Exportar Publicação

Shen, Y., Gao, Z., Wang, J., Zhang, Z., Zhang, Y., Postolache, O....Mi, C. (2025). Point cloud surface registration for cargo hold material matching via sparse line-based sensing method. IEEE Sensors Journal. 25 (23), 43305-43318

Y. Yang et al.,  "Point cloud surface registration for cargo hold material matching via sparse line-based sensing method", in IEEE Sensors Journal, vol. 25, no. 23, pp. 43305-43318, 2025

@article{yang2025_1777281563668,
	author = "Shen, Y. and Gao, Z. and Wang, J. and Zhang, Z. and Zhang, Y. and Postolache, O. and Mi, C.",
	title = "Point cloud surface registration for cargo hold material matching via sparse line-based sensing method",
	journal = "IEEE Sensors Journal",
	year = "2025",
	volume = "25",
	number = "23",
	doi = "10.1109/jsen.2025.3620816",
	pages = "43305-43318",
	url = "https://ieeexplore.ieee.org/xpl/RecentIssue.jsp?punumber=7361"
}

TY  - JOUR
TI  - Point cloud surface registration for cargo hold material matching via sparse line-based sensing method
T2  - IEEE Sensors Journal
VL  - 25
IS  - 23
AU  - Shen, Y.
AU  - Gao, Z.
AU  - Wang, J.
AU  - Zhang, Z.
AU  - Zhang, Y.
AU  - Postolache, O.
AU  - Mi, C.
PY  - 2025
SP  - 43305-43318
SN  - 1530-437X
DO  - 10.1109/jsen.2025.3620816
UR  - https://ieeexplore.ieee.org/xpl/RecentIssue.jsp?punumber=7361
AB  - In the automated unloading operations at port terminals, portal cranes move through a multiline laser radar to dynamically sense the material surface in cargo holds. The laser radar sensors provide detailed 3-D point cloud data, enabling precise surface detection. However, a single scan is prone to interference from equipment vibration and grab bucket occlusion, leading to sparse point cloud data and blind spots in the scan. To construct a complete material surface model, multiple frame registration is required. Through the integration of sensor data from multiple scans, a more complete and accurate material surface model can be reconstructed. By merging data from multiple sensing scans, the system gains a deeper understanding of the material surface structure, significantly enhancing the overall sensing accuracy. Traditional registration methods, which rely on manually designed feature extraction rules, often fail under the conditions of irregular surfaces and sparse, line-shaped point clouds. Such methods, lacking dynamic adaptability to sensor data, struggle to process real-world complexities efficiently. While the deep learning algorithm RPMNet can handle sparse data, it is insensitive to rotational errors around the vertical axis (yaw angle), resulting in significant yaw angle errors. This insensitivity reduces the system’s ability to sense precise spatial transformations, leading to errors in material alignment. To tackle this issue, an improved RPMNet + iterative closest point (ICP) fusion algorithm is considered in this work. The initial registration of the material surface point cloud is performed by RPMNet, and then edge features are used to build a local coordinate system. High-discriminative edge points are extracted via sector segmentation to generate posture vectors, and a masking mechanism is employed to filter out corner interference. The edge features, as sensed through sector segmentation, play a crucial role in defining accurate positional orientation. Finally, the ICP algorithm is used for precise registration to optimize errors. Through iterative refinements based on sensor inputs, ICP enhances the system’s alignment capabilities, significantly reducing residual errors. In simulated conditions, experimental results indicate that, regarding accuracy and robustness, the proposed method remarkably surpasses mainstream registration algorithms. Moreover, it demonstrates stable performance in real scenarios, and the corresponding data comes from different types of ships and under various weather conditions. The robustness of the method is validated under diverse real-world sensing conditions, showcasing its practical application in dynamic, sensor-based environments.
ER  -