The Worldwide Journal of Robotics Analysis, Forward of Print.
In monocular visual-inertial navigation, it’s fascinating to initialize the system as shortly and robustly as potential. A state-of-the-art initialization technique usually constructs a linear system to discover a closed-form answer utilizing the picture options and inertial measurements after which refines the states with a nonlinear optimization. These strategies typically require a number of seconds of knowledge, which nevertheless might be expedited (lower than a second) by including constraints from a sturdy however solely up-to-scale monocular depth community within the nonlinear optimization. To additional speed up this course of, on this work, we leverage the scale-less depth measurements as an alternative within the linear initialization step that’s carried out previous to the nonlinear one, which solely requires a single depth picture for the primary body. Importantly, we present that the everyday estimation of all characteristic states independently within the closed-form answer might be modeled as estimating solely the size and bias parameters of the realized depth map. As such, our formulation permits constructing a smaller minimal drawback than the cutting-edge, which might be seamlessly built-in into RANSAC for sturdy estimation. Experiments present that our technique has state-of-the-art initialization efficiency in simulation in addition to on standard real-world datasets (TUM-VI, and EuRoC MAV). For the TUM-VI dataset in simulation in addition to real-world, we reveal the superior initialization efficiency with solely a 0.3 s window of knowledge, which is the smallest ever reported, and validate that our technique can initialize extra typically, robustly, and precisely in numerous difficult situations.
发布者:Nathaniel Merrill,转转请注明出处:https://robotalks.cn/fast-and-robust-learned-single-view-depth-aided-monocular-visual-inertial-initialization/