While the Iterative Closest Point (ICP) algorithm is widely used for 3D point cloud registration, relying on a fixed distance threshold causes unstable performance and demands extensive manual parameter tuning when aligning complex workpieces. This paper aims to address these limitations by developing an adaptive ICP approach to achieve robust multi-view registration in dynamic industrial scenarios.
An adaptive ICP algorithm based on Random Forest Regression (RF-ICP) is proposed. First, the PointNet++ network is used to extract deep geometric features from multi-view point clouds. These features are subsequently concatenated with the spatial overlap ratio to construct a high-dimensional feature vector. Finally, a Random Forest (RF) model uses this comprehensive vector to predict an optimal, adaptive distance threshold to guide the registration process.
The proposed RF-ICP method significantly outperforms the standard ICP and its advanced variants. Specifically, it reduces rotation and translation errors by 31.43% and 55.56%, respectively. Furthermore, the approach consistently maintains the maximum dimensional deviation within 4 mm in industrial applications, while substantially decreasing the effort typically required for manual parameter tuning.
This research introduces a novel data-driven framework that dynamically determines the ICP distance threshold. By effectively integrating deep learning-based spatial feature extraction with classical registration algorithms, the proposed method meets stringent industrial precision requirements and offers a highly reliable, automated solution for complex workpiece alignment.
