Resilient Vision-Inertial Fusion Algorithm Based on the Depth of Feature Point

Visual simultaneous localization and mapping (VSLAM) as an efficient positioning and navigation solution has been widely used in fields such as unmanned aerial vehicles, robots, and autonomous driving. A key problem of VSLAM is how to model the representation of noise in visual observations. In current mainstream technology, the sensor noise level of the visual-inertial fusion framework is typically fixed as empirical values. However, it is difficult for fixed noise to reflect anomalies in visual observations because of changing light conditions, motion blur, and other conditions in dynamic environments, which results in the visual divergence problem. To address this problem, we propose a resilient visual inertial odometry (VIO) algorithm with an adaptive adjustment of fusion weights to suppress visual observation anomalies. First, we derive the qualitative characterization of the depth and depth error of visual feature points. Based on this, we design a method for quantifying the credibility of visual feature points. Then we dynamically adjust the weight of visual observations that participate in fusion estimation according to the credibility of the feature points. Using the European Robotics Challenge (EuRoC) dataset, we demonstrated that the results of the proposed method effectively inhibited the influence of abnormal visual observations, restrained the divergence of motion trajectories, and improved the precision of the VIO system.