This thesis surveys several motion measurement methods and moving image segmentation methods and compares the performance of some typical dynamic scene segmentation methods in view of the application to the automatic target tracking system.
There are two major types in the mesurement of visual motion according to the motion primitives: intensity-based scheme and feature-based scheme. The intensity-based scheme is divided again into pixel-based matching scheme called template matching methods, the gradient-based scheme called spatio-temporal gradient methods, and motion detection scheme such as difference picture methods.
There are various methods separating a moving object region from the background region in two dimensional time-varying imagery. But they are fallen into three categories according to the motion information and intensity values used in their segmentation criteria: global-motion-based segmentation scheme which use motion information only to merge adjacent pixels, local-motion-based segmentation scheme using both motion information and intensity values, and intensity-based dynamic segmentation scheme such as region-based difference picture methods.
This thesis analyzes these segmentation methods quantitatively examining geometrical feature errors of the object mask separated by segmentation procedure. The methods, such as global-motion-based segmentation and local-motion-based segmentation, give relatively good results compared to intensity-based dynamic segmentation scheme. This implies that merging criterion must not be heavily dependent on image intensity values.