AWL-VIO: Adaptive window length–based visual-inertial odometry using statistical measures

Abstract

Visual-inertial odometry (VIO) plays a prominent role in computer vision, robotics, and augmented and virtual reality applications to estimate the pose and velocity of an agent. This paper focuses on an important aspect of the recently used optimization-based VIO architectures, that is, the optimization window length. This window length is generally fixed arbitrarily by the users, and there is a strong need for its adaptation as not one window size will suit the entire trajectory. In this work, two different statistical measures, one based on change point detection and the other based on standard deviation, are proposed as the means for adapting the optimization window length. This adaptive windowing approach has shown promising results and is benchmarked against the ground truth and the contemporary approach with fixed window length. The experiments conducted on the publicly available EuRoC MAV dataset and the TUM-VI dataset demonstrate the effectiveness of the proposed approach.