Photogrammetric triangulation and dynamic modeling of airborne video imagery
Abstract
Video is becoming a popular image source for a wide range of remote sensing applications, particularly when timely acquisition and low cost are the objectives. Specific photogrammetric applications include close range machine vision, rapid orthomosaic construction, and georegistration of tactical video sequences acquired using UAV's (unmanned aircraft vehicles). Accurate position determination requires rigorous mathematical modeling of uncalibrated video cameras, which contain relatively large systematic errors due to lens distortions and the unknown principal point location relative to the image coordinate system. A photogrammetric bundle adjustment algorithm with 10 Interior Orientation (IO) camera parameters is implemented using unified least squares. Due to high correlations among the IO parameters and between the IO and Exterior Orientation (EO) camera parameters, and sometimes also due to weak geometry, it is practical to recover only a subset of the 10 IO camera parameters in a self-calibration solution. Experiments with real video data are used to determine which IO parameters to recover, and to analyze the triangulation results using blocks and sequences of video with varying geometry. Both cases with fixed and varying zoom are considered. Tests with simulated data are run to study the use of GPS observations of the camera perspective center as the only source of control, provided that the aircraft trajectory does not approximate a straight line. Since the rigorous photogrammetric condition equations are nonlinear with respect to the unknown parameters, they require close initial approximations and iterations. Linear image invariance techniques are presented in this thesis that compute estimates for the camera parameters, of a single frame or an overlapping pair of frames, as a function of image coordinates and ground coordinates only. Optional constrained nonlinear refinements to these techniques that improve accuracy and reduce control requirements are discussed. In another application of photogrammetric invariance, four techniques to solve the image transfer problem are applied to triplets of images, and analyzed. A sequential Kalman filtering technique, which implements the First Order Gauss-Markov stochastic constraints, is developed and analyzed as an efficient alternative to the simultaneous bundle adjustment. The Kalman filter is designed to use the optimum combination of the platform trajectory and camera parameter values at the previous frame, and the observations on the current frame, to compute the updated elements of the current state vector. Experiments with simulated and real data show its near equivalence to the bundle, adjustment, with potential significant benefit for video sequences with zoom.
Degree
Ph.D.
Advisors
Mikhail, Purdue University.
Subject Area
Civil engineering|Remote sensing
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.