How Can a Point‑Cloud Algorithm Boost Safety in Aluminum Workshops?

In aluminum electrolysis workshops, real‑time monitoring of shelling head position is crucial for process accuracy and machine safety. High temperatures, smoke, dust and metal obstructions make traditional pose estimation methods struggle. Moreover, the head’s continuous motion and similar surrounding geometries complicate 3D data alignment. To tackle these issues, researchers developed a multi‑stage point‑cloud algorithm. First stage: weighted 3D Hough voting and an adaptive threshold boost initial matching, using an enhanced FPFH feature matcher. Second stage: merging FPFH and PCA features, the RANSAC‑IA coarse registration provides a stable start. Third stage: an improved ICP accelerates convergence and refines the final pose. Tests in a real electrolysis plant show the method withstands noise, occlusions and complex backgrounds, delivering reliable pose estimates even under harsh conditions. This approach could make industrial processes safer and more precise where visual data are limited.

The proposed method leverages three classic 3D registration algorithms, each tuned for the typical challenges of an electrolysis plant. The 3D Hough voting, traditionally used for geometric shape detection, is weighted by local point density, reducing the impact of noisy regions. The Fast Point Feature Histogram (FPFH) is enhanced with dynamic normalization of feature vectors, enabling more robust matching between clouds with scale differences. In the coarse registration phase, the RANSAC‑IA (Iterative Alignment) framework combines FPFH correspondences with principal component analysis (PCA) of point clusters, ensuring candidate transformations respect both local structure and global distribution. The final ICP (Iterative Closest Point) is modified by adding a convergence criterion based on the percentage change of mean error rather than an absolute threshold, improving stability under partial occlusions. Additionally, a curvature‑based penalty weight is introduced to prevent metallic edges from overly influencing the pose estimate. Experiments show a 35% reduction in average pose error compared to standard pipelines, while maintaining robustness at noise levels up to 20% of the original cloud density. These innovations pave the way for more reliable 3D vision systems in extreme industrial settings, where safety and precision are non‑negotiable.