Visual Inertial 3D Reconstruction System Based on KLT Optical Flow and Voxel Hash

In this paper, front-end pyramid optical flow tracking, visual inertia fusion, depth estimation and 3D reconstruction techniques are studied, and a 3D reconstruction system based on Visual-Inertial Odometry is proposed. Using the pyramid KLT optical flow method based on bilateral filter denoising, the problem of blurred image edge information is effectively solved. The 3D reconstruction system can be applied to scenes of various scales, especially to multi-layer scenes with large longitudinal span. Experiments show that the proposed method can effectively reduce the cases of split-level reconstruction and reconstruction failure, and improve the completeness and accuracy of 3D reconstruction.