Human Region Segmentation Based on Bayesian Estimation Using Chroma and Thermal Video

Abstract

We have been developing a thermo-key system that segments human regions from video in real-time using thermal information. We developed a new segmentation method for this system based on Bayesian estimation, which uses both chroma and thermal information integratively. Segmentation algorithms were previously based on a simple threshold processing of thermal images, which often results in wrong segmentation of objects/environments whose temperatures are close to body temperature. Trained with only a few sets of chroma and thermal images, our algorithm can extract human regions from video with much higher accuracy compared to previous algorithms.