Robust Visual Inertial Odometry Estimation Based on Adaptive Interactive Multiple Model Algorithm

摘要

In this paper, we focus on the problem of motion tracking in unknown environments using visual and inertial sensors, commonly known as visual-Inertial odometer (VIO) tasks. Currently, there are two main types of estimation methods to achieve VIO estimation, the filter-based method and the optimization-based method. We combine multi-state-constraint Kalman filter (MSCKF) algorithm with interactive multi-model algorithm and propose a novel filter-based VIO method. Compared with the VIO algorithm based on extended Kalman filter (EKF), the MSCKF algorithm has less strict probability assumption and better accuracy and consistency. However, traditional EKF and MSCKF algorithms both adopt a single fixed system model, which is difficult to adapt to complex and changeable application scenarios. To solve this problem, we introduce the adaptive multi-model method into the MSCKF algorithm, and combine the two to build an interactive multi-model MSCKF (IMM-MSCKF) algorithm. In the proposed IMM-MSCKF algorithm, several model sub-filters are designed, and their results are fused by transition probability to obtain the optimal state estimation. The common data set KITTI is used to verify the proposed IMM-MSCKF algorithm. Experiment results show that the proposed novel algorithm has better estimation accuracy and robustness compared with other solutions based on multi-state constraint Kalman filter. The IMM-MSCKF algorithm can achieve long-term, high-precision and consistent real-time VIO tasks.