Foreground segmentation based on thermo-visible fusion

In this paper, we present a background subtraction (BS) technique based on the fusion of thermal and visible imagery using an adaptive Gaussian mixture models (GMM). We investigate how to effectively combine thermal and visible information to optimize the segmentation accuracy. Pixel-level fusion strategies combining different color spaces and image representations are addressed. The standard GMM implementation is extended to integrate additional information consisting in the thermal imagery. Tests were carried out on challenging real-world video sequences. Quantitative as well as qualitative results are shown demonstrating the improvements introduced with respect to the use of a single spectral band sensor.