Direct Sparse Monocular Visual-Inertial Odometry With Covisibility Constraints

In this paper, we propose a direct sparse monocular visual-inertial odometry system, which based on two techniques called adaptive direct motion refinement and photometric inertial bundle adjustment. To ensure efficient system launch in various scenarios, we design a novel multi-stage initialization strategy based on direct method. The sparse image alignment algorithm fail in cases of severe motion and large light variations. Therefore, we present adaptive direct motion refinement, which ensures robust tracking by dynamically adjusting the inertial residual scale. Although photometric bundle adjustment is efficient for estimating trajectory, there is still a relatively large accumulated error. To further reduce accumulated errors, we propose using photometric inertial bundle adjustment to improve trajectory estimation accuracy. This is achieved by constructing accurate covisibility relations and effectively combining inertial constraints. The performance of proposed method was evaluated on the EuRoC dataset, and the results demonstrate that the trajectory estimation accuracy outperforms existing methods.