Asynchronous Adaptive Conditioning for Visual-Inertial SLAM

This paper is concerned with real-time monocular visual inertial simultaneous localization and mapping (VI-SLAM). In particular a tightly coupled nonlinear-optimization based solution that can match the global optimal result in real time is proposed. The methodology is motivated by the requirement to produce a scale-correct visual map, in an optimization framework that is able to incorporate relocalization and loop closure constraints. Special attention is paid to achieve robustness to many real world difficulties, including degenerate motions and unobservablity. A variety of helpful techniques are used, including: a relative manifold representation, a minimal-state inverse depth parameterization, and robust non-metric initialization and tracking. Importantly, to enable real-time operation and robustness, a novel numerical dog leg solver [16] is presented that employs multithreaded, asynchronous, adaptive conditioning. In this approach, the conditioning edges of the SLAM graph are adaptively identified and solved for both synchronously and asynchronously. In this way some threads focus on a small number of temporally immediate parameters and hence constitute a natural “front-end”; other threads adaptively focus on larger portions of the SLAM problem, and hence are able to capture functional constraints that are only observable over long periods of time—an ability which is useful for self-calibration, during degenerate motions, or when bias and gravity are poorly observed. Experiments with real and simulated data for both indoor and outdoor robots demonstrate that asynchronous adaptive conditioning is able to closely track the full-SLAM maximum likelihood solution in real-time, even during challenging non-observable and degenerate cases.