Autonomous vehicles require precise localization across all environments. Complex lighting conditions and dynamic environments pose significant challenges to the accuracy and robustness of visual-inertial odometry systems. In response to these challenges, this paper proposed a robust visual-inertial odometry system named Robust-VIO, which utilizes weighted deep learning features. This novel algorithm combines a learning-based front-end with a tightly-coupled optimization back-end, achieving robust, accurate, and real-time localization in challenging scenarios. Initially, the SuperPoint deep learning network extracts reliable features under various lighting conditions. These features are then efficiently tracked by integrating them with sparse optical flow. To address highly dynamic environments, scene flow detection with IMU pre-integration is used to identify and weigh dynamic features. Subsequently, a tightly-coupled optimization approach is employed to jointly optimize the weights of features and the pose, effectively minimizing the influence of dynamic features and providing precise pose estimates. The performance of the proposed algorithm has been assessed using two publicly available benchmark datasets. The experimental results show that Robust-VIO surpasses other leading visual-inertial odometry systems in accuracy and robustness in complex scenarios, while also fulfilling real-time computational demands, confirming its practical applicability.
Wu et al. (Fri,) studied this question.