Segmenting People in Meeting Videos Using Mixture Background and Object Models

摘要

We have developed a meeting recorder system which captures a panoramic video of a meeting room. Segmentation of people from this video is required for tracking and retrieval applications. However, the application scenario makes it difficult to rely on the usual solution of static background initialization and purely motion-based tracking for segmenting people. In this paper, we describe a novel framework for segmenting people in these videos using adaptive Gaussian mixtures for both background and object modeling. Based on a Bayesian formulation of the problem, results of object segmentation provide feedback to the background segmentation module. Experimental results on real meeting videos are presented.