Motion Segmentation in a Dynamic Scene based on Parametric Motion Modeling using Generalized Principal Component Analysis

This thesis addresses the problem of segmenting motions generated by rigid and non-rigid objects in a dynamic scene. The difficulties posed by this problem becomes apparent when trying to segment multiple moving objects simultaneously in a dynamic environment. The proposed method constructs a parametric motion model based on feature points in the scene, and uses the parametric motion models as labels to the feature points. The labeled feature points are then used to segment the scene using a k-nearest neighbors algorithm. Experimental results were obtained by analyzing the problem for rigid and non-rigid objects with various parametric motion models in a dynamic scene. The contribution of this work demonstrate a method to track rigid and non-rigid objects in a dynamic scene without being restricted to a specific motion model