map; then the point cloud segmentation model is trained by artificial semantic annotation to obtain the semantic information of a single frame of point cloud; finally, the positioning accuracy is optimized based on the point cloud semantics. The proposed system is validated on the local platform in an underground garage, without involving GPS or expensive measuring equipment.