Hand-Held 3D Scanning Using Computer Vision
The goal of this project is a hand-held camera which is moved freely around an object in order to compute a 3D model of that object. The device actually consists of multiple cameras, typically two cameras. The main research problem being addressed is the robust recovery of camera motion during unconstrained motion. The initial work is on scanning objects with dimensions up to about 0.5m. Future work will target larger objects, for example automobile-sized objects.
Background & Objective: Existing 3D scanner systems which have the most reliable performance include (a) turntable-based systems in which a fixed camera or cameras scan an object rotating on a turntable, and (b) cameras with attached motion-sensors (such as ultrasound). Problems are expense and the inflexibility of a fixed installation. The objective of this work is to do 3D scanning with a hand-held two-camera device. The advantages are cheapness, and ease-of-use in an arbitrary setting, say around a factory, or in an office, or around the home.
Technical Discussion: The hand-held scanner has two radically-directed cameras, one of which is used purely to determine the motion of the device, and one of which is used purely to acquire images of the object of interest. Dividing the vision functionality between different cameras in this way, and making use of distinctive points in the environment, enables robust and accurate recovery of camera motion regardless of the type of object being scanned. This contrasts with a vision system which attempts to recover camera motion from images of the object being scanned - in which case recovery of camera motion can be adversely affected by conditions such as homogeneity, transparency, or specularity of the object. The complete scanning system involves, in addition to recovery of camera motion, two other main components (a) foreground/background segmentation for arbitrary backgrounds, which is the subject of current work, (b) model-building (which has been the research focus of the MERL 3D Images project).
Technology Area: Computer Vision
Modification Date: June 13, 2008