Processing Dynamic Image Sequences from a Moving Sensor.

摘要

A fundamental problem in motion processing research has been the discrepancy between the precision and reliability with which image displacements can be determined and the sensitivity of inference procedures to noise and resolution errors. There are also indications that these inference procedures are inherently unstable and, in some cases, ambiguous. The approach of this thesis has been to deal with restricted cases of motion for which the inference of the motion parameters, image displacements, and environmental depth, can be combined into a single, uniform, and mutually constraining computation. These restricted cases of motion are sufficient for a wide range of real-world tasks, especially since other associated sensing devices can be used to ascertain the other parameters of motion. The author then apply the procedure developed for translational motion to local portions of image sequences to process general sensor motion as if it were composed of independent local environmental translations. The resulting representation can considerably simplify the processing of less restricted and general motion.