Efficient AR Object Alignment Using Deep Closest Point and Cylinder-Angle Interpolation

Abstract

Accurate alignment of virtual objects in Augmented Reality (AR) is essential for precision-critical applications such as surgery, infrastructure inspection, and digital twin systems. However, conventional markerless approaches relying on planar detection or external sensors often incur errors, high cost, and reduced portability. This study presents a lightweight AR framework that operates solely on mobile devices, minimizing reliance on external hardware. The framework integrates Deep Closest Point (DCP)-based point cloud registration with a novel Cylinder-Angle Interpolation strategy. The DCP network, trained on synthetic data, employs attention-based feature embedding and SVD-based soft matching to achieve robust registration under noise and partial overlaps. The interpolation method estimates alignment errors from sparse angular samples, reducing computational cost and data usage by approximately 85%. Comparative experiments against ICP, Go-ICP, FGR, TEASER++, and DCP-only show that the proposed method achieves superior accuracy (average error rate 0.029), higher success rates (96%), and stable performance across diverse environments, including low-light and GPS-degraded scenarios. These results demonstrate a practical and scalable solution for mobile AR, with strong potential for future extensions to non-rigid registration, multi-user synchronization, and SLAM-based mapping.