Motion-based segmentation and classification of video objects

In this thesis novel algorithms for the segmentation and classification of video objects are developed. The segmentation procedure is based on motion and is able to extract moving objects acquired by either a static or a moving camera. The classification of those objects is performed by matching their outlines gathered from a number of consecutive frames of the video with preprocessed views of prototypical objects stored in a database. This thesis contributes to four areas of image processing and computer vision: motion analysis, implicit active contour models, motion-based segmentation, and object classification. In detail, in the field of motion analysis, the tensor-based motion estimation approach is extended by a non-maximum suppression scheme, which improves the identification of relevant image structures significantly. In order to analyze videos that contain large image displacements, a feature-based motion estimation method is developed. In addition, to include camera operations into the segmentation process, a robust camera motion estimator based on least trimmed squares regression is presented. In the area of implicit active contour models, a model that unifies geometric and geodesic active contours is developed. For this model an efficient numerical implementation based on a new narrow-band method and a semi-implicit discretization is provided. Compared to standard algorithms these optimizations reduce the computational complexity significantly. Integrating the results of the motion analysis into the fast active contour implementation, novel algorithms for motion-based segmentation are developed. Since both static and mobile cameras are taken into account, motion-based segmentation is possible even when camera motion is present. In the field of object classification, a shape-based classification approach is extended and adapted to image sequence processing. Finally, a system for video object classification is derived by combining the proposed motion-based segmentation algorithms with the shape-based classification approach. For each contribution experimental results are reported that demonstrate the effectiveness of the proposed algorithms.