Stochastic constraints for efficient image correspondence search

The navigation state (position, velocity, and attitude) can be determined using optical measurements from an imaging sensor pointed toward the ground. Extracting navigation information from an image sequence depends on tracking the location of stationary objects in multiple images, which is generally termed the correspondence problem. This is an active area of research and many algorithms exist which attempt to solve this problem by identifying a unique feature in one image and then searching subsequent images for a feature match. In general, the correspondence problem is plagued by feature ambiguity, temporal feature changes, and occlusions which are difficult for a computer to address. Constraining the correspondence search to a subset of the image plane has the dual advantage of increasing robustness by limiting false matches and improving search speed. A number of ad hoc methods to constrain the correspondence search have been proposed in the literature. A rigorous stochastic projection method is developed here which constrains the correspondence search space by incorporating a priori knowledge of the aircraft navigation state using inertial measurements and a statistical terrain model. The stochastic projection algorithm is verified using Monte Carlo simulation and flight data. The constrained correspondence search area is shown to accurately predict the pixel location of a feature with an arbitrary level of confidence, thus promising improved speed and robustness of conventional algorithms