A recent paper by members of the DCIST alliance develops unified representations for metric-semantic localization and mapping which allow robots to build meaningful object-level maps from online visual-inertial observations. The work makes two contributions to the state of the art in SLAM. First, it generalizes purely geometric models by introducing semantically meaningful objects, represented as structured models of mid-level parts (e.g., doors, windshield, wheels for a car). Second, instead of making hard, potentially wrong data associations among semantic features and objects, it shows that SLAM inference can be performed efficiently with probabilistic data association. This approach allows global loop closure, leading to consistent object-level maps (e.g., containing doors, chairs, tables, etc.) and also offers significant advantages over existing techniques in ambiguous environments.
Source: N. Atanasov, S. Bowman, K. Daniilidis and G. Pappas, “A Unifying View of Geometry, Semantics, and Data Association in SLAM,” International Joint Conference on Artificial Intelligence (IJCAI), 2018.