ElasticFusion for Indoor 3D Reconstruction with an Improved Matching Points Selection Strategy

Real-time and high-quality 3D reconstruction of indoor scenes has been a research focus in the field of augmented reality and robotics. However, the 3D reconstruction methods using RGB-D sensors suffer from weaknesses such as poor local model quality and producing holes in points cloud models. The two key factors that affect the quality of 3D reconstruction are the accuracy of the pose parameters derived from points cloud registration and the accuracy of loop closure. Firstly, an improved ElasticFusion algorithm is proposed to achieve better reconstruction quality in real time, which is achieved by improving the strategy of searching matching points in the iterative closest point (ICP) algorithm. Furthermore, radial basis functions are used to construct implicit surfaces in order to fill holes in the point clouds models that are generated in the previous step. In the end, benchmarks such as ICL-NUIM are used to evaluate the proposed algorithm. The experimental results have shown that our algorithm can significantly improve the quality of model reconstruction and the accuracy of the camera trajectory estimation.