ScholarWorks@성균관대학교

초록

Vehicle localization in GPS-denied environments relies heavily on Visual-Inertial Odometry (VIO). However, a single VIO system is weak to drift and instability. This paper presents a training-free late fusion to alleviate these issues. Our method generates a unified and robust trajectory by fusing the outputs of multiple independent VINS-Mono runs using Gaussian Process Regression (GPR). The GPR model captures smooth motion of vehicle and sensor noise, resulting in a more robust trajectory. To ensure fair evaluation under causal conditions, all experiments were conducted with loop closure disabled. We evaluated the method using Absolute Pose Error (APE) and Relative Pose Error (RPE) against a geometric reference and also measured anchored drift during simulated and real GPS outages. Experimental results show that the GPR-based fused trajectory consistently achieves lower APE and RPE than any single VINS-Mono trajectory. Moreover, during long duration outages, it effectively suppresses drift growth. This work presents a simple and practical approach to enhance the accuracy and reliability of existing VIO systems for real-time applications.

키워드