객체 경계 강화 기반 개선된 깊이 추정 모델을 활용한 단안 VSLAM

Abstract

This study proposes an enhanced integration method of a state-of-the-art monocular depth estimation (MDE) model to overcome the fundamental limitation of monocular VSLAM systems— the lack of depth information. The proposed approach incorporates an edge loss function to sharpen object boundaries, a depth refinement strategy based on ground truth interpolation, and image preprocessing using CLAHE and inverse saturation to enhance depth quality. These components aim to generate more accurate and structurally consistent depth maps. The refined depth information is seamlessly integrated into a KeyFrame-based VSLAM pipeline to improve pose initialization, local map tracking, and loop closing efficiency. By combining these improvements, the system achieves better spatial awareness and robust localization performance. Experimental results on the TUM RGB-D dataset demonstrate that the proposed method achieves notable improvements over existing approaches, with significant gains in depth estimation accuracy, tracking stability, and real-time processing. Particularly, the RMSE is remarkably reduced, proving the effectiveness of the approach in enhancing overall VSLAM performance and reliability.