DS-SLAM: A Semantic Visual SLAM towards Dynamic Environments

Abstract

Simultaneous Localization and Mapping (SLAM) is considered to be a\nfundamental capability for intelligent mobile robots. Over the past decades,\nmany impressed SLAM systems have been developed and achieved good performance\nunder certain circumstances. However, some problems are still not well solved,\nfor example, how to tackle the moving objects in the dynamic environments, how\nto make the robots truly understand the surroundings and accomplish advanced\ntasks. In this paper, a robust semantic visual SLAM towards dynamic\nenvironments named DS-SLAM is proposed. Five threads run in parallel in\nDS-SLAM: tracking, semantic segmentation, local mapping, loop closing, and\ndense semantic map creation. DS-SLAM combines semantic segmentation network\nwith moving consistency check method to reduce the impact of dynamic objects,\nand thus the localization accuracy is highly improved in dynamic environments.\nMeanwhile, a dense semantic octo-tree map is produced, which could be employed\nfor high-level tasks. We conduct experiments both on TUM RGB-D dataset and in\nthe real-world environment. The results demonstrate the absolute trajectory\naccuracy in DS-SLAM can be improved by one order of magnitude compared with\nORB-SLAM2. It is one of the state-of-the-art SLAM systems in high-dynamic\nenvironments. Now the code is available at our github:\nhttps://github.com/ivipsourcecode/DS-SLAM\n