Dynamic Deformable Object Reconstruction Based on Non-Rigid Registration

As a core task in computer vision and graphics, 3D reconstruction holds significant application value in fields such as virtual reality and augmented reality. Implicit representations model target scenes through continuous functions and have been widely applied to 3D reconstruction and novel view synthesis. As a representative example, Neural Radiance Fields (NeRF) excel in static scene modeling, dynamic neural radiance fields aim to model scenes with dynamic objects, which include both rigid and non-rigid deformations, making it difficult to precisely represent non-rigid objects with complex geometric structures and deformation features. To address these challenges, this paper proposes a method combining non-rigid nearest neighbor position registration and a lightweight deformation network. The method first matches 3D positional coordinates across different timestamps with those at a canonical space. Then maps these matched points to the canonical space of the standard timestamp, and achieves high-quality dynamic scene rendering through an enhanced NeRF framework, effectively capturing complex geometric deformations in dynamic scenes. Experimental results demonstrate that the proposed approach outperforms existing methods in both deformation modeling accuracy and rendering quality.