Semantic visual SLAM in dynamic environment

Human-computer interaction requires accurate localization and effective mapping, while dynamic objects can influence the accuracy of localization and mapping. State-of-the-art SLAM algorithms assume that the environment is static. This paper proposes a new SLAM method that uses mask R-CNN to detect dynamic ob-jects in the environment and build a map containing semantic information. In our method, the reprojection error, photometric error and depth error are used to assign a robust weight to each keypoint. Thus, the dynamic points and the static points can be separated, and the geometric segmentation of the dynamic objects can be realized by using the dynamic keypoints. Each pixel is assigned a semantic label to rebuild a semantic map. Finally, our proposed method is tested on the TUM RGB-D dataset, and the experimental results show that the proposed method outperforms state-of-the-art SLAM algorithms in dynamic environments.