Fusion of thermal and visible images for day/night moving objects detection

Abstract

A background subtraction (BS) technique based on the fusion of thermal and visible imagery using Gaussian mixture models (GMM) is presented in this work. An automatic daytime/night-time detection is introduced that can be used to dynamically adapting the fusion scheme. Three fusion schemes are investigated and coined as early, late and image fusion. The first consists in augmenting the GMM model with thermal information prior to foreground segmentation. The second, as it name indicates, consists in the fusion of the outputs of BS applied to each sensor separately. The last one considers different linear combinations of both images forming a hybrid image. Most approaches improve the performance of the combined system by compensating the failures of individual sensors. Quantitative as well as qualitative results are shown to demonstrate the accuracy of each fusion approach with respect to foreground segmentation.