DP-SLAM: A visual SLAM with moving probability towards dynamic environments

Abstract

Simultaneous Localization and Mapping (SLAM) systems are proposed to estimate robot poses and reconstruct 3-D map of surrounding environment. Since most of the existing SLAM systems rely on the static-world assumption, to perform them in dynamic environments still remains challenging. In this paper, we propose a novel sparse feature based visual SLAM, named as DP-SLAM, which is based on a moving probability propagation model for dynamic keypoints detection. The probability indicates the likelihood of one keypoint being located on the moving objects. Our approach combines the results of geometry constraints and semantic segmentation to track the dynamic keypoints in a Bayesian probability estimation framework. We integrate our method into the front-end of the ORB-SLAM2 system, which acts as a pre-processing stage to filter out keypoints that are associated with moving objects. Furthermore, we inpaint the frame background that has been occluded by the detected dynamic objects, which benefit some applications such as virtual and augmented reality. Experimental results on the TUM RGB-D dataset and our own sequences demonstrate that our approach can improve performance of state-of-the-art SLAM system in various challenging scenarios.