A new method is proposed to improve trajectory estimation accuracy by visual inertial odometry (VIO) subject to information sparsification. Current practices assume that the result of sparsification, i.e. a subgraph of a factor graph, takes the form of a tree and impose mutually independencies upon its nodes. However, this oversimplification may undermine a close approximation to the complete information matrix and eventually results in excessive estimation errors. Therefore, we propose to use a cliquey subgraph with preserved mutual correlation between selected pairs of the nodes. Sparsification is further accelerated by the new discovery that connectivity between arbitrary nodes can be directly read from Bayes tree, by transforming its surjective mapping to the underlying factor graph to a bijective mapping through a downstream-upstream traversing operation. Results from public dataset EuRoC indicate that the proposed method, while maintaining low computational profile, achieves lower absolute trajectory error (ATE) than that of the existing ones in the easy- and hard-level sequences and performs equally well in medium-level sequences. This result renders the proposed method a potent candidate for power saving in visual navigation-purposed application specific integrated circuit (ASIC) chip designs.
Read full abstract