Open Access

ARTICLE

A Point Cloud Segmentation Method for Adhering Coal Piles in Industrial Coal Sheds

Ping Zhang¹, Chungang Li², Jian Song¹, Jianjun Tan¹, Han Tian¹, Le Chen³, Yongkun Song^3,*

1 Fuel Business Division, Hunan Datang Xianyi Technology Co., Ltd., Changsha, China
2 Business Management Department, China Datang Corporation Ltd., Beijing, China
3 School of Computer Science and Technology, Changsha University of Science and Technology, Changsha, China

* Corresponding Author: Yongkun Song. Email:

Journal on Artificial Intelligence 2026, 8, 323-334. https://doi.org/10.32604/jai.2026.080463

Received 10 February 2026; Accepted 27 March 2026; Issue published 02 July 2026

Abstract

Accurate segmentation of coal pile point clouds is essential for automated inventory management and safety supervision in industrial coal sheds. With the widespread adoption of LiDAR technology, massive point cloud data are available, yet traditional clustering algorithms often fail to distinguish physically adhering coal piles in high-density storage scenarios, resulting in severe under-segmentation. To address this challenge, this paper proposes an automated point cloud segmentation framework tailored for complex industrial environments. The proposed framework consists of three core stages: data preprocessing, ground segmentation, and instance segmentation. First, relative elevation mapping and selective statistical filtering are employed to normalize coordinate baselines and suppress noise. Then, a Normal-Constrained Random Sample Consensus (NC-RANSAC) algorithm is introduced to separate horizontal ground surfaces from pile slopes using geometric angular constraints. Finally, an unsupervised strategy based on Principal Component Analysis (PCA) and sliding-window concavity detection is developed to identify saddle points and decouple adhering instances. Experimental results on real-world industrial datasets demonstrate that the NC-RANSAC algorithm exhibits high parameter stability and effectively resists interference from non-ground planes. Furthermore, the proposed instance segmentation strategy successfully decouples adhering coal piles with varying morphologies without manual intervention or pre-trained models.

---

Keywords

---