Continuous-Time Stereo-Inertial Odometry

Abstract

With its considerable potential to address asynchronous multi-modal sensor fusion, the emerging paradigm of Continuous-Time (CT) Simultaneous Localization And Mapping (SLAM) has recently become a competitive alternative to conventional discrete-time approaches. Many fundamental research problems, such as robotic navigation and manipulation rely on accurate fusion of multiple, potentially unsynchronized sensing modalities, rendering the importance of Continuous-Time Simultaneous Localization And Mapping (CTSLAM) seminal. In this spirit, this work expands upon existing continuous-time concepts and strives towards an extensive, generic, and modular continuous-time framework. A particular focus lies on developing a robust, open-sourced, and extensibleCTSLAMimplementation as well as evaluating the applicability of continuous-time parametrizations to common stereo and stereo-inertial configurations. An experimental analysis on established datasets records the performance of our approach in an online stereo-inertial setup against the state-of-the-art indiscrete-time SLAM, achieving competitive results and providing a direct comparison between online discrete- and continuous-time approaches for the first time. Furthermore, targeting the absence of open-sourced, continuous-time pipelines and their associated, oftentimes prohibitive, initial development overhead, the complete implementation is made available underhttps://github.com/VIS4ROB-lab/HyperSLAM.