3D LiDAR 센서를 이용한 선체의 실시간 3차원 형상 재구성

다국어 초록 (Multilingual Abstract)

This study presents a method for real-time estimation of the relative pose of a ship’s side hull during berthing and the continuous reconstruction of its 3D geometry based on the pose. Because an automatic mooring system must attach suction pads to the ship’s outer shell, it is crucial to provide reliable hull-shape
information throughout the berthing process. In real-world conditions, however, the side hull surface is often close to planar, offering limited distinctive geometric
features, and a dock-mounted fixed sensor suffers from a restricted field of view, causing the observable region to change rapidly as the vessel approaches. Under
these conditions, the correspondence search and optimization in ICP(Iterative Closest Point) become unstable, making registration errors accumulate easily and
leading to reconstruction distortions such as drift and multi-surface artifacts Accordingly, this study formulates the above phenomena as two core issues—degradation of registration stability and contamination in reconstruction and designs a reconstruction pipeline that combines two improvements in the ICP stage with one enhancement in the TSDF(Truncated Signed Distance Field) stage. First, a virtual sensor frustum is used to reduce the likelihood that unobservable regions are included in correspondence candidates, thereby lowering the risk of false matches. Second, ORB(Oriented FAST and Rotated BRIEF) feature–based 2D
matching is performed using the LiDAR ambient image, and the matched features are back-projected into 3D to provide an initial guess for ICP, improving convergence stability. Finally, in the TSDF integration stage, an update strategy is applied to ensure that registration errors do not directly propagate into the
reconstruction result, enabling more stable updates near the surface. The proposed method was validated using data collected in accordance with the actual berthing schedule of the training ship “Segero-ho” at Mokpo National Maritime University, and its real-time feasibility was examined on a system based on an NVIDIA Jetson Orin NX 16GB and an SOSLAB ML-X120 sensor. Experimental results show that the proposed approach reduced the average RMSE from 0.367m(traditional method) to 0.152m(ORB+Frustum), corresponding to an improvement of approximately 58%. The registration failure rate was also reduced to less than half, significantly improving tracking stability over the entire berthing sequence. In addition, the reconstructed surface exhibited fewer distortions such as irregular thickness and surface tearing yielding a substantially more consistent model. Nevertheless, because the evaluation was conducted mainly during relatively stable morning conditions dictated by the training ship’s operating schedule, the method was not sufficiently validated under extreme conditions such as afternoon scenarios, severe weather, fog, or strong backlighting. Moreover, the experiments focused primarily on small-to-medium-sized vessels, and further investigation is required to confirm performance on larger ships. Future work will therefore include long-term data accumulation across diverse weather and illumination conditions, the application of environment-adaptive algorithms, and the integration of multiple sensors to broaden applicability, with the ultimate goal of enabling safe and rapid automatic mooring in real operational settings.