Abstract
Autonomous driving represents a transformative advancement with the potential to significantly impact daily mobility, including enabling independent vehicle operation for individuals with visual disabilities. The commercialization of autonomous driving requires guaranteed safety and accuracy, underscoring the need for robust localization and environmental perception algorithms. In cost-sensitive platforms such as delivery robots and electric vehicles, cameras are increasingly favored for their ability to provide rich visual information at low cost. Despite recent progress, existing visual-inertial odometry systems still suffer from degraded accuracy in challenging conditions, which limits their reliability in real-world autonomous navigation scenarios. Estimating 3D positional changes using only 2D image sequences remains a fundamental challenge primarily due to inherent scale ambiguity and the presence of dynamic scene elements. In this paper, we present a visual-inertial odometry framework incorporating a depth estimation model trained without ground-truth depth supervision. Our approach leverages a self-supervised learning pipeline enhanced with knowledge distillation via foundation models, including both self-distillation and geometry-aware distillation. The proposed method improves depth estimation performance and consequently enhances odometry estimation without modifying the network architecture or increasing the number of parameters. The effectiveness of the proposed method is demonstrated through comparative evaluations on both the public KITTI dataset and a custom campus driving dataset, showing performance improvements over existing approaches.