工程科学学报 (Jun 2021)
Survey of simultaneous localization and mapping based on environmental semantic information
Abstract
The simultaneous localization and mapping (SLAM) technique is an important research direction in robotics. Although the traditional SLAM has reached a high level of real-time performance, major shortcomings still remain in its positioning accuracy and robustness. Using traditional SLAM, a geometric environment map can be constructed that can satisfy the pose estimation of robots. However, the interactive performance of this map is insufficient to support a robot in completing self-navigation and obstacle avoidance. One popular practical application of SLAM is to add semantic information by combining deep learning methods with SLAM. Systems that introduce environmental semantic information belong to semantic SLAM systems. Introduction of semantic information is of great significance for improving the positioning performance of a robot, optimizing the robustness of the robot system, and improving the scene-understanding ability of the robot. Semantic information improves recognition accuracy in complex scenes, which brings more optimization conditions for an odometer, pose estimation, and loop detection, etc. Therefore, positioning accuracy and robustness is improved. Moreover, semantic information aids in the promotion of data association from the traditional pixel level to the object level so that the perceived geometric environmental information can be assigned with semantic tags to obtain a high-level semantic map. This then aids a robot in understanding an autonomous environment and human–computer interaction. This paper summarized the latest researches that apply semantic information to SLAM. The prominent achievements of semantics combined with the traditional visual SLAM of localization and mapping were also discussed. In addition, the semantic SLAM was compared with the traditional SLAM in detail. Finally, future research topics of advanced semantic SLAM were explored. This study aims to serve as a guide for future researchers in applying semantic information to tackle localization and mapping problems.
Keywords
