DP-VINS: Dynamics Adaptive Plane-Based Visual-Inertial SLAM for Autonomous Vehicles

Traditional static visual-inertial navigation systems (VINSs) confront substantial challenges in dynamic environments, while current dynamic VINS solutions struggle to maintain high-accuracy performance in static environments. Achieving higher localization accuracy in static environments often requires incorporating additional geometric structural information as constraints. However, effectively leveraging this information in dynamic environments to improve robustness performance poses a significant challenge. This article proposes a novel dynamics adaptive plane-based stereo VINS (DP-VINS), incorporating dynamic factors improved with reprojection and coplanarity constraints, along with an augmentation dynamics adaptive loss function. DP-VINS utilizes the epipolar distance residuals for dynamic assessment and filters out unstable feature points in front-end visual-inertial odometry. It simultaneously extracts, merges, and tracks planar features. In addition to reprojection visual constraints, our method incorporates coplanarity constraints and adaptive weighting for back-end optimization. The design of adaptive weighting and augmented dynamic factors ensure faster attenuation of target weights with strong dynamic characteristics, thereby enhancing the accuracy performance of the system in static environments while maintaining robustness in dynamic environments. The proposed approach is assessed on public static and dynamic datasets and compared with state-of-the-art (SOTA) algorithms. The results demonstrate that the dynamic factors and augmented dynamics adaptive loss function proposed in this article enhance the traditional VINS performance, resulting in significantly higher positioning accuracy and robustness.