Autonomous Mapping of Uncharted Ports: Comparison of SIFT and Semi-Global Block Matching Based Stereo Vision on Real Port Data

Unmanned Surface Vehicles (USVs) require robust autonomous navigation capabilities to operate safely in uncharted and Global Navigation Satellite System (GNSS)-denied maritime environments. Stereo vision is a promising passive sensing method for real-time 3D scene reconstruction, enabling Simultaneous Localization and Mapping (SLAM). This paper compares two stereo vision approaches: dense Semi-Global Block Matching (SGBM) and sparse Scale-Invariant Feature Transform (SIFT), evaluating their performance on real port data from the Pohang Canal Dataset. Key metrics analyzed include processing time, point cloud density, and valid depth ratio. Results show that SGBM produces dense, detailed reconstructions but incurs higher computational costs and is more sensitive to reflective surfaces. In contrast, SIFT yields sparse but highly reliable depth estimates with significantly lower processing times. The trade-off between completeness and efficiency suggests that SIFT-based methods may be more suitable for embedded USV systems with limited computational resources. At the same time, SGBM is better suited for applications requiring detailed environmental mapping. This study offers valuable insights and results to support the selection and development of stereo vision strategies for autonomous USV navigation in complex port environments.