INTEGRATED APPROACH FOR 3D POINT CLOUD SEGMENTATION IN TANK CALIBRATION

ISSN (p) 0321-2211, ISSN (e) 2663-3450 Автоматизація та інтелектуалізація приладобудування DOI: 10.20535/1970.69(1).2025.333512 UDC 004.93, 531.7 INTEGRATED APPROACH FOR 3D POINT CLOUD SEGMENTATION IN TANK CALIBRATION D. M. Proskurenko, M. O. Bezuglyi National Technical University of Ukraine “Igor Sikorsky Kyiv Polytechnic Institute” Kyiv, Ukraine E-mail: , The paper presents a hybrid method for segmenting 3D point clouds for the calibration of cylindrical horizontal tanks, combining RANSAC and DBSCAN algorithms with subsequent boundary refinement based on local geometric features. Analysis of prior research indicates that RANSAC is effective for detecting cylindrical surfaces but sensitive to noise, while DBSCAN excels in clustering noisy data but requires parameter optimization. Hybrid methods combining these algorithms demonstrate improved results; however, their robustness to low-density point clouds and accuracy in transition zones remain underexplored. The objective of this study is to develop and evaluate a hybrid 3D point cloud segmentation method integrating RANSAC, DBSCAN, and boundary refinement to achieve automated tank calibration with high accuracy across densities levels ranging from ~1 million to ~18 million points. The research results are based on a comparison of a scanned model (18,012,345 points at maximum density) and an ideal model (17,986,543 points) of the tank. The hybrid method enabled precise estimation of geometric parameters: radius (R ≈ 1.5 m, error ±0.03 m) and length (L ≈ 10.8 m, error ±0.05 m). The segmentation identified the front bottom (372,890 points, ~2.07 %), rear bottom (411,230 points, ~2.28 %), and noise (2,181,240 points, ~12.1 %). The proportionality of point reduction for bottoms with decreasing density was confirmed by linear approximation (Fig. 1): slopes of ~20,700–22,800 points/million for the scanned model and ~20,900–21,100 for the ideal model, with R² ≈ 0.999. Relative segmentation errors range from 0.1–0.7 % for the front bottom and 8.3–8.9 % for the rear bottom, indicating higher accuracy for the front bottom and a need for improvement in the rear bottom. The stability of noise (~12.1–12.2 %) confirms the effectiveness of DBSCAN. The method maintained accuracy even at low density (~1 million points), although the increased error for the rear bottom (~8.75 %) suggests potential loss of detail. In conclusion, the developed hybrid method is robust to noise, scalable for densities levels of 1–18 million points, and suitable for automated tank calibration. The proportionality of components and stable noise level highlight the method’s reliability, while visualization (cylinder – red, front bottom – green, rear bottom – blue) illustrates clear component separation. Future research may focus on optimizing DBSCAN for low-density point clouds and reducing errors for the rear bottom in transition zones. Keywords: point cloud; hybrid algorithm; geometric modeling; segmentation; tank calibration; laser scanning. Introduction Processing 3D point clouds is a critical task in modern computer vision, robotics, and engineering. Point clouds obtained through laser scanning or photogrammetry often represent complex objects with diverse geometric shapes, such as cylinders, planes, or spheres. In particular, accurate modeling of cylindrical structures, which are common in technical constructions (e.g., pipes, tanks, shafts), plays a vital role in tasks such as reconstruction, quality control, and automated design. However, the presence of noise, data heterogeneity, and the complexity of transition zones between different object parts complicate the segmentation and analysis process. Currently, several methods exist for processing 3D point clouds. Among them, RANSAC (Random Sample Consensus) effectively estimates the parameters of geometric primitives, while clustering algorithms like DBSCAN enable grouping points based on spatial proximity. However, these methods, when used independently, do not always provide sufficient accuracy and robustness to noise, particularly when dealing with objects composed of cylindrical parts and adjacent bottoms. This creates a need for hybrid approaches that combine the strengths of different algorithms. The objective of this article is to develop and demonstrate a hybrid method for processing 3D point clouds, which integrates RANSAC for initial estimation of the cylindrical part’s parameters (radius, axis), followed by clustering of the remaining points using DBSCAN to identify bottoms and remove noise, and boundary refinement between the cylinder and bottoms based on local geometric features. The proposed approach aims to enhance segmentation accuracy and robustness to noisy data Literature review Processing 3D point clouds for modeling geometric objects, such as cylindrical horizontal tanks, Вісник КПІ. Серія ПРИЛАДОБУДУВАННЯ, Вип. 69(1), 2025. 75 ISSN (p) 0321-2211, ISSN (e) 2663-3450 Автоматизація та інтелектуалізація приладобудування is critical for calibration and volume estimation [1]. The literature describes numerous methods for point cloud segmentation and analysis, including approaches based on geometric primitive estimation, clustering, and hybrid strategies [2]. These methods are often compared in terms of accuracy, robustness to noise, and computational efficiency, as shown in Table 1. Table 1: Comparison of Point Cloud Segmentation Methods Method Advantages Disadvantages Robust to outliers, effective for primitives Handles noise, clusters arbitrary shapes Effective for shapes, no initial assumptions Struggles with complex shapes, sensitive to threshold Requires parameter tuning, issues with varying density Computationally expensive, sensitive to noise Region Growing Based on smoothness, intuitive Fails in noisy data, connectivity issues Deep Learning High accuracy, automatic feature learning Requires large datasets, high computational cost RANSAC DBSCAN Hough Transform RANSAC, introduced by Fischler and Bolles [3], is a robust algorithm for fitting geometric primitives, such as planes, cylinders, and spheres, to point clouds, even in the presence of noise. It is widely used for detecting cylindrical shapes, such as those in industrial components [4]. However, RANSAC may struggle with complex shapes or multiple similar forms, and its performance depends on the choice of threshold and number of iterations [5]. DBSCAN, proposed by Ester et al. [6], is a densitybased clustering algorithm that groups points by spatial proximity, making it suitable for identifying arbitrarily shaped clusters and removing noise. It has been applied in point cloud processing for scene segmentation tasks [7]. Its main limitation is the need for careful parameter tuning, which can affect performance in scenarios with varying densities [2]. Other methods include the Hough Transform, which is effective for shape detection but computationally expensive and less robust to noise compared to RANSAC (...truncated)