Three Dimensional Scene Reconstruction Through Stochastic Structure From Motion

We propose to design and develop an algorithmic framework that reconstructs three-dimensional scene models from a video sequence collected by a single airborne camera. Compensating for the uncertainties introduced by measurement error and feature extraction, the system uses stochastic geometry to model observations of scene features and navigation information as multi-dimensional probability distributions rather than deterministic entities. We employ Extended Kalman Filters to produce optimal estimates of camera pose and 3D scene geometry. In order to maximize feature association and tracking performance, the system augments 2D image observations with information-rich, quasi-invariant feature vectors. This approach affords both computationally tractable, incremental operation and robustness to uncertainty inherent in operational systems. System inputs consist of camera calibration, a video sequence of arbitrary length, and sensor position and pointing angle for each image. The output of the system is a continually updated three-dimensional scene model. In Phase I, we will develop the core algorithms in detail and demonstrate key system concepts; we will then assess system performance (accuracy and computational complexity) as a function of free parameters such as sensor resolution, frame rate, and altitude. We will also analyze the hardware requirements of a real-time prototype.