Long-Term Visual Localization Using Robust SuperPoint Matching and GVL-VIO Fusion for Intelligent Vehicles

In order to achieve accurate positioning for intelligent vehicles in GNSS-denied environments, this paper proposes a long-term visual localization system by fusing robust global landmark matching localization and visual-inertial odometry (VIO). Firstly, this paper improves SuperPoint to enhance its robustness to diverse weather, seasonal, and illumination changes and also reduces computational complexity. Then, this paper builds a visual feature database through dynamic objects mask, global landmark feature extraction, and keyframe selection. Thirdly, a three-stage localization policy via robust SuperPoint matching is developed to achieve high precise global visual localization (GVL), and the estimated global positions are input into VIO to suppress the accumulated error of VIO. As a result, the proposed system can maintain excellent short-term continuity and long-term accuracy for intelligent vehicle localization under GNSS-denied environment. The experimental results show that our proposed system not only significantly improves the absolute positioning accuracy and robustness with the reduction of the absolute track error by 66%, the root mean square error by 56.7%, and the standard deviation by 38.1%, but also effectively reduces the cumulative error of VIO for long-term visual localization of intelligent vehicles.