Incremental 3D Model Generation using Depth Cameras

The ability to directly acquire 3D scene information from a single viewpoint could have a large impact on many computer vision applications. Depth cameras are a new sensing device facilitating such data acquisition, but are limitated by occlu-sion, noise, and poor sampling of obliquely viewed surfaces. To extend the ap-plicability of a single depth camera, we propose a method of retaining knowledge of surfaces from depth camera images acquired over time. This method draws on techniques from computer vision and computer graphics to reconstruct a 3D point cloud from a depth image, smooth and consolidate the point cloud into a good model, estimate surface normals of the model points, and register each new depth image to the model for tracking and model update. 1 Introduction an