Robust Visual SLAM in Dynamic Environment Based on Moving Detection and Segmentation

Abstract

In this study, a robust and accurate SLAM method for dynamic environments is proposed. Sparse optical flow and epipolar geometric constraint are combined to conduct moving detection by judging whether a priori dynamic object is in motion. Semantic segmentation is combined with moving detection to perform dynamic keypoints removal by removing dynamic objects. The dynamic objects removal method is integrated into ORB-SLAM2, enabling robust, accurate localization and mapping. Experiments on TUM datasets show that compared with ORB-SLAM2, the proposed system can significantly reduce the pose estimation error, and the RMSE and S.D. of ORB-SLAM2 are reduced by up to 97.78% and 97.91% respectively under high dynamic sequences, improving the robustness in dynamic environments. Compared with other similar SLAM methods, the RMSE and S.D. of the proposed method are reduced by up to 69.26% and 73.03% respectively. Dense semantic maps built with our method are also much closer to the groundtruth.