QISO-SLAM: Object-Oriented SLAM Using Dual Quadrics as Landmarks Based on Instance Segmentation

Abstract

Dual quadrics as landmarks in object-oriented SLAM have recently attracted much attention due to the advantages in the mathematical completeness of projective geometry. Current researches suffer from a lack of either robustness or practicability. This letter introduces a full SLAM framework with pre-processing, data association, single-frame ellipsoid initialization, and a multi-step bundle adjustment process. The cost functions in bundle adjustment are built with the approximated geometric error using contour points extracted from 2D instance segmentation results. The variables of dual quadrics and camera poses are optimized repeatedly through the multi-step bundle adjustment process, namely object optimization, pose optimization, and a local bundle adjustment based on covisibility. It is demonstrated in the experiments that our system can reconstruct a precise high-level 3D map. Besides, superior localization performance is presented with and without accurate odometry inputs.