Fast and Robust Semidirect Monocular Visual-Inertial Odometry for UAV
Qingxi Zeng, Haonan Yu, Xufang Ji, Xiaodong Tao, Yixuan Hu
Abstract
State estimation is critical for achieving autonomous navigation of the unmanned aerial vehicle (UAV). We present a fast and robust monocular visual-inertial odometry (VIO) algorithm, which achieves real-time pose estimation for UAV. A semidirect method for tracking is used in the front end to address the problem of the slow running speed of traditional VIO on onboard computing platforms, which preserves the fast performance of direct methods and the robustness of feature-based methods. Edgelet features are added to the Shi–Tomasi feature extraction to enhance the robustness of VIO in low-texture environments. Furthermore, a quadtree-based feature extraction algorithm is employed to filter features, which effectively addresses the issue of feature accumulation. In the back end of the VIO system, a sliding-window-based nonlinear optimization method is used to fuse image and inertial information. We evaluate our approach on the public European robotics challenges micro aerial vehicle (EuRoC MAV) dataset and compare it with the state-of-the-art VIO algorithms. The experimental results demonstrate that our proposed VIO system is capable of operating in challenging scenarios with fast motion, image blur, and weak texture features. Moreover, our proposed VIO system achieves comparable accuracy to the state-of-the-art VIO algorithms while improving tracking speed by 54.28% compared to VINS-Mono.