Cognitive Mapping for Indoor Spaces

Simultaneous Localization and Mapping (SLAM) is essential for robots to efficiently understand environments, navigate, and interact within spaces. Object-level SLAM, which utilizes sparse data to provide a high-level understanding, faces two main challenges in real-time applications: under-constrained optimization and difficulties in online association. Our work presents a real-time object SLAM framework that addresses these issues.

First, we introduce a novel approach that leverages prior knowledge about an object, such as size and orientation. Second, we integrate a data association method that balances both short-term on-image tracking and long-term global matching, exclusively utilizing information at the object level. Our SLAM system fairly initializes the positions of objects, optimizes within a factor graph framework, and improves data association. We have validated our algorithm through tests on real-world datasets such as 3RScan and TUM-RGBD datasets, showing 50% improvement in object mapping error and 91% success in online data association.