Improved foreground-background segmentation using Dempster-Shafer fusion

Abstract

Popular foreground-background segmentation algorithms are based of background subtraction. In complex indoor environments, if an object in motion initially remains stationary for a certain period, it can be absorbed into the background, becoming invisible to the system. Aiming at solving this problem, this paper presents a flexible and robust foreground-background segmentation algorithm based on accurate moving objects classification. Our algorithm combines low level and high level information, i.e. the data belonging to single pixels and the result of accurate object classification respectively, to improve the background management. Accurate object classification is obtained by combining classification evidence from different object recognisers using the Dempster-Shafer rule. The proposed algorithm has been tested with a large amount of acquired images; moreover, real test cases are reported. Reported experimental results include object classification accuracies obtained with a proposed Basic Belief Assignments and measurements of the quality of the background image such as Recall-Precision and F-measure computed with different background management algorithms. The experimental results show the superiority of the proposed segmentation algorithm over popular algorithms.