Analysis And Comparison Of Background Subtraction Methods Based On Static Camera

Background subtraction(BS)is a crucial step in many computer vision systems,as it can detect significantly moving objects or regions from a video stream.BS techniques are widely used in the applications of video surveillance,remote sensing and human computer interaction.In this thesis,we focus on the BS techniques for static camera.First,we present the main challenges of BS for video surveillance which include gradual illumination changes,sudden illumination changes,dynamic background,camouflage,shadows,bootstrapping and ghost phenomenon.Then,we study the basic principle and characteristics of background subtraction,and analyze the strengths and weaknesses of some state-of-the-art BS algorithmsGenerally,in the BS algorithm based on the simple frame difference,the background model is updated with a linear moving average scheme.In this thesis,we propose a novel background updating scheme based on nonlinear moving average.When a pixel is determined as "background",then it's learning rate is adaptively adjusted according to the difference between its input and its previous background value.Under this scheme,the foreground objects or regions will not easily be merged into the background model,and the background model will become stable with time going.We also integrate the nonlinear background updating scheme to the ViBe algorithm,and extend the original discrete updating scheme in the ViBe algorithm to a continuous updating scheme,so the original one is a special case of the new nonlinear background update scheme.In order to quantitatively compare different BS algorithms under different scenarios,we test the BS algorithms on the Change Detection 2012 and 2014 datasets.We find that the performances of the two representative algorithms that adopted the nonlinear background updating scheme have been greatly improved on some of the video categories.However,to design a BS algorithm that is appropriate for all scenarios and variety illumination conditions is still a difficult problem due to the challenges of background subtraction is facing with.