PMO-SLAM: Enhancing Incremental Optimization With Probabilistic Metrics for Object SLAM

Object-level environmental representation and reconstruction are essential for robots to achieve higher-level scene understanding and perform a broader range of physical interaction tasks. Among various geometric representations of object concepts, dual quadrics have garnered significant attention due to their compact parametric form and well-established projective geometry theory. However, monocular 3-D reconstruction of enclosing dual quadrics remains challenging in robotic scenarios, where viewpoint trajectories are constrained by motion patterns characterized by small baselines and limited observations. Existing parameterizations involve a high-dimensional spaces and lacks explicit ellipsoid constraints, hindering convergence with limited observations, ultimately leading to suboptimal estimations of the enclosing quadrics. To address these challenges, this article presents Probabilistic Metrics-based Object-Simultaneous Localization and Mapping (PMO-SLAM), which introduces an explicit parameterization method for 7-DoF dual ellipsoid representations. In addition, we propose a novel observation model based on the Wasserstein distance to facilitate incremental optimization of the object map. The algebraic properties of quadrics are further analyzed to ensure ellipsoid-closed parameter updates throughout the optimization process. Extensive experimental evaluations in both simulated environments and public datasets demonstrate that our method achieves superior convergence with fewer observations.