SUI-SLAM: A Semantics and Uncertainty Incorporated Visual SLAM Algorithm towards Dynamic Indoor Environments

In order to solve the interference of moving objects on visual SLAM (simultaneous localization and mapping) system in dynamic environment, SUI-SLAM (semantics and uncertainty incorporated SLAM), a dynamic visual SLAM algorithm integrating deep semantics and uncertainty, is proposed. Firstly, the dynamic prior information of the scene is obtained based on semantic segmentation using Mask R-CNN (region-based convolutional neural network). Then, the edge of the segmented area is corrected using the image depth information, to further distinguish the foreground and background feature points in dynamic environment. Finally, the semantic segmentation results, movement priors and geometric errors are used to calculate the uncertainty of the correspondence between the feature points and the 3D map points. Regularization items are added in the process of optimizing the camera pose to improve the accuracy and the robustness. In order to verify the algorithm effectiveness, experiments are carried out on the TUM dynamic datasets. Results show that the pose estimation accuracy of SUI-SLAM algorithm can be increased by up to 98.41% in indoor high dynamic scenes compared with ORBSLAM2 algorithm. While compared with other SOTA (state-of-the-art) dynamic SLAM algorithms, the pose estimation accuracy and the robustness of SUI-SLAM algorithm are also improved to a certain extent.