Kernel-VIO : An Optimization-based Tightly Coupled Indirect Visual-Inertial Odometry

This paper proposes an optimization-based tightly coupled Indirect Visual-Inertial Odometry (Kernel-VIO), which weights outliers with hybrid kernel functions. The novelty is that the hybrid weighting method based on kernel functions is firstly extended to the indirect-based VIO, which reduces the weight of outliers by dynamically weighting reprojection residual items, thereby the method suppresses adverse effects of outliers on the state estimator. Specifically, the proposed system processes the visual measurements and the IMU pre-integration in a tightly coupled way, next aligns the two information to complete the system initialization. Then, a nonlinear optimization framework is established by jointly minimizing the IMU pre-integration residuals, the reprojection residuals and prior information of sliding window marginalization to estimate the state of drones. The comparative experiments on the public dataset show that the accuracy and robustness of Kernel-VIO is better than the VINS-Mono without loop closure, and simulations demonstrate the effectiveness of the proposed Kernel-VIO.