A quadrilateral-based method for object segmentation and tracking
Ronald Chung Hing-yip · 2003
This thesis presents a quadrilateral-based methodology for segmentation and tracking based on the principle that regions and their descriptors should be generated simultaneously in a single segmentation process. In essence, the methodology first obtains a content-based quadrilateral mesh from an edge map for each image frame, and treats quadrilaterals as the basic units for merging. Regions are therefore represented by groups of quadrilaterals with boundaries also approximated by the sides of quadrilaterals. This method uses less data than pixel-wise region representation, and gives a better approximation of boundaries than block-based representation. As these boundaries are available as a product of segmentation, they can be used in constructing the region correspondences usually required by different types of multimedia and computer vision applications without the need for further processing. The methodology also offers a flexible and unified approach for the design of different segmentation methods, so that a tailor-made segmentation method can be developed with different specifications of edge detection method, features of interest, merging criteria and merging algorithms. A quadrilateral-based segmentation method is proposed to demonstrate the effectiveness of this methodology. The segmentation method was evaluated by eight test images and its performance was compared with results achieved by K-means (KMC), Constrained Gravitation Clustering (CGC) and Seeded Region Growing (SRG). Objective measurement demonstrated that our method performed 8.54% to 93.33% better than KMC, 7.55% to 89.60% better than SRG, and 32.58% to 84.66% better than CGC except in one exceptional case. It also appeared to perform better consistently, with reasonably good and accurate boundaries, and the segmented image was composed of relevant regions with low complexity, an advantage for region correspondences construction. We also propose a novel method for tracking regions obtained from the quadrilateral-based segmentation method. By using this method, we could track two target regions in all 40 frames of the “Table Tennis” sequence, demonstrating that quadrilateral-based representation of regions can be applied in constructing region correspondences. The success achieved in both segmentation and tracking has demonstrated the usefulness of our methodology and its potential for practical application.