GRVIO: Semantic-Aware Visual-Inertial Odometry for Ground Robot Platforms

Visual-inertial odometry (VIO) has emerged as a key technology for state estimation of mobile robots by complementarily fusing data from a lightweight, low-cost inertial measurement unit and a camera. However, ground robots face significant challenges that degrade VIO performance, including textureless image regions from ground surfaces, dynamic environments, and motion-induced vibrations. To address these issues, this paper proposes ground robot visual-inertial odometry (GRVIO), a semantic-aware VIO framework specifically designed for ground robots. The proposed approach introduces a novel camera-plane distance measurement model, a zero-motion update model, and a dynamic object removal strategy. Our measurement formulation process leverages deep neural network-based semantic information to effectively remove dynamic point features and enhance the accuracy of plane features through high-level scene understanding. All measurement models are systematically integrated into a multi-state constraint Kalman filter (MSCKF) framework for accurate state estimation with a lightweight filter design. The effectiveness of the proposed method is validated through Monte Carlo simulations and real-world experiments using a publicly available dataset collected from a quadruped robot operating in diverse indoor and outdoor environments. Simulation results demonstrate that each proposed measurement model enhances localization accuracy and show the robustness of the proposed method against semantic segmentation errors and extrinsic calibration uncertainties. Furthermore, experimental results validate that GRVIO outperforms state-of-the-art VIO methods relying solely on point features, achieving an average reduction of approximately 40% in position root mean squared error compared to MSCKF, thus highlighting its capability to impose effective measurements and enhance both robustness and accuracy for ground robots.