Semantic 3D mapping for dynamic environments

The development of a semantic 3D mapping for dynamic environments is presented in this study. It is composed of the visual SLAM (Simultaneous Localization and Mapping) part and the semantic point cloud 3D reconstruction. For the visual SLAM part, the feature based visual SLAM, ORB-SLAM2 RGB-D, is modified with dynamic point rejection using information from semantic segmentation. The semantic segmentation is used to label the scene then keypoints that belong in labels that are dynamic such as person is removed. This allows the SLAM to estimate the agents pose based on the static environment only, which makes the SLAM more robust. The semantic 3D point cloud is generated from the depth map, semantic labels and estimated pose. The developed algorithm was tested on the TUM RGB-D Dataset and it was evaluated based on the ATE and RPE. The developed algorithms is compared to the base algorithm. It was then compared to the other algorithms based on ATE-RMSE. In a self made dataset the performance on the algorithm was tested in indoors and outdoor scenarios in real time and non real time evaluation