Research On Video Background Subtraction On Pixel Level,Feature Level And Semantics Level

Intelligent video surveillance is equipped with computer techniques to process,analyze,and understand video signal.As a key one,background subtraction has direct impact on the performances of intelligent video surveillance system.This thesis mainly addresses the problems as follows:changes on background areas result in false classification as foreground areas;motionless foreground objects are regarded as backgrounds;and a pixel-wise method brings about complex computing.Particularly,the thesis carries out research from the perspective of pixel level,feature level and semantics level,respectively.The existing background subtraction methods use pixel-wise strategy and thus lead to false classification and heavy computational cost.To this end,we proposed a prejudgment method prior to background subtraction to release analysis burden via similarity among pixels.This method adopts information in compressed domain to classify pixel state,eliminate impossible foreground pixels and just detect possible foreground areas.The experiments on CDNET show that comparing to classic GMM.F-score is improved from 0.6 to 0.75 and running time is reduced to 50%.Since background subtraction methods on the basis of pixel-wise analysis lead to false classifications,we thus proposed a superpixel background subtraction method.By analyzing parts of superpixel to decide the whole superpixel classification,this method can promote background subtraction and simultaneously reduce running time.Experimental results on CDNET2014 show that the proposed method boosts the overall performance and ranks 5th in this dataset.The current background subtraction methods fail to classify background motion and moving object,we constructed a robust background change image feature.This feature is a characteristic of foreground objects,and yet irrelevant to changes in background change scenarios,which thus can efficiently classify background changes and foreground changes.Comparative experiments show the proposed method improves F-score from 0.58 to 0.69 against RcurGMM in OTCBVS dataset,and F-score from 0.49 to 0.68 against RcurGMM in PETS dataset.Finally,we proposed a background subtraction method using deep learning without involving fine tuning.Fine tuning is a necessary step in deep learning based background subtraction techniques,and whereas it is hard to obtain ground truth for training in real videos.We combined fully convolutional network and traditional background subtraction methods to address false classifications in background change and intermittent object motion.Experimental results on CDNET show that compared to state-of-the-art method(TUTIS-5),it can improve precision from 0.63 to 0.75 and F-score from 0.71 to 0.77.In summary,this thesis conducted abundant theoretical and experimental works in the field of background subtraction,leading to novel methods.The proposed methods not only improve recall and precision of background subtraction,but also have promising practicality in surveillance videos.