VIO中的信息稀疏化研究

Visual Inertial Odometry (VIO) is a simultaneous localization and mapping (SLAM) method using vision and inertial navigation sensors for pose estimation. The scale of the problem increases linearly with time. In order to limit the scale of VIO problem, sliding window algorithm based on incremental optimization of factor graph and marginalization strategy is introduced. However, the marginalization strategy leads to the disappearance of the sparsity of VIO, which increases the time complexity of the problem. In order to restore the sparsity of VIO after marginalization, in this paper, we extend the information sparsification to sliding window algorithm, and propose a novel sparse factor graph structure composed of nonlinear measurement factors. The sparse factorsare restored to approximate the dense marginal prior induced by successive marginalization in terms of minimizing Kullback-Leibler Divergence. By introducing sparsification into the sliding window, the marginal factors are transformed into sparse nonlinear factors, which can be re-linearized in the subsequent optimization, thus significantly reducing the accumulated linearization error and computational complexity. We implement the proposed algorithm on real datasets and compare it with other state-of-art algorithms to prove the effectiveness and accuracy.