Research On Moving Foreground Objects Detection Based On Graph Cuts

Foreground detection of moving object is an important component in the field ofcomputer vision, because many subsequence processing stages are largely dependent onit, the result of the detection greatly affect the correctness and quality of the follow-upwork. For many current algorithms of foreground detection, because of the variabilityof the scene environment, multiple confounders, the poor quality of the Image, the inputframes contain much noise, or the characteristics of the algorithm itself, the result of thedetection can’t achieve a good quality that the result is clean and the object in theforeground is integral, and meanwhile the contour of the objects is smooth.This thesis departs from the standard practice by using an algorithm based upon theminimum graph cut to separate the foreground from the background. Two methodswould be proposed in this thesis, one is the method based on graph cuts, and the other isthe method base on grab cut. In the graph cuts method, a GMM background model willbe built according to the distribution history of each pixel, and a background probabilitydensity map will be built based on the model, the map will be used to Initialize theweights of the T-link edge on the graph, meanwhile the N-link edge will be build andinitialized on full account of the relationship between adjacent pixels, aimed at getting asmooth contour of the objects. Finally, a new Min-Cut/Max-Flow algorithm will be usedon the graph, to get a high-quality detection result. In the grab cut method, the first stepis to extract a rough foreground, and then, a combination of morphological operationson the foreground expands the foreground along it’s contour, the result of this step willbe used to construct a trimap, then, the information of the trimap will be used to buildthe GMM color model of background and foreground. After that, the s-t graph will beconstructed and according to the two GMM model and the color information betweenadjacent pixels, then, the same new Min-Cut/Max-Flow algorithm will be used on thegraph to separate the foreground from the background. Finally, this method also needsto build a background probability density map to amend the segmentation result.Experimental results show that the minimum graph cut method can correct local errorswithout introducing larger global distortions. Qualitatively, the results using the new technique look cleaner and more correct, and the method produces fewer errors than docurrent practices.In order to solve the problem of the long processing time when using the minimumgraph cut method, this thesis proposed a hierarchical segmentation model based on localarea to reduce this kind of complexity, the first step is to obtain a rough foregroundregion, then a combination of morphological operations will be applied on this region toremove isolated foreground and background pixels, the result cluster of the operationwill be used to determine the location of the moving objects, then，the minimum graphcut method will be used in these local area. Finally, merge these detection results. Theexperiments show that the model can greatly improve the efficiency of the algorithm.The model also includes a method for the shadow detection.