| Video object segmentation and tracking is essential for content-based video processing. This problem is challenging because of the need to segment the semantic video object from a video sequence that has been the 2D projection of a 3D natural scene involving non-rigid motion and occlusion or self-occlusion. This dissertation presents a framework for a semiautomatic approach to this problem. First, a semantic video object is initialized with human assistance in a key frame. The video object is then tracked and segmented automatically in the subsequent frames. A new active contour model, VSnakes, is introduced as the segmentation method in this framework. The contour energy is defined to directly reflect the energy difference between two contours instead of the energy of a single contour. The Viterbi algorithm is used to update the contour path during contour relaxation. A novel template matching technique, Correlation-based Adaptive Predictive Search (CAPS), has been developed during this research. By utilizing the information extracted from a template, CAPS provides a significant improvement in the computation time while maintaining the accuracy of traditional template matching. CAPS has been used as the primary object tracker in the segmentation framework. Since a video sequence might contain multiple shots, a novel scene change detector, Modified Continuously-Valued Adaptive Resonant Theory (M-ART2), is also considered in this framework to partition the sequence into separate segments. It is shown that the segmentation performance of the edge-based VSnakes algorithm can be further improved by contour propagation using region-based information, such as background subtraction and image motion. Various new techniques and concepts, such as contour simplification, background extraction, bidirectional motion extraction, and quasi-connected region, are also introduced in this dissertation as components of the segmentation algorithms. The experimental results suggest that the segmentation framework and techniques could become useful tools in many content-based or MPEG-4-related video applications in the future. |
