Object segmentation, registration, and tracking based on contour detection

• Main Authors: Cheng-Hung Chuang, 莊政宏
• Other Authors: Wen-Nung Lie
• Format: Others
• Language: en_US
• Published: 2003
• Online Access: http://ndltd.ncl.edu.tw/handle/23142408510513602744

博士 === 國立中正大學 === 電機工程研究所 === 91 === To support object-based multimedia coding, manipulation, interaction, searching, and retrieving in the MPEG-4 and MPEG-7 standards, segmentation, registration, and tracking of objects in images or video sequences play an essential and important role. Among many kinds of investigated object features, shape (i.e., object contour in 2-D space) is significant to be used in human visual processing, scene analysis, 2-D pattern recognition and 3-D object description, representation, and recognition. In this thesis, we proposed some algorithms based on the object contour information to achieve the above purposes. The proposed algorithms include: 1) a two-phase snake model based on an adaptive gradient vector flow (AGVF) field, 2) a downstream algorithm based on an extended gradient vector flow (E-GVF) field, 3) contour-based object registration, and 4) video object tracking using contour information. For object segmentation, traditional snake algorithms, including greedy and GVF snake models, often require to have manually-drawn initial snakes and adjust weighting parameters in snake models. The two-phase snake algorithm, which is composed of an active point phase and an active contour phase, was proposed for the automatic segmentation of multiple objects from noisy or cluttered backgrounds. In this model, initial snakes are automatically obtained following the active point processing and fast deformed by the active contour model which employs a no-search movement scheme with space-varying weighting parameters. The proposed AGVF field adopted in this model is capable of automatically adjusting weighting parameters according to local gradient information such that human interaction is no longer necessary. Among other techniques for object segmentation, region growing methods are considerably dependent on the selected homogeneity criterion and initial seeds. Watershed algorithms however have the drawback of over-segmentation. A new downstream algorithm based on the E-GVF field is proposed in this thesis for multi-object segmentation. The proposed E-GVF field provides more effective gradient information for object segmentation than traditional GVF field does. The downstream algorithm starts with seeds, which are automatically generated, and then grows them to segment objects via the guidance of E-GVF field. Since the E-GVF field is capable of preserving gradients near object boundaries and diffusing them to elsewhere, less over-segmentation exists in our downstream processing results. One important application of object segmentation is image registration. The contour-based object registration is focused on the study of local-type image registration techniques, which concern about pixel mapping between two correlated images with both global and local deformations. Our algorithm is object-contour-based, meaning that control points which conduct the transformation are adaptively selected from the matched contour points. There are two characteristics in this algorithm. First, we propose to match object contours in the source and target images by using locally maximal curvature (LMC) points and curve projection. Second, the selection of control points is capable of adapting to both global and local deformations, normally denser in contour segments with substantial local deformations. The proposed scheme is also hybrid in the sense that local contour matching and global surface-spline fitting of control points are combined. To extend our research to video object tracking, we proposed an algorithm based on contour information by combining change detection and snake processing. The moving objects are first analyzed in the temporal domain using change detection technique. In this step, a correlation-based local thresholding scheme is employed to find out object candidate regions. A distance-based clustering process is then used to group pixels for multiple object detection. With the AGVF field computed within these candidate regions, object contours can be detected after snake processing. Experiments in this thesis show good performance of our proposed algorithms for object segmentation, registration, and tracking.