Landmark-based Visual SLAM using Object Detection

Visual Simultaneous Localization and Mapping (V-SLAM) has been successfully deployed on mobile robots with various applications. However, these solutions rarely provide user-friendly information about the mapped environment. The proposed system is a landmark-based V-SLAM algorithm, which utilises Object Detection to include landmarks in a pose-graph. These “ordinary object” landmarks also contribute to loop closure detection through scene comparison with Bag-of-Visual-Words. Therefore, our system achieves robust loop closure and provides qualitative information about the navigated space. In contrast to laser based solutions, the proposed approach excludes dynamics objects (tables, chairs, and the like) and results in a significantly more cost effective system. The system has been deployed and tested on a mobile industrial robot in an indoors environment showing promising results. Our implementation is available on GitHub https://github.com/anastasiapan/LandmarksvSLAMI.