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

In this study, we propose a robust and accurate SLAM method for dynamic environments. Our approach combines sparse optical flow with epipolar geometric constraints to detect motion, determining whether a priori dynamic objects are moving. By integrating semantic segmentation with this motion detection, we can effectively remove dynamic keypoints, eliminating the influence of dynamic objects. This dynamic object removal technique is integrated into ORB-SLAM2, en-hancing its robustness and accuracy for localization and mapping. Experimental results on the TUM dataset demonstrate that our proposed system significantly reduces pose estimation error compared to ORB-SLAM2. Specifically, the RMSE and standard deviation (S.D.) of ORB-SLAM2 are reduced by up to 97.78% and 97.91%, respectively, in highly dynamic se-quences, markedly improving robustness in dynamic environments. Furthermore, compared to other similar SLAM methods, our method reduces RMSE and S.D. by up to 69.26% and 73.03%, respectively. Dense semantic maps generated by our method also closely align with the ground truth.