DCS-SLAM: A Semantic SLAM with Moving Cluster towards Dynamic Environments

Abstract

Semantic SLAM, which can greatly improve tracking performance in highly dynamic scenes and construct dense maps with semantic information, has gained a lot of attention recently. However, it remains a challenge to obtain rich semantic information while ensuring real-time tracking. In this paper, we propose a new semantic SLAM framework, called DCS-SLAM, which uses the Moving Cluster association and MapPoint's motion probability propagation to solve this problem. Specifically, a novel data communication method is devised to obtain the Moving Cluster for each frame in a low-cost way by fully integrating geometric information and asynchronous semantic information. Besides, the Bayesian probability transfer model of MapPoint is applied to further remove the motion points. Experimental results will demon-strate that our framework has better computational efficiency while guaranteeing absolute trajectory accuracy. Moreover, our DCS-SLAM can achieve dense static map construction with rich semantic information by effectively removing the dynamic objects.