A Novel Visual Inertial Odometry Based on Interactive Multiple Model and Multistate Constrained Kalman Filter

High-accuracy positioning information plays an important role in the field of autonomous driving, where both the visual and inertial sensors have been widely noticed because of the virtue that they do not rely on external information. However, the accumulation of sensor errors degrades system positioning accuracy in complex environments due to the fixed system model and variable measurement noise in traditional visual inertial odometry (VIO). A novel VIO based on interactive multiple model (IMM) and multistate constrained Kalman filter (MSCKF) is proposed. First, the trifocal tensor model constructed from three consecutive images is used as the measurement model of the system, and the corresponding position and orientation information are added to the filter state vector to form an MSCKF, which is then combined with the IMM to form an IMM-MSCKF algorithm to interactively fuse the inputs and outputs of multiple subfilters to improve the positioning accuracy of the VIO. The proposed method is validated by selected urban environment and highway area data from publicly available datasets. The experimental results show that the proposed algorithm effectively improves the positioning accuracy of integrated navigation while reducing the positioning error of a single sensor compared to the conventional VIO.