Spatio-temporal Video SegmentationUsing Clustering

• Main Authors: Bo-Yun Lin, 林柏筠
• Other Authors: Gwo Giun Lee
• Format: Others
• Language: en_US
• Published: 2007
• Online Access: http://ndltd.ncl.edu.tw/handle/20182002962703541652

碩士 === 國立成功大學 === 電機工程學系碩博士班 === 95 === Video segmentation is a very important technology in video signal processing. It is an indispensable front-end procedure in the domain of second generation video compression, computer vision, machine recognitions, and many other multimedia applications. Better video segmentation algorithm needs less computation with its result obtaining better goodness-of-fit to the true object. In this thesis, we introduce a classification-based infrastructure for video segmentation algorithm. After analyzing the clustering algorithms, we proposed a video segmentation algorithm based on k-means clustering. By studying the spatial and temporal correlation of a video sequences, we develop a method by assigning proper initial partitions to improve the accuracy and the efficiency of the traditional k-means algorithm. In addition to the clustering, we applied the principle of morphological watersheds to split the unconnected regions. Since the video signals are already labeled by the clustering mechanism before object detection, over-segmentation problem of the original morphological watersheds will not occur and the procedure becomes simpler. Finally, experimental results demonstrated in this thesis confirm that the proposed clustering algorithm requires less number of iterations and achieve good segmentation results.