Robust Monocular Visual-Inertial Odometry for Agricultural Vehicles Based on IMU-Augmented 3D Feature Point Correction

This paper presents a new monocular visual-inertial odometry (VIO) system designed to achieve precise and robust localization for autonomous vehicles in challenging agricultural environments, where unstable and low-texture features often degrade performance. Unlike existing VIO methods, which are primarily optimized for urban environments and structured settings, the proposed approach specifically addresses the unique challenges of agricultural landscapes, such as illumination, vegetation occlusions, and irregular terrain. To enhance robustness and precision, the system integrates an error state Kalman filter (ESKF)-based attitude estimation, improving IMU preintegration accuracy even in visually degraded conditions. Additionally, a map point correction technique is introduced, leveraging IMU motion data to eliminate distortions in 3D map points and mitigate the effects of low-texture environments. By tightly coupling these advancements, the proposed system achieves robustness and accuracy in pose estimation. The effectiveness of the method is validated through real-time experiments across various scenarios, with the proposed system achieving approximately an 80% reduction in translation error and about a 60% reduction in rotation error compared to state-of-the-art methods. These results demonstrate its superiority over existing VIO algorithms in terms of localization precision and stability.