PRECISE OBJECT CONTOUR DETECTION USING VARIABLE BLOCK-BASED VIDEO FOREGROUND SEGMENTATION IN TRANSFORM DOMAIN

A precise object contour detection based on foreground segmentation in a video sequence is proposed. Our method employs a set of block-based Gaussian mixture models (GMM) on different block sizes, followed by foreground object contour extraction, which is based on Walsh transform (WT) domain. Every background model has different spatial and temporal correlations, and exploits WT-based feature parameters. This scheme can reduce both false positive errors and false negative errors, caused by dynamic background motions, such as global lighting changes and scenery in heavy snow. Moreover, a set of GMMs on every block size mode can be calculated by less than 10 % steps, compared with a conventional pixel-wise GMM scheme. The contour detection uses the result of the foreground segmentation for extracting precise edges along the actual contour by using WT horizontal/vertical spectra feature parameters of edge blocks. In addition to simple computation, our method requires no initialization with respect to boundaries of objects by using information of foreground regions. Experimental results show robust performance of the proposed method under outdoor conditions. Index Terms — foreground segmentation, object contour detection, Variable block size Gaussian mixture model, Walsh transform 1