Speedy Sparse Bundle Adjustment for Video/Image Sequences

ABSTRACT: Bundle adjustment is a commonly employed technique for backing out camera parameters and 3D image models from a set of images. The generic case of bundle adjustment (overlapping shots from arbitrary cameras at arbitrary locations and times) has been highly optimized, and is used to good effect in such fields as 3D reconstruction and photo tourism. However, most existing implementations stick to this generic case, forgoing potential optimizations specific to video data streams. Given a video stream, the spatiotemporal locality of frames and the relative consistency of several camera parameters can be exploited to greatly improve on the execution speed of the generic case. This effort seeks to explore several techniques to improve on the generic state of the art for video data feeds, including feature tracking rather than simple correspondence, better initial estimates of camera pose, and restructuring the algorithm into a multi-pass, guided version. Such approaches will enable better reconstruction of camera parameters for UAVs in GPS-denied airspace. BENEFIT: Support navigation in denied environments, Enable mapping of aircraft coating damage back to 3D CAD models from video data