Abstract: The reconstruction of three-dimensional scenes is currently a research hotspot in computer graphics and visual fields. Firstly, this paper analyzes the principle of depth image acquisition by Kinect v2 (Kinect for Windows v2 sensor), and explain the source of depth image noise. Then according to the principle of obtaining the depth image, an algorithm is designed to crop the point cloud sampling range. Secondly, the point cloud outliers are removed to fill the voids of the cloud to improve the reconstruction quality. Common three-dimensional scene reconstruction mostly uses a global three-dimensional body scheme of Kinect Fusion. This scheme can only reconstruct scenes in a small area. An ICP algorithm for point cloud matching of large scenes is designed for this purpose. Finally, surface reconstruction is performed on the point cloud to realize a low-cost and accurate three-dimensional reconstruction system for large scenes.

Key words:  Kinect v2, 3D scene reconstruction, point cloud denoising, outlier removal, ICP