Recovery of Egomotion and Segmentation of Independent Object Motion Using the EM Algorithm

Motion in image sequences can result from the motion of the observer (egomotion) and from the presence of independently moving objects (IMOs) within the eld of view of the observer. Any vision system intended for an observer capable of motion needs the ability to distinguish between these two possibilities in order to successfully perform navigation and collision avoidance tasks. One approach to motion segmentation is to perform a statistical clustering on a set of local constraints on 3-D motion in the image. This thesis proposes two new methods, based on the EM algorithm, to perform robust motion segmentation on image sequences that contain IMOs. The rst method uses statistical clustering of linear and bilinear constraints (derived from computed optical ow using subspace methods) on 3-D translation and rotation. The problems of outlier detection, and determining number of processes and their initial parameters for the EM algorithm are considered. Also, analysis of the e ects of IMO boundaries on linear constraints, as well as a derivation for the removal of bias inherent in translation estimates from linear constraints, are presented. E ects of xation on detection of IMOs are considered. A framework for hypothesizing about motions underlying a set of constraint clusters is detailed. There exist situations in which 3-D motion constraints are not su cient to perform segmentation. The second method tracks depth-structure over time and evaluates