Robustness in view-graph SLAM

This paper presents a new robustification procedure for nonlinear least-squares optimisation problems. In particular, we focus on the robustness of view-graph SLAM against outlier correspondences in the images and outlier geometries in the graph. Our method utilises revised measurements model linearisation and decision making to detect and remove outliers during data fusion. We utilise innovations and residuals gating to decide which observations were affected, given the most recent model linearisation point. By exploiting the inherited locality of measurements and states and the sparsity structure of nonlinear least-squares formulation we can check which measurement is affected before and after data fusion. This locality is the whole basis of robustification. To be efficient, we carry out the estimation using the information form. By doing so, we can include and remove information by individual measurements at each step using simple information addition and subtraction operations. Our results demonstrate the robustness of our method against outliers with respect to the use of kinematics alone and RANSAC with Levenberg-Marquardt algorithm.