Enhancing Point Cloud Registration for Pipe Fittings: A Coarse-to-Fine Approach with DANIP Keypoint Detection and ICP...

In 3D reconstruction, loss of depth data caused by highly reflective surfaces often undermines the accuracy of point cloud registration. Traditional registration methods suffer from reduced accuracy and computational efficiency under such conditions. This paper presents a novel coarse-to-fine point cloud registration approach that combines a density-aware keypoint detection method with iterative closest point optimization to enhance both precision and computational performance. The proposed keypoint detection method optimizes registration by progressively refining the initial pose estimate through multi-scale geometric feature detection. This process includes a density-aware mechanism for removing edge outliers and an adaptive threshold based on normal vector inner products. This improves both keypoint identification accuracy and matching efficiency, providing better initial registration for the iterative closest point algorithm in scenarios with significant data loss. The approach prevents the iterative closest point algorithm from converging to local optima, which improves both convergence speed and overall computational performance. Experimental results show that, under optimal conditions, the runtime is reduced by up to 78.01% across several datasets, including those from Stanford, Kinect, Queen, and ASL-LRD. Compared to other traditional methods, the proposed approach delivers higher registration accuracy, even for multi-view point clouds with severe data loss, which demonstrates its robustness and potential for engineering applications.