A Semi-dense Direct Visual Inertial Odometry for State Estimator

The Visual-inertial based navigation method increases popularity in recent years. In this paper, we present a method using the semi-dense direct methods by minimizing photometric error and the inertial information for localization with a multi-state constraint Kalman filter (MSCKF). The introduction of semi-direct method can better adapt to complex environments and reduce the computational costs. With the fusion of the visual-inertial odometry and the inertial information by the filter, a high frequency and stable estimated state is available. Furthermore, we design a keyframe window for global optimization and loop closure detection to correct the cumulative error in large-scale experiments. The method obtains a balance between the accuracy of the state estimation and the computational complexity. The performance of the system with stereo cameras is demonstrated in large-scale experiments without GPS.